From 066daacdb92b72a3242ae732317012a6de5c2307 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 30 Jul 2020 09:20:50 +0300
Subject: [PATCH 001/173] Commit.

---
 scripts/demo/car.png | Bin 310725 -> 131 bytes
 1 file changed, 0 insertions(+), 0 deletions(-)

diff --git a/scripts/demo/car.png b/scripts/demo/car.png
index f22d8d66fa83c4a51f76cfc44be8f261d1c10084..bdceaccab7fb392caf3b07648ded397a43d3b214 100644
GIT binary patch
literal 131
zcmWN?NfN>!5CFhCuiyiQEyI%Ez%V0JsiXvR@b%i4zUV#sc*(ZbIS-}oecc{BxBu-^
zHXKhiPeOGWF?yA=WefhO5%(0U1Cs-T3xy+DLwiY(BgRy$B|-p;6*-&g2n%Vhku`gt
Ni<tdO(it5rh#x+hDIfp<

literal 310725
zcmV(|K+(U6P)<h;3K|Lk000e1NJLTq00R>M009FC1^@s6P7_(;00001b5ch_0Itp)
z=>Px#1ZP1_K>z@;j|==^1poj532;bRa{vGi!~g&e!~vBn4jTXf|D{PpK~#8Nl>O(E
zZP|6$39fWE&AsWpdES;cs47%}Bmja?vn6WO-Rh3+`E&HA%%_<T6EPj?8EQfkRJTZq
z5(yBfLZJ%fo;T^;O>@&{&hNMOJ@35&p@^B2@4R#FIs5Fr_S(yD*YV!niHpwYh@Q^Q
z7_8T$-q#x|i)*pJzZrM$PQ;)6{EuU+xfQ?s)o<eIvt}F}_r_V&I6I@OyC=%g8J!)T
z9pn^oet!PqytpX8PmWJwe|IN-^uv$h&wl(TadlxfHkKFT7ys#R;?cctqwDN6hI?x<
zMEQdQy)iV@h_efxjg7_4JFi8pZy-MV%~$c4|IhyrJzWu3X2xQ$w=4R(E@Gg!JNo*&
z;~c1VTYGVQeimoIbJo!vC(#v07hTah>xhlkacu4%#{LPgb=TtL{36acFNB*jg!Swq
zI>E*Mek)<`?d>J+-stG+j8pR6+1ZWO!FhCcb;Llu#=CyfckyXQerG-%iHlC^d`v$3
z<a-faq<Mjx#F@thzt8fU-dZjC`)g67F1?)HJ>~-d7o_)W<5CR!=e_@3@Xm4z?s;C}
zowKvEBf5dRq_>`y*a2Q6PWX=`)IVs`)7V(w1jf#2G#bFsPv+-wd~_IRXW;Vm6ugjM
z8_s%PJ<d)}Y0te<4!K@Xj|*_wMcWLH4#is^{vh7{=!3X&{YKQGtHYzC%xh?9D0+H&
z;`HQ%JfR}s>!ZytPS2v%Y9-!JXdh}s`a(D2@r3lpr`%KSG3VjoG3{`W=?@Q&$nPwU
z4i95{Ym;*;PLB_%PsHKDKKYzPt-A}@#cdZjq1{i9Vq|nUE?>SB-L%avw7I{xmt`Iu
z9LC;WD^7t)+;(-*9^Kv9V&`p~i~g>j?id;#$!n{%PaeB*46GoEx3>QoI2JeH9DJXG
z$5ZOi)lG|ZoJDV6cMOlzV|-#HCZ>m?G149VjUMu8#`DF+*xcBp+zx00oYx1kU4-YD
zd~E;endulF9*&F7i`d=YjrGm-*xufZUhv!7Q%f+cyLcEG8jktdxtN=s2Ui!dwYeQD
z%PX<5u?Ag$=hMR&9HH&&eMu`vC)5MUP^(D`o#fXa^?^q8)cWF}bpTvlQEv=DkmuCt
z7@9hZQ*d_-Ep>JEM6IWfvSwp;YBKKM`zF@cmSbt@dD2^N{{ZdY1I`ZO{PZIF;q9@J
ziI|+2NLtz7-;0f{jo93xKFw|NJ&C@7KFX^{54gBQo7qll%e>!c(w^NhI5-Gx4O7<+
z=<1aE90AJ#c$Vi*q5~YA^85@s?Se>c7wa-GFu=7pPHEF!%H4+s`}_Km&iD8B;rC{A
zf{!!WZDn~2S}F8i<D0qJ$(Wg*fu4Kg$&+WX_-rL<W)Ruc0Zl!Aya^qh!BZX38ts_$
z*PC?Q+&xB~oP!H_y(f0|!6|gq$NR&h(`@Ilu_5piaj<tB>+8ZM{X^3<4QxSPL7oxA
zgFV1977Me}v{7F?dHg7LnoVf`IB9osay)L`yd(LN$6@^L%ZKs!!D@h(AbG$J@G?!w
zwLVZwnj9V(h?&{(7@HhK#?_KHSC-deWo12fn@92RKK$3bh#}~8ViLN9-qzPPVr2>1
z0X{k4>ecD^AOFukkALy6|GQ|>7GHh!b^PM*|2`hyzlY4O#VImjAO2ig*^62~`ApYS
zo-QoR$DjV}r^wfOe0T3&eDTE>vAnVpt7~hqxp@rzod;qbTnG3H+5{PP2$gd6P_CUL
z?|q*#kBZ52UErz{HbyP*JHW3z)X%jJFV?^{vJSdBl>c`Q;loqfui*RE?ONQtbv34^
zr;?5qmzQFhQ{J$=8jUeDI1E3IMgzD92OH6Zc2`%HfpstOaejJ^HtK@D>d9NJgCojl
zQJ)j)bpWlGIzSG{#4hMren;kltKJ$s&9zbQho0)lmIL(HA?=L3+uY?mi{|!4(7Z7*
z1;0-qhsV(A!|2n&I<liTw$?Y|*vf#E<D-t)+(w3zeis$hf!wb3D%+s97QX?QYQUFX
zk%7RaoW6VKc3ir2Ii5XTiboHh#?jFkb)Eoko$<wQzl){kC((kIVA<$K2iBpTvy*P<
z3pv!)m%MV8GF07r3h(u#-adt2FD{OdE9gUb_3qu7n4TLz|98a%ZL|O_-MoF1^yqnb
zwFzH-`}Mc+cYphfSb7TVz&A0{88@!Z#?{L=qN}4u+a1Q*`WETYgYavIWzp^j$9u@e
z6XeKrT$-Ojessh#vT1u`1A3~#FJpYW6W@OGkT!0{<tyXyFaF~H1V2p2<7W@!*Ps4l
zJbCsMnb8q%zWH_xHU_gz*H+i#um9uU#m;ubhacXB9vZQUUKLl{+seAqwjPKf^a40S
zuS|gC<74;+zCVCx<=YeJumd?IUG^XsY`+2QB6S8|#{m2yE%fqzA9d?TmYp1(#L><{
zoFQ{Mj=%|oUeiWX=ptvkQ|>9R(dXI?e4F|V{!q88FQn5Bi9o&Ol_!}7U5!wM_RnGz
z9rEFie;Rjx^y6q;x&hHrMu!!n5v-Jl8I}JyeD1aVJ1;(X_4KH&uN*RP9$wpaPady4
ze>DpZpsjAG2f7-ed4i2_e5{?I?yTNp4N=cd(sbD%GP1A8Qd#KxGi1Usa!PweJLEL&
zkrS@kB<PdMCgHcI`rYrnKg1q6fFW%cvj@W|U$mXf^YHM1e9e>mIWN>xwhQR$La2Ej
zo*l*!ym4YWzLPJd8P6`5lB5L_i6`0$8FPvbRu5^9b;8){mo5b#c5wbi_-S{geVLzu
z2Y3!eNA-aZU9KIcjr*biPRq_R$qQ+k`DJs1CuyYP{X0|GfoIVJa03X$iLvomUtK|A
zti-Kblku}Z`=i*|-Hu=U@;C9|(FTGQBc-Q7jd~OJ8mgs*1qgI0b$F)IHeV~GkQQ-A
z7!NnEUyTpmeJ3tsXzXlm#HauG%UFUj#|8&tvVp+>0&C>a+uxT)w~j$K3?Z=1tsRur
zdW;VB#l-MX=sv(6(buEV(+$&gqErr|bwWe&@iCRgw24Cu%>xXm-NVy7TN=`6mQ$S#
z(w)(I8rd?K#<93Np>an?hSR9fFgYYoVVInp1Re|x>avS5ZKF}~M6d=xVx3c5q)`UZ
zXb5{>F{kOg*OBn5!Qb241N;gP4T>J{NKzWr#ubmkmV|EJ>8#+K(7GD7#ew1Ez4>LE
zm@+S=ZSpzaWj!%&g&VpfUla?(w~o;5gvdrmMi5jOSJoNhs*_C5fIkVDtomsemF6Le
z(Dq0nWC?0++FIjxczh&ozxQUm_wk2u{q`+z*qaPz*$OlnK_w}C68ZoJgwDV|T21HR
z48S^2wx<~53ceGJ)#DSG5=K>+Nn3|7v<}JvLPCZ;LP=^sZEbADE?W77cXIIIQ7h|k
z3}fp^G#b4ymX1;@W@l$&c5a5>&<Ku+&d3Sn9Z)y(OgLz(s$GRwl%{hh-WB!({n5`U
zoH~w2(2ay^eaxpDI<u`bTE`}EfCu|A`j>%w8yXJ`l*@7B_GM)K0W`Lcp;KsPb8|C|
zx6KU<ty)LS%uJxnRmSiFc;2Fp>WIr%uK;5g#@1@At*s}#8oIsRSc#-NJi?j6Q5YE+
zjQ8F{@nKk~EX32&^X1sU08FEvXA{u&==f+l6dGLOU3?!NNWUG?7#PIx8Y>1b7Dqa*
z{H3x1#|kOa>+A{RAdczO)D#M>J03iE7^|zR+3p$;XMB4C4N8+;6sAKVosb;@mo@;G
zD)b#3E8Bel$7ObYF3X>y-43x>n_GLdv$Bi)uwHZ=kB>02pp`QPGnv&;#L}qklQIcd
z^Gf+W7{FuHSEE%&NN3^zLsHmveC4|%=xc9h2N;oQ&|C)=fzGwGq2u$`TW@BaSHb-*
z4yeHiVey+*3)zL0Gdez$ieUo>4!uI1sU)X{!q}MQH{$;N2c+$RmIh*P|0r6}z{qF=
zV+X_S!7~h%U0^w*Z80u5H>f-IHO`gBUPp|M^<aeS9M8w{;!|j+i4l7W9q?|r2ZsZv
z7YA++{#setkLCtOC9u%c7#*GP9&~pI4NNuSt#_}+-8XNh;dFG|in;lDa9m4gY4PcD
z>@^SL$-_N#1P*ZzbkRVroOfcloy7JIw2w$seqFgb8UN~E|0MqV|LK3gaNUg0fBQvz
z`st@Qm5UfaJ$!@GJ2}9a0KYYi#;GCN8HaYD9`C;YUQFVMZf<Vjls$=64B<7MiS_Mh
z;dmXkXanF^-SqQq4ft$l&<}2*DRAVij_3t)w1+gkeaKj@UFeB!+M_Ypi-Xz^{yFig
zYQUr(<b&ji-5ne_3{ovn4Y{GA!Kh;(c9G^7-q2V;5`fpDu=B*pPAMUsfe4L<1Vn1&
zejgg?!|?0v?c=**%pD_N4Tv18j7$lSq?|+_aGT#}03-E<S9PieLor^`(9x*jSq5#+
zkKi+Czqhm07nn3gHA>|_GnB`UU&v6ugXg$EN+akDIx!G~Lo$SsKG7M&L){p!gK_)T
z9i0AI`13TLKV6N7_nyV$hpQOl7wHt89O1Z8LJ#FwKXnQw4KOM*tsC%greRB66qdH3
zdM*P1)J4Z;e4-Jv3nMYJFc8!8eX($53_8R^LZhD|&kkDq*a9o@)$i`b^T*_k`k@))
zjn{9)TW`Dxu9a6qaqqhav9_|AI!+^L?*O{p$I(VUHwL?6{?Y_GU;-Rb-@_J;(NIhy
zvuxXMzJ3tjefKo3U7Lu%_?Q1O#&I(4;Ryfg*T0G<PoKrk&IRdfaplSaPWS-yxfS1i
z`z#&Zd0-wL9n7|B?O~I&j*$=0B315%{ty}9mpY$x;E>nikAC!%=<ls3tt%^)MdGK1
zzEyT8U;Clasi_GZ+VPkeM-JPz;!g}i8+$ldC-BA5_Hk@2Z^tomdk7p14m9Gpb%YFq
z?rjfbpYpIW6jK*-N~Y4I6k;$mugWvhY7i?E#pB4>WPJF^AIF<N{#lINx=SLEr$b@O
zt1J1In>O8Z*zoXz4v8)9<6@7r9hml6fNB2_LD2LC^myJj-lcQxkaODZ9NXXm2du+D
z9d$R*D}3^SG<<}9NW-2(K2(ky5INAk;Co?G2Uxd~ug-w<j}5?m%6MdlIzWgG7#nO=
z#&~EG6uW>tNSDUG>61r=MLIPPf4A$JcrF3w$~h4kX+!a@*it1;mPNaFBEL`0i>x`O
z4gA)CR1I09u`dM51oJCxSMo_?Kik3laXxILQ=X&8ZCho6^2*?x!B921xXVDLG)37q
zhcu|uZ=EXmkxJ-7Xabun@sn-FzaQM4x==C7+<jy|IX)hn2<@Hba@@Q*5kLR=pFr3<
z=}g>zxE`%TJfd3LnGi5ek(EX*nHC^4SYdEZ2}lAspdn9b;1PuJ6pS#00eN<Gh;mp3
zu;ZA-*crjw?(aev0azy$JR{FNj68*%1fkdFyRo503^(f0N!}e7r_qBn>fwcDA7LaO
zNK8=e2}snLIHrCF5a0IRQS2Znk3mK!4RQ?Z$2t}u_gt7kqjfQ|DQnGPv}kNd7)E`K
zJiqt;`x%+p+S-n<zy3CsmRD0~>R|9`p2-Yt97qI}Gf@aiVltX#IS?9YB#MkY=}2H;
z^r7@RI)H`-)S)={Bwe-`r^czoU7Ql1TO98a0Df@h@AeNL=TE+sugywBPCch%aRPH`
z==fe?I-nzmU_Zh*(-5}}WiD+XovVR93{vW_xv`y$sZ*OCHD#z6h9*Yi`kQy-U7U&A
zuiedvy3UJr_E7oi;R>V1u|~2Ubuzm~yh@d%8j7hbQJN=bGOH04NgV~OfS{dK!Y#hh
zxiRv#!MU?loC#r4L0mv1I)@FMRh@|ol<i?_H->Q*r!ZKIM5p0Me$uM2?(ZJZ?(&5O
z0)`K88*R`C7VpxUj&mnYZ1!4EcLmlFG@Vfa=;{~+?a|lQkFhrz^9wVva0x^BWG|jR
zzlQ+a=h<Xjy*>{^U&Ll}9h&zWjhR-`n8rsPEjf-6Jw)4R0Pmsc7nhcz+1!iESFUE%
zZ5^X)-N+;D<TuAH8Ummd{%WwvsXm;E*D&z=vpsj4yXg!b9~@&S4Q6C@a$+(ao*FcG
zh!MF1juc!*%(lR(5nFpmMuv4JTa@3#soqceR;ua5B%vu7q0e57raFpjpnm{)vO~Jl
zbVhY1#HCSk4I1H<ziZS@gU+Z;Yi}3Bejg*}Fpa87l<3u~S7UN|D*ACOwzqdN_K+zQ
z?YH(m?ChXGaV9iUHKugZHK3%oq6ka$NM7nxm~xJcjwBsuAnqe;40`NfL^iRCHCm5Q
z$kO4+5C)(^7yR0DB_mCME9Jq~_I9?FoM_Z_YfA@h2j}u6hDYl$Jv&A6^LX&^X>4PF
zd(`Xp?DWBJ!)aLFzyA=Ub^wEAk~F)qxV#B2YjNZHbZo9{WTae!tQ#EJBfPl{4Lu8F
zV`hiZ*AdsRH{u8HUjz1ySbn}9>#Ik=9?|IUh_Mk3$*z>88|xSz&5P&*KeN!qG{%9n
z^!(W_d7sDYZ%@a2AH4+(Pm><Ukk_3Zl)1J|p4-`Ty|H!#4H*TaxoP`eoTlF10rUg1
za)UO-k(#?S7XR|U`#ApMU;RZaudc+WpM8#Tc8_-3$+qDi?S*5ryB8P8z`^0c?1>&8
z8Nooi8V!thgErz;CrG*%_sX*_jF%oyWN$C+s6#2Q+SazM3Q~Qfp`ncQxl!#-u9B$A
zRK+hZm*57<myAoG2Kfx{8xRoZ9XJym$Xt2PUczqjL3Nb8xk^XMCI9?hVUV)j{1hA%
z=%4u;)Tqjo<|^aDG^zZ(ninwVC(x?SYB~=XE5*oZ>q7-OY5A{whR#SMCLJi{6nqF?
zJ(XADK<R0{J>`i!xj*4M<=+VeqntG$u!j-c(SbwS*8y*O_NMNe7@xtQ>&(F3S6|+b
z&wq0-o<5>o07qSb{Zn)@cq$H4(K9*~)>mAUBw<zSN<XA%%X#64p886U#QfEI%v|b)
zF1ll8b_6;zFi?6O_jdN<>EqS7_bqvdX3D>Ibtc|_=LQD)AZdH!>#x3v)#Wvj10H&k
z*48)oVy~s%hc||LVv2I6rX3m38{^t0cu>l^Cr=jR+4Hr`^Jjnd)9f96_WW5qUwoc=
zeQ{|O8L^j+wzxH7-P*UeNm&Rz59)yRMIZd#fi2R)pf5fAovE+vt!V9EAX|FkU;K-I
z6*sQk!qHyF;r=YXz4sti*LIP6LpWC*IC=(1?Dee2<;!z1Gd&6YL&sFQ(O<)H>O+>p
zBbz&Mx_=(cHRRXM0cmPD6N3SFK#NpZlA%GQN9pj??n#akg|=17(Wcm>F`eKv$T4Ph
zu<g;kbN9`7>%))Z^4-^w$0OKN{n%3IZ1T^53a54TJ9Y6fvh_F({9?-#N3y^ouM|Bm
zeYbm#<sS`Xo}FuWW4Ib9)&Tci3fp#FGQema<eRiCuOy!oSu2mopY{m&y}G$LOVrSJ
z<l$`zR<<1iPhd<Nma9CPMmpcgS5?n|3~t)<w%(S7vnM>!S@5YGQtoJDOaCOZSkA(?
z);IH$FK8oqzTgP>VW>3d;foAR0H`*Tc_n{ynuqDNalD$3G%24MBs0*eBVs$s2i7H>
zdh(DYbi{Px)s5b(pNr$5ZC}ZVVn>Ka{-d3^*Qw*akQF%k!JTR9<S-n7B=bpx`Su2e
z^Uf+B`FQ*{fBsVh`d<9v7oWv<_g7>8kV^E_kr*}V#XUI{gqH?UXcZs|krbhPl95Id
zX181ogjU|P&VWi<2VqcS8ksIEN{17^-{rfV9gP6KHOe<s!yy@r1_oe;ra-d(i(Qqr
zlg8r#M2nh0LDdi#-El;vk4`V5xqpa(grEfGP7rn|LS!faKd~`%ET~3woMohHets^8
z2uNH@D@(Dtxe-76*`LLmZ@d*NtLyRi|L}`60(2xY0^_#|Oeo4*8{t4EqLDmoJWoo0
z!362JP+bLYIubZk4#(-DJ~|XOU^W%kf?}V8KX5OeeV)XT6r~dK%;Ibg4hgor&o<#*
z8FC|$_(^dljC7xXj|PTXeV{kn=j;T<XHCG3%1k2H!|lUpP_b^TZDrdUwN+_4M9iYc
z5P;X;(wX=u?!NIxM&&I}L*4fhyUIMhZ|M%^5Fg@6!dHpQ@Ln}kQ_w)nsl2Tb71O9c
zB)t(tjnM<zc5{6lhr(|EEtuyp9Z!2EPN|2As(~S`GZ7auS1Y5CMjIWLpkNR_4PFb7
z`K*sdM0OE^35QjA^g??Y5W=HEN-v+hPQj7U4Pln{GcO?4k(inukLj5SoQEmc<}{uz
zK7dY}7?9BZ_)v_D4dU<~WIct&JREjq#6tt#b6}vJG-X?@LVr8FpP8M<_!>>;YJ+;O
zuC8Geu0#`CR5r4WC5`Fn-gxuPYdH)?;l7LUXqRn2ZD~YIgLYzK0%LYOdjS+=hv3aJ
zcXVE!FK2Iw^>HY`9)|no)>Z}yY<vxV4QYj?-!vLHKJZ2M?tnWZV)aI&C|1%Opp2iB
z=i*`$oE<>>-8d<{@9V`F-iGIQ$lGYxISx=Q<`%}|(&dF1)v$$!>_Kp7z{Q2VN+<~A
z$jb6kIw1D&=@8q)=P@=qM%%SAvaJIoKKgOc6q*AldvR-&CJh<dOr<I9ot@|qB0I?6
z2)yd2TX}(ek#;Ww&s%T370+?(bSUhpvi+7<7AcS4(D~H#L`K+kn(jY%gxqN2Se-@x
z0C>tFPBqH1XK4pRZ742XxfHv!>%+&(v9aMWC1@><!QWZRIAtzbMRVr_13l7kk*5*c
zaraI=-hb~VymS_8yj)-1rJYc|)TM^F?nk)pV`w-O$ze^S7%{Kk9*x&--AyOzn{OV*
z>c&P)Ox5F!x39!SCvp&4=mVDrI901lI4O%8@$B&)^;hZJ8{jYuhxpiQHJ(o1!$-@p
zwy_&ylcVvopMMyi{P7=A@74J7cNliir%bHVX#3c;@6mywSA%YbN5^9vr*?2?C}pZj
z&bBsaV6;&ORp(ykAsq;Ngi|W&oG4E?(IF~fP?a)EJ~c(9$i<h6UVc@)+UurF)R6JF
zhKAD9d+GNChg>~I2Gq4Xqj?4r3~Hp&Px7MtOTMN{eaJumOUhA3`jbRON^<Z_g)coT
z6O&;rH~)%Gtqgzj7oegCIDto*>kzzhaLod${LAlrq`}jEmcf|JgWo>#weX`c?alKl
zn$$fxECjh}`ZVA;?3J!#oZ1t$+ib<Nr^^`gkK@Us^)!%=X)ASwdZXxZGZUXa@hV=e
z7kP4~J;oc27J1kR{DH#?_;ko2PP53zneG@EyojM;dxttP(hWLvBgc+$nz!Q7gX2K$
zV{k@HO$@~Cn`1FPIts1y#M0s-&hr+i?TW@=gYvuKwXJx%bOeu{M<1rl)Jzw0xj#l6
ziO}1d^mW)e%mCZoZY$;XJMX@oJy=INTj(Qky0UEE8|m<>Lp4?p4^NOW(vsM67+z26
zoH}yewocv0nZcq>hsK^JU;g;Ve;jYT{wA{E96tOG$NP)8cYhh#ZqGzF{C*tUP2>Tx
z?dsLRxP%@VB;O%)n#01ouvvDtH{%ppHPkyCN4v+dx4j?Qxb@x!HdhaFfCfil8O={>
z<ezdrX(Zc$!^hGpr@TX>B%PFC-vM<)a9nhvE2d-N^0k<ryF}ZJ#6@>+v=B@O$WDh%
zD?igA@Pr;JCm`uh9w6@|KlAn(@6>I{XZ}uJgyt-t^y1b)oz8TE37mzh<%x3VL*dSs
zSs9Kjr2MRmbZsXi`v$MnOWIRXcN)z!oH$ATFjr~6{awly+Rp4VkSKk_C&`b%VUM$R
zn>5{%!A5PpQUyO!{wO1qGv;TxK1;iheC3PMQ=&5j3@HPsZ@Vt$<7v4CKJzeZf8#WO
zAMFM07oB+9*r2gNF^A7trn1gYwH^KL6ubBcFYpjOXIg6_+zD?fwOugn;(ewl&+~En
z-ltu26ODgUh^H9wDtisBlam7oc_41wxRyfq`Qi%3gi%NlP=lV@h|SD0>%5we|CRE2
zN(IEA0!iad$00pD5|Ap9ua{MWNxE^hJGmOAI_3QljeA7d2ZvM?VXvdoKTrg@rK$Yw
zy>QOM1+aT>uS82@K!(u>Oy#SPvNT{4Fb?bL6C-3P@3>D7lM_0loL+>_K-3Thr)Mzc
z{!V&+SFT*nDDfT&e-S~otr0<9)nP+RbmG*;6L+5C!Fm^GLVTC^60k7wGx1QoSsT=}
zz{c-F#Q9bJrSjytr+D}J0)$e%#D8X$X}vG$W#sCt=tD+MLmr|Mv$bwLZygq{oOW4f
z?*r*ATPEy5`5!{0WK`Odtf*#2Ci(}uV;m!DW?>%VbP~beTNo9AB%!A>1hHGM6e_}l
z5J_TB*t|+MK2rcmsyr81VQGJ>p=ob`#)gW;sHJV>kR`h<Wp0^K;h<;TX|ER~(Z$s<
zf+{fe9AwOa@^omI(GR{Yj3@R)5qJhJ6^zF~AI=npvI@j#i^E66j1Ewe7iUIfpd07F
zjE*+or8>rR4I>t#J5ab6N4&ExTKi`hY$!1lq_nMLlExN>wUK;>Ch6Q7@zAl~-`hp0
z9j3A55Gx&`9?Efu*b0i$;aZ(24|}+3H6yY(K>ZQJ!}Y{-&Ns2AkoLA7I&jt{g&S=t
zy=u7GR?Z30soKr?1H0+q?CtKv+RA#Y8~xfp%9g>O;#+ZBhFK{`#GY+oluu=3FV@!9
zZmh32`F0zpXEmNa-2%6#$!j`w!r=T971s&MPkymSX?O^n%Zn&--%Gzm86b<emlO8(
znn`E&G!0{nI*+24ifiQRysWHjWZ9|g(ovKSj9wSR$!Md(T-ZlP#$t@qp&dpz9X4Sb
zI~+?#+8&p=x%rryno0R$U_b<@1X_pt(9!|r4#gyVFp4u%92N{9<m%>@!3AiWw7VDy
zYimb1P=PU1oS>Y?1tFcEcgMqrd(fSV3<ZTlqvCY<l+EKa>{{yD5z|xMF*`dDBhbR=
z@Gvw)yLTyb!C}g`PGBx{3i_bii~N24>Ui9_b1OzhhoILzVBdv@!C~FuStB{DM+Vu)
zSzq7Sj^*Wz*w{LUmb+tOdL(8sYUVG^#r);VF+IBwy#v$}TC5NCQTMT!nVaQ)DE2Tk
z9ZJ*JKNu5}GdNiWYG!eA=VEUDDhAt?m>9nd&K6>Fd;z*2rETlfd5HEHp&f_8{TS^s
zOPfsOeQ#}q_k*MxgiZ!&OXu+PgQFU_gCH*)PUjG~ZrT-RAE)i)*ttvwP-yp4oIjtT
z(@uHr?{m_gpTVO(!)7#yrID;McG%((29vx}PNVr4PZ*HL@KVYHXv@?2QkjP9F;{s_
zdOka~Z1dGomF^6p$dgEgLx)&8<jH)?Z*k^(mA*Rd0z(dSh;~F*81liH&S9&NpYr#d
z_ZrL{IC)5FxB?@RI-Q<UKk6zj;#{8LyfCoRPuhOOs>80(ciqr?&oDeZ68*iS;CVdy
zYNI?iIDmuK;d~A9JMPK*{=N~~V-y)Ql5{>eI0QWoLR<Z51S?=2oA13kSjVme(Ui$W
zw#&2m4BAHIz&p^sq8>feU$0@T*4qYk6I>o3E9xoplobc9qvT_U=vn{H4nu<m7swY!
zD95DJkPdeog~rFn$0<wMa*{fr-ss?2EoHPktt_#v>@{sc^U6bemQyZeV6Yf>MQ2~6
z!=kJ<f1e*i3w!$j??2m${Lat=!lHgquau#1+*4=aR2KCO&|1Ab2c`o&qqVgSY_8ox
z3((0XJh!=Zku=oN;oK6w>8l~1?N#oHoxSaN{NxdIgWMU>k+SWvBM=e>>x3r_)YPlg
z5!*?fFYb$;EwG9BqRoqOO;YI1=(WLC(?VUeh1b3P{T$x8w73*cpFF{-dk~K~pFMmS
z&mTXGl_!s5eeqdrtt`d%ngOo0*xg!>ovn>@e0I0Da?Z~l_j?FH&wbjs<@`m?19Xi#
z-T4ISS@m+Uqikn@!XW8z4jZWNfmUiYbdPgR`-|>-q5pWN&QBT2z0U`*Bm2Q$8Tu%F
z8r0MVwFk1BXI;n~-`g8m(m1@-9?xp{Q92Y{)x%Ca%XtSn7wGfiz(a%Rwv75~TTqXj
z=XtOnd+6W&gI4M#uRFWDu>n5THa0o8V{^M1O`dJyylwC7#rE!gUboP}TRdCeLWgcN
zV|9}&=USfIv9`6tyB(hI#4@^WX>}`>)<~z`#ra>u!Cc$f=bmeGFIM@zy48w}=3#8@
zqKEd6DTDj=xy{++`Oe;P?DB4}r44cty9Z^;z#eDbA7DRxaC=(nOvcc#A43uC?QFqF
zn{jzzJpSy@ehe`m<S?gu4_2sn>6u7p0uNs9m$8bGE<$bO@+=K92~x^~V8yvz(%osC
zLIfZ>C!LV4zpp#H>UwEJ4FeUt9-6}(9J6g+y)?9u)oeSBZ3xQl)IJ*7F}SE&lpcl%
zLO=oA-Q&b*jNT&*9t9aGj#L0JmCvn%(Nvj5&-4&!)PiU`A<8cDYoG-CF#;4u+cdOB
z*4@`$$B0;<oP&7q@Nq0IE%TflIy3|w@=_+dhz<polU3800)7o64Urr+(|%uiEgZiM
zfnz*M#Lhd&9ts#krb(Q>V0dtCanuKGLOyz@wx4OlyGd<}A`<d>9_3jIZQh?Lu$3L0
z3K@qG82zw2dt}hzDdHIh#Nf^esNl_RTq8XF{q+=93Ot#-lu14y(W1vD`{MSyugANe
z{4j36_8NxNAoa8hvv}hQBTt>nb7;Vil7;|A0j84HQ7AnV8fyNgd`{a@rnW8`&zmri
z4&4#LKMh@&ZDVaM4eoRjj55I7j>n}qgotxARKiBRPoTSEu!|?7Y6rBd<3pV^?L3&C
zE{uFm>*ug8ouF}y(_V~9C+pgCR&NX@qNR@_dU0woC=YPT9rr#xgAp`7$o)V}Obt_y
z17LC@c5!SJU=DFP0aoj(Q>#Ezm}v+)#@jhHI|y4N;L^A>Z3PEVfJ4KhY3zUf&6D`%
zn>7@Az)Khg3O>HqAyd#zPK?uj@(o55?WPlP0UhZ0IIdFT+la8r&N1PZfBn`?YSWR!
zs-8c4mN1+l)c1D9w{@i~6!PeJZw?u%*<+XltI_7%hl8Xw+tSIV>+6RIPJ5YDN}Rwr
zJu?`yb5rEgNQJS#w*x)yP^!HW{V_5;l+K2D5<i}O(BMG-5cuh(ZQ)xi<R(0-(_JQG
z4`BdlNERdi$>S&S-MuHVfsieD@4%>Sa5`rJ0~8z?*^x&_a3bm$y`)NG(4GnBmgr1t
zJV>w38(FwCpYv{xf!T@T>l^FjzY)#7%{VyT$-E~f$76D83P)}<dHwUxzlh&{{v3S6
zhZqqW_B1C-CdkX-O8uz}hs9-YBmV9`eHSZ>r!h2m5#tjMnZwA%FtpdBzaKo44=99R
zooL(0c`8n1-&6V6a|B&T=i<Og`BXC&7f~Qf2kA6@@cyOv<cFVt!;|>-+i&9j!w0ne
zNnE}%6Sr>L_P_^UY;U1>aK;v&Zy?ja(eOY_%}i0Z8DJX&*BE@Xg~o8s6QfNGh}3~;
zcyttFM0qDAsigbiyZ)42_O|HgR0g(2v%NI{oH{_$rn=Xl(Wz55_&Wm-DSXTW*+;vi
zK}sb(B}0DEw#rToBY-fy6iv28Z~2T8Jb6~6CI`<mKfaZQY&(ZQ=zJ=Fk~fpTYy;9~
zekIRxlzd+PGIN>HgzXO!x{@Zp1Avs}l#%d>{dUp}Cbge?HR#im`4#xe@F(iUy9`Qk
zRy{G|P}!~AH?t~ekiUV&I%I?tzR#f>91eHLAQ8XCQOO_K#3xQVHIQf<s?U69fI?aA
zu%F{Y3?__1C(LhbZbDDUZe)zT^vZ1Ij(E*+-jw6LaH{w1nK%V5gIR+^J#qQE6R+o^
zzi}R`8xN7Yr*ZB2416_6S-Uy!>jZ;lb*UBG8)s<~Opf)&G`wf8fW0KjD`6eOk*L-B
z;t__f#+^<}2fEfE&&b$lwyO?+!&A34=9<U~<vrgykIaCBy*CCubCDFjb)gvX<M3RE
z<1H_)$Ft}5pdlCQWrRUR680t;*it@exZ3OE+^@a8Q+R1OZd|_sFO8tncH@iRK8#=d
z!#ZjBkUA+J%mzOJdEbEzQjdj&$+&cB8hKTR-?#bZ^XNs+zjNo!815U2qkRlmjKIKg
zKGj*>LucU3D(h```9~W>0HrVK#`gDWp*nlK$Uv?12{oi`llBXIXi!FHSDww`(&U$9
z%-04Ty&9~ofiLB1+F&HONcxkm{7xR!;nKm8PYq@|%)6u!|BjeR!##f&7>oT^3M+ik
zJ|$nDaqm@G<gnS|j8In&=kQ6t0;}$`Y)2q?@?XN^Ax{+gFR<wpm`A0ft09ES69at)
zoplf$0_;-(FfgQjRb&>tll-a!S<X_PHdwK*GSEev`%Dl`;^FT+&6Uh$8C9FtQeRH$
z-mZ2ji+s{1GlTZw_lh9X{QluHPpkBzEKX7Cv-FkPiOeJO;bohBbtYzPTaF^whavQR
z2=lPD9Sieg@#8=F5rW`2e*NihaVA!ybx33O4CL@2yY=lh$($%bLD2p|0TLGXAd@!`
zAq2p4jemF$!oavTGHm1+A?=(8C;r+!WKFW$hpWm^K`wD<9GQ<yP~~ZNr;*4y2BEzT
zPO!0Cr!w$`CZmSDOGU$Pi9{jjA^15w!LkrKEd<69upjfyxde=nuSzpwZFNRQ$I{r~
zqnvQ*SaKZ;dqH{;SPCl{PNI>36`(m^gjTgq4kt0vA=4zGQ=S<c-7qRyR&r8U8Bw<T
z3#I`#3M{`#=JU*oi0z_4l6l%~@xqzNmiH*%DljrSp&&6Y&{7R~GQ!i`K7a<O<6wU}
zc3qMo6e)R}Yv4oW=IO8oBifr7wlaamNwlKbl1C<Mapygpi68zjZr-_zp;J#|(()7z
z=~O8s5E{bdP%o8=O40n%fgr=egj{{+bD5a@%L-bZIgZ*DWL%l(_z>?I1+a%VzrMDT
z#;5|@XixTLL09#9Z+1EN^w2h^(8%#&rlFb1zf6cqhXw=E?(Jigl->DV9S*0cfgcQ9
zXv!#@0@b-1MyhhC3f~F4^|8H9&S=|{Rt${{#N5IJ#&%gD$6-#CPWh#5S7vb-iB9u&
z)2O-8i+%FZu*47>D!mf+lsSaQ9<egXvPAQ8;y>jZA^PIW=W*}e5#PXPd|w}MVs5=e
zub?j-K!s;rBLMh~`nY10^CxV7>#kEXXitY}ar8<grac5m#}_(Oh|7Pr8`VW=*HFkf
zm;)NZ{n>uvMT2%@1LJxd1A|Ww51mgCw2y`(#`9bU2J2`zv2k*0I-~c;w8<{=M_J@*
zCF4fN$yY;09(2_$BVy8=!wY+Qib1o!z77qY$0&xY!?o-k*`hGVAUm{ab#;^9PBcdu
z4;5$Kc^J-#Qid4O+rt5suLc`K)J<N|sKMFn9msLl_NJ<cBtmgH?e}BIgRzlpw|n>Q
z#Wxt_TRWS%a*%R$c(|W7pNpB9Ir6@UFTVJa`s3*G?&v@nQ6g=7d&n;EBMT3Y^<Y3P
z#QglF`0a1Mj{p2O8=(CxW~RkitqhAB8O$MC`xsRA{s^~|vz>_Sgim{@bl@FhUa!|O
zlAEh|K)=mxjIc$Ax%I@GZ!W|~AAX!e!0tbIfU~*<t@Qyf&cvNNz=e#1?wyE?cmxi6
zPmEmmMK`icnb+0bhkQ`E>-=hH^A5Ukl`)+PdAOs~`3$}AkHMlYVCl&q3!Im2u46^>
zCCsTikTXTmE4!3nMNX9=hf<$sk%x*Y{!bhzpE5nNrpUjx>@=vr)s#gn@=STc^ZeS6
zd7rRyuYAcNDJiG=+He>UQs;o@J}1JLIT|*gye~f%L!`*wG}0l0bU5-^ku#>r(?1)W
zcgl@4n4o*VP1yP8^P<Hw8K?R4h;I$b`AB)64n@{Sa43%#2?{sj+#X!{(r5V<`Fk%)
z3=pZWa!4ojQBNiw@^jO*;eCNe^;KzSgC69PaQdwJPMs|+*_)Y;1oCZ@-^%MIMz_7E
z%6w&R!HM>QwBSM~_IwRw55&aO2u}NSOwSvrIO6#^&(-BR%d5G1pG~;eL>H{Mf?roG
z%nsp9OvW)5xH3&g!d}mmdB~8}wbkVN>AC4Nh_-fil8%()_Q>uum9vh{C<`9Uz<5r;
zx5r?=wF@nuBU9_-eG;>?)94siJ3EMFWX=8iPmno>$VB<Ej6zUGCvH+MmYx*nM;aiq
zy_IV<V4Rt8p@ZI7LsotF?P7fM^*ORjLsPwpJRc0?Ir3?^GiGLokzrjp9|JfQQ`zhC
z?YCdWK05I2Td&2Xsk!LB=uW)sZtce27Ior;^U&|e+_4GjR8RSAAVfo!21s~qbMNIr
z3+@JX=RG=CuHlUQbhuop(p3)+)Bdsk@`y(<{G|ekvCyf{<OSJPKGE(F=fz>rS&)vP
zTX|O+ZS##+vyva=y*9;X{yIJW@~B2HNSjWyx7~dvZmZt$(%Y`Q)rN%D_+9uy#~B#3
zM|el^d~04LrvPoqlw-+fIsqr#x93IBE>L6Y%H%-}c%1=lA?hMZApd;k)x15-S8RcX
z7tcz~UjFu|ew+5y=ZRjg<&ycOsPI(W<inCq)PlN9oG|m>i|<L3uU@<>-z!7?-go}?
zxxcMX4wX{u^%x8+v}v#T3unSHk{a?Z51W%8TbnU8-iROl=%Xy;^Dn-MhYwd{|A5A?
z4Wzf4la?u<kf6|5P-#%lYz^PGfs&{+Jj&`DI291u1>W4z(S8bWdl*tclHSOll~(X5
zm^6koN{!%aOlgQ~U|NnuWh7RiVVyNxRZ7_{4Z+w=rE*hw2v0S{Ng|U-U6n<~*@w{&
z@H}@gBG0=nF!@kuq*dJ<9xLoHBKmqzw!+by5fmc}4)t)pfvd9g_YaUK!Wr0%z*vUE
za8CKHppz+`Lm)BdxNnGAqE{#zA<wQk9Eu$F1&*9ZngrX%Z7N#I6|){W9yse`&5MFa
z%JYdIab#O?$ea><{^V5Q+FPKKmZ^4eCNx5aom+rGAxB(bK{$~)BU^lH#LMWjf_Hs$
zCq3zuA$%ucNlT5<o_PKJx8t1;KZ<KNZly7nQE<xB2zO|TSKBxzFVZgd{^<3WvvMeO
zG87InYjHGWa-9i#HWXm#OkfZn(bi7#b-exB>N1XA6L>l>Ze6~$fq`6)nYqbSPTjS0
zaK96~JKLbaiN6|17>(2m*&f?CCmN4C7+pFP5Cb^q#A(@~Jx|jabZo7W_wA<h8rpa#
zH<d#V4vci-<i(555_Ov4oSGSpt5@eR&^vPefsUpT5^<rienj0hO4o3PHr96$J{>VJ
z-W`_~CNWrtfX|*bd9FAv+bBVYP*5r`oWYyYdvUZ~npM#&zi1pfCqPF;M|^B#Bxa{)
zG6MAc+4FdS5o`}y3TO=^aO+vdhvR%vuJau%Unz40+zwYXVwBNgp5@9!<6{Zi21dN|
zDY6%3&q@4pKgKf7*W&ZjSXtG`L^(D(Fq-=^!hiY76`ZXR@Y~8IWP6|sC!SA|X9Fk0
z=(<k9%IZq?L@5)@Q+jZs=+e@0Oi#PoiNj-DUe}(sJ#bsYpm!ys(lu`z!Jfc;icE1H
zM?dXZgO3l&a=i9jfMcpb8|%zWg0G1*6znP03D)@WbQLmtK(1W98sih=v4%tQ%{O1i
z^6GPVn4&N`a_L!cuR-Owe*+*B<hi*4el~ZL|CDV`YFwkuZNGs=EhcBiFdl~D`P1e2
z&9BN(r>V(+p9bQ}<%t*>#X0EiNXJt}sesq{vG-to&0ZjvkG#O?=tQoRw$ykCoMV>;
zc5=8;*eh4A#r119;hBvbBQ71A&K|FsSskFjd2^*RhmSx@{ew8b7&72SnJG>ct4Vii
z81h%{r13zrtJob1loK*}XOEwnj}&EGH1Cy_%B5mlkSFw{f#JR77vuJVy0+S~PYN&Z
z%IPz1c;>%y_`AqF={n!~n*Y+ML7vz%>N^krFK1hlX1elDx#g{SiARk`+ssJ5iHrO$
zd?Q=}WnK87ln;+*UILMIQJxh!Uh-|%)xUP7DtOb$Fu;~h1Sik<m+#6@VUO}%>IR?r
zk~e1MVSTFrP4&z+NP~%QOrP*(W_;JqM;NO}tOcgy<fqhEnA5leKFd$|dG6|Z9vS)N
znSnbe#9OQrysO}vM!h<`WiKo3s*&vw)#H;R<a=)#@UAQ;k3!kVF!MRYfjtJ#22&gg
z+034hQ@%Mwezo9v?TXQ1hiY{ryARX(uS3hut?{=#)~TaN^TCH7#wdL2oTrD6p2RwO
zWEDO*%%M}I=f$4dao|vwsEeJ0*E-mvJ|`LEaB1<e@sX6rPoF-GFTVH+S#gM*kS8=E
zHG1JG4S#fy`d{7TLMK<QUV=Y|Vtw5noMz%kJ~oiGy1c>l9KMW{m5>*5!9@ys(#dkj
zuE8{y#J_pta_05?`I9)<-Hkg}Zp6&UByK}z9AfM@H<~#i+g|9&@#(m7>1xc*;A|kv
znm8XWyX<O!>QDJ>px-%f4jpv0OK1Uvx#Hc}=s0i>CyWm5)8TTCl)*gb@oDQwFKC3s
zvH0?8kFqwG!Xk&eLWlA@DM&*Cu9lnjQqqsK=1{#t-Rge%m4fp-aV<S3z4M-2llBT}
z%L|2mJ=GK1g7%6!2*PLPZ}39AIxo(ibDtHGEqqfs6FjCI+OcHLL3-a?4!;C3p(L@_
za_@iT@O@U1!~8^pY$RSZn?MD1O4O8-hncS@=|D|6d64k<li%h6X(wH1cX?kORXWR-
z_H=4(1Iu*9)6_Tchh-R~DzGc>$d~`5>k^PIaHI_4<CK-Oy>j27i`szNZ)vZ}rw((n
z5WQ`Uss_fe&cxpCRtyjJ#U~%Xlg$0iw-0e9)?#-b)1|u~C^fPbNMxQxdK3ibRbaAK
zyyapFBbxmP;sKPm!<s0c2G%>==Pd#mBR~VlvmYZy>>6#B_>AC=U~IVRjN{83mSZ&c
z00LAYJ4DF!o*mz8#a&O^sEFx|Dj41Eg1Dg~;5myfVABDao0*B(+1WHuj}V&MFz9o<
z^~V@A>y&c<;U1Ir*ainKz4X35*U7i{f#(=;hZvPk+Ux^&j;Z#Yic-(lao8&RVu&~_
zrR*1w-C?{_pcW&&G_re5XqY+NOJl*dDa2TqBkyc{Z$YX<5{ffnqQVIIq9lEBElSNH
z5>&*4c9WsRt)92bULGPGoE)zs(Fd_P0o#?fXiVUT$|*GGwLnN3na@`dtVb??SqwwR
zS&NSj1n=wbz7_8wtgqj=4NmGgR7AS1b?ZQ&j3Gpg_Z)xOR$8f0+RlVRz<VhwjR}no
zue6bZM&sB?(i;Ao>khqe<sBUgoC%Ea?d=WHJCx`G0bI{>{?bg$Uzv`f(SGo4*R&Dy
z7H}OWj*Ela0nOoPG<S27^5H)8g}}Ku>m={8RAdg<0^cbBv+avP#cc`;R|s*L%icbO
z`SC8Wp2zrPBd%Yci(9v@C1V>gUtC;*&)l=YeIyz<9!@Mas=SJkzlVU<sqVnguEmWT
zqZrGhNe_;*w6~@hU9}W$CrUeBX3rOfp36^)L!D_^QU0rpa91q3b;B<8!FclMaXfr*
z9|Lra6t4O-6;snwiQ`?I4=dCIJ?8uac+0sdyC@x5Ngj8(LFXqpFF}LcaiI>8vOtFo
zIg#2(P8uS#-^RvXE_Zu?GIGe_$Y?EQ7v^JnW;Tsjov9p31>Y&VhH+wCqS?q^_Bvs(
zo2LrHXpqCn>>*rRbAqM@fb%LGmwt``Qo~WX(&pE$lne5T!m*l5Q+oZ*P|SO=vc8Hl
z(M%&zheUq1mqMJbF0Vi*E<k``MuWS$($OLKMvh&&G!eJ%;55V6_J(Y1uF}T#fRwS`
z!mFd9Lw@<v<?Pvf_~0Q%#|pAR8tuu29`=wEd(bhgiNh|RbSJ$%f7C+UUc~&|5d1V9
zvvXs}LC5=-xi9ttIp*FTkMF)+OL;dqIEqoAEU|}EO=DCagSS&Th@;t|A}HzV10&QQ
znu3sAwaH#bdmc2Vq$z_lMd2Djv^S9}r;e?%NL(K~X%=Te*`4wUcyo0j+Q>+x!4{Qf
z)iY7_gNn5DL|AE!UUf<K5R!-TO6BefC&~b6*CX|f1noP%k$KZ$&EG{{+OE>G?Oq&{
zBL58Ns1!9yDw^<+j-8Y#&UN@KKV>60)>zFt32+uiKG2AB_=LlE9%)!--cE*eSXAi=
zNT&5`>RjZY`mtT-HcVL_Fj=NDuNYF^amZ}>s~AoBito~y;K%n)V8}YQ>z1&hTP#~6
z&_8pkdRn}ipFH6kkK_^f|3%$MkUXKoM7~xkvB5Rlj&zlyVI2%?ID|u}v&UA!&#!6)
zHSsGhEwAbkc6g)1ejHLQJPwa@&R9Bp;uVMD<P>>*y3d!=fw(o0h5iFCr^gtw$NStL
z!J}Fx_NY@g@MYU71YPOv@L)IjVG?KM!;e2gK3opRq5t-aFXJEn;g_V*u@x?z&7O2>
z=H_OA33-bH(Q54?!;nF{J25>qjv;?7dpV>xpMU+$ck%VtE6|K{nOq$Tn+3VqLPp6m
z$<xT1D_1YahaY}`?3m9U!Owp4McONt>#6<XJe(ef4+2Lzu?~GkzB%0H;p2y~x_(I6
zjkt5?GV*f<`MwpXySO%YDLRk|&5i9iu!jY`c-%UP26}kf`AU=1z~+3To@@mj0r$yy
z_H;3WFfOn%G=u}CVn=2etdf5Wq8TjJh?Fc0wCPY7bjxLwbqsM54GbErIXf>y1EDrs
zIt^9Z)_oA%pF|uvJc<N~bLR%R;Dxv}D3>hEkpV?c$>@ZFr77is-&IDvKfL;hXW>`w
z<rnYy0oj2|pOx~=Px7q}zxOw!n`ge_JBrR4yn0k+38SP>^W^Eh<y1$OYeEk`s^Dmg
zggo+@?Bj1v=WiHDw&&CRj~r^1y33P4<&Jcu0q;l_ui{J|l=m}3$`VB$$~zxBeAe`3
zD80N)zL|(Js&w4EkP?)ga+=I6TfR+s$yMxX`&*RuOga-E-YL#R8GDNdiXn2azYBqE
zr&0Ow$FGG;Cq8)aEbc#Cjphza>KgVKl_DE)O`%zIHsFh<6-WW`nXSgv{GHp-=Q4s+
z$n}<8+RO>xdY+Eu78xq?0Tj3`>lk<q3k?}3n=0hoP)o1eRhwLH%}Gkmr!a!4A>}4y
z3aGsumlJesuyZZCQLZ3R#3AHI@U%aC|Gl_#^JYeA`*BpxF<2bxWS9S6{?%W_@9y7=
zH4K4$j2?6~L=58W^26w{>(adrc2V|6PLPB8jHK$pJJi7XJN8}$sK=hp1cKCMGBu1;
zCJIpBr(&ZOB_Jc-4r$5ZPDvNw&{MpMQx36hE2=z>5WuNK)N=e^lC-b+iOnvd6~*$0
zoC%F%mtX6o4)_5$5-4B3J22`4bGtmOM-T1R2OhHxkOYhCr!g?phcP%BBVz+;I6r;9
zn)}yWzJ5L4dguMPef#wkObQSUa=USj5=qQXWLCKdztIC7t5hoBquf`{1WFHsPmegC
z0h@}=D90}J;f8x|pl9!jj_~H@Is$Toc4<NQ=P@`^2W}L@{8Y>=OhkX98zw!%P-}vd
zb8u}>8CaqW*~BPUE>Fup*cOI!<&d-3+H6L1Yd4&elMV%pqtd2n)97=x4VCNo#6VoT
zK92*_p8#_}jK2!-@Tdvy&SQ37<#RO~P($+B6NLB90Y*$ehVWP_jVAQAu~CNPY&S13
zpvySf4?dWPD_17@_B<7tIMM-eSy)W^aDjpneUuIxC4$axBjU-IG@)}mG;B4_U%PuJ
z7Ut*3e+P%+A&$*+-tV<no|;bl?toK|+%OKhww-oyw46WSDsCz~o#vszp_rJM%nc(o
zZl&MhQI}hF(xN>RI-AbnD7_=w?WI0luGwCviK*#a!AB*fGi4si4x^5ad9Gum+bfwg
z0iB%kOaZ21jShQ<n&1p&-BpH1ItlfHHXA@*Oij<GjF7%;D|t^|b9qUf38OYDKbN>Y
zIL4TF0)xCnU8Qpsu6&a{EWGIl-|~jhL!Ba-%%OAkPK=FDpoAPE*Gz+c6dWkm#Jk3&
zG(R>vnq#q_KYfPN=;~6)hO@4in!;#Mc|bp@EWueC!j5k~>W+rPvL<laFdiJ<aCYW|
zW{o3_T_<Iq0?#3|e}V%vK-m*HB*A5ubI1{OwdbNUl#lXVKQf^gcvTR+7${0}4QU+?
z4LDBZvC2gGq(WED7fGNZAw9Gxlc<=n8A=2{)A^i~yytBiry7_lA5>-Xm5zAn6)1`<
zU-_G}lvjo$q!RUy=fx@Dp7t<LuO3;NG;W!hkq=8g=9l@a7zN4SoJEfK%o-shx$-#M
zTDWu^v$s@37rM)6J$1`rK3O*l(ZDIrl83`(4GQ_YOxD#Y@`YsrD?g4g>=ITTip++L
zxQY)0CCX##E^PJ;i03RX|M$<@@|`#kH{>nM=>Q0O`%Hs9@4;L340^Ur%h4`BVbh2p
zy^r}Xb%nS#PiasGE9+MAkMD}U@lU?ZJ#b8?Tqa-XrBH!51GmCKzRHBmzg>>O8})lJ
zK#>O!l1`ZK5`N(Iv@JDC9TF#Ad@pYTIw$x?o}Zt~yN*{kK;&qLlMIOEvbyM%lcW8V
z$ELq>{c61X{`+z5#?1_B{OZ%s;{WsS|B81;#Ji9^&h<KtnW@Q`o1dZH(vuTVoAB5M
z@_8qRVcxiLC1z(`eN6du7>^!3!kKtT+D_!_RNTFLH%3M#Vr6+f&!wfMlm)kL-Hbo}
zlRx79?RfrtG5-2*{x0WV8mvh!6Nzp7X)_o|tvl(Qdb^6>XTSMPEUmVXyCcY+nVgg5
z(NFtaot=-}&Fy&h@NsN!Y-OGo@OKY9-P@xab_9eA%#<ET`P$9>o;_PaCK+%dox>Q@
z070I@OQ4)~unmz9$T8bp9eLsc0oXaYVIHzb9aM%6(w4R-4xgi5oU@mtM;<z`8Gy*1
z8|p~j(jv+E#&YDt_BCrIJ(YL-6FZb%t_4fwcc$kAMHZ9KJWb?p$QRhT=Xas~^2};d
ztwed1U%mtgdCX^L;HdOgloYm+WjP!-Cw@a8PVQC*wOON-pV#tNhJTqS&-hrrlpoS5
zK$j_}lRwC(TmcFUI6uy7yXd5S`PmUUI+?apIs)XG{GVl*zI;>U7V=c%E^Rq&7kkXi
zqwr!?AM>$}>G=7yEhEjZk|X9|fWV-kBPtRP+HpU8Z8n{Wa~Mx!S7X086FU%IZ+!6n
z^_ZERjc3nS@SqoCv*{`_C{)ts{tQN$K(qx`f-i-o0A$rET1udB=S%27=v4H2>LGzA
zm+{OAEg(&&p$_v5;02$;1P)uV`zw(JB_xkgAvf63m^Z?y2cHvtU?debs7$4H!aI#T
zq2v_?NiAv+b)9l=U%eFn{Ga_iu3lb1U`)jThCm0(S5Ws?R^tEifBV12-~Rd^;~@f|
zDZ^0g3H5X{J$o@ku00Z-=dKuH{V;I!CXHy=5JvPI8)^@Qy$?D$PTWf&NIh~Ko`Oz6
z3A}beIn2rRu3aj#MAsCeg@yc;O>9HkXk0bwJ*wxR-5+^1Pp{%bBJdlPRY?y~@IIXh
z%Lk4D;I0F|4%g8^D-JQNbSCW4k>HK=I&A0^;I|QIdq)>BcV!~(zI8oD5NgjBpT_;i
zk4b}aK}cP>ay#zcc_+q4=P3(f!*CHd=`f;rY-=40;Xb8&6}(QF%G7(KR4SJ2=1j^`
z_>^>hZ!KIFc4K`V$8kTM0KeJV+6;G3J~+bI%jGR;vwAI8pBbJ&cy=Gh{y`HbbSL_8
z+MK*<+jUTm1m<qsMzb7?KhW2J9*2O_9g&@g+T>bx%9eo?0OH;uI4(J;5ivSC5V!AK
z#R+Ht;t9`OcDIG#b}af-I+h#j+qs$2!M;YOlfX^enY{-Z;GhI!u_K;7md+!7{Nu5B
z{q;-P4$d2J!mKMTIc8mFqD7)6j+OfbI0VmOS>jPz%Spqunf%#oZlxl*ee*_K!vIzg
zxRbBFIqst(edYu~7$%J*-sdi?{B~HL6CR&ESx!0O>P$LSy|w=QMq}S|cw{IIY3Ts6
z<6RvG1{{Jbd*gQSkQ}n2(JT*)O^oBjjHS?Yj|ZJKd-z;EYirZRAX?C|Jm7Ev=SVnQ
z(5Sl0yLRS2I2!0>r4@OwFC8>jozr06SjPZBsp?QjH{$jbW866hgF_BW#DUbQhrl#c
ztaNVRg8?FMq~RjHf=eej>!|6tT4s(F1_y&f)R(p}{Rsxo*ytd29?9We8V?ScvZq6a
zT*8se4Hj{#Hn(szXy2jXPUvMghnb0YBT~w<!TMkz)uMr8GzJcZ*~xzHk}j>4$eyc2
zIgVLA*GcH7P2Bu#ka7o7uFC768ZvNPaBt#8r^}(B)$k)_LOKzo&!{<nY0%1(z*8Iw
zkOZCPe1o=J%D@D-=NJfOh>Xe#)<b@IbvaA<R!LOS@+HVt85%$l-_}>hG7VesUCQ?j
zr_VTTyCMb3;hklPud+viiz<*0PSfdex%$j@wEQ&Itq-SiMVYK~U|sT^&y)e?Y+c)F
ztv|(N-8Gc5SD_7u@;jXd?$ZDkHzY3bkRQ(r{7PZ-6+b1fETes&xUI@-pTa6VDf>!!
zqyZrXC(3`z6UVl@FUoy&dQ|U<Gl9_|Ly|+4CXElDkv&hy(*OBcKF@DFeJBK_cfwOT
z;alF>i<Ld@(7(%O+jA{^zO&6zCvv9I15H?5wr75;zNTJ@sd9+F3Vz*R?F-x0@={-c
zs}ki`&zeX%<$N{xyUWSp>Raba93i9Jz1XF|_u)eqQE~Of<G6nFO5D7CBd*`Lj+`2Z
zFMs!S{M&!~m$AQ(g$rM3bJ(^cBaJlJ-I4q}HxIKX6#EE$qw2kSbu!#P$i)_<<*n__
zc>44SGO91GUbz-;zWFZSPsbwm)9=3gIzIdKH`%r~u3wK2KKMbr_WIo%{`nvN_rK14
ztE543DJ@B}&f`+{PESu~kjIf34#)bNzx`z_<4lZ=55(-;a5!hp4KK&g3DaW}I1|ly
z`shilB5Uli?}Wcw=$hSa19i^%>WYzZ_b+PXu*EV&64)QRN*^eov9--?73Cr*FSM$h
zknSV~9?HiO232Y&NK#3?A{%>1OXk%UD<v<|L)!VBq%4U^7c_Go`904R4?dS1d?j_J
zEy$H@%E3)3uh^ER&xe`7^u-ZQBi(nvX?~f%tKGHF;*k3s#{fN>8^bAc0JxejAx|dn
zS|(|8wWu`c<wZFa-c4t!9J#lLP|FX(!AHshc}m`pCme=vu+39Gu_rS1UBX@CTZ2L3
zN_;CDO5nkq&ClL)>r)JW!~52w1m<;&{Z<>wz_UI7%0uhmNF?VWn@`F#6H){7PRD^v
zQ$}zKueL<;9Q6hY^mhpY_)QLL{p5|g3mM$*LL0)N9!CdzI1_u+sVCliV<E0xyB;fR
z8*v{`cYSL&orw!)UvQ?^YDGOoB>-8F1z5q-@Z4KTPMxC~2y~@|YUL({u@w=qMj+jA
z#x3F<-)Iz6X2_lUP;{2#Doi;BneSYdQF`u3;bv4a0ZA`hq#m^}TGZAcOWWLF=KR!9
z{P5ko@h6}BFqcxCo0*QmiK)2g9mvShHud~p|J}cfzy18PxQBt*L=T?hd0qftf4!c}
zt!Xcz<va-#zCvH)LW3&>pEUyxhu;_(;Dc=R6s{Pl5^mKikVZ81(2=l5!ez_SnZRDo
zwYMS8f=hF?fh7{Tq9YFZR>CNXAs;2~6<C&^gbXxtv8XX3W)g|TLxn0qbdr8VVdPSf
z{C0UZ+sk<^c3HajMK1=NOQ7wzizp6Y|7dSaUz(1qH?Kwmp|rTV7`q1;q&QqI0XkHl
zh#OZCrlU78ck7809af{tl~-I5y9_Iis3oV$!=W3yIE!|hs`PnHLhGd=S|VKCiC<Th
za@8R3rKOFHwH$|faJ)};r%^)@G=>Lac)St44GhIV5uEP>hdmJ}F7i!gO56Y&xY0A0
zab3-6WOzK6u<YxtCp->~GSZU5#&+R9C&I36EMpvQrZBp8Z3Y8b0~#X_1~%eud)RYk
zSGtjHH^y@;x@{vK953n+CXJh34F2PjuK4b|omgGfnHY$Ve|RUx#)iquNVglvVWgut
zTt#DZYbVFN@3-vH!8ks2nB7S+<S}d&mVgT^IM-+bH)Nci*0DLm(8j5OPgL$s1}|%^
zEBK6%B;mnpM>#*`;lpR~=+QDpbsy<<oO_`Sm&z>@x@{jPd^<Nv$H(5fv9XbaLzBT?
zMQK)fgUJU>T``Wb9>P#oF*#X#AEoCcU-9icl}C>rgM$@t(i21E)8mFv?YRyP&&=pQ
zI8heJ1UhDYIDr%eOsBECbC|m~udi={6UXGz5RjQ02u@CqCN0^^>E}75fV#^Qt;0R?
zIO2CV&H{Yi>qK~sa~ZQ3#qxZv{za}iY>`vNGCn?>0g-RN`8L}|gUjI#4g)O1_@Dut
zB!?H#eB|4i28n5OWK>=b`*M$r6WX-5yOv8nI{!mj(Gh7OtDJzY<J8%S1IgLlJpi9L
z(Ox5r2FKBGnN%l~XVeuqON6@2pAu%_QGO^BlocsL+VpR|Y&ZF{1d9qEsxPJI(krcO
zDed69l=&)e_&b#n{2*PV5#slmmp#?+5APke;c!8R>RE!%lv>Ir`BEL>o7^>4$3Xgl
zN74YNT%C!Wo0hz3Z-M~_Wv0svJI1{VN{|)<E$b`enkz7;!JheX<;;q4B>k(DsSDtW
zl6)up)=^pRVX#QUGv#pB(45V?)ThW`zc2V`KPpfCml=CjhhIab;@Q(*0+4S!N`9G5
zW|9B3-<6}fuZ;5l+Tr_5lbO9tpZdcx+c0F?w_#Av+P)ys54bll<d8a@c$aR6zIgAL
ze2pWAv)a395JSDM@l_coe&-yNGMtHa=hR>kx7Bauk3|bp;-pPK`J0m;fG<5^9Hra>
zNZ|Eqb;?jGdx+O*E96M;am-(yim{2oxO`<UW@e{j4cYL&{y+XQ?%jI?Pw1Ew!?eF2
znKNYXp5+$0QzjVja0l#}-0*HBCZ|Tj<&~9D&E_`zT+am*u3f!>{us+(#;2ct9>4z8
zuM#iwb8~S6N9NL{1$0e!{2%|v-^QoE{vwBh%`MDEV{kBgAPh!G-{a#GacSWSa&#zn
z$o|X!*Qc@FJfhsrn4WXAgo`2c#0(CHLtIY|j$><m8~bUKd-OYcZDaK)R#xN{14K^b
z?g0LQn3|fRjT*7Nu@m>dS&AnQp=WGFn>Y|5=im_(C(mOFRHRZ){`0p3%M3b-bJ+*6
z4F1Yh`KH~6s4waH?HrU`U<TQM$OF&za2doYH0BU*<-Ds)dJoF@P#Z_P0lG?uRy#;2
z3PkFxidU5b+E&u4!%y9zT>2~U76TZw8Tv7(pbb@m-6SU;1KrqQ*gMjcZLW++XM(Ht
zi1!{YP?7;`>LGtPf+FQN?^8bMSCBMi0r$#RdwzwpIK*73pL6$8Z}!(nZjcQa;V%P)
zE+mw)6DQX?sXKBgsO2Y}7Mmr9mI12>aE^+2s>+dn4{@vw!e`zb_9kk>C}WE}Ej(WA
z@zPexKOK%7G1CqfsMqCBhq;>2HgFhZ$B*8|nV?|17wxjqz&|}Y$noMvlHYj!3f}j%
zSk;+$@Hp48E&EJBWi+TBt?h3E7QF&ey;?|~95TF(rEun_<rvv??u3=rTRKBQ2%rsb
zL)wFWh$1jL>2Qt291UTPPW;aK3N)(SEID>dYU6nd7t%Up$_S!!BTkQ8nddN056`0x
z;a;5`ig)kcig#XnEsYa*{kAtmr)6nz1q1GW{NnS^;}ML!hT*qMhNp5vPUNoTP$!Y<
zlAvxFHjKizVH|VlICzKN$S`(Knx{jxRDebrt)C5?D?UMR5~jw8#*2-l6JZZPpTnHA
zS#oTsiddN{rzVp5t)#^9gK0#v4Jsp>PQppT00v)r%?N@Cxi6k3#L3T!kh-`|kBmeE
zqPaN5d4n*gCdP2Irf}-EF<|%k6z}#7qjGyaPLUE<ZeNRjjP%Z4h`Arw8}3v(jPgFj
z=zL0P12I2yH^zppBjkrU#a|KjG<$)WcB-5t;mS6qjw*F0No$1X7Vp-9wv=Hs8cGzQ
z7~u96=uEgCyu)ZVH#QK$?$*8sy3S%?urCIHt=11ScD3THvqhW96d0R_j<W?8=t;v`
zq3ci<6;SSgOxqZFpPrgYfu0JHy1Fuy?Id8fojb3)6~FtD98#_u<jtcL?2WJvxm&am
zb_7a}DMl!qYdqO52np%IsJv~uhhyUKmzC9X3~7faMSSvc#7}<w8cx7C`Mbp8UOE#C
z3zxH-UW0Wn_k3`5kkfenY&pljYRJzoOhb3BJMWUJ_IkOPL4Wdx<8&WAe3Z_*yW2WO
z*hq;x;VKtg*~zHJ(#mo?efBIJ3Rmwi(r!DsE3T)UEuGrSC@+X79R&rE_hqOJI9A9C
zC3zdkaaAQJA=)EiTQ+x$Agb_AQQ~9S9*c|4FIkMC#sGXy9?++~5;{pac^bSsOeuSk
zaPHk>!=-E2R=2>t!}My%vldR~7C701juiMuaTEG=B^1A3=bc8kbrNp*Lj|F*R~{)W
zH5d(I$iKOg4F<l-O@VAMKq2gRu6p9?No#n4uHLk`xK3R=p@~}T;6xrCIZ@SKw!)Im
ztr;1q#n|{HdD^?d^9$!!4aF#qY^`%3qy8z_fkk6Yr2;a5)j48Di&U=MccF}MB;2RK
z{B<JC&t8a}bcZ8ZqQev{&08m6S4ny($32ACceYt|dZd8|pZe1cx*VRPbCR-zhXx2V
z62(zTr-E{&z@pTYKbB=$^U|4dv&H>Acnu{fzojw;rZSFRqoCZEm0955w!(yH74<2@
z{H9FeExnuaf#v1*>I9CInIhb`${cR}l_x)xKf)?c3!m@(LOl82?>x&pWvJTB9%jn$
z8}5~3<dpA~SLKk_%b9s_$#m^!_Q?6$e;&n&;6{1QpLnVc%P#Lr&rbe+Q9tvr4lmow
z!}q1^@;(3M{fqMb#*(;NR?!VS<2z{}ovn5mFJR%l(D|JQY?CtD{Bt-*8#h&+(q5h@
zwGybXoZ?VO(_ZuB-hf69o8ouQ0YhgRxKM_ad`lgj<5TpUMx*73dFxi%P(GG-#gmI+
zoSiz^dxLNGQvUbhz>iN3VT@0ZwmW|Fn=j(;|Na-TzHS;v8g!<t*O}0;%?-p9;N*RL
zc!rEU0Zxr_S6OpD(|laMJfCnFG%*-3IWdJEohGl|To(G%U;i3-x^p74OI|x~4QnO-
z?mzuweEH={8qGIu&c)Sh*TX%w?&CP9FGjF4CJcC<cf>#b<FDhFzuZAipX0~`a-%zH
z*bS-2!M#C@G(3?R$_><ZE(^Q5><-qn7)@v=slB?nB<;;xz%h#R(Tv~y_F;Va*$RB5
zoyO_3i8QUC4^V4wUy_gQ%~i6{uI}mUg8!g19E!9tMHA@4E-(2=Hc20p-PaLgquta=
z`3>yY5868pH+IrCRY*F~9x~vNbnYMDt83JIuKZ`uclLNG*V=R`ts2l(ujQg5JU0O2
zFkBbQkZc{E<3xXL8drUEBn}@^mVg)fQ9SBYSav0^wAYlCE&!B^=U~%mP)olW+tRCx
zCFuOAGfVK}G>3Q_Ks3nVAx!4!sUzYlsP=?7jNefJIv3(eN5$Zv!<D*d2=$HgSaQQ+
z8c{xyZgWv1cti{ccfpnIROD3hvQ_gXWm~>IjKW>X>55PNZ9u>es^8k>rZv!2Ej&?p
zk-YM|lBVO2-<?Z&pm$~?I{e5<(XG9G5Zn=W?_7@S*RSJDY(m&ia+8b;l$Se8UqEHR
z{IVO&Kk~9+3xH%zh4_Udfk5HR?a^&G3UhTB8D2$Ubl&F@y2F9YRD|R(m^S5Fh6!v$
z$It6fxS~=%=DF9Z8>)Cz2q$5oOgsv_<BLEx!~}-kb&RH4moLQ>j=%(`={MJ@<Nb&6
z%{TYr(aL)4BS7{MC{4Ix7Xs_1vF+kb!AV)dVqJScvU@K`04|5QhruG_8HKjH)X7&W
zL>a@i-81q+gXyhGU=9tD@KX4}RQ&^G!xh?#awI^BLjJ7^mTb7^YInLXbA>92g#;3E
zDp0#5`O5|_03!<vk-c;#N;Iw;Avuf?8pinTfELbA4r5|`I6nB`9h{;XhN_#Stj5yX
zQmkyQM-R_#zJ42=V9cEz#(CFyjF5hIeg-3WEtkma>}bT*OK->g^cykQP@qDj969Nd
zQydAV?Ib+H=Wm6Ok;J2;y)?vaZ;fu1s1v%XZaS+Zv$_;!?hnDe%eOiaZF6fQwwhZw
zn}<31x{mSL(T#EBiXi<b(O(x=3Rj&8hw8u(v@Ps|!l6u*#Yt``MVTjh`7Yy%@SPfr
z$*GZW$<=|nf=-y6D{)5KDO8GqX&qIHu0!rn7@Y}+>*UZN4QlGCBPHER7lUXdH%c=?
z{^;>z9NI-3iTzkv(O8Li=k17(KD>ePJI(jjle%Fz%`aR+7^BgETc_G#cPr4Y!>F7e
z;XVg9Zd}Ew7@{6H2(Xrz7T#(fh?V74@Oz%ZHkI7Sa83xGo|#OA_~`M|RBBEl7IwSg
zT^`q*(m<pk9Ll3%Xs?NUV0&iN1R-g}MWG~ro5o>cF0-h77($__G=#^9i$fNV5WIW|
zPc?wqomX3#*8<MjImF?b>N*rELnlB>-UcLcB^hw)Y442P`niiX?dN#q#YJ&ajAqMz
z``v}Oc4Ic@pxl4(6r5_jQY#RI?99A1YBXr%8{ya^PorsWr01d#4tvTC>A-`^#Jw`a
zqlZhyI{!-JVq(%|WJluhlgIJo$@AFWalRFJ;@#9#UrbHvc)B*ej@1xXC%g}!IQruJ
zxFe2nu5!K;xK-wecY9IF_+@1uKqozcXXRqbPm+^fJ}JrvT&dAH(`e?6X;o^cX*5cg
zrWOY}^}<(0w~9>nkn?hEKT6K@MgcYC<;luWEc8`qLiy&GDZe<Bg{4R21jn>&hUnZj
zC!^<L5Wr&^9RX=3VWz#5RmW%cQk$1h2y03=cu9EqmNdDRy#-})lnUI|KYK~C9?(ep
zsFbexqy!>uic#+^hj*#Ncn-Em4~{6O$RF@Px=I#RXLWr4Sv%9;KY0A=H@|-*&f4$`
zyE>-?T}qCny*M*Xd0sg&{uVZIS>Cjtm*knJS7}KsmR_cE6nRWS^0z!`K>Q`GaHbL8
z&ZC`q9wnc4z#yN9&#XlrzAL8&R4?)q`hsTEzh#29x>0?ggUYvUW2QJ8%Gc6!>X5>W
zlp<rZH-tu$R|oj5p<_=|(j;&egG)X^mh2kT*^HUlQ4DaP>pCHiPV_@x>J0awSYKZy
zjYCNfNZ|Y|@#Z2D_U?i+ejT*VVy`KmiK{?ncE+X4gBa$QVg|XX&T<H@Hh>Pr(9kgQ
zWj8+m&2RYTv_#6$70Pm*@_XMsjc>ojp~h6XdTl)3dgmti>5nhI{2g-5fR;VMwG8Au
zdbAd+D-NxUz)p?<^insn(5QIkX|M{z-Yx|LcrI<ey|EW7OXtWhfTsd|I1~CY{q^qn
z)1Uo3rl;m&b#XmD{l_ojH@|*B-AYdaO_KHV-lmovB{}KF_V?bnhIsIF?y9z3?x~T(
z&nb`Rh@;#RpbpMQN4mh(Xw1z`#~}8G%Ziq{2YieDsa;rB^@OizKLZN}n$&l;wY}WK
z$kL&)VGM3{jEiOTp)X3$gTr;7K7XDoRB3#>GL`ek)Me6|c7j1imoDDgbX3D$^zvaJ
zdPrJTcFF6qTg7v-l=8}9xrJ{yd6%^7DSyZ_i6dmGawugOC+%UG@{V+>Eg?-hWKCMm
zAwuMpPBnFr53GpukGi^h$x~i1f4=E~{&im99cjxE3<m352{fH`&SKkv2%EYMXYwgL
z*1`MSAC59iVGx!7v%IWu`<wF2Z`$AUZP728;$<3Tgm=8?_~ZBH$^>i((P1r4L_J4Q
zw|1R`8aNa4I1@K=2!#rFbz=t_)u4u9%{&E>^iqtyDhedhQh1ZO^kfmv951^RI|o7n
zEx?I*BhD%k8_npJOS?JLNkhv56q+gY$<Iiv8Dx2!`M6Xfg`D^ip&13C0G$Xsk0wVU
zD2TO&lA40BF5|6Uou7|scY^}#wbk`__z*)5hvMnRdbBVmJTYCOd43Yzb;oFH5SOxb
zWUQ}c%k09TL8PM9P;}hAJuBI};xlliur#V$BD*MJjIiQNAhd)HhSHHRBBH0<R#%Cq
zY)8r{Lf`sQ9^cuJ)>oXX*3vVUvq>jx$q;;<iUq><YT8UM?rG>wt_=uz2!^~sVH>e`
z#jV@7F2jU97;CL)V#IAUcVlbkIO@ZlF?V$?hH)V5!$f>V59w#;r{gik=aVPPd^ZuV
z-Tq<Ry76(04bNxYZ6Bl84%3n8gfEwfg%&hubBF>$NXE5HqmiY{(st%s9Nir6timEM
zX-2PK9BX*G)SW|5oFuolv4-(+h|v!n0oz3<3ZeHn8Y3Mj52Jcz#V>bY&W#V#DDMnc
z5NmEbxjQf*5t8r$jz7jAj_5>pj88hb)}5(cCCt?|9NJQr-gMsqS2psuYk0p9CR`PY
z3Y4s;LW#EYkj9<7Yh>D9ou%dFSXx?*?WTsg4k%6@_#PhW$H8-`%Md(zRt!}I3~=kr
z%8M?C>Tn*Tymq(i+)PZmys9gRVe~<3;@w^mPza55rjQm!VO5~b;%HyLag92<Y~f}s
zE;*rnH)+itIVTZ2epgzo1gzhhk3A+i;TRs5&*c*h?pz`j`T_`WYR`*v7>-bg^@On)
ztd38ORG6;yE<edvLnwlU`MG%Gjk`EX&hJ=8!R}GKLs@W8jvXol9Z&}c#W<YDoq^rM
zVSQ~Y9f|cdqfoYmGJ<-;61VS8#ogC#q+<E%tM72Y%3caIPu5MvQt_k_v+OZBg1?{#
zd8&WFMIY>?a@j}M#drB!=W29ho=)2K-nzU0g$r4Tn{6Ck9aV!eM%qd@{3tSSkbH6G
z!=4^{l29z9Im0lpy)^d4Ny`7^XYWgWlFz}RJg*EWDaw6`-tvt42#@^lcUei&_)U?U
z<+;*DnPMGP-Ud)SH0<R)WtGZNnXHoErLJxon6i*`DpC2n%ojtKP-kVl`ap%~;q!Df
zl1|K%@)cPLby4WjlzPETQ(hJ8o}Wb)q+w4w?=9DZr|o(c+*psyz^m2u;6Bk&a%$rO
zoxw*%5f^7jJ{8GoIE@|g=4n2Dr-77TT10ivktLExQNrd&NtVwkoty5}wEouyfy12J
zcRZ}xr7G7v@*9rwyrg@1__m$?m81LyPwmI15+;8pXsJ&G#JA;HruRO71H1U4K8Y{C
z$qx%&lxz7Zzw=wK?eF+s05WgWR;T&MW9mim^y<}bGM$RqEcnlNI=7|A!X?)0@LxTh
z2;XU(<svBdx~4-?bYD6n8fY}S3w5;io5<QhWY}mflF`!(3|-3PBjg-<-(lXn7@EyB
zXxH9N9FT(p9F+@rSY6eLerUzkI`Yr>1*o%!Vr*O^8HeJ=rCc$uM8^+-p&dlT8E`3X
z>!l30*IH*<d4K=DyUT0W1bK|ao!2gtZ%=&o+1JPmhn*A}l23F#)Jv)>=UNRkbi~lX
z7iFyb;lUYp7RWv?BTbz9y0_yp>gT{@;H5Yb*KZCWM-Sp>Kl^8K`O+1fjrI7&fB7^%
z|MflWKMa5DSZsCLNd7C$D9u-G;Z+Kem((B1m~1MVOt$s6y%cS>&3w#rda?_fawKM^
zCV&qo&k5QjaC5sHVnka>7a%Hm+(2svs?ZA=VAeKtWWeA+^f%PQ^_(-}=A*t-_N?XT
zgl+i8Jw1@k4d<J=oOUhcN^UQpT>;%XSJGkL&Pz)=X){U6D`>~MTC*a5tw@m(#nHtv
zpbqM=>?Pruz44XPoU&J?W)7gN&|-0vMV8J)RYtYY0WUetpXb=X?qS9cY<X~nb0IxR
z*Yb&{0V#W%6DIEc{MExL2H8mDT}JS<%kapmT3+>$59B|9%}4F$)qC?12OjN99z5&#
z(Yv!3h3RZ~$B2_2<7gLVIXS$DJGZCf)~(yQdcpmNIulKtiCQW`$H^H*;A82Ba%4z^
zFbI?wM2tfza)O93&~S(QEr{bdi9`UB0HsSHbGvrOt{JJe`>%np$R$?On6!i<{IWNK
zy5;WG<f8%sh19naKpM*`1`tj9eh9Q1*k&h2qhaHtz{W9Brs|EjzOWGU)3Z1_I1_6d
z@$A`RJbL^r9&fD1F5dnDMtu{5WgWxz0)tdz%SILs302%$Zi#kLzP7)3Aq@~9y!P`{
z0Slv%#Ehyq;g-XRoH`RKFNs&fCA<A7%dvu#lBZ>W#cW4U%kZ>IUjnKGr*)TE(wQ&?
z$n`X>#z6Jl=&JOgVQG7GoE_)>AN@E67d&(23cDjGry2-CS1iKmL2~9YxCn|~gprQK
zI0B&$hKA+y9G@JG2M-@07`I|%csAaA<B#I*tslj3V-EPh2{39z9v-?W&~X|oxxWK#
zl}5<HF~YE{a^*&13I@miSK|~<fJ^4oaIx)umIBsq5VE<~1m`-!WPNiTY(S;p(cXiT
zivxtvMNH2P0JqDK?&n0-oTH$Ef+ob3znzG;yX#OV>C`Ar2aW_r;aD$*V;6$S`30q&
zG<Y;3HM(4e)Gk%m@pj2QS1!`<vdz*!(n(T~+2*u|WhoF`tz)1u2we;$A6SQ7j5Gz7
z#(2^<?PdgGWXQ;MABOdEj-8hr`rsAwaT2z3D0F7aQn2po41Iw2<S}?Dc@ae;-gMGj
z!N*J-p69ASE&;f(Fdh?AW8k5MGIi*P0;`OjR1xYh*+XQ{kw$?+RL4T&Qb8>ZRy{J-
zM|iErFoM6o&xo9ph@CWjl>9I3Ib;gqX)n~{M^C9+3nyYG-hS(~xO4Y*Ov1l95)U8V
zM=5Uc9XvzZN|!oP_Eb_O%BIz!QM=uj7FXlZ10#)i0M?rmFjH3ExjP?kzw>$;{;m(d
zwdpVj9WK(KaH&Eb#t(Xt1@e%R#(bqapHnPt5-~nDoZZy6sq<c3vf9n!9OImdY(R&q
z0dG6u8>N8^b%%ZxgR7i5x1pnh40C>ij&6A-ULc%OHn1ymtfz34h*#kYUS%3-q>_DJ
zt#ryrc|~DmKBh~Vr6E_z2kNVQXrEaahiT+J9RqvTT;|%SYj)uy3p7S_7F6~b6yn>Q
z@XUMfg-2SjmqFQVe&(F;D5tolql2u>o;u*M&Ze?frb`9Rz12y*;TTw+`R8*3A|-EL
zWqtg*eCtt_X*T5x4IpJVsq$HY%hi~&w})#<U-X#c@GIj~I@ciZOy$jc4H$c-5_W%^
z54_^Agkm809N;a3)4X3jc=X~B9+4(Y)+u#N;=nd6&+?vYS*${OkRFTPO{0V}gDu&X
z?Qcz+X-YbA?z0Rm@Z0Bn^fLeaHZ8)$dA0#(mA60^T>IF-n$N%^&rE71#YNFSrGos*
z|9$5(VSDwcdPWM}W-?C7<}63`t@yO9ier&_-u9pkYzqxLBmX%m8RmT9SO@`ibdip^
zx!IVSnMmg&mpg~e)N!uHbD9&acVlmVJN9rUcABfP*>tGlLDc#R&YH+3$xvs;Z(A)F
zYdDFC$)31-XEtu$ypcWRt^iog-x0+wp(=h{oWb4mb2#MMT6~Jr@btMu_*_QZ%}vLV
z74_V;e`VRds>Hp$_yu=v09wNd8lN!g-kW;y$RIFq?C%+rOA&`&Gca32rm0g-k((ER
ze}fY>5I^|fQtFG#m#@U!%zW(cwBom)eHCAP_Ao#}+D_VoIsw`IT&-&^kN_@0W!6b@
zNXK4sF`>=HX=|F&vk$mIb&L&n#ll=YE?=2U-0=mprw*j7ntWtox+rb>PZ}+9C3|K$
zY)dzdbRLciXf%d&C>?qZztR5Ysf@}X7rLZY!@)!b_Ze`b9UPje46rv$y3szhHzl1=
z5G>tV>9RUo+xSk1ya1@cawflf@@?UHf6M=+4)T$(c;+Y_o@aW$=b81Bb_*Z1`J~M&
zmX}UMKC?XUwY7!CQ#`j%ao#Q`53}%_ymD3zRXu1Y%Olgh`rQA7&u1L`6jyoceI*NC
z&4k;`;fL?eT<BftkYRHN_9V{4LF{bp$8qZ%#+ZuRx9?_T(qT@^>v(%q>H=km&Q51S
zgGYoY;2htc-KqeXQ3M&D25~~Gdl@*<(+F(>#c4f^C<u`ZAz>*D#zw|y+<J*5I7|ql
z$}W7$C{09uAV?u*qgkG3R+K!e#!)Q{IM;zRM&A3Wqh|aJihUfz@7mm4T%Es^8@xCk
z-1YyTJX?xK>nqX3us&+kVz=u&){pk$2!c?FNPuMCE(;=}fkmzG75eF*&JZ^B-`{QJ
zu%}b%E?^R;y&I10R(NJ~4g<`_NG5@yYQ0A5+s=f`-dS79gADnGUwQh4Zz{*B5VG$o
zgTQ;s^c2>7#+gw*(hIC>dmCA|8>%1V^w(^+Lj1MbSq$*ZAsvndw4HNU1!v-&d>RPD
z;nBh9>2vtm0q+B$_$)Ryc4KUGKHhx&lem53N6}ZCptuqxvIS2lQv*ZC@#W@qIGe2_
zXatzD$EHNM9YUkwBn=pS$y&GTqjRLNcC2-4&mlBMew#TY#~oMY6_vzpYZqbLilNb3
zT)R0Q)3eU)*v80greSZJm5D~K*fWrVS|ytb4?NXs&`*5;<Iu6TDlNyb^BifLQ3(vg
z(ktagVPkRo)}1uoT#ZQhv(3oQ_I4iyBT_1#<Q4GfYDm(R8-=;*PwuWxJuKgz3x%UX
zPshtKqvFhAO!mOoW6%Q(8etmpcIhfVa-PN>&ILGB0g5XPb!k&O!3p5ID3HpKbByWM
zw!KpBK-(45(>ON>>|$UQx>xw+cuN$NW$E-uf3~Z?OK`y7J=E0=_}mO_7ZP;;lK$LK
zDtlLk8i@;6KT#$)kHC2Xc5yrD(S02@HV#nAm*eA)-$f4HLRoaj=Ehol{q>hPBHP*H
z<mPR%f&m75V{(NxoJuFQJ7h{_{^ZFPWgG#!!dz!Sxz!c#zV~`uyM7hEUyrYT_ceUl
z0$&cdvByxlk<{#d2k*d?QL~&k1#f}}4G*J_j`f|wsn!v)PVRprpF6LkULQn`xvG*I
zT{;P{ec$b*X(L$0$SA#Wwogh4{!|$9d*S_*H`K8*fPJSinTin@m4}IYzm`uWe2!u$
zs+8?jAb-A9PT6MCUCFx`{5tMB2l9i?fc)ii+ecd4X}T=3vbZSsG*F?boZQUs+yz$o
zpNb2rYRe_dFhBE3=OO>H*mCgRzhqWU+o;@we)$Ie^J+e&TH*u@kj~TEo7OgQOs|0)
z)2D&Vix&#Jop0jGH$F?bX<L#;`quDF<Fnw%GskX2?+zizAqMbT4oT1$X_uKpYUL5q
zTfX<jh~*h&QgW7=L!`J0VE*L~kC*r5`HL5M|KeTV7dj}q57Et4+}gh4$Z|!HPL}PV
zQI*a?@;|3E(f*y^A&1oi)g<%`cw~^v9`z!lJWKmYr!Vh|jBL|D8y@M<YcggVW~!91
z=~Pu`c~{=21C!sW%L~oe#zsEW7fBP)yw9sLD%$pYWt1ts&o-lta)l|J3ERqc%E`VO
zIvlRF<gWV$fixf;CZmk3Y4v8L9C?K9Jvz3hbvt(V?XlQ`w_W67C*^lv?Ko~+U!c76
zc=miF)>fPFd{<n*;nLf;lV_BF>LpMOk2{9mp`P|kG)QwZ-hA_o#P{F+?cc=L-@4iA
zarPQa;3NzUkH^ZY0Zvzsve%&lgS_y{e#^bLdgIcSS+4ciYHr3hGO^ioUn}r%F0P?{
z(55W$w}GM-Hl0JM#wJGMPk!>g%@TV!IOchP{CjeLF&^LF$RSDRq|;2$eiiQ;>SfY&
z+gVL|qRq8=xSkwQt}I8Uwyt|5jg9uk_}B>c+(1l@_r>&7BW9*Ya+6Qt%Y7)a{oos9
zvaKv_7Tzv0!2il``B?je8qtpOrIXI(OC$4D9>PcFeCPK*>xCY08oPTTN%Xq%FXtcd
z9lt%RVMk>pGQRb+;+|QXm-*#s{yutfh)EGq>;)=Ou5D}e1yb_9x^hCpNzgC&f%`nA
zf66A0vJsqcv|-_0ri1qFb@1L(aCn+<S<eD1`S>v3mv8d)-ZVbVtB?G~oATQtOjg;N
znbxwn^Z)9zl8Cn-zB`LEA<@~*)DZ(TM)pi>?d7=f*Y3{69h`~Ht!8|4?@=tRZpK;Y
zO!Q<FP^FoT0<&6}6QWNMG#(0!97{}^8mC>`8Y~iFjzfhg?4oq#0u8>y;}*naw7Mro
zhDH$Fy*arymxi<~^3p)FM}i+kkSWm2>qW(YqC%4CNK?l^LEpzIin0S3hJg3-$Z*UJ
zkH(Gpg}A<O8E2ytVX&5yJs&@L77sVpV;5#P!RtT7P~OE=-#c+!Fga7UMQ}?g5;jVM
zJVm-aD8g#A*ZCgq6zRN)WV(b7&|w5udoZ2GugW`>k_27#OuTR=6l^w}CB5oQv@;{u
zO6iq)7v-XWFKNVsXi)G<xER5y;B`6_ZWs<+JG)M4hb{zI8M@Pr0~HreiVqln6r*S$
zdsO=RFob#sVmpV^Y(Re(2&Ul}9UDYQ?PJKV<+}e(=x1bPHeSE`N!+~pL)c&h!CWRB
zDb!!qD^mGcSB-6lEV%+tIumWfMk8HkLipVX%Y*9XZw*#yN@Gmp*zb+jJ8xqPqh0u1
z{i(UPokOHX5okIS6O(m*?^E}KaAQNuv6sOmJzWZLY}`mQu=2eQZdpCSJt<UFmXE;W
zox`IXCvUk9QL;@PAM3<u@13Y+8^|NJkCENw<;9Fzsi<u~>n{)JG`M-2^J$tmpqrc9
z)KB9WCkY;qA4>OlSz<JM0#96psJ(Gv8bFTEpCZ?e?AZ`EI<oc}I1kEUK90LQY@Ni)
z(nhQ-Z}H5?5IhW0kB@Pn^{vQzpzdB4y(L<cMpvO56=FJ0&=|#l2PdgnSDZ=d7@_QN
zC>6a+C~j>QWoolS;xRNh0G{D*Xxup#?yb?>wpRmO@ZJT*tcS2pPmjl~TbJYN)dlFO
z0go=n!$<d#F^AN1G-hX~p~s%w{oAp{&N*@X{lS6zuju^Xn5?Yx-8L|mNtwd$*y|sB
z@L_agbUb_dES^4DgzipauW6KB{z(+44AWS2;=02d+(W}1Z(UUjwG>aDKH+_D%+1Z=
zY&KFBc{qf@NU56&c6HhFhC|-zFghI($Mzyakr&C;;7?lC=~PY_c(GjZr!37Ng_Od7
znLY0^u)=N5X@G9>cPa|g02%q@*lE6ZKNby59Y3$6Fcbclw$D=b$|Jk`yXky-uOaET
zp7u)Edt*>XzN8uRP_}EpOB3d$L0JMs=o4j33gwbIomt8*%BcK_tHeMmY(R&fGPA6o
zim>Eiei?P9o*I3{u+VrY#=K=2w6fREh_tJ28Q>9FelLINY?mQc*4wP%NgV=CjScV9
zSTR5H)u>RXX=n(uw4Y`1PC7_ul4og*ScdO_w=zyj8zy4FiFj5oAcNtnJpE0+UMqt*
z^_6F$tl-TI#F6+Y{0Ck*gdaaLe~TY+sV?;Ml&2|EOIn?Tgrl98hp5h0NZ4>J&;bf8
z`Cg+}Ipy#{<%^4>=x``QY=46W(n#evXHOiwU+4zo(E31!)qFtdE*DYZ0NTW}24U)}
zHXH(8oW?>w;x69{F2Lq9@y(eJ`6Tg{xbquyqmRwUXZ)$BOPeVNY#W~$OzB3hEMVAb
zhz~(G133gyr-Rl2e`OWAK6oAM@Qu#QDF*x|I%Ng9x`r-uY4OdtecOG5PU6X<C-Ho7
ziEqzie7q+vU7Ab0x{A~KnhS~O?2;519cnZ*)En=-b3H!#=%bY9zx;<^#^;}Z8}5Q{
z;BIWJp8M}O)b-i3P2}TYww;@D8n9|MPmn{6xPwE|Xbi-&#b>d&vP3zywJBAQ+9U8K
za@6?~M)qw6er2!oCqMcmckq7n-~s2uSbDk?8_PS`e&C4D9j#$nM>*OH@3PePO#4*l
z()q7cGJAX6%i}nmn9^k0FEi7Pc<=p%c;oHYVtQ^Gyo~2?PkGdQQYQGlyeQ8})8I^6
zNS;mJln-t9^8O{CP>?)C`<M?9kxsc&(j{M!%(gE^IZ%1!H<?p8Wr0;mfyoW;+wvg~
z8%iAJ?~3NC-&GpFYoF!=3;;!T=K1Pr8m`(d)ji*2I`C#aKmyN9PK}V=W+|=9vBF{<
zvL4cTW}_Svru-~rig?WIImHZ#^UrUDIOS2BU-F%wlPqy+deVvK#4m^VdwH1NTwWgj
zEP0qG_aD4Bf8oR;SMPC{%ODKo1Y9?#I%*x~YEC*6YcR!E-`tPq%j-GJDL9X<G-jd=
zmfb-d5<^i~7GIj78zeaY!6j0i+?t*zfE!(vz=np0F;2H&(lRVV038#lG42kIMrAW6
zz9}npcq?M9uoVIRc14RmDo9LW1ECp>a6f^Q1B}Mq!{|i{UmEU@*Dha<*REZQO9+*7
zgrI9OFD<XccQDpNjNyZRNOTaTKiD4^{k7QIHzHA#sS}=(_=N`|b7+W;PR{#~Hf$Jh
zqa4qB34t%2BqfnJaZIDeXrIbj=5hRB?tl%fMmt}1CQ^}DvP{8|#FTaSeQD_G^m*k>
zNN{=OKN(CC6mH4QF`^D<@^@doGX}@HVr+6CMn;XGYM`9q)YTDMW87<mSy$V;pSE%$
z*ki8Ej1Estjl}7>W8K%1xs7-?>XUK%#z#04pFjg6D4S!-)%jAvm&iiyj+}KXb#b?8
zg<r0YbEGrjm}-rpqJR^P94bo7(Gk$$RB<`fNm=Cf{_Y}eWZt!vTZc}ZwP)fedT7&;
z(XJRC?ac^KW6=E+T+;M79DBL2Fc*`PDi#Hs5!x~Y4Ybl0I;}cHWurTHbMM74H`;G+
zh;-^OCntQ_KI?0n{4O-E5Os2#%K7o5M<|(f>Qt6SRhj63I04&<&34Uh!hbsOt#p=-
zQ(@~&Xaq|SMzj1@VSA!e$O#4!z`ZzjW4SVuGD?MSlrdMj!a)(gU7fwLvD%DpzJ3x9
z@2}uoApHA`<khp+LuG2uQFF&tyc}C9?MTP89&9L08{PLFIS;&+14VI7((>SHdTGz}
ziU-Ao@8u6ggyW3S7kMv>n-*P=xHlUox0{EnsZ35bGT3o&ut!-(I4m8=jk28H0FExT
zLgCdhZcf47E@@7{`54ObS}dXD?|t_m+&Nz*Yn_x^*Kgbaziu`NZJ>ltj?Qz{r(Bf;
zLMzcz3`gWgZYl)LDR!nRCkFhlEH44C`-jZoAWZ;+n^L;Sf-=!4XE*s7jnrVvT2WDn
zSYxUPc`NRH8s^j_BgW9bJe$T7v;@j{171?8@;!)A*q7ln#W6WN#96S1Lw@1=3~Yc~
z@ml7yIBx{pU;sK4sVsZ}ZNX=IEms2*Z}8l)!_-QH*b>61A?ztEDT^q-8eZg~+|LcF
z_>-R{09QCt-g%^KqVfG!yqa?=K%etF<sbSWC*OiQukII;zrBZ8H2Mr^<a^pgIxVNe
z3;S~z-JoNqjW&ya0M`()H^HElc$4RHXdUGmVNbl1wi;5P&M1aCuzS)LWw~!0E9$7c
z%3c<JXLO%yM((*f=`>|nzT@8Kd|yTNl_S>KG|Dn%^h<-d;7EK5k9A7iauyuD#CvtB
zdy2tKUBx-iJPcM9ed4mK4jr_YTzygkf4P(--z1HiCx~(dH-iR#R~E!5Mp@>_e<f$(
zMD_r;^i!OQgA7`1ZEtgKrHqlkbVeNB>poGtSkJjUFG8ajO4bW`0S=^W^OWu_%_>-K
zI%aLys(P6&(<aG@j|vwC*$NJ;Bhyvib7WcE%EKja)QEw8m$0_?upX0B6S+L6L4Z1R
zq)nnscAUKd9QOb?IX%ST-omNaf@g3NHXf1QW$LLXdT(oM1v$Ng-a3L-JJ5e+*yF~=
zA-cexe#<YiZepSjM`V`sR+h7|x)Hzo@~e3AY%S%#j)EI*s%PE*p}F0Hh8#lUs&{Ts
zd4gW4$F1Afa>c#nl@%O{=Sj2n8ru_*a)2}E8{$kTBb0ygg!Q?4{Yw1klOLilH{!RS
zeIDO@@gO!=T5+a$gl%;wlsnD=9vZ?v7?dxa+z$Q12L>thXmj~jWjtx^jTjv6iur}n
zc>Db~<EKCWleqQP-RMUaN0-4AkN^yA0w`^p?OQb-&+>Ui&!zod9Pmy)QrQ4Q5(Bnt
z9!2cf_&npxJUD^TJb_f=DtY*A=EZN74!B9-F5P5aeydH(FW*tS;v^S7%;$v=Ik=h^
zvd&@pInOKWQ|L9{wSNPMPf0_%lo_QrF?9`TQ_gy=PKyREdE8)1S(aRbUs<J$GEYD;
zkA%x?eCCJ3Sop*DDU%XTQx#tGc#+z);(}i#kMAG8uMYDu&5Jxp-tocv3l|y$7Z{ZT
zD9o{;0Sv$MXkw7<ZtumLZ(NDHcV3It_09O**Z0zyaEMb!H^wkTTDH`-`>9kU1tbM!
z1q71{OTsl3E0r<2VaW<BBf2xwGca=_9zT7Iz}ey3(j9KYN%*<?koC{3913M)(6k~}
zz}vN{kZ(fRB@vs9!|O8yUnh--?vE1;vx~i>m~GVJ{W~|~_T?)v-WZBQ1mC*R?9HvX
z_wZ3X+XoJqr(?7~I>5;}j@IhdX6)=7B0y>|`zQjVKO7fs1RU{~4O@oe6xeejsPqah
zO5_&+fXA*lg}gl$o10CHMGQq82#rm5&UH9Pt}gU@&V+xZuAWuBB~;6>9m?q;p{1g-
zK#uZ_za)S{%yw<tm8%0Y=v;uQ?wGqg786rL;O!hEeI0{r1Ve5Te!alBbhv}#reOeR
zz#Tg`Hgz6c4%CUJj{OH$Dp;C%us#)+=iiCzmp-I!Qy3`{e?kWia9}7`z+^r*;!-$e
zTc6SHT%D+GRO^KAs#RW$+6ceKgi7092q(_x`0*nJ=Xpl-S7}R)C)ZYXO#Kc{&+h&q
zPT+ZrjvJNKU~n~zju;y&V=tXVYqWcKq>ez>iFJODj)u#Dmfi)2y!7>XcH>N7{B+b}
zb$Jy5Uv^ne;{(HSea&TR&*Sa4=VESd61)tj(xP@aVGmP)TOXs|(oq^pj_n+1fZHAv
zwG+^JkFs+PhD*blzvFA&L&F{=4Y;aXH`lk;faGdK)g4i#F?$f42&^+<MBGjI6#VzT
zeG*^(_G#>FUtk<};mC}~FiOq28xG@e-;C#rPbtd@?HW|H8uD8<=XLV0G_P~7lUdp*
ztCzH<Y@~@&K2uHVr;+Gyg(nF(Tp7jV1>O1~v>{Gh)kCAib{F@q_+j}vZhbg}cFqh9
z_e29n!yye$EUOO=(`K$%a)4|)hvo(npYoX-#cgch@BycMVB4!03%>{?I6Usa7&=IJ
z<yp(iak`|@dDOwqo{2FhZ%%MdrVPtfHE28ML8Jmgak~+mt5TI!mH3@_frff|^O+8$
zJgu=&h#n|FRz_sySs#f<;E>15m~`b{cL1wERKE^4@NfvA@QD+LvN%N0$?I;eXarfF
zb;)1LwFlHFsZNGNnfBcA`=G^lu8s!nVnLiCdo48E!=|BJm7LB&wv_OA;aY+fM*leZ
zQ_>V^Q{;_@zr~#fzdcnM@rQ15693p}Oe6ap-lxOhoElG!S!KI(fE)_ig<&M_z*4qx
z4hbcX(#z0xyxN3k{u;_UInD!fSfDsgqgsy{z~zPA9p_1jVCRE4jLYGX`{*ARaX=1{
zPT8rk?wNrNo{OuD&?7giUYJw(Eqi<@FAW6CCXH}q+mfc}rGgR9#7q7zhq^Q6J7=Ls
z4!)_L$!p1PwhL#`b?_2xqr6kr`#y(Wfv*>3aHIoAI`N)C8_#@W(-Z?lJ=)V}?@SrK
zXb+DvNjac{=e?Un+HUE%@vLv4Kjn*U=TI@*NZl?z#jQON>1g1<@8L{1AIpszUEIRq
zKdzv2#Cv;u4IsY6p-yyrSQD^Ss8ZJWztD>!uL-+#<C=~x2Thvz$%uEmewA%f965VG
zYv^q)W@%fRl)nuuS+7!OhbE4WO(GlTaWrPBzq>_)mwGQW3BR4}#mTw7c=A*$orvuw
zGG=oT-rJ3BmrsWVs1kg30gVMR-5{w;#s_6}pj%xWV>m{KN8%KD?|wh_<jUhi$YIC&
zKVMwt*%5g-FVMU?GA~!F8yPi^0puZ03k}dBL7|Ov@QE%&CmKzsmcR&2D~l2n#ReJ2
zhPi$FW~?nQWl(8zc|XpN;3E<h2@~k7YSkz7wZl7Tm&}85vd8lX*?&mOfPQ<FC#ULh
z<JP74!AI}J$3Ol!rfy%)z-ws(do?_5OAnt(jq;+Wj|yMs@05+)+u)uOneCj-$<>J+
zZR0(8!!k0Rc7$@Q^!R49-;@0O{bFfs%Bbw5Vy_ms%qL})^VuLM!GOp~?l;mOX}J|P
za`lE&_>qbE$2~gEG*x?KJ=$;cK7VT)x2q#8>Kfh`Cz-cC^9FddAG2pd7`<jMNjsbN
zVObuPoJhMISYH%xY3664g5V<#Th#k>WcW*d;sf)jj&fVRFt75il;a^DJ3jcqr3<_N
z^pN^!^vSV$4A*;N55`}CkgqIE#5-@lk3q2)Uw-{vJbu2M!<@Q%hY_S2bRw{1?q7kb
zAgy?EJg~9&Z8XG4nB_a!#U&ISKBIDT{cw%j9KVTBYU73P`4VKLP?vaB0vweyk;vjt
z1>n(!B3z5N#Q9(jj24YmL@)Gr5?wIaB<0>&xD;0r6vKT3I1_u}PRg6qTc=?iVK;r_
zQp{Yx9Pn@a_TK&Y<^gr5{F&LIc<Ze@v2gh^LSrX>`O9BnXz$So8plQ|jQZOsr7<Or
z>@t@~_$%G@{Rjpe5qJ1KY#}&suE2?FC%ZwMQTrSYl#NP#+I7uCoVE|dcGkCuS>Lyb
zCkey0(Rfp^niY7+&8$c(EDqJ_If}meNsLbR#@I{)XJQ&fKa$^XZEfYIfjKc5Vs=cr
zLrC^9%3TRT1uG-lZVu-<CGY7O7!ZZ(#ImW$xmcLJ%lTH+`=)4nVbfqL(O#oO!l5_s
zkRRJVncudw-GS4|*Ba6GJP4D9qwjTCR6wrY<nT1{Ze&(W7|~zG(J@-4QgO|1C;RST
zD0cUr$LxGB=Xln?w~rIy-d_j*gE&>LDURXlFdCO1HHxY6;1DzoZRbcB#n0u~jFy5M
zmlJkkokNOtcaFiEJ58U*)KtXXy9)^IVHAj)3b{g0#LDVww!4bVG4<kCL9Y|x$`cw&
zN#~$p2m{%5z{kcWG7`9kBk}n0lX&#_5ytZ%1|UiXr%RGZBa?=<SLk3D1FP+n>0nr{
zbSG-Jw|3&`6P>U_U~pWk^F)NPo+~G5EZC#r1aI@&#({Rf24!8LMX5}BL(oF!M<^6P
z08hMw1O?r6M&hs!Z4&M=Dtl>no&iRoXDTRMQn7VS0mh$3QRc1iPf4x<P!X3o3!UBe
z4qag6xF&rqohs>~jyAKsb2B&$YUh5OLUTH{-nUxPl7^R2Sd4nUE0J?J2l_cfk?up^
zDjS*u8rMk2_@LaeQCHF$kNNreJg29ob0}st??H!39?%I?Mp&UT@m~JbkW(5_TWDWd
zsd1Z8Q_AtTvcmR}j;(Sk{#0B=As*r!6a(Jpo<%0wtEthV!7Ke|FBH$*Ps5e-G-~A^
zBg76rbU2GLQM{|Hit)_%%17ExMIuba@KSCRW7ZFhNW1H)@5?Y0ZrkZJRE_@I4z{22
z%?<wgakfV6m2;S+y*4@xT;08WV%%tO1LF=OR{4k?vIoWdY{#lT&IwZ9Bu&ce;*0dw
zU!KTC0$Qa9+i!Eo1v=%Fccw{MLwa?Q2CI0?I*CN`lE!krh|zOs!n&Ei@6r$vzXnM3
z!l`^3Ae3Y82RS7DRdUNeac(->TO5hJaxJ*Zh^AI=KIaPV@|ua{3X}SO-W6PXBfQ!6
zC2!w*=VR%|;dWJ?;xFM$xC#!eLndy2oyFyq8)4wI2VJ_%VL{m@g*NTA(3!{~mOOJw
zN*zI$lGf?dG1uuZqVF=-wz2J6dfTK`4Qm(4a7@3$<us5DOts)kH>%tVcb3-CbI%x8
z2lLsU0U_F^^z4Hca4Y(B6ibhGiMW>D8S7qjq_nE@DTZEDH1S4N`BNvP5lJ5WOgxEC
z^C|OU%97Cief4lC)X>N%&c$d3*4(kbY|Pou7wTAb7xg&Y0^T+1wFR%a&(9`xXr@Ee
zZ0?&qa@K&tMNEu26m~9kt#*gKf@>>f(@~dm%|$mldSYv1C+D&4x7@5%XCk2UB7^lC
zm5J)nTpAuduhXEjk-ZVNHFejHaDkmR3cwdPRJR`+V{*KaD|a@x;58f;aVwl+ada5x
zU;@2}ek8}lb-w1mDqWVrp_4Q_KN_#Sc_ZHa!8`HVJ8#A0WgH1~3ea-uB*=4I#5L`h
z-{$AGQn{DxEjtbJ>b>OYfBt+?xb5W$@3mpb#?AlQb?2+9j7*W&<QPuUWzsAeH2I^@
zPLa`;$Zz<xP)*@8ui7sj!kG4L=1}s@@)FnOE3ZDwQrczZyUfh9)U)Kp={x_*#}uEx
z3QS@_KFIRh2{_EEDy`JR@7q`Y&oZ+f)$#J%mq&Rktf>S3;PEntKa@wu58hh<UaCMv
z`f(<vCK@q5+!y=J?O1&BEM_K$;s@`2l!n+>-|9@P#L2}g&IFVx0Hw;W6rd1>$+K%E
zskih-$g~o_(T_5Fz8l5AL%I@?5Hl*?GNH8?l1BQCz?6w!3VRT$;Iq<gtSg(dA@c{0
z3u){s6+S2C=sM@pV=4iXjX?cV2)+4{iRj0OJ4PYN^sVE=*haXm?QF+u@4prA{P2S~
z;G2K@@Bb=(`I{#hZFL>!H{QG*bMp%b!2S5mZ+;ugD;pGCm`1~*EZJIMES|3S013Rc
z?%o`lVR!2e!qb(F#EFrQGW^H->xfvlLMGzQ--+XPJ$<b+k%xu{hKR;U@t{qkpc4)R
zLDEMT0zGrf2)~Bjz~Ds;j)If^lNcJW$NZ()R0!hRAy}V(_S?j_hK)v}!$@+HJ?+)g
z+fDnGNZ{U{hW!Ns*C>x;-N$l8qq*suF*ox%4(c3j+{1UJ4jDD1PAX*U=<op}TN*n?
z_oy9@jYJ~kFwitI94=7xNaRohe!H)Sy%RE;2A9U2!=9c$Uy9Am%~Yx?j+{V>p@`ic
zH*Sr@+wa~6Rs>xK#@F#SM#O&NV-VW!@5h+Okx<e3ophmtH#=*D%7|Utk>He8zyA80
zaD%&6>kxP}VjbJsi1~#^;vv^2hsQj1m>h0nG*qLuALpUB*EtYI;$6$!p$28I3VB7`
zyWFhCmrbhS<iu=O;cz#0g^j;max~`?!Ef8{BE5B#Q6uxP?Tro_ZLtT%@wbk-)dA3P
z%3W!@YP3liE}}9O#|}>r_XkFwPIMfM)M_{t=fn-sRBXo>;WP)&lPwd@QYI%##wPCV
z!<oRLa7}mgHE9%DB;;Qzb?KZ;lfI!(d9}?hq)}#|+{QFoB#m}YQPxQwM|SXMa6yN6
z8;5Rt%f%jmR|khoE`Sx(13$mX#OR-@NX4WvoxNf>4}LG7o&q0_D5pDyM;p1S)B^nO
zIXONRgE)b<)d~FIWN&*zN>7TsW^D@Y<y+gc=E^?Y%avB?<@nxmCQQ(-@QHJWX*o<m
zTq^g{q1M@^eQje6R{6!@BwllzJ8kQ|MzK-gT#_|;^(+mugQHgBStC#5P~0kmoTrxf
zlP42)@dS@aJHiUx2x~&lzcPfT%-=C+py88&p@Dij62+Ook$^t?Fg^z`I;|g>B#mZo
z3WcbYvlqZd;ygQn=YZLvjYhQ%6v%tl%U&3JfgRT9A<hy`U{nt1Y<R77$OUC&q+6ph
ziAZ`awCjf&%hh4sr6J#E&yVfGS?IBRS9FUs3Vwt++qUG9Zz$KRNAjHC7NPsfQE0rB
z=r{bwgM_cV$Y=T3lC4MkK^h<JGvD)CxYFq{HDyVU(uVh@@>}7u&g$9X44Im2lK4F|
zG)p^_>`UGJKJ#fO6fcfpSB_YICA;$(r@gj16J?(jcf2=H*vNTA8Aw59X73YDh>o;!
z+`ObW9W#3;_H-U_C=4!cyJ775S}vrK!&Q%QDy3EV09rVa*12Z;qMLFFV}*~X%>R%J
z?Yfx1K&NBmTdpOLKporkF73!;;wGJ2WpLJC{5Z$arJr4FWHg8Uj*d;l$j}(i8riEo
zPzMJ!a_Z1obvO2H=yePFby@9oWb!8RXb1ZNA%tvj!?sKF4ym2YI)C-`m+|DWy}+kw
zw+z;GY7C~?!vn9=bOtLsyZUlon2xmvV5j_5wi$b*y5OY9IS)Y3V83mfzd?~=S9`n<
zmh3~$+Toy#JR+Y%WTyJ+%9X*Gn;nd4>_LZ!=Pu#^A<o5>lx?q+y;sgB8yg>r%U2iT
z#%nj@(v9meJUJa*=)f|pNdWjlTTott4yz%O;=AI{LazW;0qdo_;?jP`k!kp?^OWCI
z@7ifSs?=3hzyC6SQEv701q?pP=N0c(SNXnNU&33;&kU<P)hZ%4A@D3mm9AVfH~xS5
zz51f!weKIoV$RwSmJK|f?b1BVLb%g0<|ygQ-OHnbpSw2v9u+*V9<RRt|L5p<@7;OW
zg$$AI{e4|Aji)-+b|#)beiBps{^5r|!l-J+*Z1zn)1{R-KJSi>o*`%hAtWM7ZAyk#
zz@Ox&;*;<s9O3g+02$d$<3@#wF`%MT@QX}|bI?ct!l;Qc>k^qNRSgb%CM0M#7prEV
zXT>?QIY{H5Z*rGFt{QLRUIV?3vEGXUr~&T)Tvmt6ChNmdBf*K|JYmpd2z?8tUv2Kk
zdmq0SfAq5-#~Fg}|Ng)IW&Fqg{1xy=EL<LqH{Q4n!(Bya9K|pH@z=5Vd<`%t6N60W
zM}u2D*!UVbT{s7oq85IKyr|?44)#-!=I-Mdb;U?3QG1W#h$#eA7AlT{W9ybqV735<
z|FxY7hgl@$B(w3o&ID;4Z|!)?jFeK3`e1iV&Gg6QTwgTzR&odHnYmd6-DnDE4fX%>
zcYmMJS(g#bxeR3I$~pW)h&WDNg<>TCoOV=5_GQ<ryAux&jmOx?C1~naj1Di5XJ6(a
zy%<T+P*Z5zghtIu&s2#VT1$tDxPOGb6UAUUqMAl8kJ6YjorZ=+hJ5I-5*4bwMbDov
z<^*w-$sy^D$~ykkxgoE;F%>`i`6n?uH;x0+#Q0oEhoZaZ45fkM=yz_z0Y>6M_RhG<
zm%`=B<@ucd;N(1`Rw^BxozH*sIdu{T2z_8xxh~AlqA1335S_#aoEUOWZr#OkcD%8M
zqwS#c(}Qza>*+_Ss4Q{nby#pztfQ+1Y491fHu7lw9E)#+SHo8&Z++bd#!1@_acS+N
zXb%<IML>yPX-67#IG9d^ilb@^C=UFQ2-?XZEzqoZOGn<ew}*pwC9+F9f%oJsX+b9f
zTA(>?vOFj@pGmGd6Y`A4GkywAk&$T)S%<1&C~E{JoaQMJ)9$cNvbDe3usG~*;nE0s
zUQqXLly`f069sEfx^kC}mJ_jenrGmoMCd?uy7&nia9PisZzYRRh~;F2p2}poJa<ly
z_%#aacx0DxbocYo;Zc<Q5X#jdO&G@TgwgX1deH7-vieRQ)<6^QwwuOxHAdY-xQoy(
zwKf>w@F$mW)G6__rz3~4fN$Ze_U;h>(u50Bq+!LW?9z#_x6a<Q9Kr-#m4eD%KjeU=
zr?Cti@{+Wuvn{+Lql!uko61RAHt6Gg8h7+hgAv(f9i5vaKa7oz;aH5|kRtnNiwq!%
zaq{qLFOSOuJ3mE7TI0~3LWgAK29Xft>G1))<31l{xS(~AujOxbNaZwCN?!OhXmbf}
z;8pf#*_>G_4cW$Q(s&w`$QZ8%&cwS%I+Fh8XBpx{+mdg7!`UVp&a6(UQ#-#`m4Bve
zKP#n{Bl$~Q{{9J`P3QYOOL8CN#WVsX&*gg=>eY<>!Sg~-1t$g`bS4u1>MvpE8Tv#z
zm$%K?DqECv?A4%~FDy54#R-(j*gmUzD*U$6z2>cRSn#0Ht^DXpT6dKs=|nvz{pYHE
zZD&KCGMF}mGg|IRgYL^Zn>A++#Y=f<4@GsB-d~4JZEbEw6WMH!l(-QCRglTw%4YRL
zC9{$jg|A(Yq!r-Ky714m>KN@nag+-ikY2jVU<c21s?&f*HtQtoJUcH-Iq#f{iHWfo
z8AUz}U_irba2NEvk8W`u$xdu;EQc#W9UbkHpQ|zTp)ZGlwKKl`_N#dO(9P)NU1S`x
z(cw^ln(#T~*@aN%7A^sM#Ft-w10TWnEgS&S<A>8Av@6XhNpMRUd}mLva<(1pq59$I
zY4K$lgOo8kgmcFOb-3;J`fE4hqaR++g-9C8GU}}i%7BGEfXEp0ws%U|&`(<qQ}=Nk
zh~bIJs39YBqD4yH(%Y(Yf001QgE|t@j4&4Z@{rCGaxWhKF0a0SWV*uV<u4y{lDM5d
z<xl%AlfO79&urUeR!9EKQ_OIKJT4DBEXim2A$R3@mdvx{aqrttg{>SfpH+1$SF_=o
z-<tO2JO9d2<i@M8q)hW!1tUj0H&eezU*44ezkm4MKRqg#s^j-@^It!{pAKhY0Z^#|
z6gtq?9TQ_HiNW4z?QF;5<7fP?$B+N`rx;^<@y&Pl;|b2hk?W#&*I`DZdj$!36r5yr
z#8F1YI2G6_;2<ImF(=6xC3W(z?^*~64L1*!l-=Jd@70ynSY6$K`O5GW^OWG7ghY)~
zAm?-buLOBUOkjA0NTpa6AUYC*8iCw*leUlVdik({wbP5>FyfIbQlQY*G2pfj&f^C^
zdMAGL(@&xk;qmYP_rH$+__v<|uMyI|T>JjD*IvWWK8?TntN$3wE1NJ~i7G4FG>{ZR
z9vuoH4GGE?PGPoZM!ab}J5)$v;ZPa{fxQw{#HK{js;FT(Gf%2s+B{o;3qNehbSS(p
z@5)UvC{+w9X!Gu-@*`utap~$9&e2djd44Y?mj;5)gcFMo_7CG%|M+R_?l`7Z7>W^M
z@5Jst0?(cejB{6jsyFN!b~(}Un3|d*D~GD|Mqkf(3^x`KSo1h|{nS_3XeVI+CpuOV
zZw{A28RT-Qsq}%GsvGTB$Y*4P_Ou71OyV~>dzw8!PBKel3%J~r#tF6V7h+V)HgNK?
zt0(D9xLfF(@6N@)`tSc)Iul#lE3v-$3`4Dn;foTo+X_MNrhUR`cl#PNXq5ZfmFZYm
zn9r3zq%jRW4f$XF<7d#c8{#0H!My@@W_l_Wi}PzVesvmGSC@ckH5;z%Al;EhsLG+%
zGXTCF_kNs)nUnH#CRTEZRGb+Mh1|2ih~I$Y%zI;Lc`c1u4bL4MKj+`9Z|ssMS%W*s
z4wxtflF*pfVAOd5z2LHJ!shUwGus53m2ag_%XO~?@ehh=4<V&$pvG@)8%RC#m2PvT
z51dNNFr_^i?(3ie(wQ)w4vc$r%wHNqNx93j&eR3z+zd~Lu|(B%rW_x<(>zJ|vloN%
zoTGB}s{3%%pwC7s@5azjj-NLwojY{n^yOTW!;9F%vEAFDeKzc&08U^~ep9tPgj*+B
zLpLiYS@PfDu)z!*Y-o6XegRmgve!nZ_5k`xrvV%&FUt@-Xb!&3Rq3Dsdyq7ebadoj
z6@W6q$gEf8phmhhUksQ+!^J6f2nh5|zPXVbz&h{2Rg*M`az&+#+!tp-V=E0Pcte9+
z15A0Rv7G$w$x-BU)ibPNt}-u*E2Fm{#tomk95WA`8hX`+VAZbZIiJR!h>SS9v@J50
z0(556$%U?TtnE2)60U&=odtQ%xd;wN+Ck1YcbeI2k`uvsm&2^!>C%gUF=r#SkKz!t
zjZ|RI9vrUOQdBS}5i2uEmr+ylI8A+{tSYb-y#q~y>&i&3PLFDclIa}gSLQXrAK=uy
zk_l+X%J?Y}{Bk<PQ22z!-<hxJ_+6FZJLSHzK||X{$+~FO(LPnWlIP2Evi;!6EC*V9
z=}f5a<W2M-`K41*nwM`1zBG8vr!v~i%d2%29@ClL0EqJk<U94S>4eKX3LN6F$b4?|
zx%a~3?@AX{b<nYJ7?Ql<@Fsh4?WGcDwvVz{gWMIl<QIFcjP}bfL&#Je3x~bwbV!fd
z0;V(QBp#%5>B*r`Iuo17#qH*HIu!dlsoUGJgPhml=%$WBmjO-WgD?xPLCg+>fj9%|
zHcAsg-enyfVrVah?Ulc+7w^TNy>&Vw*<SofFY<)@?##^0V(iZ&GmuRU=Q5Nzke(X#
z`z_~-Y#^&~gt#6Y>{6n*LOves#oF?6EH5>Y_wu(r6h+P|109Nc_4-Ww;DdK@DkkH9
z`?r4?5AJUxFKAyuH%LNzDmy__)(e(I?x<TzTbFqgS!8}={&Uk2BiMV+uNX%5{qW<v
z@jv`e|0-U4{SIvhV9<f6q2&{Y&c({r1Ib2xqHgPk{{zP&Cs=prMA#c)p?(jp;XNIR
z3-}u>W}f`V3Cw2HJ_-+r`|3ga`cICMuFVPGOO~IKsc-WKFHQC8)wV5dZest-hxy2Q
zmJiFB`Eba7)uFt}qm7#vdE~+MMg6Kncni%`&ogI1$U47FpD<Qw@_pXrA0_jQtL1Yo
zsXTmVV(ZX`?e`Az$!|05_m2YCA3n-AzkgKEzW=V{2k&2|Jzy&c#T}2whI%pndg1`@
ze+h$Za;!i8>}P+D7rYl=e|InEOthTD*wv3BFO%acquJV(YEx03sSFk3A~&O5ByhzX
z#~V64rmQA%2!lD)2O}52jSM*<N#=B7>-xGo<s#HGPYoLlJ|mJMwjec`ojO>CYc4yM
z(<q)BmuU><@|`*{cDM3fCvf$c58rmcM1AC^;p6zsZ4}|=9*PM({oqG$#Rnh19~UTs
zzy7=b62JZOE8v9K;gN}nQG~|~N}?7|pD(5%Cyc`8cI8e?7pDp#C#Y3XNd=KOQK%Iq
zqN63lck|3hn?^Q5jPgs=?4`2v`$~YTEKLKJ`32%La;%Wdhbb`1P5Y|!;Fb7r%w~V1
zE5@h#<MP$<n7=$i9Ufo=VugZJ$M9=Fxx&rz(i(<=D*>HRCWZsTqlLk<ymrdH&IH0J
zCz1_<k8zCc8H_#0lB2}WdT3J&>e?iZ+%&>#09d30{g$>;wKs<{t*cCBeKn9&K1RjM
z&=Di78n`NWqaz1gOSiN}mPUOUUoBp1NvqEm7jqJ}3Qj}K9tOwKH-@`$Cg$Ux|I44o
z^&4*5vl`#r`<%3oQCyu_kHf<~gysQ?wiqJ27)HB0N3nqOF*i4xN>BQBvfSGGcHH~!
z3H6>uo<f(&&??Vk6eYAgHWbqszCAcv?)tp8wvzRphIWL_$>L7>b>D%t<!xxDle{kB
zkWFX(o;-OPyUiWwcL*ojjedr*9o!V{G0y&%zx#%CXV4Ap#dU3MC$QPF$+cIqjRx%X
zqAub&dB{^orWpPD@zRp>8WiO66Qs-GQnq9TulP;cR^+SlbG4G8VK?%tVR#RQ6ZNHO
zCviIDXJuuBx`9U&7<wdH`)4=<Gx7SHH{#NzNod?5TYJ=_OazuLH89<u<jE7hU%DV~
z^QWvnm$P-k;2^vWPDckRb23K8$5W9tX@}o^^%dW`<Ee6?Hx=qm^C+G^UPq1rH1$z_
zq@zV0u~EQRsgISe^>w7PbLHAx%*@WmmCIM-@};X7(W5ve`|;$_Q}S7koh|SSJ?uea
zE#!s#=&*+|U~zb$#)ZqxYS?FFA9_@#xyqM@x_nW|E9F4xad1alBe6OI7+v7Inky!N
zUGTO}L{o=&*JUOT;_%pERR^?_Lwjf(Xs)6=?}bZ4Q{%QWq&4t#6ji^rl`A9JUfsPt
zF)}s^K3#laDyGsA9?KpXhpCjY={jV}Cmb8#wHIi4bvZWSD`E3@P5|F+fp?t-mzj0>
zP=`@zgq2`H4rfh^M#LxTllGL|F1czj$U`Nj9880b`|?`{R7i^alaK7-R2CRu$RS(Y
ztB|d`Mzw}02uistFTcP&Fv!D|BDG!YZPa-6S<y|!Ar{BBLpmPq^p=r#<!82aUV*R3
z`I5i)>D0&@z$-nFD%&h0<L1X{wUm3FJ}Y%Dd6wwEUzNi;<opM&9@*PqFDF-Vl8#>H
zEdiA@(m)R7s&Igu5;li5s;3g3EE_uEdvPrsrmt{RbeMWu9NT8<B(=J**^rbbt#R5e
z4r>yp;>@`(I+e#M%gYxjVmxeH+R91Y4kxk=CMG9hcGeZr+`YdJ@55u*1Opi9rq$*X
z2bJu2{><f_pX7CIWd%CigAFd?*uWDsk@beJ5>n`_^iY(uYP)nGz@hLZjQ-;rzZFJX
zTso4TE1nXEspA1i9yb^#oWg6bo%<<FOgb#e0FkhF=6*-+lhbqs%UvAn-K{t|Y{hYF
z4?VMuuG-9@yUixL2f1mFrM<dWZw$vrKfE2Uzwt)=yMOopjz<qS;vD;>hI7>AhGrMt
zI9U!6bC|U}fJv`y0m(VH!ob8)3))6RSYv$!*gl6l4Wk!zB>u<$<zK`n|IMFfTS7IQ
z-k93Pl{ee1mhZ)9MPu-pJrNf=69`uw3SS+S;Dt-_<^yF_(&fve&_f&NTq_>;&}{kr
zBOkTXzWlCzB;T~taFn*pyBFySZ?((F_u{HNE6)o(SV>zp^C%V0|NqIi?VldWbFc7S
zbuDmI&tJlk_dGAKy?pnF5BZ#LY*?=CT=M_CrtML|{`-gNs<i+0!~B!p+sBK#=27l_
zmf*f}__O1~53V5FVUCmY7{E{*!I>D*nLueRKYN}-o__xGKS$BE;@bxga3+>w?|}3j
zZe)f>`$EvQ8^T;U6{H&e2^h#KhNDJ^lU(cRwM$&?Mk(_symQGqPUlM)<x^=pw!Dnh
zR=_)EuNaI901Z(~&WcdB5swL&R|a;>?bD~vGFmZ!a_>iJ4k4@>a8@@UovMhjCR}Ao
zC28-+Hg(>_aM-}$z!r%QKDZq}`0(xM!U*`)r@x7>zWuf+dn%k0wH)sbm%GI4+S+Dz
z2pECXXh4l*-4!|#qUlwXK9i_KmG3kFbRygp*h$w;LQFviO_;7SvK3kia*4&dmQ&o9
zXp?x5*pgR)uLw@@lK<y3mu@^f+688W$mBpw&NO0jR>Q0wJA0d)J1FI&?2&M&%;_;k
z7H!;DYsAu$!xLJ`bhZWdEe;_i8vVc!8^)oSo*qdj%igDq@(m2fU<1Qsa3Sjb6F_Hg
zh0#R}2b62pi8fTZ6{ZAO>XeaIdm)_EO1t>G7)>fdqv`GCPPsa)$hwwcAr5n@#SV&O
zad9b@S5|@5c^=1U^yo|skM+e5KDrwJ?7#h!xOMvq%59gn{m=31Uw<0=7<4yoIOZ3p
zN8?9}lp~cWkZWXV@Yo~bgzM$iEsW+C_3X+WPaWowLrhR^PF^;e?&Rz2F2)G%MtVC5
zTV<(0(13Ip-%iujmZ;ah_3Olly%Lj?Q{;1$%kSzaxhjz>200`O6r>}g1N)oL9>AMN
z=?Fkxd?U^P9~zg2R6;rvj)OhcaOK_peg{g-<y;HzYV50U49XhCvRxYu9INp<d3K<n
zG!Vr{F~UbsK%*mlF*&V6Jz3fuC*Gk94k3bqN#BZ%EuAnYzaPbS-#!A97xB?Y@5Cpc
zyq_y(J$if}L;3(1Q9!i2&V(Dtt*oxY`#4w{VBHvt;M6%R^#RA}g9q@cLC_c+jsecC
zrW<NLOefO$N^zkv)q`<+9*-Yv#3Igw_Ml=`ytGA<Y=LpomAz_X)ZxahYjOAWyXnY@
zYw2%in|5AYjm@<!oNH(f-q_y2DTKD%VO$z^5ea$Tc?e!>ASomE+-LL^SytN5g$#`F
zX&{_sgfFAK0>UNbW&13&4<FF>EgEZoXD5v&S0>Zo6kqa#1{Gg~8-BVfQ=8q}yhNRo
z*MLi>0holrUbLy{nV7${5Th78gQLSaoGCY@LRLDA!+mfxO69e!bsWQub=tI<Mw$^}
z$FuL~*toGI{AEuHfMkTT3{6sQ0RShEzJRIB6~XvHj^w)kDZD*3Zc7iYhj2UerTXom
z4oP{c5iQ}GGmUBQhCN$Fo|>=n4jHR~z^VK-d6CCOhWJ|=D6+F0wx#9y%)GKkl&k3z
zZr~HvELoYAUwKv6DXS@o^320t3lj)453_7#TxF;iM@bn@7Wq!NN}6;s_!FKIFe|X-
zZ|NuN(bfyRS08qD*t@QNgdU(D;gfg$rUYm_r4|0fh5vDsp}EfS>ZKTPNxKPn%3t)Z
zN8$zjRpCmQvR<S+MJU({?|XY|eP(_Rt19*E$?v6$l-aaX4xgkB8R5;g5~`A~X|rBj
zrB$5~m-@9gSou3MH-p1Fk<Rbn$WU&Ys<STc?76eoLZ^EZnswfV&V{QqI{(HV40}$j
zo8_s4vb~`-k6flxIIIJI%6w_bdY5*wzAmoP?zQ5*J#w~JX?xm&{H&Arjy|y0#o(H&
zm(9-4#_Zf2j>Ty9yeQwy_vipSXRir<9LBYabJ2visKeG;JbV5));3y6Ph%53=%=x`
za_w?__uY4~y}6q~KZjoq4~`;ryW`o@6?BHqh&-9qqHU0K=mCc}IULG>igtpvw#OE<
z#-IQB+wrge=l?c7{PD-pgTwaHk>CKvkQQjeyjn+ilgjf_8u64KDo29rIlsa6ne<3r
zl8*G1e3{rc+j4#NsA#ObtKNg#?>#Svvg6g?dDjNHI(%Q{<u&23eY|OZoA2AtJZ=4Q
zl>3rL`R-)_-%nK?)-T_^NcOVqaxL%6(S|91zs#c?<#&Ge;%LKN@r@8DUrF<-6H(x-
z4)5Fb`hI3Fj_;>e90_B)tUSu~)pS)}RgPXeKKkJ5g^VGSX>^Zf&qNKwW+#@OK8>Mz
zZ~XKpe~QvSjfYPj$77c&YawX#KwZ7)#bH5I$jSkW!mNq{`5+@no(Q;pj0cxkbY6ff
z<v2M>I5S$EX;6A7;`F5TY$0NqIL8l@r%JX&McT%7`x}MI$fyy3p&SS7GMXykUL6#m
z8!~dxXv6@*_XKC6xw(}BSV8N0{yR7>yN7472WCXhd+*(dH{ZD%9SF(a{_c12@bS|u
z+co1Y+wNF57&Af$SScV=fRbKyAiOyq9NgqscfL`u7iGptDvi@z;~&9e^wpAe5<I=H
zqIe2d1(b-gQA@*`rU+h@g33FUTq=)>s3kJTmBZWM`tW!s`sy7qyMW<8<!URP2sQ+2
zXEAWdIUS14^(Mlq71NWG2=gI?@>(n}yNQ}mlQTu@%W1o*V+Y8y?xZmsQg@2ByLxhB
zCN9n2in*Cv(ce2ld#c<_qd|>R2TqbaG@yXnI=LxbZeWNJW7VtS3q=tYWt&O~oR`U`
zMmvi$;Sdp>6L*hxDNy&CaD^97Cu8sKAE1Q#;-imm#-IQ5pTxqYDeAl*zx(>v@vC3|
zI#$=$<MNfhxPJXIN>2mYNcTC$my-yMh>xVvB5X!tS5_QObri3^F&mc_=7Aq$e&4m#
zF*a9Mzz0U^*hm`APSSJI-Q?s%(ypu0Y(RhRy6o_)?e>oIU0uBy6}K0{9*g_;A7Mlu
zlGhOQ(T^fL1+G0#_X}{99vqQF@B$t!UwnG%G{8a_vt{D5cym~f<luYsDc^%%1KE71
zBhhF$WCe9V8;(y5$F-~T7{RqzURuieC(@O5SXh|O9v649cD-`DyQ`?GbZ%72{7+Bp
z+0nT>j=%Zqf5=s=KKk&T`01biC@1XQzyBRD?BS4hBt2-18zpsov_{Q7#(_~SE2rV9
z!>tx{S=_zWUYt{z?%w{=BeT908yNY!;L}x5diw^`8UE&Xi?R3|KBJB*RqN}JD{LHO
z#bC_N&S0cW$MozJ#`S#mE=W&KsNcn5+uGQUl_i{s)y>@V#^tg#T9WpVo6@XuEQdPP
z%TOnyt44kuil#A?^J_Ht($LUYDgry@F7++^TLJ;O+@c0K^jwBLwc-?jy(*_{KiexK
zs$AVzEPF7)k<owoAMVF7a0ySBJRZtPo(nh*I;c~V)3kLNN@6594S3QXP=-CTyVOl5
zNq*6ow)bTZXC_xm(y<4DZeHi=GdeKxBw)}kg~n_*dq1qByq-oU@)@a=f2Q?UIlV83
z-#UI>9Qn-QFlA_%bA${;iO;s-2Hw&*rW~W`Wq904*1hN%aYG3bx`&7&R}DVDS5ABQ
zz3Ie3rV(Zhq;@?FCY)bD)6j~;b#mAZa$2ctFNiWWm8xZB-qx_>q1?~BX|U{>AuoSh
zjCHY0qu@TP1gf{n_{4>BoqVJX7t7G_bjXpq4h&@d(?JA3y^bY3??}gYuep!G2!5a=
zWqCSQ@R0OYhBAo@VJ>j`eYLPemd~Gc*g`fbKb3V^#R6h}>W{38Y4f4>rOt#jUSQM|
zFu&qt$hXRavWcC;;gXgUmXe2UQpvD9;CJUeiog5=2hxK<ivi@ZPK3_H*u)s}btKzn
zfOjqkp#IkZ(aCLYZN~<(%XvDv^52TXM%OXaH^W6M)Gft$N=UMj4%33LtY*@ZFi1D1
zux>@CdJ4NbsyLk_O(#l>sN*Y#N*d06P=<!X4Mp8~+un<ubL16Asoz=9@bB@l3wgkA
z2M+&qCB5}nS;OYq+oIJEaj-8knBpElF0<<fd?V!B-#d`9YVjG4#HNd(C`XY0kcP{)
z8{$#^`RuTT{6_{N=Hu#>h)+IwJ^tb^{xm-L;}2sPXD!!xwT|!>72_k?2At#&Hh9dS
zLx+ljYo<#~aeu+76XBV?7^Q8~)={{D!))JtR_Ml~x-axl9WUR1|Gs_j&8zg6SHFKH
z3!j;1NmHG_R}SxdTG3*r&rC`h@7qOiyz-r2W&S>Wk$^{qmda5*FKNFIM>#9_y)XGx
zzh55TD<6DUuogHy%J1r^-dE4QSKn8Tmu0_tyqfmq{i}K8SJtPxZyy~WeTXw5Ghh|g
zp*)v`^*nMC^5XMo^w;7?pZpQucEqFS&*Jg(#b_F}zJQo;CMu68qfi#^n0>qGDJ&z-
zo*=k*_@(%28+5z;6pW6W){s_?>rBXi=H=={3It(0Cr``LSaf1`3S$+O`I1_p>n^6Q
zuwz6}=2zj{&Dmx7FvAe<CWl61bZ{u#;8`Qd)q0#5*rJ>kMhSjGq(cuZufKUUZrr(+
zKKhs6d>f0)D`~hHLClB=7MM=V5%^P}h~892z@(wzq(co83BZWC%C#5~Dq(~VfQyR(
zCo%zBS$AI?Ns#4Qd9v0oKggpHeK{rQ!~n^_yF!_7l6f-^l@m@3w<@%Q<7PC5d*a&7
zsTiLept+n=Vg%bsa%F7y0fyA#)8$xNT#HGBoib%@br*vYqozH~OJmF>^{t)_;4^7c
z*`sqZHa37!d^zsidYkJTQR^PW7%7wI?YS_To8xuCxm8R9gg<*AbS83{N1o?!Co=W9
zSL^2HhZ>yKvY0APolYb98gcfbx$E&V#%qqt<~t1}=M^}=<GuH<#Gm}+!*nJN4x91S
zx1Ywh_r8wf)5ExNV*-a_HfeHWa|7Xil*5-?E>(qP`C?l}-fkYoy?bl%)|=O<?-=yz
z(xQ6^^L3nq4V)<_?w0QD;gLo<7K0ezD%91rb(9>&@1DDFV{9Pg5BBB54xE=^+QA9g
zhZw@E7<9+b35FjC*Qwvf`OwgGnNy8y8$jAB!{aoX%8)dTZuh*f?FWZCV|L!vM~oav
zA6!rXa2xHfL+co4i~-2Vo-*fqICQHAD|~t7d90(*bfg?|;Cv93x)TS*hjSk6VU)IQ
z|E$j`#$XMl-ak$~PU1iO)qjeW<<0o$hwtG`e3C09-MjY<P65tse@_m7tadlnNV8p=
zICBE3p)-t;Fg#R;UJi4p*y0j$h<1kQqX%3aV}Pe~25x$L>oGh$o;dyD^Kaw%(<W_0
zz7CtiVZ*)3@y+gYFg`JbQlCgCLIYdAb<Y7eKGX<VU-ev%CXR~3{?ZVFCZ&Lk{K7oh
z<CFms4D^1StD57ym2U<JT!F<1Bs5M#lwMY7ma{T^btsaK$v5Z!K>N<euq^4X1H;6&
z<V)Hc>f@g~yUQ^;aJIQdoX(hzj7F8qvF1=E&Iz1}8Js5P;<%rQu;_fd6sPjf`r30>
zmge+XSw%@Yb@-_hFG~OfhUDYy$yO-H_rRY+k;pS;8}F3;DU+yQI!S`uhMRwhJO6oK
z{Z?0alz>UO+S8@6t3FWHdv7*LXF3#l7JMn&lznQ0qHEv-VK(139Of%*=}d4h&OM5=
zQh3MuNdu;{H=>4-*M~fVoibRZv#4%y*LIyr(>Z*}Jc>DB#T=hr@FOi7$tD>z-pDAj
z&RGTn&`+XNnP?tB?k5Ef|0(i;p$DrkdugCE132kSP&a!yg~`2RoX4}avJu-GJ2;=<
z<cL%XXX>ALA<tAH<X1VLM$R3g_miLgH0QYqt30mkU3|6}pMCaOu8Joeq#+M&37Bvz
zS4+T3ctx=ePgx~Q*v(%Zt;3_P^u6?*&V)K0I!XP_VSC#CwpZ?_BC-;j{&}1KXC0)A
zS0_R{>?vuUQzv44Vj_nhJ4C8cFTEDUMo}l9pa&dkx|Ti1E30Kbnmpi&idlC~cl_32
z6>mCxwoSGxM>_tTLT8>8&AiaN{&7v5h;R;R#KzQCP(IY5A?J$NV=+B7le^~+yDF4*
zp-o%>!j*~)%sJQM;K;xc=ZVYYy0=BU_oAsxI>@uRwHwc#Y~w^6qaQD_2SQtaWZ0fl
z_(q<E-kq$iBLNx2#8_AS<R>@cFaG7v;_VOKk5L?nPTD!~Z~fK5@<_Hb&qRT|B&})y
zVAG@%VXs6N-<^pI&P>jg4{X~!_*ogtDM#8z;^q$?uRi|+X?T<+i2EuJk?zAhe*c?4
zkjIOo-Hz@1(zfNO+W8NH*>2}oQ@^aw%WFP+0cSajE$s2?_4_5i%&TCoz*%T2OHG(x
zo>iK1khZ#O*R8r%`M%8m)#LlP_(Mna)%U+f2Q4g&{+nb3+P4QHF*x4UiLlZ@XzfFg
zT_~ctWFQ5o%TOMm=*}_p9nxUI84)Mv3b;~@mkrRicS;%`fDh9u5G6K=E2At}GA>K0
zP*UJ1VC<4qA1HV<K3yq7K~yOiGfX9>V&S=01)V}mrRT1jNBanEuHB^VBagv87#gP5
z@YC>gu7Jh}wnhqTVU*mBTskDn;UUW!f1A4*-IY-^98}Im{%lkgMhY6ZRArSr%<@bF
z!Xt5|aB0`u_R@$m;^ce@4Z)%WNn8=B_kPPY1ryg71d#7b`l?4lzU6;jA^<YF!0*h&
zll&0`7igW~(Ry6DIvbNyW2ulFsx{c4E}Z?fM(pn%;b^VL&Q=S9<|0;?F*Y`Ew)Ro9
zWSwXyORpN6Mr({jw+>1~Tt(8`+kvng&P{*DM`xg^8Z=cDjLQRhDqN+zYMg4-ei{k>
zmT_JxERBQ0VAj?4O~ZxPDoxR17k?^5lva*GH=0TtSbgcpvFZA$IutI$>)s<zpFNIm
zzj+wz7>kWYPfp5q0&1Zx+Zs5)y~dhHI<aJ;QD(2yCQd`n5uq(L3bW_JXfnpu?93GJ
z>`{<bVKz>qgbu-Tl_q!AHD=^q6YhTPCTl8PqvDSL-&l8CE%fQ6+(VtElW-k!_e*HC
z$fs56?NA_>v#r<dRv(CoiQyQVF#6n;L)$vw=jqu-y!+nGxO{mk>pwiw7t=FCapl@%
zT)jFSSFWglCv!f+FnKsUYi4=^JlTV28HG0;4q?xt6lL$w&W>YyH-UM#t*p0m;e?gt
z4QOa5o;+#BgL|v-%{NOJSjRCxH<gq0jVL=A-N?7NHHe_$?5bPNZ&}3=)Ty<-d%?4Y
zjOQ@?W;z234qH2+?cH%&N5JJrH7eZX&mj>ytnOgm(Rl*D8mT8A9NGDWnRxy6*W=Qa
ztJHUbca3ZZ$N0N)*T&i=u&>1PXU}m&mh!jrW^(d1OlCV~4_w1&V*}^B9(A0devAQk
zwl@-6gFd8@TyLM>GHMRhgtk33*u<d(?6Q~Qp@lqhh+MUqmj;d96>uHjd({Y(H_qb3
zO&+<*N5bIHk1N-%#B0>+&9~o*yRW?-w{PExtJkl`6poXI*Dhr|UtEeO&z{D2_wVJB
z#t$Dp%KLlwzKf?%p2qUB^<B^5A}ySZT)pK;$F;1qlz|Y+EwTrB*A5;@h|;2UQsh)8
zPg5SK12bZ60uPjee3iN)^D3uL%HfwUkLrE(XIa_PKzqn{mC;hZF7K-}=978xeZKRz
z|K)Fwl6Ug31}(H{FNnN5H9H>T)593_I5&OW7_WWEc%0IsJ$MUT&@LML-8dbl=^>4K
z1@+W0dfjM|_d0=YhBh<n0uFN&cY#q@=6ERc1lpP>B*0PNEqGI&BXY3YN`4(Eu?x2x
z?!*|UyvE=_jEoJ@zP->+S^cW4%*JU@PyNHavR|Dk4;V4GuFC#l>iFZI{21BrZd|{9
zBjzt%O5;|2<N_8BWlP5cedoL&d-75~^1ejrIc>+3;lKrAb<&JttIr(zsZ*$2lXt9d
z(gWZXX9D=Zhw%9fpc6KGA`&G&YnS7_-)6uGSt4CJe{K(1T6!SXaj4dEuBF#4WV%k9
z!xM!^ePgeNJsA%3bJu$32w5+kZyiO4pt@?AHVCC=eG;(hd_@n-f8Oh$33D2HBu}c=
zkl;?;tln}cv^;G9%$`l<<Q9(1`o=2qaygzodxnDp%se+BVtVyU-IXneMv=!;F)=xf
zY@3P6$yxO3Ow7+;#yK6M%nop%LxKEszX+TO)s=f%3=g|cguyU$DL6PrpX+4IOm)X=
zuZ_jefBtcN^hY1X_`)oX1P+1+bvVcV0?{@{9$9^xM161OX(*AAL1l+EdA;C!<$*}y
zL)$x>H~++&&v}q%Ig)-}Isbn?gr~Z#zVnjzFZ2KYeRWjNUgW@M72F<~QTz0$uK)C~
zxbMUMhw`p|Y15Q%{^?O3lKREHN69<y^8Q7-SK)n;>8ocs@~&j?y+iu_{!zYrS&x7E
z&F>#CVJ&CZrMcovz!?xA0;kjNMI2KIU!zEdpPie}ZgVHqZKgAULgQm2ZUXAPpcU}=
z_?3#K@#kay%fj12!>5%8@0n{>Q-<BiR!D=&v6UK#BCe~`2vtuy6#iB)D#Qw-mZ2;f
zq8T-VX%t#sjbgVj$SyE2YvemX86%B)jAQ7w_FMtNala^T0aVDsEUImaMd&#3Yb5Kw
zX1DL^9xx|Kq;MS{D_yLP1Jxi9ZtJQt5+*mV^LN4z9;-+ND9HP)GmKe{pY?gcqKXzg
zrI0S)m%K~!dH6d8Fn<M?HlcU4W!sVP^c!!?F9ku@d5lkt#I@^}5vUka2u{ay57wPL
zI)u=@h^6Pt@%YhmjJY#}@d-lDNyr%M;G~978Xk3Qcn?am10|_a!HGC_qO)+7arc9b
z-f)wzsi`>(<;ip=RQk4|tH0RR)oK|=(WMiWkOpZ{us&A-p=h%HWvsB>t-jCb7U^yO
zbS9`}PRQn+R}FQ0FYMW{44nywX>_89>-Dmv?cu>*Ja}-Ab0eH!KG^7rvC#(Q*(PP(
zf9cKWmXrGX+yT|-GpEbpouRCsJY6nP!?t477%3AFBPOYTowP>NbMh?CcxgX3mfHpX
z^TfSf-ory9kY9~@VUTa_#M;_+tgr85tXo%|WtFIH;O0dx7U0mX(k9~0HgJ-)G<)OL
zRoZqe#;1n(-5sN2T^KQAarf@^7{Ty$SwivP5{4tg4V;Tc8rk*~x>=vS3}+aP8nSLO
z<s?T9cn!8}3;t{aacZ4hy+<cbM}KS6wbI=HaxWe}*pAOXTaB;2T8p)1d!<~sKqI>=
z9T7JQ%gDMYp#hz<q7nb(@zZ$p;0f*La-n5BzH?)QL3*`UX4|<Ylq(&)pl|5OD5y&|
zmS`x?fw%OA;UwX(O=Cb@yLKxsU%H-3+TI=`tr{b)ob~MKvshYONjP+hwl*-T4>aJx
zZw)6Br(~esn-0Y2#ApnUk4Aln^1z>SCAxc|ZJhJ$m2e4J=v+Qd#^BHxGLo)AZcrf&
z49d?4Be-_Bj(lPOM));S3;<}fiYW8y1`cV_VG_=JxOU@u+<oKqc>CS=;*Gc7qWy2@
za0-J2Ixrekj~_jW&p-cdeD>LI;&;FM66fw~=>9%5ww^s{&Mi~kr^BG^!dNqUCg?Vf
zaz&WEOSz+LN`sZR@L**~ro-SjT;$2>e+@1Vl9J5yqQqxXQ@O8t6!PSmWmQTq12sI)
z2tHRckRG#l0w$3Dgd=68FnE|QZ@qfce$Lfw%N&TTVCLzW20Ze!AN-CXC#Pn|VraY{
z8|ei3i6I8<o_CzaKC;7}OgBC3(#R((doNsmxz>m5fR{=yKyd{7arCEWCZWM0^2aUY
zJ6G09gEZ^OwT%~fE0;{BZf+)+lUF%~OFB5`O3iUDA{x-%sH^WmdpZ)%cT0IlcIK0`
zl+{MBqu-pX=$;e8EWGmkTkjN_kvCi^XcWU*=l1dA$MD5g=Ht**VZVOsdW=twMK7`=
zQAb@2G`PToPPTB_F4B&@Zt6Yf-be=KqkL9odsNOK&?e0!4@2+S8)X0{ot`oR!FdhR
z-OID+1S!Spsc#J+DUUYSH<On=os|95;n~#X`SWKP5Or}D7fR5Hn8ZMz#@U{k#X!fw
za0rsT<H~t%Y$=_qLyHpwjpr{8{xdLC=stgP@<)30na)Jcfyz3T0HU?_v|a6K%wS9t
zxwBe^<Ss5P#p2>(4hMCZnZXkEShbVBy~w>aoX6pDbOw&;oWo&fVrXy-ct>$82H~L+
za2*=zNk^i;Uw$}Gy`B1%d<F+Q<IbI-`0&F!@uMGo7?W2oL)ZOOfLd5Z`H3<Ou-FS>
z8`#zcN@)ylIJNz7AUbg*JRRC(4~2(3R2Ut2N;{RxR~YatN5yxQZR8;sg=RhE>5`_L
z<>`yp-k3iB^T+qom!q91&jLoit6p=Iw56?Iq~lsf-2CCAWM5!<nW`$kyf5#{Z;uMD
zQeHl3pVd+BUw%{UzVf4-iO;<E;Jx|zJ6~7V_Srtl?JEzym&iO{&9_Qd?qA~M`|nCC
zzI@g-H{JL*S}BOg=z&SQITf-pr_1{tV_Z7k)*VcZ^t#h?)s+pUkf99#5Hfe0!Lt#_
zTp?k)6wq814hkQcP^!&|co1d^9T-r>UhO99*zgt}y8H{Nm6N=6V2t#+q@i7+8g2?C
zyOljVWI72zVM;z17+;PdMJ7j$@`j8MO^(Es`KdVE-$^F67oigixD!W47E;iYEe=OL
z=I5qSECca;d7TqukSfTSMrU(LL@Xqkw9K0*g0WJ7-fQ3}OpRv05QJo6opmBg4+{+m
z!yECEr@$D$O9S&?=9LYW*_Q9~<X#*mw$p(qc$NW*Kr6#5R6_X<CD>i-$RR9_p8%ae
zV!w80PZ=mDPm;v(t*#uj{CqK^vCeTga{F)cJIUok%S7v&cgJGx@(6}hZw&I=jmXR&
z-I#a=q!<|*h^fhmn3|YIsEv}h8=>hm_GTN}p4Fs1@u2Y0@YiTgWkZ`PU{qA8OrVRG
z%W>M{qkt_2nawS}s$GKZnaE{+b4=_>_Izk`IGGhup2Bo&td20(Xy1sf^`kh(Fz-cy
zjSkl1cpsWMkdBP@W9&OrX=E@45!MY(cR(Fz48jWs@w=~{;}DdHpF5Y@3(;rSKJ{`F
zt`2Czb<vIXx}bngiTkL41Qc{XxO1f+oRj`K^nQdPiz3<D4R<|VTRRUvOUG7PQlW~P
zBl&~&I6YH&cVKkuNXbv`^HIwU6>q+FB^E9lp&!Mm8IB3se5BDIJ<ysvpEmm455X>V
zc>N4Te6SCH?~<3pUiQ+6czFL_Jbw5%mY0`Pxruk_HGzOGT{HgZ0EYtS!Ii`Y`Uhiu
zwMiR+Kj7KlJ;yO3IdB@N5Z<%%W*i)@rvt9RFCEyX-Ccd)s{=f3#q#2MY;W#!zn6|x
zF2s<|N9nC}^0}*>xNw83p}4ekra}4I#yPL);E1x$QL1B6Z;W6h&Bx5l)tH{Viqf5?
zeHz+$Ij3rMc@bm#VLW>HASZG^dHkHbcc2{%HfS#+y0r1w)JROsj>qWCSd2`K#t_a#
zZLmK&?TtW2bm_q0M0D2rp+}W$83K8L@!}>~%1b9nyGf@#COQwotTUnUsB<hI7S5*4
zQ*I*L`1Xi4au18a5&3Q+E?vJK?|kq<+`)Nx`v*UW*WZ2z!*?!Dz>jlj>|y!#oBQ$E
zuYMDs{_@lK;?v*8gKzJ}+KMYuZBZ_ATn8H%<P8Ny88Qc*aZSUauzZn4$}QzP*L<F@
zOkWH~<zhbPnvrGFDvPXB&J`(HXE~Wn<>e>*_d-YTTO7&Ho*rpz^4`;Lj2wHF2D9f`
z!-q1e-c04IFe>vjY}GdfUasZ?e!W%S_@b=7ru-^$Uk5;Epw^FE>f?D2a;6j7JKx84
z0`{IFM=|$lSDbE2@wX0wHF(ix?PYGO(_y>=&`SSMAC9LzR~XpH{FAdo-U}zPncq5-
z?k^%eD7%z#9bM=z_bedKQ}XWUhVQA9@TyyK(my)Uy66~-iySIP+g47Yy0_NfoxLmF
z=&vBp(~G0%Yaq{GyA`vSuSeft+0k7cy0Nqx-+lR2tS_&U=RS1Z86)V;x$E=L|0uNw
z&hu8BcOH}9Irc$U)Q8apb@Z95!qt%5e3NYg4|!F#t1ET<($KakY;@ZP{HrU|e#j<6
zw_zLDv!R~QCXuG9Q=Ophliu?CycRk-LkYRC!|vX8PGVkKejZPsK8i>3IQ;DpF6V%`
zjIm|eGdVgo32!gpoL`N}xdrMv9gWd3a59t&u$0jT#liEVERJT=X;W8OKg*Xi<(-_=
z-F5;^T$~I~>i7cS%q4kVr_Y_fo!jB?%9WKBbb(97uH-OFhp=s|Zsll!)*fwlq&}>n
zYevztSK`{WH{#A~@5ejuzLUMo6O%*eC|BHsCQe$k-$l-qbf2Y}`9{3`{?+)SKlwQB
zyzyG};)vwlJAj(QGmf;WfKFH|ewSaIBin(~krS;s)0u#d<sWr|%;o9F{PeTpY)Rj3
zUgY?v2KCFsXNk)fN9I!QGJX3y)0KxW@AIhgeVMj8%i(<+?=Q~syGSN)UrqZ@^LTOZ
z&-TSb7<^32SaF`Wb8Od#uU`2p_N*UN`If``muLC3yf=L**Y94bLzd-T_4ne<i_?5c
zIhI%QdG#!*INIeUh4Z}1`TJ*en7Gu<=Up>Xjelc<r*W$i@G-xQ0=g-Z#{H#Bvlvwa
z7y)+CVW1%p5;zF#3<0J=V`lzU(IpEp-;9nxT&V!^;6t8Q1f%d*S!<XZm3D}d5l_2p
z9Wz`iUW}+T6k+C6(Cu{g0#ph|VJgE4N5UgAECynRKcb84_)tAA&CSN-*jOr6*XloT
zb11*j>xO|8+8S^Zlfzl}p1$6A{Omc#-5zin0WL<c0?k7s$1ZCPzAU<3Ho%ZZXQ8C^
zOhG7<=Ii1NR6}5jV*t3|CmWf+?=u!Tn;_3pmdu_Jdw%Lf6x^Gx;3CW9UeRq!<@jEb
zxW|Wz#_smb&Goi2)&V=qNVCIM`cMet;{zDpHJp#0aA*7)3U+30m=i|<qiS}3I_7Z9
zrl;+Wbt13Bcx?AF*I;-U!)tf~e7WL=ZCy*4s<|;52ma1_@=qKn#Iilx!bbyA!_M|L
zqEpSED3LCm4+Iv6s~%O3rF7&ZX%(`@jY>=d)*&PHhI0Z2`Su8-eGi3$^Koc<fS(?F
z12AAQtpC59{bzS>$(7#;Zg}8<_a4`q^p;6iq^e3KX-aBo^_%_{z1Dn%{6M|BdwO2=
zj8sxh70Du*NhX;-^LhjCy}Q8tfB$^~c!0b~Qi=zGb55K%VRr1;CSpefuw{xKK%?B+
zOEKNq0Y8r<MDKYPXJ%>)fJzvUlBxQ5S0m*#IM{Cn!9iW%2qrH~tuvzRgXdZqy;l9z
zj}WM5pU>Gw*(X@<Vh$sCh=wbQJu_Lr26$mc*5B6_g9F&=pF~&ZsoJSs?Zj!>-O;1^
z?Wt|Iw>H(L%3BH!Hn|8j2=<uAuyC3!OtFh+mVi-mgmKPP*j(Qd)18x0EqH9~7(cwG
zM*S4UkL&QzklL`x@M8C@=YqxJA_ESX_+f$*!i}c5`$}hAztIv`uQF{f3M07EzB{|@
zzQkq<2zKF7^85E6$FpbKv9fZiyhDo2iSUydj6K2W*m<q4A_#27hTzIpvTN(iRP0EA
z-c)~H75q3QxkWG;Q2A<7a+~9LUu~#1#-4aj?YAHy>B-|K@%Zs0<@Yq^=H_C4X(?8e
zmZ~N=_Q%xpSX{j}9XD=Yi(9v@$IQ%>cVxlN-236zIwYXUt8tk5KqzUEGC@B?$zuHj
zqs3mtytyTO*@zt}>nzegR5<IoI9~^kkm7zSxL8TXiw^K(z>hUKSFT=(o40SptFONv
zKl<U1<K6e(_gNU2XJ8Wc;QoX7?XQ0uzx?N4#_vA<I6nXEGtWLepPjWb2rinrSN#fS
zI%DrAwT=F(eG2_BJ@tKh_g!d+@GyNcHC5T)6~{Mc&go~}E<3{N!^?LUzONi%g;p@5
zWnmk-Rnh+{lKlB?x)MN`>S3Oud&=uqqBW`pw1<O3G5^WYd)2>LY_(Z!A>kSN?!2Yk
zA++*2bUSH!!hvuEL+B0Zmh*G?ma|8^m|AIjCm)nV(Ww)4bWDs{7le6Xe?Nz~HL7ol
z)`<qRwo4hX5<@w_69x{*j??egJ@B);y&ugD;xqbz3%gspqJxw{g%oJW(;FSs+H8hJ
zeOa&%FS~j3c1%poL{E1w8cM7#F2>hiei8S+`9|rl#IEpDxqCKtbYhHJnP;#~G3Wm<
zSOaZhTXyi20T|9HVGRRe@Je7#ThV5?<aGt_C}Or>X~#Mgb(`VN-s$<w8!r<_@kRbP
z$J4}Nw~y!+%GR3b$MeTeMc<x@4$OO}GG;WqUPAp28rj#|Ym=_2$;lX(ynyK!yF|8k
zc1oEOy;2@*e!!U(M}i~xkJ5|>H<6_*Pq}>f>@T)JkOyyUkpLW2HtObbx?$5Smo-e7
zZlU-rEiuS+z6~=wyFpi&eP@P)KERCZK)>kL)MVVac|%IY6|aS1#sxFbscF?^d@ycY
zpNjY1eK+2H=l!S+kGovblrS$~+C*iMte`SdPeUHKtG-4W;+}cnpfD;A4<;Uh8jcSu
z?@zfeuKxae+=cj+qg}txmG=Kj<Gjr4UtZ4~_jT`iy6`^T<uCXL6kZbbf-^p!#i`&{
z=g6zzJnpO<PWWQtS+0xWXYs51beHRR(b1HOpzHU?$viH5;s5w(*N1dKgiJb|k`Zwb
zO%YNtGr)fI<Kx&)cbYLWJ7Hx4%}NNxg4a@A@;|SD!i89jLLlG~V!+7zqA)rx4q<|h
zgffu~LZn;56{8n9Dj}p*K7=zmz$j%BrR=0!5jZh1p<~h^i?9!|!!y0tDdBi*pf^TE
z@LE(LOdFO(NH~Pe^-#qDCeQGKnd#xUdFzUZd0X6j@L0mewuok>Oo$MeDJefH<Y{&T
zBI}<c5rSzmBkx(3Goq&-u%I78J3RXC5%~PpQ-q`h921oG{<egjIUa})ERt~I>s|Q{
zm@}=;$S8~e^BaYF6f0*Xw&KatXWq%h-;kE5I-}Pxqi*c+r8IQ+)pXC|H`S|!t$M}q
z1lv{t6+3e5zIS%8QEeA<8I;h;84R7OGek7o+BiqLW;B$K?^#QM&3Uy@osl@<A+Swl
zLIRkGY*8KwPr*yLbc_O=IKw@P1%wv`fIQEXiPNL#>fv0BK7D&&V!X4*yRbsBwaLit
ziG}&iEw*l5mvFigi}Mn6?=8ms>_QCmia-xeM0;~bj7gyG74vM9l9YltCPJ98FtgGW
zUEOSDik&r#0}~n6Bn(LDU@KKg+b0qlcBRy;vY>9$qh%~W_tdl$SI%>QcVYeqE`gxn
z8M7G^Vsh87U5%?(rsBqpE4sfDLlSf`OZx3cpTt*Re64cNsW0tG_?Pk`9OL|qt?iB2
zVg^FOJPL=Ea+Mi;rVdLBtMTyuQapQdET-BN<6}b>5?CCF!o-e4v|+71*Lc|GU$920
z<)mt6NDzj+enSd*-#~Y?)g;_YksqI$h?y%h(c2IF)!&sb0vNOfMS-$+ba42ils+l1
zwHkY%qx`iBmYq_Fx?ENS9O5#tpyCMT^Z^cclX87b8CYX6=!O5BH?g|H;e(s8xWKlq
zD^@1vp3h1FdLHxGDsODZG3$x?+ND5D#<iPQ;@XX?HV2%VR$F(r$D#1$P;GMnJpl(9
zIhB$OZnF(?M@NtPQJZIFnAP<`ci@-W)axN=Tfr)KMiI?{{;97E`*kk`n8W*&{?Uor
zUv1Sp&=)gm(=jP9AN=@-@x!0|Bwl;{&6u2;j&2El;QO<u&*O{FzKD;0^HF^K+mGU#
zufFz(=DKKG*>#G61N90kk0uwU71_W2+=$}Nv%;Zi?vKSM{U`!anN5YJaV0*2;Kgr}
zK%L#cEUdbJnJWym{z}{;XqCQS1{4at#3}XsU*&U_hoW$ww5QNVy{kTT-6hyzvWWMY
zKN7yxWvxuu%uGVG6$uHCB#S~#pD2g+;V6%Jm~flCtgy%dzn1nE;mx7bz^sW~>DVO;
zlR{|TfOym7lmyqwVfB@E<t9F-wrlHPgACR<P(BInjQH!V*WXAHL1|?+gV{OrZ1n|Z
z<>|}pR_NWvOv4$#NG0jCG3#-3ydNWDqvG*Xo<(j`o3bVR?>_n{zWCycSdr4ePG1dz
zM-OKu^!HlHg@2#~gTpr2s}v?c%A#m?Rb;_AgBDr;CJ$EFv{CifM+&4(vI%HYx3`07
zMZK_m%G>kz<@u~#bt&!YYjqE<6rStn$~W_!%1nPd*x!{5uxj(CM-Lvv)5njkU?9M!
ziMME2caIdjK`RvF64Hkyn6oV{n;ZZKANr_bur?Cegg!|bQA}(;P7(DC#<|Zae-EUo
zy{%LtN1%9|^(&QIu%aF)rEVk94DbYHf<X^29-osE<Utjb2{uY#vxdx{?!ZoRon5^u
z`)G`fO~vrYc+7Bi&djyAef!mT<<&Q1c=DQfig*yTlaItFMa)u(fF`&xETHJ>j-blm
z>4*5U#>?;BHef+sac(D?%I%cT4*LsS#?5PSf9bdXeJ<0iu&%n#=ko4l?(*Ame_q6l
zab|{e79-uodZjO}yes&QOL(|d&feqB@40?AynL75ArAk3cQKv6)HO|vb$>_(*@1P{
zKM)+j943fq0t-4>;KlZ->^_5F=geRv5ax;jrQ=#vT7a+(Ih{)P^?*rbB85VTJIVk8
zUor_smtnLB+nDQ6o+O@(;BId-BS9JDD1*{S`}xXr%AG;sdWLAgxKR$eqy!AISW?2>
zp@;>G3|a8WjyLot6bZ?qu`6McYTUR!9yf1aRi|%_hmW7A6K|-ai;0M^l20-vMx$~x
z(e<RM4H}3Tgo3F;;FFY-pSzodP5N&|NKE)3hEyt<XT5tFG5}7V>+SkVymSe3@+yqi
z3k*>nlqWm$;K>_Pi${;2$2Z?BN$A?uH*7sfPHIDak6rkvgi2;4`Ubj0PfkQo*`eyh
zM6pMVqe}`Wj6HWW!eFd5Mn^d4ZN?@rEDSVERB;GBCP5aGbx%8EzJ_u^u8id8!cP>6
zG-vRlYcbw}Maj5xv<sX;aR86F_lU3rdjw_jLn%QCWoDvBjCz1Y%7fiv(o#52j$*$e
z!W|M`C8w*+fEzG7a#{9v4#apw)VB{}VpNQGU`#TY7`)28z9QjUiU!;At*&A!gkXJO
zhQT6zY;QpgZx`C!roP0&V2KJ!lSRMVoAO&M7H9TD41p;LwIO(ilILB3guk3heRv?j
z!8^5ZiZ%<k*)b|l9q&_K^aJdKKl}Wd+F(nH1PbG^cm!vev{)!QIu`Re*p1!2oEe}E
zKz?Q(ICSj!(^U!DJ1T2K+_^m!6XRpi)75R00~j2N129FQ0q?!-qC6w7L&JkM3maCu
z4-T_?RF{bgN-XP5#>A8nUf2p2)2cGlB;-(EVSC|@7#Z$YeK^BnAjZb~CFl&q=qNiA
z^-9nh7aVBsW2ZMXG92v4)YOPT$1Y(Uwl=Qv4hyGhs?TApt$>fL<Jgcgu@dvPr=E}5
zxooc^DG8n3EfN++#AI*AtFPbnUhWf<W5Sbu@zI7@T2cE-F|#?NaF1D8c4$MHfFWb6
z%q%YEW~}#OpaMmL(L*n&X4LaYxGT-)RAo=|9Pl*-F^^tLamQq3Knm)OTQ}pK_ujRC
z{jE3R`i+}X&LqeQkyw1r89IOX<PY)5$G?xyKl?l$K6n_*Qi_<-P1=eQj?kN<jF{Az
zESL<5MnG%ilQt<>S9|E1!wS7Q_ojJzmqJ?hW4)<p%q5!TZ*)~2<z4al`r>oFgO1|)
z4o94l@0rsp1ld$3s9Tj5f390q4+0dolpzH=2RRS!N_>XHp~ol^(6n3^CfeMl4@?h4
z3&|5^q6FwSUS#h9kQ@}D$S#xc85&D|Sr73hrLCGt0hUrKn!<=OgBb{(&?ja^hJ^cL
z6C=XwKAYAxwW{BAsqab|XI%otR7so%^eT5cOn9h^S+O`0jX;rL(J|rXEBcOgASf2l
z0M_9MOJYZSkOkE!r>qZSyY#QV_*{zVZ#`?bqdt0|vNed7b@ip0Cpd(e90E5B^P$P~
zVFX{}9Xv*Tl))cn!{9BnT`Ci-y~+jrE*ux8Ap|2Ug)Xn!j!cT9y?Adp*14)3FS{I^
zN?-7xgfTc~g@b`Z0nBAo9e2Ew8QY<+#Ol&=&IGKih>z^)yFBEvr-zL#pl^dx_IgCy
zSoq$dd*q`Im8HS$VtI!40gpv5!Owz!v|Uw5mrH?uM}OiiHByk_#cU5eB=IT-{g`q0
z8m|q^JkzhL&GL{t)j{&vrf4`Db<q5*-D+grUyYshdZK?|B*sQ3MN3DNe!FL~8c-C_
zn%SMn3PToA2&-p>TY-@G&P^a-1$eGRuG$TER=>(o7aUDj?)jv&&&yoCs~)Ptd|%vp
z_m{+ZxhvmY{+{RWPqY53`dH`X(6!u`NR@YRb<Ow*w~KW+hh>Hrn3Xv7adGwUbf>t*
zUv&9=v88Ze;KelldtB4xc=w0gl;oL}U=$PPjxDnAH~RWIMO<6NNZE@V(-C1Q9fq%T
zwV-@ffV#rdnW~Nip(q`SAC6~gL|78wOhsAiJE}^DO9nzZ5OI%0ODH8v1a$}{&Qo(r
z$J)I*K~q%*2$#>QP?3iRdg9uZshE;L!-6t|qV<i<*xHw%DlGy#RK{2Kjsyo~driv3
zwQEzhe|-4(S<El2I?dJ=7+Yba%0poYwXSDXHwp-`@aP|51j|z^%F0)aP<R2hm>0y&
z1W#|y9G%tms1Miu*@IG(0AqNQcO5+awhA-s5LnchIRf3_QKMlAVwkG11J1sPDu=vn
zZtWXpO-&+-5<pnL!gi!E*Um0RJ36F9U=NOvLi?gLQ1)gquhTf#J&u*7btzsfNN=_R
zGBb5eLSByut3FeE7DCB(gD9!od-n^K9Yp{E7oiP>00vaTAp9tl7LHU0aK_4n%9nA#
z4MA3I3A4_bi8^IMf)+wscXxZ-xH%nfz4K}ekFp?=nXWC50A9V`5m&AbM0ZcC6%EkO
z%8cY)@lG*swZDini$2-c`}vdk`0Dfd`0@+38_WZ|np=(q(<Yn2AWX9Wm~t^YrlwGv
zu)%=(l88Rn{XHow>^Pz}%hQk<O@_e;rXt=fy5?+?y*<n)P?}DKx4M_$in0UzPy{dw
zeD-WEKL7l&(m0TU*B0HK?Bs!P1kM8AwB?48X5SQxP=D#}Wbqz{&9Po$TQI~F>P1XV
zPIxu|Wv?jw2n39t>C<3WD%&4RDB`e51YYm=u4@nbr%O2iAL$vGS8(_gGYrhap~SU%
zCPRvht|%CIgsNc)@jf}WzgxJI2g20^&*70Fn`V(7Tc1u%UzPGPq4KhAXnXYcN&)B}
zQ(3wsARNTP+)AvhY)F|{kM%WniV_1B^IcurGK|_fT4G{qFmB(y5_ey_Eq^zrXQraN
zx68sLcIoSz>+x)M&a?t0qP?Tbvk+D!q{N(@HVHRajGk*eP$ah3H*zN=F@JD{(MdzW
zIGg<tTm*;d+`x?OYOh_t9&fz)X1xEw58{ot-jYHy5&eS$o)La5CGneYzKw6c{x-h)
z@~imr^DpDcqbISvBqc&Z1i~YIF=>;MRePtx03Ad@g|1b?RVB>i_40js$cSl4xjGD*
z#x?IvThy@`P|0>qx<hB?zr*}orBPkWJ+3|+N1FAHaA*d!z1$NMLCx^1OCO(fx4M#M
z^&aPEf1PyEVHNEqj6npH0PxNDXtN+G(ZT~}JW&Gi@H6@Vvl1Mx#-d^KDrHsDrJOj*
zh_DSrC>L~%zKT%j#rKRVtADFUL|0eMG;<Vd0MXgT<`dDPns9)F&T1-bZbZRMr*d%B
zRC?3jTvvudp)#`NHX~ztoBZZfI4gx)yvz3Mj6BOT3(Cwhw0T{9<k9{6!mH1GC>jF<
zRt%I$qw3eg*)jvz*Efh3F(Qs4l*)uFF8p$Q(HQuN%_oKb;5+7@Ha}H;QKE@Y>B^uS
z(071bX*YiPLa*xD5SQ)9&G|d7)mQobcf4Oz3bP*gljoUP3lsnpBxcsv*H*>HmTUsW
z`mxoORr3_q#X%1-31gx5(9lQ>4i0+*jgI!3aRR!HVg`<QF}%R<tSO_*segXTX95Je
zd{Nbf@`3N*P=N>aD|NM%i26K-o<h%-Foj}m#nMVF>bjz~+}hf+soefPypz5Hrf{|s
zYiq<~#d}@|Pf#W}fEcC2L<6}4rNZmJP$r76lnKgC9G4lNb$OXH%<EjXLEhu2*Y~;0
z@BWgwe=5%RyYkKDZ+ZUy^e%sud8;jT{r+O)3&l9Q$9-S?Oe^DNobp}qxXbtH&cEf}
zm&zrbF7EukKArD#=jFNVZ01C2nK1;ivJ+PiU~27Hf_1lq5=P8e*yzz-37BQKf>R+n
z2u3%?DV6*w&FmZ50sNM$ZtxbIx%zumP8~19^jv+LN{S!PJaN5<42|6*KNaDT$3~G6
z1UjHQq2jWT4YMUunwc4o_uhRg-h2POn4X#n4z^oZl(4b6AG@j(Bo2i_{mV)@+i0?l
zB&7KM!zWfIC<D{X5MY(KQp$5QJj;huc?8FEB5)S0XI=GN7R6Csn3o&KM8B%5|0x}p
z9LGKNtkNJc(y4yj!(0=_j$o!%JmrJX=Zu3GLvN3m)@Y6rOGtyDGE&_?&=#Ykof6!z
z>BYwMDAv|bBrG1rj)dGJ2|#Un*D6NqZSJ-h_1%fZ`Hh%=zAD0d7_(2=%bRI_Ok@V)
z>eZVfzWvId@>Wys!ZC!#+#(icmS#$fyr8slg$d;ds{}9!N=zfelv6>YevH-}D_qQ4
zxL6i%gKsDqxoG#m%nBuhbQuM#b+*S_Z@&`nz4w;QwW5&)hP%->&=far^~c>;uLwQ}
zuo46$@OtVPX1gXvEoEX|g6Q1yrFi;yQHsRV_~vV1kb-&tK#PRYZqMLgV~g<2j0*}x
znqZ}8TG~pk>QxjW6lE6AZi~pHK)8?5cJiEWvM(iqEtwCG8YMu!knqYHCKf!4v25=6
zOp33+dK&lctxLdekRre|w*+6c6-=AaNMH)iF{+y8ciGN^1DsHfnb3(9+|F8y+IGuj
zz|jtTi$`Gsr%)gk)qmC`q*KP+WI##>c&RpZAJrA~MA$??BX49b1ew-0cADz-HP<0g
zM%F{^NXeA)(ABPQnl13Kwg;T?E?Vj*J|$l`)zdR1VU9CQ=mU!1+@-QMdELpvf|#oq
zD+<KAg!^3y3J5x0b92%t0j*0)#H5tUyYbqaug3J%>F6ICh!)``=2Kg{QVztd*QGGQ
zkeN~H?Ch0rHxRXsUR6s<rHID96bTf|+>&xJHda@?t@g3ng+1#TnKZ_#&tulbB5Yuy
z-Wbh-$?(XC*FwCF9rOF|3pd~Kmac64!1)hPpFWL0{Na!B$?rdjFQr60e)!PqYcPvr
z5jFiET9X1XoebPb+6o<FW|4lM?TsJ>T}pw<-`Oh_yguYxlV0CY@N%DVe3hr489x!|
zOPrTD)Boz5dY9$Mu*>go<(u;|LOZ}Q70&byB|ioG`nXPuXF;7?N0zUG?FwHaz~4|W
z!4BHO;12mfC%t94aF88<5c)k}0iA*O>2pq=_8ApUGig#$tY{DNpo}PvIPxk3Ab1~|
z3QbV6u~jv@EU~*9heEbU!Da@LF4*1EExIM8Ugc8i?zi;KJRbuAP~Ve|_%$0yF=Iix
zz{b@V?pZ)pn(W}j`U%c6U~q_j%9@M$XLG`>M?N*$2X?CLP4r9g{?5Lh7#bf{{}@za
z$OdVv&y1S~sFb$Xn5bz`&XlwAKQ>(lmAYzOrX9@G<w0@4q8j|8O@)_k9F?`SrGbj)
zQiYwplOqqRcGi|Se|8?{Jgy%S#{n1F=^pyTU<_W5rfaNap^TK3gW*sr=AO;EZ)2{A
z`2r@nY^1?%P<@<T)!QEfoa@rt7aSJJ`UOGb9R7Bc1=!@3@^IyM!4WR+&q}>2NBM{k
zo|%@2H|>cpZ;1~xsKa5XMVVOjzzJu@uwli9Y1oc<%c^$^T+6k2n8hK_2_Pt>_RLVA
zJa{HT-iic7n|oR_@y&t}u-An$xgB(mbB(Bm@AN;bl&)MWSK8=h?((ab#l0N%r_y`5
zEAjJM|Ns2@UH7sWFU$J7<*RU=V85!M_MLA_1mJfTxw@W*r>9*my7I1i^8d`K{=T@9
z=AUxs<v2^W&Pkc*{7|pVWKbj^Iu?}0fDq_7fT~kWozZKAp}oDs;M@fe%8OqmBvGhy
zBue6pGhvkn$sim?Jb4O2p$IF)KcD$7krHL3I9Gr;FhYbBIzHd&nRm*?<t06cKLQB%
z>}!ued8{&^A6=EOz=+?Ee*ELOe(RRYcJJZim|b3vO?5Q0L=}PRp-iw=ghg0{fAipB
zEG(^4JmpV0P$a7DN`ZmAlCyWwDQ{IDhv*NXHa+#t1+^($5f3J$wEWAj3S3<O>e-_>
z_>5s#;i`-4oue!W=h(uhFs|Zxnw!INuvcfqyw9gI_qxoCupu0>c<b7=YZ5{wFxA?W
z?va$J)0mro5sS-dFFr8D2y(mY_x$;6eEHe)`09(5c=l*L9^G4&0Cy}QS8292#`wfQ
z+_-)xdVBiKfGM-<0lvwjSWsar5@Jk@{uI-tGyzFroBFV;P2y6vJA{>yauf~P0Y*=l
z{;XgyDpMBmlMW+gzy!0SsY!O}YL<euE@mlVdZZRZBOPMoi~{zU*;=uYV4I_gAp*>Z
zPz1Xatt=hJ{5<=5r-FfiJUZSSuf8@Euf2XPW`t|EBp?nAviKaK6Xr_U!1HR9U-h(k
zoZFpqI;2D>oU-~{5|k|JY#|4{V^^)y*i`;%Dvv^P4hXXeb92k__|Zzt&F;HAEST@<
zW`VbuwfX}}n-_;;nxVeTRV5KznTc$dBHrrlfyYM1qpx?sWnnS+>Y5r-8PR<a8nDC1
zb{F#u7#Qz&q-atn&Vc~7jlyf%i_IEx3LZvwpz`N7sf_YcCzO_)r9eSP!DLWE0-}WN
zgTsB}27(U6Ae9jmPi)JFqOTw2r$c4nL~t?5ZSY-6*2;D~y#LG#xtGLPvC&2m!5jnS
zo%}kx)y6}eQX*c7cYpA9+`4nq>l2RD4x9nP$+##Ht136<0+<8z_YX^0kl@feY(*kY
zggaaNQXIBoU5doA@>*V6jCE|x*_$7MNs3E);WLb1VF;xNF`P%iY{c}HE22j~@p_0i
z-+nu8-n?a@kwvv%eeqR%`iIZrqmO<Ye~=>a<k4gAw&sQM@}6yw1CF+Z7k&U&NCPk~
z^aNS~-DD;(`wp~3Ugh9w+N>K3)VhL;z?*B*>#EOI{f;Z%0~84J74abqN1XaAuB1`#
zOM3X?%6-YByvOrgXg_(9H}uT{Tb)yE)0F(K3Omn24!T+1U(AD9I?s61;rK=uAi*FS
zGl(eL=4k?R`cxJ$D?-VJ9zo~Iz*xd2VQ59pX^VDQ@XxY?`)#6?&@N^UIlqFLUue+$
z;%q#8^w3+HDiOt#@Za3#*<#jUQ5NV*?p&yLljp#?vs18D1_o=$pFV|x0v~`ba6ZQL
z^pq6jPQ_Q>-bA=evn$%4evCQK=+t;zy>%n5+_+|?CY1uVWfr|+wpjhugOF0f88C2u
z)S7l79SNcx?QD_FcWgEQM@{(H)K%Zn&naW#9WcAlVBc}TH6g2?VVv!0xI69h?;W@D
zYCR_n9dcWMV;ReBO<Q=TO@M?Np;*|wXm8hMPs<DJuCyd2W8G_1_?lCwy9LA1;o-P4
zbH(QiL2I#oD7cD;N0EtP<MJg83ZCnq>%=w7#4X{HM?B+E;;Ee#lm1PZaaQoRIWF`|
zZR0I+S5~cHFvGMe<${?Elm}+R@tp9_Hs0)xx3i^o0jH!eu<q;RM8J5F{e(4MKD-Xl
z6@>zUf_l*hvY;6xPx<o7KQ6i!?|{72yZZLJ=zvjm`#tXazx_*H;!C}!jCuX;JBRzL
zbmjV|+~sdCI0XSiy(Qvz();c=JeRc87Iv58oPWoau2>u#;oOxpzWe<}ce(8K&s;^{
zCcS0?a~mOvF^5IZAj^qIp7-}y2-z0By}kOFkrGCaS!5*K5+Zp45(q_Mg+V0)p_u~-
zCj<ghglK2&dyqBbAppC>AzaC4xRY``It?B~rX#`l5Dv;n8{f0gmE9wHy1Jr6jFW|;
zbRf<^K-s=^`=*tN+poPEO^O3y_~U0^#*?|F*b+0KtCJ-%FW}$=<KdBE&%QjHpO5v;
zwd^D=2SkVSC?wD#=*rVzsa)uW`W5EPvmNyyfpkTfru@VyikG*xExoexnFdVBNhVD6
z^GHP%*NRF-AQ3hY&}j#Do|&ASjO#bA$E1`64o^avzyyWSUYG;32%NbAG2zg!nQ4gP
zgm;7(9%37~ZRNSIv>GLhwFRf@qPRh@R+ib8Q>7CCnehf6l(pKUF(#x;+`jdSl?l&e
zQNOAVfFgx|F+WD>V8r^i;1GiJu^2V)se6HADrxyG?>#di=1e`f5}oj3UrKtU1(QUe
z1IJi|zO<q?-I9jTcqCy`%s|Q`BXz)ROG5L`t{AH-f|7!X5du6}OJ<h9C*`a??+1q2
z?>*w}cciTT0DJps3A;TWx#RFP4o_MUoDgmiz-ed7j=2tpLb34J3z*f8v;%@avo>D0
zA>mq0Y9*(yKi41Z?og(LH&@BZfeDtX=P<hO_T~#S?xv<wDKwmBJrx|}h9Zd~fx@=2
zE<soVGpCTVPk61PQ-VgP`uJfydGa(C7FI0eT)8qGqxu$M6o;KSHp0vlt*jiy;u4Bl
zgJ-k`r95GOP2b>r1B7yBG&w_J--;Z&<p^+U69jA83_-i}1xAGtcG#_iwI{SA${1Vc
z(&v&M^$JGNqfRlwQz=o)@%2|vq|`p}iPsMwJh$C2f<Z?|tMWrw!R(CLj6C4&)i+*?
zAN=qGui@AiPAssKlZ1uE<>gpdl=q_SEy2IZip1dHgp{Se=&lWjt~A8TA~SJ|G5?Ik
zkP9}2I1m#*RR=JIQk_^>>mAcD`H)8$=A`+tsfn1Hnej}--~8>rj-UMOXX=+T#@+e(
zrTG2tK8b()Z~qj(_~&26mtTJA84q@YVitnM{Vd8wS@4cA5;!GFivCEDl>mf*QMPr3
z7O;-RN(6LMGz$FkAdC>%^r@mz_(G+U79|~{@6^*@yq}dG!n|i>t9ZP_m2dFH)!*Yv
z{E}XAm(!@fhlfC~$d5jZkjIsR*MC=k1rBHTatu<H(P9L-Gp9JFpW-RRCtc?u4}LNz
z0Y8BR(Z^sGx>cMU6!0LMXhqIg&~KF|SBQ}!B5-o(Bc^(sJ%ghU5*JGV7IMEh&5>sM
zd~+-9$m}2IBCxJV^hdPJYZg@w6!T4Zo0N19(uj7jnM30#GmR=2GZX?1b;8V25OiOo
zA9QwcN;(HBw#S5Uk<C?_8!?koy_GftI+TkUiN3+XxOV%Nc;+iHJ~<(t+hSprUE&b1
zSq}tVML9vyAnjZ~A~@<tIx3ZDPP=J5W?eZOgQ6o@LvfouORD_UhQL&x*iogtsRHB-
z+xlmIdZx|JyncRH#Vc1r5z6TYc()_r>I38r{b3*#b1-)F@_^(E6h3D4nta7{aeZ|y
zIK$%6{Rc5WI~Tj6>8DZx+f|R@;lY>|kHuVVV4z>Ls%AMwAaGis9QiT$2~SF&eTXYK
z$2H5vKi5#ONKZQSS!kj+&8hf3vK47?rDDu}?(Ji;wQK&eu_<~b8HOG6Hr6*Pr4e&t
zcAH}j#+G>bMyxH1AFZi<*LUMca^60sxcYuyiqi|lZ#DgBQ5nRPXyXcY!1k=Y>K)~(
zuCx~3x5d>xl=gd>ds+COi~oJ^F1n0&h9i8VTfWqr>b+l`IX|AoDgXX^UBTPx@hq`~
zQVFlGL#b0$(&}BgUW{AZd073s%lF^??xH)-<FX^Xyla{q@9}O<AS8qj<YvbfI(qIl
z0rPJ4DZh-C+5|<u!fG+EtelA_u4<I3#+ex%wIDvvC0yQ}N1l!g0jfT7JjzrU2%b?%
zMzk5tXRqy23G(8WJP;sA3;SX3<1S)%0AbbpyRY7f_uhY3!qH3|st)X-|B)E?V<{6m
z>L{d<<r494Z4fcw(61iPOe{&WUy+6nA*arCB*Vb>Sr!umD7YNYtb-t}0+UQT<GJJO
zbLwcOjsQ}Kh<+ZSslMZy-}vz?os6j~!l_`$_YP4|j$EFIA+$<}9F5y|ZpZDrccdJR
z=Bb}UERO9?LM{Ohrr6cl6J0e<Y-UCxXA4juIkXMapyp<kq2b7a@{9ybMnu^Ydv%4y
zc>3TtqE$2oMNHWXhEgUb2hWrVp;nez1vCK!0CG9H2kwDI8BU!UrNcZ)xf-3kRpuE?
zVB{>IW5ieStc>6@9qz$X6bkY~=)vTPH3=+mX1aHNVKbJ+7!Hp&qr10B3R0gMM1)vE
zAj$;mgGifZVYd{?nju*{dU~40fLn7Rx0DdpSIk^x(d?wI6M|2-(%(}0bHQOr4~1uo
z5>6=tQ{ymi$_{SQPB_d6fep-z^B*upup%KCvev*0nuq$;7MNbpC)vVK3}5h;?jZq%
zg94dlK#@wX;;Zb;EONF;w@>t?A0u#LXUkTu&!0=d5u>4O>ry7~i9!DHkDQFWsN&?J
zP!>L;jG~YXONm6;WHb{~BSs}RH}~|da8SY!r)V?qfD+FcE*w0^R*~QrvlA!?v>^g6
z_-Vmpd&>e7w%4>3TV&GUl!>-Fl(K`(FKxhDj-Bm&DJuI?&bH&h{rULhcVEZ1-#(44
zb#}6Wc1S4^Eno(bk=UA)W)2LSxq2n0gtshM@9LGZr?!1QHyg8JvU!*l%BHk4wF5hC
zfwMzHQoN<S937sjtcOw})|Kbexc|+An0r3w(Mm=sPZ*ummn{hRpta!6Opa$^R9{A!
zIj7{#-MjIVzxrvs|HB_znV7k9&5GB92al{o{OXtg5+8r`aXh&9z(UKm6%+(!`aeD4
z#OuMRpBcHy0JKE;3K{6j9B7v|{Lt}2N13@$$As4Khj;hO{#YDk<+*siExyDn^oYKR
znO^-j<-Rzci>nW>zpr<MlQypYTk^M%s;l!?|DliDgsiS?IPM9<nctLm`i}U}o1`ci
z4%JCm>YFNm($2LYG^Jw*=Hfs4@x4cx^)6`_Gi?lrC66ON^06W!d;*Vg@FnOd3I)Qf
z*MsR>!Wo!hv3#Qxh>p${(Uf+ZJZ-V_o)qg{(J{<S_oUR2E(0hj6bly2axjP@&n6el
zPMk1P!7QlC4ZOe|<A*%6b3J`*ZQPc^uFB2K#-S7<Heg`Z(RGlAWcT*>$1vuw;sel7
z4mjk%S4^CknPA-nd>7MK=nZfMmcWs%#<3}CSARkoLs4OSZPwyZF6bSz63ntVPvA`(
z>3ZI7`Xhkb>U!oBCBLh1kN~;w#KRFs{2G|Bovq_4%-`}_UJ!M=+tNXO8Kt60<;&TM
z<Ug4DvXdN(&{tQNVs#nwMz+PCi<Pxye+SL%lkXqsQ(Jaf79b697{x>-Hmty7oMjv1
zPs%)8LC35{ma(|B!T?u!QZ2xLo`pxWZ!R*252q}|0NU2pmOQ%!LT~XqHc|F@Ji?<F
zrmw6kJoHQjn|JVDNF<o)y=a+dn1ln3EG|@el{mo9E5#FRvrzKoURVE0Ju+gpZE4^7
z`@g`soiD|&hfn>x%kgmK`}7WR`1w_?l%EK7!8kuEyR+a+@BjN;Dff9I&N2&j<}N(z
zKA+Q<y1dMthu6O^->Tg%gnysAnEv<W*EBuZ`C+3+Kv_r&qY@H=93oOjLasw0Tdkwn
zrz<j2N+-e`1-m-MRy1`AH)JA6687w<Y5qe;fS8_&V6sq@^%#X<@EtMhE0*8*nvY57
zDVU<r0V!?PMe!6j*LpBJk>5gO8MX4%vk3N{1Oei|`Wkl0Z%D8hGn0Ps@KOBwHy_39
z!itHr`mY#<lGT;Y&w{s_1a2lg7qN+6UQ;?KG%d;*6DJXI>OwJrnd?ru^eh-a9I~!i
zAD;19zY1O^8t3KaRT(jZ86*t(+9_l{z=rtpywg|JJ?Ziq6CFk`PesUjVnRaTq<m+W
zlm{5HgbGGBe9nQc93;n#M0;zSXLQ)+5TQKn_Ore$Jmp)a*%=w;6wZ$JE-@4i5kr_{
z%U}t!5^9y4*PXDoWomjf?%a7L`UVCQ4k{bO>lkU*Wj;}wl7L+Y@?-P}xHNg>1QRB{
z^G`Tq;-pAs#>x;Sf?b@LJ>Z!wJ8=+vY~(%4285IsBZx5YGBQSd6r|3sW^Xdk-Gluw
z_0GeCFw<cq{9wOY#IGXvbTfkAqPlZtLWk<r9mAvjUVFhA9BhrtLiLT!wRraINoD@R
z41ffIU4-FO@*vdULfMH42Es@_Er<e$Nl!CvBY~YQC^_Y}SGb5W(NSxP?miYownSHd
zLyS$-;?>tD;@x+z$L-tf>I5TUH!Bn<@a$N)e5^8d`s8L{ytKR&_aDy3!^eB-3#y|S
z&f?-`+`lg+<Jp#o|7lE3_DfM|v!{)KAqoNb*Vi`??d_eu-`P13Q)L_7lj!R~@o)7^
z1f!_59kUn=6d-gV^t4H^=#qlrLkQ`gyW8rkQZgh3wM%*Flc1loObAQD_frn!LKta{
z<NfB?SUrq|xh)CYbMeLJkK*BjCAAZ?Fb#rTqZh9Bt6ezheQ0>dYZ8Vygi4AW3=svC
zbre>F<yqIjEL}HS&q|RQlman*^+rrzxv8)rG2z2_{&X=GXIJ9Uw@>2ny@#<hJ0JU;
zVlF{S38-g^8Y+LfJR^{M60(k@NN_TC&(L68xpgx>_~~E8-~Ii6h`X=7DTRJqibuP*
z9RB2^PvW<~`b~WE#n<teU6~da43pd%`B42c=>dHtdw3F=6vXJyqJ2_6TZEJCooX)$
zi54m#XwVkX$Fq@539LTAvZSQ<q)VSnIt3j~I#uYn5r#WmEyM^9Fx|2u;cwFM9)CuM
zF1hmlWx{{GE8mxRc=`u3Y6xTQ%Lw}t#E3~5OBy9UkJOQCdX-D>DXv2Om0u7lxWHe`
zX9_Kr_xE~+9x-zdT!`)&RtUe+p#G-YdP)VhS>Gd)CJ(aN5;O@OMSE2Oo&{|SlB_in
z%&RtJ7L8rXFgq2iQ-2f9+GpoC;gtL_{aL~}CQS_|%1iy@gjq@s6+KqlNZCdMpj-&f
z1s2I?pe-nyH7oPr4rkvWL~gTgaQh(kchl6AzKQV5A$jd6A444ACVDTN-V*QJK@ep#
z7U+R+fmt%jPTg><v4VbsBb0-~^|*(A<e5DP=9t%Fh6#U3m;ig4ACbX$puhUIE!WH5
zn999lRNr291i24z)j$2IXb$I<KjF1;Uu_K!moJLAc`x7i1Nevd6EgtfGaH-hqN$vm
zKPy_eth6|1CoLt|+=2lW`V_c=1E*46NjzbIz<iHl$$N%DdBYPE?<gqQpZ{D|!ssX9
zFKz7a=wtMM>X3M!1C5lMzE0a5cugF9XooX!Hn&A{Fk2J89h``sV$#YcDD0?+QVotW
zThn4#kam%I((9T0C=0CtvQbrK>BERAyd1@Id8h#oitAcr;Q75Bb>W*5H{Ip7E$=VK
zEAG4RFMfaS>hj>WJia8&Qq-!5b*?_2i}^a}Qta=#D#47LS9|V@uYV{m>vS={3byrL
zVW#UB;(fobcb6V7IO0l~sCjEx7H5KN2<0sxqM(T#CEHhabTHN5;}6*19YM9BsZD~k
z2!I-v4nb8Yq!~a}a2nlA%0yE_gqQ&%uqLDklf+b-1^Kv#q1ley1eiuu`V>%q<VEKq
zKUdU91zLi`N=J`{AVIjWAxzK6Z^YYgzZP%3^@fEDY+xTgdK@49_9H0?>xzF8%}S^R
zg<BmH9OrxqPAEPS<5*l-jfJH(cOp)3CRcXnp{&#~I~R4yI>G!H$pSEm{G^$%;W-(%
zLS273hf#J#qiz`_Pkf#-U02)hGL8HxkQe@rHh~FA$Pr9AA7FHJRKilHO==hgW^Dx1
z_9@)qQ6fbwL_i@HD!tNFTS|$*K9~LUS$u`?*ww{?St(hyZnbBp+O|>fg=xO<;@IxK
zcCXihfp&Jb$Hc@~T)B2VMutaJ926<>JOqCUbQVU`mQ1&2l^J0Lc7mH6Gh--&ISL4#
zQFFmSFrm!H!XYaVN{_mr0AWgkl80#mGZS3>t<tcqIlGVO8yvK+*4Y?+{T*J6+arab
ztwT7z#|}j-CKtn$BEd+XWCAl-?>ZzL&O@2HJxkNu$Id?p5fX-_Kv8dI#a5S>ZN7vO
zvAfGm1T!aWjhKv<G&l@tVtm5&diIQymv>F18C{0adwW<BZ>G@As0A;Se-yPYm3MMx
zFs|R4h}U1g9&f&RJ6^eaC8j3WX^4f=d#X#0u+}<R>(ZmX+~)HX);88-ZecwZ77s(H
zE#aSq&nRBG0CakKFs7!Y<VeT?w`m7<Zi2xeOu<;_By3l^wYeK?o6P!=#zqO&=3zOK
zi()~0(sp1vCQ^O9ES&DO0Ky_iGgIo^Cgxk~3U)?8*pM8p`l9GcLp^Dgz;hgPv+MEn
z>1sTCEW!TaQapLGYGtUqo5kE}j{z1%4#ddVPz(+aDva|n^u2JHSrKMm=o?J)V*<bo
zCa025svV}U%*5p6bj-}$kn(-irc~={j|caj#)EGj$CLX{;_1UDv9-ExJHAsXU=3=^
zVitNLxMFI=%+{&eiaq%!mFBB&y%|6H>CfV)KmY5vb>~&#r0B`~QhfW>w?3Wv_rLi#
z?tS@ntSt#&B&e}wM7>*OhOW>tpnEAeo36mj8O@c#!1^!*#AbQ!i%FLi49ca?=@W;Z
zB@x|JkZGlT(p^U)aqiS^g@!{}ct;pflKv=MeTn{F20zMoj#s(ttKW5(<Ke8J$dzx5
zD~j`d%0%HG7R>Yssbze{wI{$S@$W)@WZ`U_F4y$r#rLHwxRNhnTn#JHo=ldj+C^o6
z_Rtq78)ke6?$m)cL}+KnI>Jcbypr#@R)x12jLX0vKrTooJ=SN$DFP@5E~*XpMBC^S
z3dW?V;Z%y16bPY)8d+(J7W*D$hd5dNvwAX<ghGSTg7C?Rdy~oxFJvT|g}PQGL`PjW
zB_ir!C8AeKd8d?olEqYWXJaD{)Sq66E;Oh;KsDB2P$y>OyhD}J0zPb^P2c9;v%`AF
zDgKnNLy9jZUI?hPE410LqwH1raRm?R@9oRa_v!Nc#33Bl5>^teUfEG*Uv<GJ|IPIL
ziEa`<vJE}sDYzL|1S5oV24gVIEQe|`Gfh9C-q5=K{(<Q3>UM&Z1A1pXQS$INE0*;#
zLB%V{+!h<$%kwUKCrMTHwg-o@Qd!Ro;~kE2mzgu_hfD@9WDth+6+Yug?Yg@SzuJ?}
zS!66x;O{o)7Tv<k{e=hP7)T&pXpMv`h(4xrl#YD;th|*Ma5CPD4jJzWRv7UpD&<AC
z!?_-nqkhzxH`VLPQKzgIuGBH>ebME+dXFpD;x2}trveR3_}07nd|sB~d7ik7-+r$v
zpE!DTb0MtW@$O|WN?857^Lxt0@8bKv++BpjcfW6%nn0Q0g#aWVfoSX}FH{$@vAA*1
zBew$seMWvxo(7>#*<wteKoV1-$jm_eEU?~HodjWq*$$WiW)q%P6JeuK$Qq*SkuV`2
z@foe;%CkGH+~wv&KshV7obOx?I;s#^SBM!Jws2n^xjEi@|DAZ_jW@mP0`KnMdl(;!
zk*}<7iU^-tnP3s4&6AWjOn@@8|3Bt1ODo%%pUaFg0dYh?#IBu=k@-{R!t^O)!YE6c
zKI7$6<#?_hvo7WCqAOqIdtAmbnz-`<_fr#X1X0#}aL5j$rE~Lhxk!`+wA+kCu+||x
zL|jo+7uo>g3qhd#j07rKD-+n-B9L}U&}HhD*(nGsBf)(!J~1Mq*{`}YgMn#WZ<<vN
z4#(iYu-c_r0(~kl)ZYuLl|muR;2v>+9ivxfI)}M0IrU2TX5CO{ogx4L|MW>jK~zb@
z3Pe#PgfA?#_hM|_qhtUB+o=jZ;9G80$h2+>X1u6%HN@ahErv%~m|hceJ+avj>jhR<
z*y@!&fql3Xv^@!oCn+Rjt~4~n`T%VCrHHEinIQweF=L=EEKn@C46|f3p6PE!<0t_2
z<{XN?zJ5P{{`oiY<mr~uV~a=*X=4PDEd~+XP!1se%natiPc<osS25k0ob+LJeNt|Q
znduPY=ak%|W6bhcJJ9Glj*RxI%_N*i;-E}u-*llZWeGgGx*KEWN>AK<bt10coQ_V_
zjW%LuqwO6I(4u}Rn6p#J`Z^<8JF&916^jd7QdYsABbB{Hb&_({)1&y9ZzVp0qbL%C
z1B1~aCQCbFD#R9Q{e6S#m$|sPrMV`?rncXwy<1h*=GfTSkB1NE<J)ha#q-%EDSPa)
zbf7XJkhI3wM4#6Ju`Xm#^=G#-?7Qh}EGB2xX9;^@F-`2vG1;Iztlt=yGI#6Now#x1
zPK=IBsEorQc|M*zcpjhq;fwhCvu~ulC|~&<G2SMXyA^@Y0=|BNHH-8QS+}6Dwpz`n
z&A$G|8}ZgV@5B#2_))y|);lpgG$J@HdP~<oeEj?P=-0oAFaG#xEMRwSB?6kLw&tfa
z)wQI<qp*npu|Q70fz#7Oo0zq#F@w@8W!@$jqBGv1C%th?ZHn+f`_k6TaHKh1;<9Oz
zaF{E!(&jt#KYe48D(`wBE$9k=hmPUtgPxJ457B@`=704bm;ID?7hQReEBBYbE7vSf
zdfsz|o>FG=K2s*-O1U#0*Wxp6zAdg4%X#8eKdMmTosGEitUm354ZZh3mheA$UB$Qb
zDLfAPMA;Z9GmjJBVAPu32B8t0m`=FcLT$<S9yk&lT}Il<@UFc1rrs-kaFsz4crn|(
z>Xpj@eLj^!;d|&cpin>J>Om%b?<?u)=UE29tU<^NzJkPF7XVLZWF0$tDHF%yizGt*
z;9FpLw`kVz__&ngK9>pMb65R|ZI4+S1kS=UQBt6{ST5w2u)=4R4;<^Pb*X@jalip<
zd#oi=8I%VzxvXQNACfkMFVxX+aXSHLy^t%{dgpkR$DP6OQr!A5M8tD<9^Y-Ex^t!W
z#N&tHjS$Rd%q+EucGWQb^&l8HLI2P9q{AQ<qy7w7Pzy}SpoQRq>seJS`L%FNX);p)
z&fs$7KQ|3ww#%hex}}c3@U8KlH1N*e`EzptE`C=%gX8kt8+Vll1FjxmksN?I6tnVH
z81>BhIo1R1?ywati?(-TO9}^vBeJF<`-J+7a8KDNU8SuuW9v!%pdD;O3{P`@C5~U{
zvjTS>T)w0+RO8H1rL!>V!S_Vx-nsI^*Yod)Q1U5ps(9G|`YU0?tDY+lyl@xO&vI0E
z)wTX@-t$d~m)GJix-37}^nbb2cU75-FX89Dxcd8~Yk!G5ujhAN=J(xnnx-ae9~!yz
zKmrugj7AwB?jL((u3N$Z2T`Rm!D2!d`mqIMPQ#s>9#xgokFtOyXzjeu?I`i=Y{SSb
zLRNNMgu(JH&vYujvG!`5%QTIgibE$O9ortN;FKE=@g(D{kA!^@;oaRso80vGb;buD
z{2*R^_0{O<?u!G3fA#rSQYJo%H3>z`YKRtEnJ68O(P4JJ!IW!N3dic2i04^eloMqF
zW(|`@kUD2@r9Qv~rooTm<-6=i8OOIYv+~SuKz4rFVCSBEl}zQTxK)7v`(MVRd=Tle
z@iCijVDhuLxFABYF2RbK7!fW6APMUpQCVkXf}I|c*h55^l4dXdJ+~F*V+&6RBML*C
z=!Jx`woVC-LoqTk>HYru`i2GHffySdx9^qk*wTVIRlD+JhcJ`_MqHC=L%gw>Eu(S3
z2E~$iHcJAg70eAsK%i@yHKn&gp#s|L8{*Rz2;UG&K4Rw-PHd&V;8tmi-oB=onCgwO
zNp`qG(G|hmk>I(sDOh0^B|;}0-rFJV5R6z$bfi8}Q~ici2XJ6V@ZEM@7-d75K<HuA
zhk76oOpK4j>#yI5*I#>8@4G~x1?$~C*NeC;g8TKqd>xBR!k>l%DLIXTUAGr&9;oeX
zn|#Qg^!kq3K}owkeLWVM=?A>?$>vgoFr{Mg?~9XtmFGw}&=yy&3J(+)R6^JRud(SJ
zpWuMF)94>OQF_f%ss=3x504Ft@W1f+26MBEs^hT~!|}L!?S{~-(dSt_c<@Bm#aNM&
zy|vAvcBS3c7*kV&UONH)(Z<k#+-CY%N<>Gr3U~Izm=@<(<KEYg)SeA7IdLT>#;)jH
zP58MV^K+~hfKD_>Y1xbW4`$=jPai5T_Qe;b2xhIAUk-M}#N?35b;UCgon4(4s#*KM
zfs`w&YhDk5*%fPGP(pip`iz(0bFXN@^=mid&Yjm{ba+w>GUEA@#rWjo&*E4A^y~Qe
zUw$8t?>~<1)h(NK)S5+i)bCD&44yTRV77axboFZx0YTupbN6=q{BQm?e*UliCf<Jc
zeZgW%?RFSnfBAL%>Ysn*gXX^Y<7e^Y;Y0VuMj@jGWr&CvFK|aWv9hi7PZieQDIpbl
z)Y-0f>W-1I5%0zZz2eY4zG>~?khM1TAyXUm4dIMu!crK@?Jw28qzo2?!lUW(#ry`6
zf@k)Oeu_)xtw1?wEFL(5Or~LebM8#1>Rfpyt_5~^e^;ehuJzy5$Kwio@Z3ZHO>0#?
zzATCp$&_@9Cw_5-hS({tp+uOx<0~Bta^$Of_#JT3OVbv3mS<Fj&wN8XD;)YJHx_91
zx&eek`Y0p+qoTiTzs<sA+5q|k|AQWt^824?ewGYH4W7t5lsy(aV}i7=eyzG<G6LU~
z0-+{yn9X5zrMXNa5%8*OQ6j*HwpyFK6l~!XxXW28UELk(8<;Ka2B-HIWdg;A)<0Fe
zJ}Kv8(`=hOWV*>_0K1z~ekA}y8<~-1{0_yIHG{y&gDBMTm~~w(>Sw*iJx03O7{aDv
zN(*HI{={yP)StAuaT^HkZnHX%BfS1v!v9p<;tG`Qh*QP23?l$qUQt`0rK2#~DfxQ#
zAruMEN{jw76NoZpd{i9grexutRwyKpBhN6f$Dk1>gJZsm#6)aN%`j<&H^QrEHxH}|
zR_2p}gX=;)fCuU8YJ9E2sx0NR%pTCKa2Q`4aTJo@>A}ys_aGkpSXau5A_47OlT5I&
zxg)x%He)@FWCNZ#zXHE0#bRS^D>fw?ok$Mi<X@$xJn1u;zjN`U-UxO!m2%w@M|n;D
zeqYtqiW^nav+Gytk;S&-icdIzEA;ccCmjZNxqipd&*YV;@A>6-q|bf2OOJB-kYOeF
zboJ4=cUpSycC*X-5?;Q|=gQUH)_-@g4CzW;&p+bIHzllu<LcvpKlz+_f7g@sdAYv(
z9)F?Sbxz7ehm;8-=!pvJ!p;tA9xD@TtGntrdDK5IIa;67f;l*j)34FUlOXH060vtG
zBu7Vt+95D8f{il4?m_H)k%!mNVd#iv2&$?XU!A8j<4OgH520ckO17KxP8C85G-S*r
z=#Yizv*u%8NUm2gJvkab`r!}b&aJ!A+1_cQ_4yw^jo<#}kJhxAm1t2J(vDFd=*l8Z
zPNT)<o4w}OH`peWX-P)F8Tl3Q$*5vFsk|5#G_1<2|KO`T$C*G(pd5(qPG7-{`V&@F
zQxH6VSG>b@NnhP4Ch5X7fDOs>j<lE=;4BFk-Rjz^7X+`auIl}%1h+8>!(wt0EMW+X
zdbg%`2sfKs>mnAc8ITYr?^Ap|s*f;>U9E6Ho)IyWgpRhZXm6=W5bTJqj$SJj6B3|?
zhDM^Zvs-P{W>~Qc7^79dgED#{y<mWSFvK?N3=wGdRI7K%s*7J>SK8W4OYgxS&rVbX
z9XRAn1PkyHAQ4JM5SfV)GZU{+9`X`OJ33FKNOXAtHQR;m?6Hj=JNi)H6V(rU-$ua$
zbDj|K<n&h-m?NAn#M1Jd2(55qeN{@7lsCZ?#)uiy{2a4D8w&4@YuB!O^m}@0#&ygC
zhuFh>HJ(44jR*HwsLkk4;!_vvIQm(i!lK9Af)OT1JA?ns7GR!3+tSZyo2^aO0W7Eu
zS5?<7!L>;WfCK>vo(T4NU|e@HbP4weLWFhq_Sy^v(-@n~Na<&>`?lc4jKtFNYP7d^
z#f=-c;>wll`u0Hhv>9K0y%hKEOA(dQRqG73cY_4;fw+EsM&+;el~)?Hp%=*uCTmOU
zG5>5qg7iZC?zhik_StdtNCCKY<F@c-N;IcI@ZXK)74}Bok4KN@<Lj><$J1xa!l6?M
z%nhC~>6hX#F*PYA;;MHko0*xmnFXh6udK2w(^9Mm=h^DDFlq1$#g_JHZSAlkJU%uV
zSEjGWVE<?w?4HCoUp<U}`meu~|3&=q_n&&VrhN%=jS}2Cz*B@rx}NZdU6ppEL=sTz
zQQJ>V#9QyZ6aV(_|33caZ~jiSV?=bJ)u&be*Z=Zg<A49({*U<OKm9VEJa`nGw2QFC
z?kFff;^o3~Xq9l7pJygSFPcuJG<3!I)L0CT40vsa7v<L^wCc*N!_mo+zd=B3R+!Ha
zV3b(lnEBaMR6m4V24S#YFO0U(XwzDCFob8kXOD$%^reia{-;~_FL~lZeTp;9e2Jsq
zl7{J>JVHyknuf{Me@j@sFL8@9K{>q+L4BGTJL*+mj`H2*d&*z|TR*+=AGg5GHx>P*
z?rGYW0v3t_iVAQ-{BwSa>o|BiYXy66y)9fOZFXvdZn55p!39h?si$qw;ZN!h371sr
z7`-<?(60!Sn3rHuoNEt&J$#PAG_p{+7!8KOsQ9#vBvdO9C>A{D;ETNQhz%XsG=ohX
zTHBRpTay=I_Ye0epFNxAY_4yKwlK)Vh_iUI>et%d9z#-yMyDr)<J}&0hi>hOP@Sj_
z6pVojb`t|mw2uWtl|Rb@4P~Zsz+t;m$T{^tg-He<T14YIgN41+9Wz{-3<1;aRfikp
zzP^23zseU^-s8AW&l7R7y@_7`t$Bud2l04Uo(u2Ed)k||5GWFpGvlJCHRFu8vsT*z
zcd<3aPqVU3!HU`BF2RmNJUOSMy|d;4x-G%tRBeqpE3?+Q;f?s-5i=60=)266p9JMf
zeZW`39ZKDJB&u)lDR4TjBjoqxLI{q!^G#ZYAkSgSgyJBjK_0VE&Z9s^VlW29f&o<4
z@~~|-GSSKk=CZ4z*9=HXMiLG+H#FI-U6shPD1GN`fvM^@{82t|n6kLNs|HZydfEi+
zajw7bO4;+u`_e{9%aTT=tI{veek||K-Q|$;@6#1n`cfA=Q%{#IlAd@Ow|sli)rYCJ
z)!jw+lF!m9dcNZyWlDOUuKuHP4-V_TF2R!j%j)^3{P(%0nQ0EK%fyJLX7(wMTv?2@
zzP2Z!rP21tn9f_75P`EP3<Y$pAh6&_3c%-Z$&Ly$Lx4rAhaoW4i~Tz9h^`_E=dcNf
zm}WXA<YpogZXmWiOZww`4uru}sk5Wg@ok6cX+ju`kVtsvimM!qpb57m<V?r=@4X+_
zu3mTDo<4pWzy9UF#1~)RQ+*CRE71xO6jLa`<rH6(33fbMlAw*56K~)zWGq_ax~ga{
zmx}F9;++_jKlMOhqmm#w#L4iG$1H|;UbLxifJ4#;{ZpFO)&<m(4(<$IRlNL;Fuo}r
ziu6VFMBo?&85|N3S3axjE3vS!;6<^cBZG1G?#-B#f>rBi6JhMO%|F|So}ALQVuUaD
z_00JKOtJ5X2{7ut=ACP{8R^~HN~J=?Wse;|Hut=~hFu{V#N;4OZa<sMaYFE(XFp&j
z5LxPkSwOoKHAczXT5|hJ6bWEyWl+qDE$*yPQfI?J>_9{Z#Q_DvCJ*u~yoQLd@H*e+
zw|mtot$1wOn;S7nQW@1|jF@zEMO?cv5I66Pd%ef}!mNbRWu<v6WtDy18?00;FQyW)
zQwjT>otXR}Br{8P>Y0k`*ROfkBTRdkG2y%d*4MC07I0wNn-fYQxWEImf`xg`Mc9d_
zPgs+1D&`b1G1(IRgQ+B-40Vdx&_^J;C;}T&Y&nRLg^Rs?eM-MmWs-s`1)c3f5BAfv
z$u{&V3o{a9<Kt%No0}Zww&?xGv5BR#Fq*xxx@5u0<``xu>Qm}Z>>zV^a3WZAs*W7M
z*XA%r(hxuvrF3m>z(AW+_b#7P!U!CUf!#xFJ|U(=!x;u!2QfFh6Hip%rNt8|5+OY8
z3QmzdIcd?mm3a7QE@o$!<H^&7cs?tIKy}4b@5b$0QXGW0)6+37I8RJZdTUoqelXuS
zIu@=-aAI^A(*l$&?3I1km+%S`HH0>l2u65mmp<<@))3F0&c`Re|4cODqxj9QKaNKa
zAIFXuB?4NDaJ5<els)kgMi}kilQM}(6DJd2yLBV(y!uLf@Z%rHPyg!YaqIS7D?W_c
zeewC1@!$U8zr`>9>7V1-lc&a222-G%Ng9OE>SLy#C;?IiYAoy@>=W(jkKX?7XzP@~
z)!pIuC@=@iI4BLa4aFl|BP{w%5ea~df)A=(%s|nX%DFqtHlT!aJ*8sR2(?Y56o(yo
z5R9NF?o(?26sSZ~Do;D+Y2Mi%c`o!4XIfO}%5&1II=Wch7grze^1TIo^8?BS9i?13
z$WY>yE54*(e4)9eoQ19`q<V&iXCCF;pYknA1F(W#umK5I!aU2)Iu#T_JtGWTVbBkO
z7@YU+g3vMxm2ADaB86IQCw>EbP3zR3P&mEd7pyRd^QtHoEQNhupVnFH5xwpf{&uJ>
z1Pi@mK!-I6{8*Sxm1ri~%%(RAV)VRKeq?+|LTybrMuovS=%DZ!Iw@L3JzIrG!1(}y
zwzbJ7PAC$CqoY<VSQE6%dZR5CU~h}2ZHTUmFNhzA8XAvH_f;P}W@5}>duBrQV^)Hb
z|GjmsikFMlfnlzhV2~wgop_M|K${goskAe$+?RIcTU@~bTzPl#yY%UJpYiIS%lq<<
zHELCeaRMQqUAacbhYepUk~3H0Is2G>DGfZY^Z+lzTo;q$o40P69y1ukmh)3HSL5w>
z-i_B^dCkH%8$+;L7BEBxATi+KI*=eir93=Wj$7bMJoncOc3Jel@`|U3$!BTj9L$r{
zAe^b<q!5x`{Q#>BX2Dr+$pM;NnYmzalI_khbw#OQww|4!wzn|x+Ve~r>n<>bY6d@>
zgyZbbFL_Rlt$bSato&7u+^V{&ugY-VhI!4h<{(6d^D3{d7560=nTUA43KrCj|Eh<7
zx)PSqsKWeVrUkUqaZT^Av-jm+_R)M_1?j(2Q+Y}%xbymapCf+t-k)du?|wo62v=Tk
zO@A?ua>bp8<@<a-k9+YRclr4*bxqS#T^|AnBcgr~CicXJFdghkz+ufqXG07O_Ier<
zA!Ap<#4*Z5MU+bedSt}~Nq{DBX~@**8KHpLc_$DPQ{phn13?kd5yF@dL4Xv-h&94Z
zb}YRU@g=OM+Fp}zMpC@Pl!Pqu%kLBM$Q^;xYbLroqp!EmJNZ0$^dx@&@$cd{QYH}Y
z*m_YsT;GAXAcs-HqX}$_2mI&e7h-K=S9OT!?d`I#LY)y>af$4LBZyoz2KP*3;?kx+
zHD=zzv<mYkPe5UC$R${4{JJY8GKI>{=5OSa*yV>Vxzfk8zx>efpa~dEZ+j<4&sjGx
zIXPmp8}{I5!SM9dxcA5ImcUZ0wJ03Be@^(sp0z`DW-%CMLX3WI`0Rv5!DC@|-n$j8
zE{mvee7^{b7jTJTun-Nw7XcW>g0&K9i@Bxt&xL6$NMjq(&RUmohsCf!9)XV$Eh`ja
z#K|a<;HhdBBrNDK6C{{g;1PTowa>E)SX|Dxd5S4(LjV!8Ycp5I;>r~k=e9=gzzh4a
zNe-&(5aB$OkS(EF`GYS_BHG{==QyB5h{o$Vf<s^@C!3|HG2PuRX3gPG6LIa@Rpl|L
z_Q4c`H9JSDJ8Mna4F{A64i7>Z$CQlSqLvnSeE8k&p4wIT^V%Cjap#puD}l5#bsZWQ
za+xrjVxc69v=3Ac%w75fziw=f)!ss1NfA<>Rd4z!Oa(=lzQTGJ1e&MMp1J*&mR8k1
z2P%KfXGoxogEJ6L!%%I(lr*-n3>=^klAdRlR5$t}=V34^i5>0EHp+q&a-ln;<!Q#H
zekq)xJuEadwTV$;-UOqG=<jb)Ta0*-bejbH4fU@_kDp7?n2iOMZ%YiRO=af{hMTwV
zSdqAX>!x=ngCV0#Sy|VgPqo&MF!p#y`QR*+v9U4rRW=LgaG6;Hgfh`pV*z-p%l_!$
z)A;hUui~rEzn1^fN;kXMNvxKVtA47Av2ACrKajGmzQE2&6KeDKe(;0%@z4G$-g@^v
zwcBfIw^8BvZhZR3&*S6Yek{f0*YWwMpU1-7oMC}7g5u{Mi%@{qGcyXU>U*s<DMF$z
zeWJJcPPGq$AD%<U81c5LjOyz(G)j~5fIA4am<%H5(3T9?&<CI&^s56YKrL-*2Pw#G
zJ4&9IO$|u#z^<PWVrVb@i7g{h%F6(VCDi1e#>+zE>fN6TtG|+7{WEkD=X~S{%l9d~
znWl?B5C-4CTlkST->Tx{iC;Y5;VeLihLn7*OjO^Ihey)+1{y9;ILQ+ZOcB13_2FeS
zoT3`<l7_>pl6R3l${IX}^AzBrY=^qH1Kr&A%p5cem`WlP9<(G*suqwq`WJM<v;Fj4
zDc^+io?Y}<r^T!qHv6nSV7Ed>|A|2yJn8anffH417oA5bWwG;V<BK>t0f$vi3ICX9
za>_sTq2F{1-&#aSupH><YWHm0#MSBO9~tzbf9N4JbxVTfmXwICjT~Is-p&3;KVf#i
z*}|cCb^=a{d$&p_8YbC+nF%&nAiVmQuoK*%8%1$F!#(wn<ZHw&o^O)Z5zZBt?aOls
zBW~69x=&tto`zAtSH7#~-QdJIj^MZlPmp0y-n;s0YA&^t+8>9K2>+wapv8QNPabvb
zS>Tx~Gchqa1p;~GpX=Q_cjB!#-ZZR`P3A=_z+D!?r{JIE$Wqt;G6(<HOOq1COC`H1
za}|q6gVLc@JQXzSSJLH$?g_J3nh<qenc>`HLyvuCC$?in?awA9n2|DjflSCOHjDV#
zT9$Jw*4H*A7i?L{IK&J}f8KQyoDgp@(Ng_UBvP@E=UUw}5SJcBq<XJ!!K?afwp+R4
zsgM4-GOk=cmkL&%xQ^^^62=7w>hJY0pZz;eBVV0gxyfhl%RAj%a%pB&oC{UCstxMg
z#qYWQ?sLW|_k`oBu(LXIQGFaIpX<w4?#pw%%YK(}{*?2Y2{Owc#E1r0L+QsUnlRg1
z?NmbPfQV9?G-^h~c7w&3i3IW>J3kr+r#wr>rNh%X*!Gdpcsdu3j!A~RBYt+^ssU9F
zG*yT@&J|K4P&V?hV1uB?2&yD2roHKSW+3{G+(`%T1LgFdk>IC~pLqe~m!Ey%k-|oG
zs5X_URh_1#S@8`5vyG^Aby=Zc)O$e!7{X0^Tbd8C<tpX^EI7lJ5j}hovaA=Q7^H_X
z0dXgFMe&H0x5Z;*%f!N9Q~jy;IsB@Nt~6+Vizi;b6M-a*TRC6#58M5*u~D12uvH|d
z0I#pF=(;1}Y%Dt3QF7SHrX#M*Ohpe{h6;3;&9F70x5dR{PNuYAnmCV$Zf#32W%_en
z__1YX!pXcFYa6jCzb+=T$SI*x9B31iKX%k%8aEf%Nty6eai16C*7T0Hv|xk~tBL_T
zge4SDU~C~4L0R9TNFeB-K-la;VZfF!%xNei$^_eL?n;Sbo6U9!e+c|<zkN5}d;b*?
z^<F6wjAE-zYf^OE6E7i(5Gh~)PBFT&f54(Tm;#DKL%jCdK)m_ZblkYXVQ$#ohqrfS
zk)2v5l^`jtEL=mWKxjp&;CmDdlyuHtK)G03g#aIUu%^GiA%5_~J8|vCO!P^KWh9WT
zdr^Y1J!Tux=g-!q2mq6&7@zF7`4T2YoWNRZYZFtELZ*z^<_{Av6aoYh?q{X^&CWiL
z=g-$f?03|rO)meC6c^)>5D9p)21qTScIonL6AM3c&52;4HnM^w;hY^x;5F>GzO{uZ
zmf{KKDEA1w>I4?D7(~E~jqkQuG=0P_V|`ZYIP;{dTS~s_H#ff&PoB=jjs&H4DG~ic
zBXRZGwYYKXmUmj}9~_JhDS3=!L5Dm`N?S{L!9$1Y@s|Vms5^{?85MRa<8ZzvwMT~(
z3T%EKJ$Mq|etSQjK6w^vVvd*rVgubm-&N%d6A9}_nAS-AYHDka-od`OdFPgth@Zv>
zKlzF7@5boZr21X6l(o6|_@m!>tJlx|_-Q<T_(*+)bu6NT!Y@4io>6WK^votmz-PS!
z$^=`x*3>64pGc*kDGrYryb)dtrkK0Y9}rG_PJymTt&%>7JgV(D6N8h=Il&v#Oh!vF
z%iI?XyG6U&1>Y7aWSyexgE2la9=$!iG?x|JO-7xqOcWukqAp}1S7=SWEA;BJBm9h}
z%N53iBdqvBGo4=*7YDzvunwZhx!()FOFHm^l0Ke$JoKa76F<MH$_bH&XgUENlh2gz
zsd2_t{tWnt<|~DKraT@5La+fhRA%v_Gk&y>i4f*Yg82#k2_aLMBVmf|w_R)r{7Okz
z-pi}`%%dvBz|Q$++4%;g!?U~!Lm^>=-=pA4-&??n?(pAfsjph$(gR8Y=d28m4%ze#
zTsk=3Q*HFE1b-Avc6MUD2l3$13>=+E;piXgkD2RN(!@#1yRtD&W)0WQ=0+Sz@RwXI
z#XQ##por8u)wh@t6CzqEP@RB5cPHC2W4?M|WgB`9JhQ664|8W`qm{QRoU;<y&W5o(
zcnyT?cvpSLJwS5X`kmf5oEZ59-}+ibsds#r-_rhskQO|aa3^EOGwtu0E75Ca3|Uu+
zi4xx;UoeOUttOe06}+3ZO~7|-A`h%=7Oh@W+@<9e;ZjRXjE+k=+l+6%`dT#px#^E8
zZsD8qpdh%uN=r4n;IicS%(2uN=kQWLS44FK_jy!G;Cof|3*YhVZ}_8s$nzFChS`7Q
zE)+jz;FwKfpbC!?!OR8ti8(AZmaAtfBsZc2OOrJ{Zt#p8vlQ}N<*hK-vt2M=aq6!q
z7Uj7~X_WBlTL_?)0JlJOtrQJE^G?qUa9SDhRbjxvp0Mg3nxLx{Iy_+|ZUxg!uXyF=
z_nEW7$G^(e-|NZ$3vQ(&tlp<P|LkHo_nGEJe=#52c{o2m67rpR-*tt~;{IIxrirnR
z56J)^*hEMT?NLMlm1$H2$M%khF)yf;Kqz4WX2arRv_=3+H|OuX<7UBe2xt%mm;s}w
zgp&_fBd7|?I}?kl+}uM{Y*(s#o|87@Jpzrl+}vZ-lWF_>&>danfxwVHf;^oO!el$|
zm9=>O^tlBG_Lpaesh+M{bV*ol7lUVdmpWiC4qfMz(3*rbM!Xl<woyVLgqIOnrnOl(
zWnwR%3-``}O*vs`j1*D_Iw4oeMJz`95qzk-7hVEI0RXUNI_Hg;F5iKuMpn{MJfb?b
z(k2e|$Ak*PHO}rf<0HO5JldDwx-59@$JMJ-`UYkrhSAp}L3qe@M2KUysZ|0PBiSu&
z98{NQ9<=J7b16^)*fSdwGfb4qxgp6&vt8IibZKST3tu6WQ<JQp7;+gA^s)H|Hf&pq
zNet8WnA|YpWu;K%haq8Wim>E00*uBAhF+K|yfm|fQ7KK@fbv`Ez$*`XeZ#E8p%f{Y
z<%Dqfl~-@ZTW{X>?mz9Fw6&C*QxP~|4^*TZGHnaq7TaddM&RAZa9iBD(-m*Obv532
z^Od-H<EoUgn)29n3$e(S`m<mXp^Pcuve=L*W$dSEAD=ue#=&Uiri5$GvFPY%P#<fJ
z*WY+mjDIxRgp)Yx#`zxzDUTn{$D@avYDcxd(z|wTFnTcMk>{Yd8Z!$JdbI<mO|$D7
zI~uX(hfzOHZDk$D{QQoX!3%xYrnc`@-t0NgEFW!ysUUq5^TKxJweC??+66%lWslKo
zPRpiktsL#sPWy%teGNRu7TQ8Reap%yf{=)Pv{*Tsn3Mq9hY|z*5KM$KjB;UG)jv2C
zSFYcPo40Sr^&8i{fE9RPR>U;>3Tu0|H+`x!?QWq%zdWQ)<yN1AxTDa)=zQi1LYssd
zw%cTsjM1{^68s-OdK$}%D`s*C`pg=jL}Z_ofFnWtkOj^Xyt;Z>7cuVBvzd+f(NBLW
zJepD&8e@H9JHGt-+xXQlejUG-B60tld$FPVuuF?+osy%R#kQXH9bPc)MX{|?QbfC?
zTwspUE>C;d{HcN6v{>hpGoYr;)krAmK?A6FGCJrZYm+$hV_mqlZKm(tc8I6+C#Q#{
zM6d;Ni|Dx&w!wbkaF5#_!G^;X*Vi}HPoWRwQZY*XaQruDpf|cglQJ|Nb-nDeY#E+(
zD&y6M;Vl@-QC42z%YY15+?g^VPni<7nNCTQ{P5+QB3Qe~)a9%^g=XeA798Ypl!tV^
zjjF&6ZQ&hegI=qrGBI+_Yy-Rs0Ux|#VC(pZO&U-f#ZQEHKAAd%lPD0BQM^^nq`EK&
z0_MUKb#?ur^H~Oz0%%+rrKhjK^PrW?5WHY#5&9}^q#x2gb#-~%4{7rzue0IQcg)=m
zjyQ>4{gv4)6^BC|iG!)5*GfsC-&cR82rTlSoSBZk!GUNLZJ~6g5uz`9TU*hjN{kE*
z#O+(x;`W`ZF(uqWImPUfvK;K9cnJSh|Gu6+D@4pnK=)`z+QaZx(M{WFC+e$oJ>Yl&
zFA{I^THROg%QM{DfBwD-Q`&B7Aar(r?h<FXOa98$Ny?SDJadmXUNKuGyl8G`J8BN4
z>yZNS!n5-mYnzgr5|4}HJCq5g4#q_HhDJvuKR3sQ<fwV|H_T2sU7Fec``_LZou1E`
z1{VOl6$~hseroQNiFj3|ZATeZXRobL7;*CzxNYokMY0a`8Ar=^x=A>ubKj7sKaiKx
zK)w;(CqDg(IJ~00;E`TWAzF)x76XjnDF?)2rN96k=FL_%#P?8wz0Os5fnp7Q&>z5O
z-gzJc6FyazKI49)c<f##6p;td^*yt2o{6I@^a0A1ibBp*;MF^-d#>qI@nCQ(6$(95
zC)Kx@T;+4!yXt+;UM8LJ)z9a?ob_?$^-D+>M|cHep6cSJEAh*7wtu~=|Mp`14Ci{D
z9xmSzW~``x^=DjtIv3OYu4@_}ZT}F=1gMOjft0M7VCzGk*^xl>GbSb&LF_bhV#G_d
zKnTWEl?bB}qQb7TJeeZcleecTMHoOpgm5C53F;jU!gn-2-*WHb>K>x(k!rptj9|To
z07Hn%>A?^kz9G*X!ODDLKzN7>h`A?1k3hw@5G_W(2l{(rNW=hBo;DFu62h#8DLu)L
z(Jo<#-C>xSfYI)O=wcM)OJ`LLoEF3m0m~Y{%jsGgxhNAIlBsLH04_veq#r>X6EDi^
z-lQ1Z+v~|5T3dzT$U<OqUWPLdN-vo!?~-_U6huN8BT|?$9Eypot*@#~A%Xa&VlonT
zB!N<jf$GU(u>Awd$kaXdznxMZq(!=Y869LHA-3%!Qbw*xkiFq6XINkl%M6AW(kU&7
zEoWI^(!_#$))HV!qmmkbIDr!(5M=_JcHlr<))7z_VAdg}h0#@%54J+}EQRVwdm_*x
z+@^q_FvB<7)(S;6f{Q}IbUh=uz|E#x-L0ydgxy+*fdWS{Uu@uoYmD9>seN}itjBpG
zh@I%WrWhP(jqBG&;>wlD7#kh-$T~B1-ltpjVWBlj6wDGuVnvGU+}ymc;Hq*G<5qtW
z5-iLw#tNokYaAMsiz{ojW-<J}n4X#O4oxUenA<GMBeY{<{P@vIJbQ)`&6<s-xN>zU
zhKGkFoOjzC=Tw64zLXhEwP;UFmDo9?SIiC*iKWG*c>bK7nKDn5>AoJ6=tjS1WQ`dU
zm>~+({?1`MeDK)qLVsyRu~d33%y#HHKeymrolvGaG0#z%+S?lhtG*Z=8J4opW1$N?
zg%CG4Gn;`~)0po!5iEo!*a<U0!A#7RYuBVi-11t8$*IX07#b9vX|xi-L3^v~f~SNX
zMltCF?ni=aoA9ix5r9r0c%^+dBWIYIaN?^N8#|H+?$4jjS((^YIA^Rt?-07%l@27A
zSuLMYBCgKt!NBl9yz<&B@%DS~ct+xlx8HUjSzKCiI)D7br}5AK^vn3@w;#v+?3@@Q
zCMaoULf^KcqVlq!zN@FphX=MuX=p;2Q~IJJUJt>51_KtXYXEPVkwAf_Utq4-Bf+2T
zmoc-1o}%2dz?XD+Z{=Hf3tdEcU>yZ?5M0G94rQXPquI9a&@Ki!VDitN&c^KXdE+q~
zZ+Mj4d;w-m&vixYsgvm+SH&srGP5x~S6C4~>cepLVJR@CO97{%Zxs#YYWk0IPh8#W
z-&wrkxVONk_;{SnhEz{JDeupmKDE-IxXu?s!5`sj2iB>uLlHEeKyb8!ffn@v3-}V4
zfE7xrXYdZ8PlAq`3<cvDa~_ll2{4JX5N`%8)W`G<^vt8A3a5TnDk^x;K7`wBN{X?0
zlV}uqis(rQQ~4}J!2^H+LRNl9^YE=tM5paVBQSAePzi+sTtivy?5-JanBjyDuuV8T
zAHFiAKFrD2ZKCU(HKZ)fOWD<IQ+-wvbc{-wc<q%t@%rm`V`^r~Jcl)nm=mun4yLat
z5d;0|$ASU<knNelPtq#`fV3v^2H>ZifCJ7jF>g#qyWj|~?y9haTe)&ec;d(Tci-nq
zdR$9-`CQT|_rev?BQ~l)$>h*c@6IQI8JdnV0WOflS+CJ|DCd|OkBg>tbals`+K}~&
ztgAx_SrRQ>U0OEnU?UXzK6yfez)9n*uCAMYaOzSn)zHd^0jlSW!F|R7ze`;1NsDj5
z6Fy6L5tn$1ggbkt@5*PO?o+|HqT@eg5WQEL@ZYU%HmYGKMEK!`l!sh12M=NQ!kmp?
zkJXj+*x8UAuqQexTsf8ku_M^6EiT7`=;`LlM(i_#wJSc3-`S6S@#S58vyEAmuKS|r
zd*UG|ttj>29Oltf3Yo<PwCBFgtFL%#WxYGmy$2)}%~!=pFC~Td)ib7Fh85}gUf)9(
zO)qpcJje_0h-X(ND^H%~eR>aUT<|5lc)m-y@hsiS0e60m#2sZg%k;84PZxJNRtYce
zWzY3td4Dkuu4!_(^TYP0S~R{;&k)qSO(&x<-jAB8^=%{SgoIM2Gk4X%%Zr!-h`@`W
za_9|-)A^l+{(-FL%0!Mh!0;K(q%k3+5KGcE^J4+D7=}B9Jiy1Eo1K^ti0GW6J&K!J
zuy;*Teh_~c4Ogb}8F^uJ4%=orJb6N>eHa({xoi?Z#A9k5jS_mgBot!G)2cdS6P_dN
z2qM&p*$0GVb~spNJ5}sV#Uw0n3W2@fwmJxcEJBl!23X()4YYZHV!<D?1~Dg0OF%w4
zq%~a%<qwmPvTnm)>)y7ff}`V-j^Uf{5?FKq2O54Tgb3O&+_sL^n3x(9@nUzOR(0kr
z&rH0KpfSvBfE7l0U?fEdX4Tl(u5>W1;RM#21hkq2PIe)I*|m#F^u@s7SPTt~$Kddo
z6qT_U9h>mtUKYeoU!9H{cW%VBn^&W&OW&O!9EAm5b~D241h}nBcwrPD!4!o89KbZj
zhjke*)DVIR_O;9o<di<=6eJ8#61>9^f`#Hj1Q3q)r4(S!g>5-;*uL`^WkhXs$|&El
zh|&IHtZhAtox>#wavbtii(Q}gnOiTCE{o1tjNT?iuSW{>_}Fk<y)qpvRA*7*k@8wx
zVWIF7Zz~K8Xn(eY+z<g}svCy6D8jh1x?>#XTm#lFQ0Ds&ALu=%V%P`^Cj}&C!TJVA
zy^Sb>3yalRyw)h#V;Us+S<Fhzc>|kbY>=6mP+JW43RbMYSc~PgrI=q@jOTN+u_AoM
zo|k>iF~4G60>a3~#+vY)*$}}@f=^rPNwhS!#MJn7v^KRzv+`?#(5jDN#xVP2RXC$|
z6s_XaZ`QdmI?JX79J2T1;iAgQB5@Rx?&xA6Y8whnUvzi&N>G;q72pDkQ){9tV!TpX
z=OlzLEVDzC@^5X6!I9CpdFyuEz5A+R(AkN}Se^=wQZ_p~D+aVIeAss%Ve}s5u(QUV
z@64JplhP{OX_8WO7^}<l%@@(zI~oUjCwkr%qh5*mXUnm)uo}w?%VI)oywD;q{8s7M
zfB}WfhYAXBm<g%%)Z*6NJMr_s{_A+}{U5{+e)yA^nYk`HaT53MJ&I30{zLrcm%onR
z|Ms`>^uhhu+uD$fb`&*)cn@r(qRX$Pqa}KW`n(pRt*b3wG>D0creN-bB4d*!3#N?F
ziS|i2#pH~|{0vrLDrl3o?rtd*N~f2T=+(A;ecmjB4GqA1uB?Y(#)F-phDQ6P40T8E
zV5fvb2|vO$6a>szcQ<!qX?`^pXQhCy?Fc67Q<7Cpk4<PGO(^ZqofJf@T$t{_z(j)}
zoN~M`-{9(>i@<^_;q`GZyOJL>x0J<$H429w6oI=4oW#4FKIKHwLXkkou>vAjMalGG
z?h2MTBswR4M%_$r)oz%)VhJj`eXRb%XuTJ%tG!MeL?{|FT?8gf0HM1@prOw;N`YX#
z1he~un|>*8Z@H>+^Gz{*g(fhAh)|}2X8|dd>Wl!AMOU3f-x*B>&zmS*rA*KkC=rFH
z0;hxpXZYk=2o~jghdy~`x!{A;6Bt1sm_39Sd6B*pY0PR^b2L0MrhcZrc+wQd`^rOg
zI%c3xLgT^irs{p5baJ*Hv)|5+F7cP)XpyA7@O&XSuLt;Yjtu7yF{{EXBsAMHgYqmt
z9CgUDXPX*U>I0>1apnGHVJ~-?cEPuDEw27sm#6Z@@tVitiSJn?p~0wlVVmTdP4O5u
zwb+*Mi~?D{F0F}Ub`<N5cG1Zb!HfZ{UGa*QdC@oVlI3}J;#>D%$${W{BHUwZUN5+=
zP=ovx#G~3;;ot`4;R5|~q*B(Z6uR&}6=lWIeesuJQR;4mgL}uKO|t(KW!T?WS8&CA
zn0R{4z(_HN-;^S;w6Yq{p3lYXv)Ndb(zGPni%B$tNNak>noI8_CV7t;yjAhJg}LQ;
zaQ|_9_Qx;c55ND+v=&8eO-jh}lIYdqir~HGous_ul+r_aWlLK&I$=E#gV8pLR{vsP
zkQp_uR&2Jn(=^+&No%5gn^KTB)qYY|HaQiXtCYQs9kme~4#=xM2l|#%-;X6H`l`Ng
zqCAi<kgbmP_I*#d>4E4%&WfXKG5Cy94q?GchfQbN)bHufdah<KMTJb8jF)O2`YJHz
zC!jlyl@I7w6)!Qs$e?%JS8XKEw~0Is_G7(IcXp$&#DVWP@}q18t23K3*EBs|`w&73
zBV)86r^FexRAB^&&GlToIwpqH+1ch%)-`N^Nd=8m1h14Rl?X<Ja<nrE0l<`SuXmmX
zQalI-0+Y3G{SexVXY>}pK>Q(YWdSzC2d2uFWHb;01%f$w(wQhb+ct7fUYL|X-~c+K
zIc@@lv$zZ<Uz1=oG|(FZV#=M$SMsCHL%g-Dm;?<(Ss=EjN_Y2=-ED-F62K}RW>`{3
zH>ApH0!dk19zh7e=^fxvke$oRRe>hxdStf<nx1&MeuC0wqw-n#miKOC-rEbN2tv9C
z#t7q##53c-4j$u^W8MysQ72C1T#+!gsg6V1bSfnjotVf-7(vtag23iBDH8~1Qn19d
z$O|prHoGu9>~`J8v~?;YEcV27gT?9`pvI0qC}$`Lw9C-&NZh)8+oSG`#4Jk?XS4!X
zuyrkhphpfN&~k>M?lBpH(dMFRguFbbBDeZXC4y^?-eO`P<pa|fDG}uR0!0GCjKnTt
z*D3+4tG7kMoRp<@luRjt3g6gV_oxx3ADHt=d=LzpEflj$6N{>ufno7KQ@5)kTnh_J
zRtQk2dwMys`7k!NIL|<JRvkDwoh#+0ek^cjZ31;cI9^^xNdl*-aZ}=dtxE*;*l9l(
z@qPYGOy>Ebh$HoB7V*Qxf%+jNG{?lmpvpVq(Z+?vdE5H~6Ho+!5?64)t+m}mlo=fa
z&W&|UfM61g;Ik#96f{f{ga??NaUKLy&TJ1Hf(L|(86*S`1_c<2V}^-$th*T*8CJVb
z*#rcfWUT^<T9a@IWprEdIrWz91LqbN1-mWbWw*a!HsY2SR!<6U?Xe@g$=ySiW1f1g
zt#}6|On?|g#WQPI>+Dwl>{NaX8ngxnR4y;EJ?^@2=TJp%jKjTSF|U<){&Y6xo-3`D
zHIIlP5Ob<{u44dZC$U$~ix7=qHl%N`Ki+!#?fCcq;eUz`e)uE7deWoPjN<>xzx+CW
z`HNq9M&jFVzKNy9MdiPruvgoFf0&Ih*x}QLp*if*q&|W$PZ+!C?5n>*d!cc}MZlzQ
zdL5Vq#ACG=WW@d6^i+5+;fgke&LUiL#UapA7y4dzkAylY@oaEF-ljvk!0<_v(Q{_I
z_I3`1S1h1c`*6ZC_$a198=f%@<y~l!g$&=}Ni`ks{f=^0u0E`U*IzIDddD;00uxNT
zc&4w@pUh+nQIxCy7Mh4q4sBvaCTF9QcI5j^*J;T4D(0I5KDx!LZ6voVUfz3gzuL3#
zHgJbgblRJ?0@uNN`WQ17X1e-=H;_Lwo|q3}cE@6F*06c4lRQGao$J7?Y~lePb34vx
zpx?E&Agl||$y;q_K}G#D@j)Iwq`J`lv@agqCRu0&IKa6hEXc>yl`Hvmu_jJ@G8YeT
z3kO&@eyXr`DFCd=>a_&|Yb05ZhC;*pT`3dWqMHm*J%92fzWU+|&wN@SlOVRd$oj51
z3yml}m^M-e??xy2!E%AhTG}H)knQMe-h%S37eO=?Wp+L7Gwia*If*I_T)yLFMmSI7
zyUzud#lfdb*?0$yB^_{q{tOexd}n!S*}`a9Ga=M5Di|H)v7s0lAB|D*$WGBw`uh6n
zTFi^quS=k2r!P!<z2;0bntb8y=kO@y;z8lcg?H(>ulk>heqR1k_HxCM4sfsUpY`vH
zD{%NzaTJ>ARL=>+oFM}bnJI;rWB$ZiOPlvfIA`nbg{38Lg2QYE=1`dCGKd0SWy20;
z@Bi@0r}4#?UkQKOr4)C27e21+(q|J$WFTln5=sgrm?x~*8XTVN!<ON3z!+SIe@QuD
zc4$j@k36y|oMtc(C5QE(%&Ke($IC1WVQZ3~xNdH&*|S62(vlRU#Rbb(n3obpdS#6v
zCSGZ7i+n;E(nHJGD216u*5m;zU{;iDtX3G@pgs&}o9-*T^cS}=ZQ|#Ahp(CpAb=sJ
z8E&Q0K;P>!!pqv&j$mkIKKZ`hHBFASeaNT}1<2V1brKOCuVWCA+F09mrx+ROlfctz
z0Ssm0Nbl$}m{DLSMzU%?XO!tcut6x|=^X$fN1TxjVgj*~a4%sCZ${h6OaMYoc3LxI
zfK^q{9a*;sXeb8;pu%~LAWoj#LqIWo;2wfb985I;YqNy+j<%*4=<A6=w$_wjLL;&h
zgZK4TBSSR2n6ndegm%|=PYs{8uqs2<p0Fs2L`G&vy7Ejp+*$Pu6XM=XPTt=u60r7r
zrIkb@>*2=anY777&h@PGe$Q11gUZ3eSGER4xZ`{QwvwE=Ivuy}+zj@_#smjjOl%rS
z7^J2H_6-cgfQ0=QQWzSWux*z@CxwF9iky}M#xS_b#6{l9o$BDxKGlOVTCk@h=xRo#
zt9Le07$d5$y!vWPPR}TwaLSz%AvL`ClaX$6I-jZ0pVFo?f<KI?myX8BB&JMMf^U*w
zslAjgvl3;e8*mHe$tVwSA=%cph8P&?jF}r_F?nT7Nl^CF*w~ivxwj?8-Vzg1z(6<p
zK<?lJqm}?BK>+h02;su~ay)smE`fhH*4B5uGt&J0g8N4QK)=nuhDS!Mc%Wo))<eG(
zS8#%zl;-D_L~t;jNHn5sRB3Z_FJ@=gl=h00h;{Gu#O_D7@!w<HL`=65#^Bn>XkScE
zPq`i_L~O-M+m!Ppz<UTMkU|lHn0o|O8Sm~ww3VsS+TA_2X&198#9@6AYc!aSX4={d
ztJU7DkzrcDON5-+A`}(ZPeLmT&@peI|FA7;iv+CZ=61p0#On!`Ea0<f?@%<OEvBZY
zjT6(lVot%qRkVebh;_DMUX1lk3I5(~=7ri$${jNhQt<GU1zbIOAw^}2g?^$-D^i%3
z5oQ=M5{~bm#_G~WJbCmy76tR2O(_eii)o-dbP9nBlMjUYL*Xr>YJ($#ar4gY_|Z>(
z5<mUf&*RSByPj2J#^B*Yc7ppPe)qdi;{N>yu_6J29nNgRqILv-JxCz@LFvIvhDJuQ
z;N)sdzu3AK`mP@1_za35yfN$0uD(DUk{CbJ4L!S;i<y9j^Z^7}+5n~n-J(B3b6M*K
z%^Xsjpwxp4C@aj?xetNMQbrjq!xU#vc(A`M<|ewv7QiSNFds8#U7;~ao9ev4cjok*
zmn>^>B`#rHFS`;};^DySawUv<d&fKAA?I0F)uF`WtNQP8lsUgkTIE3&eN&W)q-W6R
zOan(>!lcPcWQ+1`RT^r)M(pB6OA-D(V<DbMT)u<`A()U(U$2C2gd5=qJ1seB%B7!`
zdwuJ*Jj_<LgU2eH6j=lz%0fSNdE|3|De0iwq2d|rfP*<6g6i3M1j1zzZs(UC8k`Cj
z_z3M(sM4U8D59xgU;+xhz}I_JZg$pVz-32$k1O1~CVXTTl3keSugs8f{s8nGXdmqE
z#J1@8#>#3e&#%Vp(;PKkoSXN~Wt?okG`}R=+;`cBhd9$H2O%&AV$cLYQvZTqE~PK$
zJjE50t6sm<)kXVW_m{@0cjQ?gR??uX&<@NhF*~GBA$YDWuf?9|C_;gn(OylKb<sVF
z<c&MG#DA`f7WAkOAH>R{1ZgQ0nCsv$YqE*0lnM9+-@09u2hYVNyehcK*Z;^dgZ1Qi
z=`P=wzF8lR<GJ*YaxHN#$Hfuk!3S>Sxf*x5*L&atF47MVg|pxh0{p6Ekrhmy1Usx1
z@SbT94`O`;JN$7j5V*3svLU*Sr2rd-jQG$|tO4BD)wlBWdC^YMKFUC;;aT*pBjGI8
z0r1X4U76|9pW>)-=vyf4YRjC_W&I!nrb<Wf5hN|^I34vXZ?2*%eUSCUC=Qq-&kOG{
zNnT!F7N5(rLXv*c`Je-al;8;+6Fk%p&tx0*OXL{Ji0tV+D5pL${DEa-4jchz%9Zlh
zc{}4S(^{ZkXn6Lk4EuhU?;O8s_sh=dJKDwfgg1>2HGaqzc6?iyL77^Hcw-(wrx+Px
zM;<8?><XmDK0Fd~lK=!!-&u>+m2YA8JOeC%L`MLz$OPexf|`H>6y4J?YZBU-iX|HM
z;t*hvuqY~2fGgjVHu(@Y8MhEZOuLk!m=octB#j3U^;(2Y3W}78b}<PK%Ni83<p4Yy
zkOhlp%7n_rwwhf%sy7F4=^0JdCR3DIprQ^g8gNiRYKU1zgyC#l!9<A3`_jRH2TTr>
zr>t}c%1(VU#i}igh?#!+3R$Xnc+zq@s#go9IZ~2(dpo1Mw-z^U-w00Q#gvFq<R_0G
zn{Y9kU}mDWW3lg$JiF;&$Bvm3i*HdR=ycX@S&VDu%;9p_rL&6_$PaT$VH*LJDeQG=
z=1lkcYzKv*b;C5@c>RqS#vDi;m))KazIi_-VHlfl>Sg>9thfjMC?micB18u!AMi;*
zh8F<fJslI}f%OXrO~yHOP|CmpY)qZl=CwsaLhoQLrml_3PZ&=zUEA8-5MJ-7jv*nu
zUG0G~p)x2PM)Da|WJKj)?<iJR*&25<=4SWf+4C2w$9^m<ZK}LG-X&*za@;3KV;8?F
zShshkG6At#Ut5!sFe{;-Q5u+gw&AI+`+IDiD@MD*0aE+QkHwbMg9vO3i$am3)d)7s
zJPi%?dm%IT%t*0qA2W+6G%Tv6J(x*kM<||w3vB_6SzAQL>^h_FraCKK;SuRGO1-qW
zA<tI75;PA}$YHBs?DZL8<V1D)CnNQkQ1x=Qglfa7sVEbTjcp?04Y8rRZtqBVQ5d65
z-95cAJ#$5P`ASSkk!0&Z%C)__WhG)|O~U#1hU&sDM8|3q)wQD{^b8P4$iZZc?E=-T
zg5#kW(snFMY1mjpu*RfBisjC6tV_9AnqLvaUW-G)4vubNn+8@NRSOHa!eb?b64Bk?
z?FH7azxhV|>|gyne)Qv?NHM(be#Zt3_wPT7k3Ranw|!;D6-=-YzQ8Zc7?~|C=RL4K
zfi)4zNZGPMLoRY<BLKpv8v`ONocH#~tbtG+5lB(KZGK3f*UuwMDk_3vDub+z*%uyd
zTlr(;O-XuJ8BCy3*^+Y7-DQ}eBq5~OzF!Q;OqByYrARO{afCUk+W43?Q-To;%OlXx
z8`H3)KOigkYOiP0<?8+{PMv#MSiO7MyTZgw3k0(qRnM6h=l~v9%23iM-zmiS4W8q3
z;{awZiy)TYIDVCVy)wbVS{BE1E%zwQo~@wEpoywI5qCXXp>kqw#^`>Uno2nkK6s$4
zqUrQC6i9Z^!u%9okhF}lq3o!Wt2~s`12A|53VE)WN}b4~JOG|RVZR)S|DZ^)o+)KZ
zL&9yrAQfpsE9uA7(frZ8hu*1J`pE0elxK^C6O;+oKXE<{qvFS+nJnaGdw4eC=w>V0
zj?Un4HSGSO^X!a>$;h^Z!5z^>PCq_9R&P?hcQ#Sxg_Gi?Y>9h(c;eZWiO~t+c$e2@
zvB)1`%cWQG-G=&e9r#B{#djQ_f7GyQUc2&5`YP~y+;`rWWQ!v$U8-!;5oZ2U0?xl7
z9C)G(dJ_`SA_i5MnegIb@d&<xbC<8(OoCtUP;cCN<&~Jedd;)Y2%Rg8%1fYzj=)Rl
zql{oflL}7Pj{s+pGUbB%{IB=b>+iS{zqpq@pQU-}9nLa|9OWt==eYTImIiT+cb08*
zFP{sVn=Ci7O+K>Jw(8BgIcBO?*QAUtF6+8!#e&%j<S}+?@*-<Nl^q6AFxdUf#}r7Y
zhd-Ci4Cv;1C4NkN20f8ukhz#)q=YIDvmRJqNLguDdbCAX7dv$Ji;oTHI-uu1;VWxF
zIfg(AcL(OZtWy(xmS+|cxd_?$Y;iZ{NBZwIh0InOKf3%49;F<`#f!QX-|IGv=UnAg
zzAPvB2COq(zTqT%zR|~qDYV(Hw3C{`VScW{s&?Qp)$`)2_up|D{<5|&?)%;m-ZVVe
z@F6=+Foj|Qq!&!9dZ!jKoR#G*S8I4!LUpaxJE&muc%a5(0ZOj&PodojiUGhRaS(nw
z8bpptK^7pgIG#;-ML_AG96W^(&K|}b%7k{v_Xw4E@Rs)o`Z(%@PzAFoTR3tGHB1!Z
z4I-y!CLk;8@`)*MFb}&ob#}DJP-W(Xrpaiz59>M*F;;#MjBb<-G3X;<>81p7rp-*G
zm5vnv0mlT}8gX{CBHX6(0CI^rnL#VQ%5DK1b22?sb{M|Ro2q(cSu4e%a=@$90q3~R
zkWuI6=;t)k{?6#>>5!0_i^}i3ayuARWhW&F6y_KU^RpIQP}qRg<kWQZ4~|#>Y?1(I
zTU;SFoidpf`9chUvpONBPAW6Z5PL-8P-hq-W@3z#qPW1Ae8z*|#R%bD36Db~V@jtn
z78V!c>C>kk+2iD5lsRBwqN6rLfbj^M(gc<s5r9F$bipEkPRaO2@j+Uqg_$0v48R^`
zf>XP9g-^`dKw7LU^tQ#oNSA~?5d<;z;}-`~BpCfhfd%_>6dd8e_$xw1-`EvnRFWca
z?2rS|LZNDwz${#<wMnS&P#?i$O#P&@1F*%?(xOMN7#;Sk#L|+Ls3UzZp5gT94(bA`
zA4)VMQyGq;K^xHCj5MPVZ%a^S7qzzbrsz}odpM7vU&KoNl5#TxgUJ){WP}tuQiw4o
z7znJi9oz6S!q1w6lM@u^9WOqfoE(VJ(P0S}y(-VC`#p=pSsMUUI@+5>WNZGOe#qz&
zBO{F9cvmmwcc|}JLjo>Qk1h7smm<+7f;u=n8iPY4%4aI3)aKavGxf|tNS=xC;_4L%
z##<78_a%IxIIu1PqAUeZWk<+GsfEZQ2y;6BvdYQ9n(I<1_V?6QM-39Rr1ULI>6=@M
zt#!4n1Q&=h3(nmK-N#a(XIC2*sSb~i#OrUp7Vm%Xf&E)=zoYUFdw}8Iy$A8-7hlEi
zfA<F~60@^&UWbEFhW)uxGF@o{Z-i5rGxe(vLL)qVO@k^=%#XOSxry+V-R%a42G16)
zmXU5~7W4uK-dOOYF9@jI)92_P&@4=7P!LcUX=g_1=!=+&(Q*jEv>(hQ6>{O2gg-`%
zw>PA0N+`v&i5Uq@pKON@V}y1YUWM+EimPF7=Z89)R9%b9?+^^DZ@(vAal|Wefq_kZ
z#0Q`W(9E*sDXyelT=|yr8h*k{uI|I?U^#2bcle|q=jBOyZl$7CG`UURdv+p~LwF3z
zqI$i^ZDj_m_)1=Imq0G1twV|+91c@c?;xkXnCVu^gwo`k2NV;8Ll3m5y-^$}8znkY
zpTf+DGfuGiKD)}>tO*m4>IzD^jueK5Q02gl+)A6-H|RNoYV;!pf-<gx6g%O>8V3|u
z%$!jWZ2Jw36>gzyFw%=~Rg*%|Ea3_G(FQ0q$IyBy5{El`(V?`v+16NB(rN~Gl#9>V
z5WOOw$%!e?-uTc=(Lk3K{-WZWkJ1n>SzfbFez@SW{b>9AF2nM<ILe93xaS_ntFz}S
z>_T?cJL5)mz2KE6ydKz~^n+u}{xb8=3?Yl^t52OCo@zF!uTNjU64!3svbiLKay#35
zHV0dkLcxjRD52hZS9k$@=zqB6jhTPqFsYxr(r3ucc~`G5$NRqTE{E5<@{Vi$ef_s3
z9OpYd7kR}zU7f(ViX4O&o**+ybg@!_EMhY#@m4nXLB3q@W+8J{F5qFglT@>6*yu&v
zC?0GwvLSrJ+?cJ-H$8B*FBy#iFlOk)r}on{p7pnf445z@dx|ihrm@I84e|_rwYDPk
z_lPc|ICQIhdn7CMM|YPL41D*1`uc!)>0tEs42f6vi~kPC(BOECj7*CEPe|4plN>gz
zaJ}#4kVy3(#p{q_(%L3PND2u`OpEfe6-aB%S9VThlN)C9TU%MPn(rMZWdcP5`5d|r
zU2GDaH-4xcp#R9oi7x^wfORvRw~gD&_r?E@x%#%QcZ4@hO|*X4AmGxmy$gW?Z7;8e
z!-jfwC1)l^M|%WJ3Ee_?w8lL(7Gryi-qXPl`aH5CqLhoq&=f^{RSD3}3ITUY%Ni@+
zi0EOKLh%?)+ug%{ZA)ok6W$_WL&VnUCE}7ErV<`y5<$j{z=w4)%Bs>p7}@U{#tz{p
zA53UKLIBK<1<mZ(Qj?&9J^WyQue3}SD09%-wzXvzBa=tYO!P*Fm<7rN`$99Cpv2s8
zz>1P_rE|Dp&HPa&h?9+3OoJTXAmA&6hDsTdkbzLF*M_A%Y1T14dmzQ1?;$eD5PDU<
zndvbRY;2@CzoDkGH^eKi+>Kke@5J!Ps7I7gCLTPvrw-F5Ms_Gz^~RN}*QG4=sJtql
z7Y0#}M9NfX5NgOR^QUa-bX6T)l5AEv>XU7Pki|$m3!<n$<}4FaGiIKQ>aw7g_8%LS
zASLCdOq1eiZv-FO9%ZyB6Zn!l!iAZk6e8dVF~T+x!UX&g6qzoDv3rz9@kNis7+a#V
zr$s{EzD-b2CM2dP4Yh~xi1jPoT_|O}7S_CtAq<gO8Fg%Mo0*@9i9W%uQ@%BB-|6=j
zt0^oB%PlOhpZb;<XpUaNaF$nAy=Fjip3-6gEHe_q>tpInpGllSh_T5YeNyj4O9bFk
z^^F%Q_es<`F;7FNKUTRg`BNN~Un%H2MHLJ&vBLHfA(Zt)5F{Lg9T?IMD0}SkH96fM
zuf6uF>T=an$iNigiL`-dYujmbcD30A5oW;lttfgBc44W?Apu$R03r_IR?=z@l)EE+
zTZ=2#ZpdGc5w$g2DS|_+gW$}I^^G+#0QD<QDwhC;V2%LRDS@0_XrK#ToZ7@leVZ8V
ziA|Z77Nji5uZTYFNqJyI&|3@Yy1ud=OY@5!@e|mpI;SuQZZJ|0TC>2KMflTKXX3|y
z_2c-Pzxy}w&ig;`%n7vS>C@*@AU=tI{BQphpGcAT_M3akd!=Ib2(`K8H4CUwe7dFF
z_4Id(d59r&2*;$Xc{2dzNA4E<R3PvXVG<fy6K;%1X<?&>rKM#HTHcyf{nKU+io?A}
zsRj2UWSF4{p-~W_*C^m9?#v1<uW~Z&hTDo6Uj!fOL%HcsC=Co0?QBVK5xf{xMhIet
zjJ8x!tNxfYPJfjH$Fi@xpPq&PX-C}RGQE5*bRFmPs%P$@mpQwD5W(Pra9@Z;{T;{u
z&(ka4B6#sF?LuDy*2khFtc`NtRJ8I99_NecnlM`Zl8rn}3xp%3KSGKzdvqUDI9q$O
zrm9nYrCa@ib6!wZIdcVbA#bm$w$9lVc?PdK*@S@+P2eq|pEi|0Rr%a2Rkwy0DkCFV
zY71qTUeDmM>t>~xodCr*)UMFL6e9RM2eC{?lpL^QP^3%cvqCQ7;Q9z}YzAhYrMf9?
zlm`?LdNd<xY>>g|>%btpvT>*+Rs)K!@)TuaUqT(bmbJP}f=5fuN-;_;vzH7&F&G1m
z&P-pC0M=`v+q546NNG`Y^Ay!Bv5Q+qYH5uMN2<2P(N^c}me-fL@~Q$&I>M^(`cKO}
zSA1rNAbgG|1;*u`{s*5!xMMdv7X896_}0~-*otqjfcEZM+<E<O%v`%FVRJz6mOQbt
z=7V3^rhac*ikWCF$^^4Z*-nBtj>KGW@R$3SyHeJ2|FUbnt51vjm&L&mUhgkE;6XUJ
z<Ll)(JQ8N_T9`+=PXS|Hkq;RthEMKD#@ZH48N5wXs8w%nf=sp}8izUVhTsjqVb+@g
zF=jLHd8ljgY`xAF?K`oyE=5IG4o2lnBs}J^m=0qWyS2HeFwW4zd<ipP@v%ddEyRDs
zcbc29+USV37U*f0_!4G~y^<BASoDjpNud}Vnuw9%sTdxciUHks_YFsfl#te1Pc*kP
zQ{AI`$(yylXqWHMeT~@<y=zn7L4m-ut6gbx-y{WtMb_{&c4^%4K<0tUBc4XTuVhBr
zuCyPmPzZAYInHhEhvNT~JAYr=y!hf?cHi{M-Tn{R+KdrLOfeX>#FT)7!tfBV5l|*3
zm;&lDBVH1d-r19uXKDZulj0yEMu9;jI+|^}RQ|G+E6faL0OG+U$e(X02=Nj85CDrH
z4Vv#t5cIM0JP6|;Ko<5TgfVRkLHEv6N{d;EzWyE&zdq4BPJ2aQXtbc*uDU@85njEx
zOhll!r&AiGG)5^K5Cm@_$+X&m1OPEs7279%*1Ef+o$X=u-MW<tY#V8~>_8Naj!#%B
z5)jGsl&jPak3&EpocJ&rW^r<6CBw=$_?J2CW>F%Fq5(|1BzR4Xp)oUhBxUQAlXAD?
z>h)_eedU^&0m8_$XV2o{y?YjFXcO$;XRcgV{!%_9C?jDR86gW*{dlG>@Zx;dPsmFC
zS<gyglrMFqeOXM#$ULS#Y!`d|`V9$fS1k}dfBxJv6U<_8a2=)@e8=~sXA>$54vdyc
zG0-#Lfsf=5LjxxelG)Q7L7B8T*MiYLkLDsgi7;82(0AY=Ye~8UqwanQd+n-cgVK~R
zjZlv84D)AJ4s(&g{viuoYs>3O<3RY*ukV?@Y?Dxcve6v_1N|{PJRo?ykYK*4e8tF>
z7BgI{YpZ71C?zmg%z`*E6a^h&y|JNH<wH<J=@!UUA9yvhA}OHKPY`5O7uBb+nFYP4
zB0>@p+615D-I$->)OUO817cQEtb4%e;X%WTleSTO*^ZI6LeON!Xj{tq$})$;Jri+V
zQ~CSjz4zV=X0S%b#)7pA9TIHmGxXW#vur1fVuA3Axf{y4v<#)obTdLeCNdiv5P8h3
zSZgC;M?{e`7+!t-^|-CHff*z6<iEPQA_BjvzJ~I*ujYgfs47TKY6rGshEY?7RK%I)
z<+Sa_c=lvbIQ%SDSEPLGiKbxE#2Sws;XLa;RByJeWagk*i84@-gA4Rj!BTm-pDI7r
zmAwAeoAGb|{_o@O{>{IOYd7zTLG;Ad#!h_k`BzpVe))@kvGR+cPuXc3Mw%^HDh*~(
zP~3Y5x?@}l{pciHX|{O{3$&6&*33@0-y<NX?;#k2r)-L_qckzk!!LTqZdOXtvhlf?
ztfWcVG2LA*X#{lCLbH4kbYQm7OJ+}(<~RJ@C7kH&>ar=pqeqWzW+LjLA3MLPe_4r8
zzY`8(r;JTLGjX07AtO6rhvx#KD!h6}pZ!y=(1mivl`CQOu7npJ0nPz$!aeY01uN6J
zoNEd9w_>#PQ{V&qm>J@|Yf)wCxTNE&^3)aFL*VvdqWab*s7jAn3vWGK(dzUPCK1+E
zPH!u$w%${FvWS`Xg=aET#bKy6ixO>MgAgkn6hM88S%3yz8-@05D1}Og(S3GJ>+DF+
z+J&0<a!oXZbv1YtPp{8l!MA82^`suq&ZNf-DkLpNA*Vm_P2p2kaGJ|vZiuPAHCz%i
zpu3o`vcbXF_=M=(h(}PF9VI@B$3DXI&TgDYSzzZk%m|@1g=QPxN*bkvg|pYLU-w29
zEJS_s<e9?R7($?d4C^M_=Ddy6niR+nj;DWk9Ca;@Z}RHPS-krAPOtLi%eo}&GJXp1
z;1IAX{i@y(=6)yG8o$I>>ANgo2S+IeNeh*H<(Wax(aGU>=Yt=_%++h6)g7Wy8!`Lz
zc`VE=#IA%y24a~Vf-lpC&<e_sJBSI8PASW1I8_xcu;5yJarN)M&%NxsKj%vR`1(9j
zF~|lj;q_(J3)KSLVlagbV3Mz3!pe*)G>4k{uvXE3);d79w$!IMe3P}h@OXBJ;*1VV
zju~9WRBBgz_22-0F5G07Iz4k<)aDj^X~zl$=Q}Z&$I0BQQaV;wH)CZ*UOawrVJ+ro
zmqhCq?B{g9$gY1&m|n3aaz)p*Sdw3sU)Q%gQapUR_pWHnf%-BBqBcmjRUeARR(Uo-
z5wCn98i1E#65@?{uPg82rQD1DLXX)glmWSIDM6d6-_rbo@q`1G8ED!OEC4N5HOzK3
zKp%mz`ki|U9m<$jd4ryJdUBf-pU&-8wbf;J9{=U;^0!T|y(wj)13O)IN2y7XpkbNT
zrE=^zvuRCedYV1$dsQe#=-1;&+z0a<M!8WOsEqB_L`dii6apM#%H2Y|5~QQJcytsc
z%IFG24h4qF8t1asPQya12wf3u3Bv?Q!6Nt~z+&XN_@-G(LrofDE<9ygdv^~F9|5vM
z3Bpv88Raw)`{B2G;WYvd8lMkMV|q;K@5vvk%$RU>_DZ-CGv*nS4C1*Hx#OrdDi@ue
zFqk)aDyr*(a7vjWJP@JmSZQZVeGp0*wX(3|cSX?<j!|cj(r`>z5|X1!gp-HzRDK*f
zR;ytOuqP#AL-1in=TM5mRVfqKC1m!S;LS;yn3WLEXeV&SjALSQS}=mqs51+x5i%$O
z`K5D<xx8KksJ5oPxzgsmElinbRahUhBh9WlF`a!!OtoFg=%f`^MqXH~^i0AmvryO@
zV~3CMh)Xyy9m&v~a8UJQG?kHbHUxm#Q8@BKXuuo}{9=j}VHulKzE50HXIH%;_&7|z
z#wP}2>dHtA4E3n}m_YzfSc{NEo3a9TPOIh21<YkOMU2|pB*n|YKoE3*Iotd(i^2AY
zD=SMDATh0?+)RPPJcoxz^i_k>pnMW=#l)wluEyg0s@Eo{nZ!hd3+fBm_XGzcLE;y;
zOwXc3OwXYFv6!3oO8Z(49t6+mze7?EIh~!gJuC>M@0mFYXMhp#@PUFNo=Zy{<ku5#
zzjIrR{;F_S{X<F@rbob$z0Q{<^e?IYtleM>S7-u+o<s1gc#2?3G!X$m)i(!f&x2sw
z;L$Pmo1gSd1cEFJ(^pp6>4KdwR#iSsp@gHDWC#HCU(Q9rV+Mk)Kls3iCRl$Y813#H
z#>&#R@OMRZI1tW>;WT#$4>*CFQJWOxIGvY~UW7r;j$tj47t)^~R4O50#SUFVL-E$T
zZ^zI7_HW|t_usdv-?s92_~40WBmU_he-VHD<kOg&o3nC=QkpXh{4qL>IgHY8ZIm*>
zh6kN4Hw&%V?sfkNb2~=FC13~_@z5UH1m)JV1ZqFbjo9XJaY6a0zc4aKHXQngaLGus
zX@+Q)x}4j9&I^Wc1wmlobXP|AP^@Vm;LCa)cD;fg+I(O`FcPfvMGAiEVk~Y%iD-n$
zNyD<-B1atcGTi~c5_ZuQMpCZyJMQZ}u3T~DzVvgRp=TvNzVxvou;IK5S=EnmB^_M3
z=So^u_{CRH9-%kT64ylk@XAGTlfGu2c?CC^U1ZR}3Wf4F-xB@tjDYabG)D~q15brO
zd5G5>NqE8h2nB(ZStnqHMBg(T4zFudomc?PmeuNz(W0`pOIdDLI7$Q0%oeb1H-lWP
zWytNjp>sJ%Lz$ee=`aHY2_F;~3$qnmfg5r87I@?gJxT+RP_Hs=fhVcnm?KdgW<~fG
zcqkmXU|jupcw}7srOWuj1_0D^o6Q>JSvZVx;Xw=48C!Nk7@X8CgL=S|wMTvZgTbyn
zj~_oaPBH=Rrqe%)#g|p_d^LY@n^M;F`7R@sxK*&9%YAVr4$fIAE#q7DbGR;~4b1Si
z%y8RNu5&ys?+ac6OO#d2)6NtqrA3xSnZUela%Mc<|Ir8W%G+;PKG_#vSz+7o`K4G@
z-``kURXQwg7mY%x5TLV6F5Ou>Ti&qCO7SF=)jRcHMTRuEtW~`(y8=UeahLmE{ri%>
zulD+%ug>67?o(b%@37+4hFtkJ<7N4rkMiba%6S<N*}}ew9Y)3DkQs?5&p^qZln~Au
zVhtP01iLF?KD@ZN<QWVGFR=h%;LaO}Y{GYUO!pWRWH&L+fnu#0Yg|}E>T`}50MtD@
zBH}q8ipA9ItT;P6Z<8y|Rl-b}-5BTR@lrgVKli}g6AAl|A3ur5kDkQ+2ak0BOfrs?
zfTy!QB>Bl>y?^*P9^88v559d6_rCcy?tS%5-2dv^`1Whve|<0RD;)ogp1=6~%lP86
zFXFRL{up0<{)N7~7qiNX%{~}FXBSe=e?gf@ASV2>N!(I`u^vp&B{vC2vhj-Nebt!X
z<MREVj#KX%{?q?FcG_^#5hq6wyqXGBHLx0y#Z(BNFQi2N@PpfN>(-QaSNZbWd$A$G
zp`kjHfUO@{C6B<)s1||-+lt!Oe0N_S0UV@ZiXPiy5l|5~2sQ<#1JNn1fvfC<2{yW-
zfWU+)67gPGz!PRkc*4Xbi9R0TjwxN32s^^C7co21uwcg{ZYmoBsYGBAZ!v%D&N=8w
zLct3N<PnVLg%8qNkrulr0>O;L#Fd#Clu)!OA?4nK2NJ?oEqKuhZLeQ72(vU+BvkP1
z90;D-X&9A7NTtKWObCZAum?D0N3o*O^`+tCwB;cWpatn!2~vJ%)BXx8BO$##Y~#ns
zncAdbUj(bovob&V>EFgr|N8$Sh2oaa2KmST@}J`${`3Dwn)#{(z+;=9y!G}6QkaGW
zf0Y*|33IBH7s~D#ChDj#6D@akF-tKy+l!lcQ1{&3gl$<7e%9BQ;_1`J>XggD;@l5D
z_>mOSE2?XgcftAn??17f{2On+5wE=ZO7!+hxK}>ZsVD-}(WB|=_y~-&6BS0NU~9Ev
z;=w4|66+Bb7i?MtQDzMV+qtbtAVv@b_M_ti@%r0W<Mta9vG-yvHnvt`dv`<M@2gxK
zPS@@&PO-7ahJ0spCl)01Z*T3#)hm;pMM(uRXT`ud+R)o0&OM8jHP&97s=nR1E}(x{
z!gQ|%w)y#GDG<wEka+r{CI0Zc$FaO1L5Ep3T3K+%mJ~ck`Nc-^SoLBJ1WapuvNP`7
zol%=jL|0dv`oXpc;f93YdthJOym>X=c;l`Vn4Wn0{7KA9IE7)KD<p>^jw>oFGdwK*
zjktb&G=BV}w<Y+Gnb_^^Af#iO@FJGP<d%dpj~_`%T6>}TqI}k3Xlz(Y`+(AAkv@3H
z$=ND_@Ta4#TT0`M`dqi(H8_7pp-?6c4&~LRC=Y2OME_uMbt(|plQV*cjei-rLa}G0
zhBGmCw%E~QM|DO4L*eBRFzCUyNB1!=p?^0(W0W>fWcCs!1%A+HQH0p;l??~3-Mkqe
z{OCu%vW4x&>UKQ5_awgl@|*bdkAI9u4<ATaS~rf<H&F}_<|wmGU@(1=7p7RLU@(Kj
zA=zG-L4y5#4glPd5-+7e?e9g-!YfD|?QKRZUl?(Nl*42PoRFPsTu@HPYiD=6f=QY?
z@f{{47AVDZF<W6_BMnXeR1b<p^)1$+jf@Qiy9cojjor@JF1Z6?N$tlXUJmd(JyKmx
zq!g;0Eb^sj(6SUfQ;;$eRprH%0R)@{fyx)*q&__9RSFrG-|>uK#<e`B5DaYxme3>N
zfN%$az_3yrQt`F&C4Ce=cmT7VC=zV-h(PNB8Q`w|30-B8wvv^jd`b(pC#AEs)*M|u
zH7Q|jQVe+^+!jqbYj@$4O^N7RYD{(rWUT?S3-E5}G<=VC0DhE<1*fL*awr01?8CL7
zr}{xNXv?gdqbZ8MBRz0JVempQhtd<kVp3%*k3f4Yd`nw<cqwg{>m}fk%%pNY31$-#
z$f`bwgMWDdh&m}{@&w!+QjSr+*~0jZx8907ue}*PgTtbg`&LR<L~EDl=VC#^&8qI%
zZy$KFiAS~vQUo&;$D+qOYKxxutH1hd(V2<(`s;7vAO7LLdiD%^c{^Z8UG)!Gxx%6?
zi4U|XVTmtz;^BY~jyUE1WzI<&h<T-q#VJJxgQF;tMPb0fi=e@{f}=R%8mE;hH0K1_
zg@II_^`1W7qVmw>!=nT7#(QtZ4}bnwF)%ejO0hC4e(~UGJbCaWo;`VPq4n^X8FF}E
zDkUz5@}MoP^tqifZx2i;6G~Ico+U%SMTSVc1tg?ZTz$Q93Ey*gT!0JTRd6nj_v%J|
zXS}Y`F5lHZ<5S+L!uVw;tG=b2iG#=!xH1?EZR0B!LEmS5K~DY8W#C!i&@hztQV-|_
zvmB+ZpdnZRoGIZ9$nZUh>Ca_UA?TCHl6q8^Od3i!p0x7YR3NR$b9FpwMOu_6J?-Lp
zxCXeYZmO5*xA4jJ0=K}OO5ULbvmLYFu=evv^twTPwdVDT?V?K@JX(t@*C*pgKY2f{
zy?Q4)`@2ajP9zV(4}q8K!9RMn-Z7YFGkm#12i+o$bM8v}6_@WX#`#mO>9u#-KV*T+
z{QR~wyFJ_fF(bi<9-U!ld(*^jOoG_JP_J4)Bxvu7*m7zn3okK^%2^E#kHSO>GZBPm
zh^p;<Q8ENDfy>jwo+4JmBN!umLfmp>SS3=PRF=58=(1?#*}#nCcQWE8K`u=~Xatc2
z0S#tp+j$9(Y(a<6WT6J(5I|QA=Z06~rVTQScg5^chV-7jh}i=6P{e^zVg%uK)n`}r
zSl`+<V`8)sq|05Tazq)TLu1f)5V}3Gh{UFvu=O4zQy?<^28QI3O%pFV2zAbeOo3hA
z%p)DmNgNnX##e`|kP+l&K{@9I>`8!O5pI{t!|2wvE7xP@>U9Yf197~69Genm=S09Z
zm)E1Cr7gzCCnaD{L~BQ<2)h`7;4G*see%xEpWX!&+^8QnS+;sY!F^dEN`Vm`v4f{x
zoMV9?PKU(2W5pW(LU2m<*!XCS3=dm?g1N$sX&-k;+N4q(+zH83KeDBBGLpuwATUUn
z>7MdKfb~0=tOR!iB<yw(`1%LBV{Bq5y8Alf+3b^;d%hrHDo^M{QRC!hPV+|aV}EN7
z&tZyG$$6c`$jFdzp4mI>6;%c)A^UsSx3lX6vqBID7As4+W1Hhzx0PvT19IkMC$=_s
zmG8RJmXf^<%xH7)0`?2B!(+n52$|}8dKf*uO@dQz3=a3i#N=>H&rI5!PhU%Mn2Xt2
zcJpb7?oOslPt5Gu$N-`wXbIQZ84BA`OlnSi5Z~BXZ;XugN?_=;5Wg&iXlrNFLZ^2b
zk)pzuuGo(@G_=@MqKC~0q#U*QWONaHD?sfwhv**|_F;3_=hNS8Rwc!5W1aoDImihm
zAQu=bLBo#?9Xh4>wswe^vw&02j6yMBz)Z)+I_F_<4#IJ4ZyZUXU^>~f(}E=jJeBP%
zu|3DuyA>hZc$8*w#}bP6bglKUCGeGa^{qGJt#{vzcYp9f+`M(y^>`%4{>evwh|fRy
zO#ah&aQ~j#Op1iQW%i1`il71GOM80gk@|qUh2Dz>(H1`MMEww$GwZ_45dsOqng<Z{
ziPz+aZhBUgg}zeK8k>d3Fr+5+K{a7VS8EK8_Qlkd@fZ=lc67H!o8ZIR42C>SD+I&r
zcXW!Rm;8XQ`>FE9PTuP!pi2T8&)(Ws^o;oQi~T+IE#bRJn!ZuGPAmJn{=0W)UR?6Z
zvohk&?(B~Ds-)u)r(FH&{8|`wc`m*@C!J3L0LD(8Xv^dg2<k6MS3mj?jGj8Pa~3qi
zLMt7(dT%9&d$pwpuMABpOS%&F)y_FGO+QSSGTXy978nIH4{98VK}(=-6uc}!LiY5`
zqGRV}I3z5br^-y(!4&8}#rD9FuD&;ZWchivU`=zl9h8oR>%{#u9Y#q;k$55gh0+b}
zWp<*ht5H2XXqUni?P3}&yvP{{6;mnOlMngvo%(dHucGXjPO?+mV03o$s9!aCR$?EM
z!YyV48SRylEBtY})CKvDG+_hKU^X;iSE%dPuUkQ8)c4_|hwdBXdB$V(oo7iZm>Yib
z=7SVb1$O-_xZr?owm<%R97#G^U`<}M<oUP$o-6MW5~UPbxFuYUwxdnp^Vy~uP7#1J
zgP<;ueZC+DN?C7zS6sb$Ena#3^|*55y69VHu)`G_L(D#%^?C^oTHBYrf*B+D$hWSy
zJTB`>ol^uPTu%f#BZ^Z!r9!A{)+xg-yHb8!xxaiRuQR7GV1qjg^^L;Hchx=Lr$2Ag
zvv>TsRxs3`(=Fdv$W^*$ilSgkc*5KvZkJ3$Z3280&cLN-bmZavn9ag_kUcOD#Wa<}
z8hskK_yzC=2fdzCyoninhp(^4n!L}K&@(e0nB%Ujr00Hlc`0W>mKJQ1IyW~PbF=W@
z*~)bLnP(|HLm`h@>4OK4<R8Yp`w!yYgZuH|;iGu;<gwmAjz{v3pFGhwPvg1LU>5wj
z(&NxsW<41E!x9EsPJLUbzu<B#IhS+NdV5))*BOnB^ch|1luXeW15|J;aL#7cz2Eyk
z0cS45&K+&|-{YFz`)SvQEO>eHgoRO<N2Ea7ByEOKe5QnG;Gto5HR|vb{=w0%n4ub(
zxjZT<;_I~*QY^4>Zf|XOL9mzf;Y$!<m?{~H$bl5OixK8Wj6jP33MuFc5Fyiepi>YG
z&kGPrOaSH$;NXBZB><7IO_s2NPzU2-WEYbXY)jEjnTkdD_H-mfjIx+$LiohYAq*hn
z5eu*p0ysUmLxNnt7<p&xidgULD-SU!PFtoNC{i%!ZWd@m&|x&HD<g%BjsXkG%}y%_
z{j?Rr1STpisskgKq(i-MG&7wiX9%iB$NM`N4a60gBvZhgqg7#~1?*UN!jx;T>f0`<
z<o1o5F?IEZ+HD{jUWh5~pTv_d@5RECISCZqapmeY2`dxPAm)3b_F_~TxSOau=X7Q?
z$}<&-B+3TA9!oisQ85C1?DAq}v7-*nh;VODkJ>s%88NqFJppDfly`ibQ?y5e^E^n)
zGf{#W=^d+nz?E4B*Oz+pPCqLWC>vEArirn^cN^=SieU!F$Tr^#{Ax$xTpP;pws8HK
zgwaiZ2XpW5>yF8ZvFPdQQ)L>16NoWyLTCkFyL&n$FtZ~ABNQ-ZMh~{dOxK+@4q*(u
zXR$V;-E0pB;ese&^2M5*wbe~A6Gn(t3zZKK(FHb`4S_qVnQ(@49l()e!KX_~U!Pzk
z=HAyOCe<SW@i-nlx*y+uyAZR_CAdnFX*k&xW7@Xa42#V;ZT#c~0z3lazVbYg;?@^$
zynZvTUYoY<F0%%VKzd$Y&B`b&r-7qHaf0-V7tInTq{s}h5OYZN>QX#0R<^&B6583-
zBY}h6!&rdcCStr{qRea(J83a1#4aUZIdDTjM6pBILuf{jjz;w-;V>-=En<hJwN)0y
zVLN&fyIZGH7}SoekH8idC1q_z@Z1o6p<YMTDb_$Tg->wjkp3*+rSji-?QXpH!}p~`
zydST<{)QB?tG1{8{L`=E*T48p{Qft;iwED_i|0?DsNS2x2}W~NG#2Kwenkowi*uPl
zV)PUbzNT_d$r%A+kU_YopH1`BABrO39ppNitw>BxjmCi5ff4Xlg`GA@>4m0XLfh6D
zwchp^n;MkzGa3COJ?fvLmr}Tzq+miXpaDhLHV!3e%>Du65kfLh!{Y8136!NTAtZBR
zwnvds!bQ&xSTry1U1)^&{OH&2<4Jv-HWy`UIz%XW5u78*a=?RkT#I*bzAs^XUp$U<
zN}O^}`Qat>3E)ee4#Ys6C(P2cD)TDeP{*W^*kE&xLvtJo{ywOYe#x);j&G8VR^k_o
zSH*FD&J!MmV43%ro4JdNps+@OS#CAA6$v%H?<-*FU)EE9zPhezf8WcShC|;)Q}~31
z;VO>v=Ob=&Mgv->_bPjuA*p|$!>FB@J+eu$URkjae3E9Sa0^{b7&7C>&M>0yq$eC;
zpvC!-Z1H-BzyeKI=~RwprOk{9a1j1rdBB4G&7EL#13Y%-oI$nAw$D-~@Ej89Ej;D9
zGDye31cQw?Zr+H#{yx+G2M-=7JQpj|V=c6(UvcFbq24oZZU^2aTxmB|(Jo<}@Fl!B
zq8hdG6Tul@+T;>`61j?pCqMURCE#}i>psIPo-9vpz#5o=n+{R?&=1L4%^mHc)zh5w
z^x7*@C~igHz@YG}EjaD@`IG1I?D5lBT39qKV+Sy&@2X~<GI!bsoHwBK?%d-_S#ajf
zeyIn?$w;Nl-{<P%)w=@I63$bWg{K5CKdxCk^2z>H?){GFx+=hT+**mQZ=^cU4_9Eo
zX9&{no4UFWakW=$V1^<N(Dm$+(x2r~uDvOc_z!#@MT0>{WJHu)4?yZ3rDJP*+XH@>
zUUS8|1=A<Cy0>B>k5aL`vSg(KONhng<-GS*N(UxVQc&1winSd!rP3AU#0T2$px}_-
zo(K2H+q;KKZl&KcQ0;O_hQquI+Q5K1wBCd8s&kXVj`nv2gAm`XMbAJ_v`hX3$0g(m
z2Sp?l2M`qG0JpN;fKR&f<~|R@6?{3zhx%{M-@VK?{rGPNKlI+ld)Q1jh-imAg7=;p
z0%1U<(BJ47hb?gg{I2(ar=i%wD^I|Nh$I4#d5$KF$moiIRWx`Itc(U>XAHq-<PagA
zZz1e-P78?@QA9Y05fTm|BrTAlm`j);06f(>N0h{1RbCdR_4N(Jz~GP%mSMV^ec_oY
z0q&HP-8cZeM|6}&3a5et0xfAs*y`@>ixw8p3h_805gU8TLjy4r1F^<loe_0*upte~
z;I$;GFQZ0u1U%ruv#=Jsw+vH5P#8JtU}grhsh(*hzmw07sv=WCd3`A%1%*g3Ry|Ot
zUa&Q&>c#Z?&6{^(^4hKF>={&f8KvxqZ+`zrFXq9PdivVc7@wL^!fcb)A||O8sEBnT
zV)hCnOY)-(FhkrKA~`+^V~&pTjdxR0XQ6FjCYVLhDVX^}%SYfu*qa!ih_Uf;n}YEj
z%7JGmlqSTQ4wj>mnJ%7iW|dG36E7;IL!A<MBPhXyt#If)i(rqB*;QvxL_$PibwPz^
zH!CTdosH4e)nXwJ1*;F`QE+OLVC+TT!ee$yLLp*dWD3Z;Vy;RLTu|Cf<1;!B69h-W
zffNQ#M1WSkco*{kM)45-_SL66waO?rR6&dvlP!<J!lWPw-hxsBp!(d@mC+a*8xlk6
zks_gdBzVuwJ&#X*|4BT3ET)WLBT0Y}`t5B-UUp+~=|Dsbd_0Wi#yl~eZA<Uoy&3Pl
z`>su!fD1DitnFarjZx0+O(_!VC?zPtT~ea@rAUnU+NyqFvl^8ZCBoZA+N7yb0zd0Z
zL^Rho1tZLw(&R#|RSDzzPW9#BuGShSrgtd3MKFRsV3*z~+&Ho+6Wh{mtO<8Si&&tJ
zxneG+-;GU=Bq<N!5p8aB5v74r4-T3(V3#QrW&J{XXGi?#r$33m{n!6G-hcmxF*9>5
zYEp<;Wc{mu{#E?v|NB43r+@f^587H@ToexNsgBSRmA6IxQ0+sX^hOCTZwhWXsG+`q
z@PX37J<5ijg{3N6jx^H8v4cmrXE&eW;o-P`<GR<WjEs%=AUj}kqPAw{p`)uUhDHZs
zY=Rv^n0?}e_6_5corD%}rK`s5FN@L(UCeTUb1sv(33m3-RnL+LkLfok@EmyO8Dr6T
zc5`A=fK+PLla;>ysq9sD{3D2qu4jE->6$iG{r@bSu=257_37EWB1j~!$o^dAVn-b4
zCv^kn#Pc@KFeW+bSW?Yz^BZ6FV-20x2M7kx4Q!-6%E^y1=;t(aUtp#n(_HW+g+|e4
zwLRvz%rKw4fOnAtxSG5oXGus`W#FC4!mkK2Fc~v=RWo6wTkyaGZI}_UlN|I8Wq@>1
z98!76SqIVc%JVMYL(ibIgy|au*esSEVM%)!&`O;3EREn{+Nt+^3(gQPX(#nbzDN2w
zkd*u!GsB%7&c0dmh%4r)Pp}1l_B<Bm7j0|4v(0wg!X=cfv$7@6Mp#2|VkYeN?K`3~
z10Gzt_w56xOSz2;@U_ZkzhbZgAwHiAzPJfWW#!J=op#kNSMyUp6Hwlj=jtlwwyM&q
zUb($^@4TvbI0V?UG{8C4kA4PDg2gm9u;&#`&V~$w>chA0+>V>KZ;QuHMy;z`@HmLs
zXR}r$eCoB}dMKe1s|sj<^Cbb&tnx27j;k+IaV4z&-OJ+DyZSWi-JeRoI6uE6ugg~)
z_h--b^{B3P#Hqj5yDID~jp~_{>OJ9j+9?%La3>W&<B#tv9HNbNrJ=w9<R)YUXe1kD
zAg|<h-t-B==qtJ^oc23YMC56I_#6Y~3|Jz2;IUkB|D<dbkMclY80bJTfQPC~gr6x#
zJX1f~-}T9F85n^c!oNiU)YsUwL^#uQ5?#GrF)%V1y#u}GJJ5V+JoFC8By~`lN{==_
zgJb2We}-KQf4R#zC5@NiP}5KTy63|d3HH~n4aMlFB4REhK>~P7j#j0Wfc(%fqmZov
zEGy(UB?wAbFwtPkO3oHY=8nC2IXc|tO%Ov>5Kkw>oP-@#*!vnr3qi(F$U+QF;MBnk
z3qYc43R!s8H!uqzWkTOVJP^zZ10qcb8p_1T=$M%HH6zi&!lJatZ3`xh<g<T08Y>M6
zP{{+q<_Lld;g)X{(`JMjgnB9h$}|*<?I;V|0P4oO#xM%YC^xI{DYn6_8!xmJqh=dW
zuk8?X&lw4PB6w*WF5FXAIzWbjXuMb9l;;^yGhj}wFu**46|*G>hXP>KD0fq9+`N7}
zCa<P4!7c*FFB;>k-+ml(stXy8NQj@DxuP^W#B^Dlu3#|}=Wd{9*1~W_r!!^;z_}dq
zy7ONu6UnH6CjuG57z~YFm#7yG?cCFem4F1H9iJGFj!t$DN@0hNjN>_#Qp!e{3Zo8#
zvS0!rn2;wO-NZnGGYYIn?c9baSXH=`EX?`#Xd@*vJR-qvh{N4#7U~fW5Li1TXbuew
z>)n{D)ui~e_o)<)nimK&!hv80QQ@1xA$HNigp0+<IkUo+#%yWFhz3f;&JG(H<mkFd
zn94RNL7Z(^TbtX=jHw==X7RKS;Zb8Wh+y>eHCve&9UqLTX|6+xbsUc#K8$a_ofVFq
z#>jBQ<OGMxp~zz*h3Ot9ifMX)&E?o=UtF1)b~_;uF~z&QvTWGpY1yd^vmG91XuObT
zZAVwvfbgTH@(7+PU`>L-=-8y{JfOa=zRy12Ynu}Cw_;gJ6T6y_ht2o|SIl`Z3uEvA
z6COs`Ip3ohAq^#hZAq1vc&db!T?zH9bJ-VMI)&ftcy*-ywe9}JPDCvF<iIur7wl!3
zb!7dD?ZBBKkP;SB{!k{ssek+L|9$+MfBW}w`_8KpW=4cQE%EfxZ2VvU=l?tY@xT6K
zJi31`)|QuT&IxTnL7?1L^bu&}&4i5O5Z@&H4y*7H!86UDnoZQfO{Utx{XEkHQ&7$r
z=#Wy@6<4LiPhFXj(lG2b9C(BY)?f?_^~LaLu5;MfT94WJx!66}GxGzVZPKDXjwuRV
z)N3*n25scM1zeRipVdc*UokuS5WAwFpnKOQgjOjOc$5eWbo4SM3@)A^CE>wYUf`;J
z58X^!q@c>1J`soSffeD7uXl<MPC)aLu_uG0ym%ZWf^ZZsltYAhvPbcuzfmU_M~<|q
zKMuxBiQxzE0ANl8`ZEFmbSBGMQBKOHXX1NASD`R*Z{{IS{xE(70B`~YlR8jt(oJ3j
zpQ7C4D>y=r3Qbaer0rP;3P|1v3yqjZa!_WY+K?Rw6(9cSWT1}>wt>r-b-`Or_LUZE
z$>^8hKt6NE6A&)#4y`=a2ZoV;IUq%vDzNegM`(9E_uvJ*7H_5HsPfz^jFXBnH)N)9
zK}z;>gkX6T3;gz$gc{)<v-pRsZ8F`>^e6{-3hX)XvahcvUb*{93=Rx=aOK{;`%W8r
z;&xHP)9#cr1;Xre7P$0`%l50Tl_$=*=Q+daD@8FVUjGai+8!9;X*+&rj(EOQa1~5J
zGN-`{-5buT8|!3RparC$q_UlrwfZn7iC3;)k2|lv8Z+0fM}Pm2dF#@`ay<C<emr~h
zB$nqFVpnucNFzEY<yCmc_ms!wb#{7^$Ca?+>YrV{%9rOq<x1T8w1MGeSO5OpLnErb
zU+3%F3RnNGxP(#F2g|dZ<5d1Ey!x)(m+<20bF@WiOP96!hV+P2?hB59XQ@mit`vB?
z-Bf<QPh8Bn1rLjWk30j7NGl7;a=JqE@jRC+2?CG$FuuWp=j>bEGx%-qffHs^IS(ey
z<I4_L@Mo+Lm?5VaC*DK^x^ir@;K9K@$smK#+=1ya5ps}2|5T1Fy9X)H!9#V^6Gk+=
zLeAa;hw^adE<b<QHT}>3%k>YtBqU)kJ}gEyFu;98zZ8jDty#<pVk%`~sOFt;IFJcN
z!oX045q?%z_oPtdNG;|{5EK<hKOv|JT96V21A>eo57B1C3E~0pxaO2E%&?433Ak)w
zCxE1)kt0Hkc<%4h(ZoC$K@zj3@e-ktr&FVBYH5vu{y{ZRYdm@UM1*nM!g&`XqbfVw
z?HOhUC`2YllXWpsWgP+2`1@kSoR_e(!zeYtCc22O2=1QXgMi8c?y8!)Qr0BOq{$2j
zi*;Eil50HF%qmB2^9r2&mC}Vh6rXWGYDO-w4dmJJiKhrTiy?&3J;F3oT$nx8#0<M@
zwYV}pBW2=7bc>h;3s%LHzxnugBIrvNwuVMWVtjH^!d+Lem0Gb`HNkKtvlT%&6L~rq
z<Wuz{JnO4}f(0F$wt>mIY-$%~Y?zhE4lkG}4on7EGD1a{n4FsO%mjiK0%8%2NEXEe
zMFIhy(Le|Czw{NCA-`d#9%g2+`S%uvij(cb+8IVQIV<8+ov&R&^<eb%=8i}RIHZAT
z&+-bpK^>@bj;iyvNNHxubIWzgEn_<chqT|$!|~GG4bzGjf|bqcI=cl=cC9+nw^9r+
zOOQY>kSKpK4Hhm-FlcR}zD-gx*wdOFrCzw6`z+8-(5sKpMlFI7>pY|!kMu`RZ->&}
zjk(!pu`s{l5#@<d+ImP#i&=twDg7-meWhRhwN(m+6q26yxO#OwhSaf9XtyLdGdqM}
znSw8-3M>%Aj6#Br{OJn_AqVW#BmnnDS8uNr8`i$C^H#gQ-IdV4CM9Cii{nw8vyP0y
zi;&`_Kp-Tu<58pPZZjPT-KT;%!jumZI%t$~5H@RCU1sKi*#<`O`}FRG_hMgKT=uRT
zoQr|b0v_778AKCdMeyI--BN!&lHy#8;jxi;<+a!1Km4cv7=QDx|1PfIxGg$S^HzU<
z{N%Ivr+@tC`1P-TEyZP40@1GWW90c%@P^1&=1eLhBm0H;<4V79KU6<v#)Ue1VIcif
zKaW@o*D!rVh+@!$(Iw!{duSkAaI<afzHoJ4G;>D)KydF?{DI*?!)bkM!;7d<0@!(q
z#n@nxr{;w(n38ZF0RmlF>tH&oD|C=iX8MKuhJ{RM1|!bGFZz$X(AVZoC>8YQvk~H~
zEPt7%?}0V3xKz&<9O*-oQ&^+FWmv{f1tNX<p19yW{V&T7!^CV86D`aH8Bx!*aMYVP
zSw`|ZF9T`v4v%ul`aRyWeh$Ge9R=}Mj#nH>5LP8+*r802CU9dg1AgYUOpK6-{!u-I
zL&MN;Cy|65kc59&=$H4&wj#FP*g`b;g%oIncD6=E5J#baMx=6aB78_eo@bN_W-SCF
zUwKAR$+UG<J~>kX?Z`nmgI;C8(GJwXBdVf<T<Jf>1dsUif4=oMl$(4AQ=L(k&J+aQ
z12<^2ZXA~@11)8s&%6`w@8F3HEFn7##j6smnVG;;<len|9#kSJ+7MXN08Ub#qPRWq
zWJNwzo0a&<>k|jkCtlG$#g%t_sXP5s`PFC$q8IWhPT}Q|EB-7mo~pW17Pj;S*t7|4
zfrBQnaDV3Nm3a4q58}q{JD$B2I8DbM-hUX6zJ1{BK6f_7R|F$y7<31kSe5FmzQv!{
z@w~A0_x0{_c>T98^VHAquj1nB^Y*j<m-oe=yOLJDds$e$D{13i7O&L%&$*IMy(eXR
z>YyuKf;M-%S6&evm%2l{-Ny0-9_29R^m@;rQgvUxDe{>FIj3h2Zy+u4BnFmvE@cts
z2<I6nulfpf&t<KCRF^dM9rt*fjVpcDHFD6buaaNM7V}!3Sv%1`)UR)qCjx<MtoL~J
z3zwJstoFHcV1}KCCtQ4YA@XJS-FHpze9-(MBMqlw#0UYLzc4XAB*CNC!ram#wqzk8
ztWDtqRs;}w*me;BZEf#L7)b$8@bmU}C;*tX2&`s!bU?C(fN<{-9#t4X0Tfo!(^xPS
z{qv|w9!Z`D-vDt>Eu*DD2-Sf(MOdFmU=?u1teNq^3lt*GimRlsS;qv0?=kJ!+Ss-I
zYmbyC%E^{|c6x8(3=xFD>e~YmLOh3rv5ta6QxGbF5zO0Vs0axj{KTDsSkEP_rpB*z
zAW(pegyhW6`REA^=ctA$-+}<lxPY)8l~*|R19rrP2=Q!q3E4Q`K){Z60i{m@e7BhK
z`1oW@UcVW=gCqJtj7)I(=64^*{QUE15n_%`NWdH&73|s58780(EDTO0$^yCbxiUaj
z9m5!pJAY5xqwo~L0KTcz1fL@_bQTd&UKleYwv5=LP*9QSnQ5EcuthM=GaKaT_7MCl
zWx~pme!z?(xMK^LbV{$OP+m+=GKz^JCv*{0KDNmX?@{8qy4qvr%0%3~dtJDJSx>Il
z0T0<0bY+FTsShOtb}R2wZ$IaqHWUy1L^xz400e2)Qcy+|2?PtK?b-TuYb!m^UYvLc
zd>W*@cGx6HMNxQ~6>?C{wmhTntnFY+nD&l_=<G(2m0;f17^CB2PFGO25U@F|lS7iQ
zRo;%{Ll#~iOYvvz4x<lgv(IAYo401-#?8sNabr5JU7HdvjM&V^GcXchH@ql~H6ti#
zO=A2As;o<pfFfmrU7A{3+hb^WBzpS$ETEzE)4nVy=b$AE1WHeZP*a}jJ;JwyYZQmd
zBbY;MA&3nK<-$YSX3ytCaI*MGtgh^N+t<yFebtH52p=VEtznzJw4(40&v<}m;D0Lv
z4B#k&f!YoQko1S89ACR}J>Guzo%rd`erB^KZ^OH=9G`vqr4)$Y#IJwx>-h8!pT^?+
zoa)KUBxoqu%Ok{dZbX@x@Tij_xK9El;hgW4HQ`?KME&$F#X<Q%DQ1ny$k?a{Ip|CD
zU*N<HS)T5X@|h><&o3_8gWs5sL4Q1RApC>=aljR3Xb9KPQt%m%X`H&4@SQ<~?4KwY
z^a*g10=cXd7lwkc#Sa2RTt>P@zf7NnzwY;<*Zk*O=#FcxN1PsAN8oOnmBDhv%KE2M
zdh}oUq;2_vi@gh^cM7B5LDzu~9^8UiGtdOS!53h@&<q4&a)mdf%d|^AuB#ja7#y62
zV}l21$mQ~78S*{_gRH9sVQ3k@GrG<j=q&A+0~9E^2;lGl=phP&*EJ{}*@=Xm@rNV9
zmwP{}kMD|C=7EmPnm`{ios18?BW-9WgCx2~kT*T6c$0aQyvmt_H!0lWQ<*5w)DPvL
z8f+1qG228jAj~F7(AetPVO1XI3TV=FW=Toc@RHBgS04Tmp6_T^9Qh|-WMD920F6L$
zzgP<LtFOKy86uU5Z@&FTa>KUX3$lW}g*P)qrCj&ydpDfk*=3uT_l{FFY~GVb#`7QP
zRo>5h<L~S8(6dXeGMBi7<BaRPGJI%<#8sPCQb)l{QT#nTubtfhC&uFSx8I7_-+DU+
zhKBXN*=yxEE5K{Ao;>pa4C}l|+hu{SfI_50f_1Kxss8?QoR_)3BwjuIF2_y0&oZ6e
z`<i{t;SNog_>pl+T8?+&+jRB!FDv7pdtV&yiZ9P4&(aU-6^Igk7N#G)-3?J)Ny_P!
zF!}OK{4?kGge!mU)Q)+hGE`ytwLJOi#>zv<Ti)N|lL;ve(n-8GO;REp?uUE_uR(@^
z$1;H4BtFlkCR_(3m$2Tmk(pc;aKRU=yxb>nbwzjKxRm_Qf~sfv{}I=8_4?t52zza<
zjnUoJC1s*Fdb+zMs5jZJ`1vyyTZteHVJgvU<j0(W#fvaCY{zII1YZ{Xu{U*lYlq|q
z30dlFm_4apK++~b5XiD!l?$yugkl&7%!JNL^%!wy^pSfbK-H1+lukN2N(c*`+igZe
z{xDxU2&2O60@c?$AObdQ#=xq3>Vs)eR|m5a?N%(v8=?W=laPvlD*l0pEC-;NXp7sj
zTL63vWuhPm#fR?vIoMN9JR@GTtruYnL4f6j#ye|p5Ga)+mrZ5Re+x6IOz=$-BI1&k
zyN2M}q89_~rX|5-d~70SZr+OC;W0%<h&zsZzyEzae)1?<1%vUa2@^btXsRE~NCglE
zt7cBuf2yYupOEUBNJlHsIkP>vCPAS*jh?+);1I57&x9wt-r)$P009<d0%k(IYuBz>
zndoGP8I>I&%8CR`4hE<z-@)KuP=*<}A%POXh&tt7@`|fAvi;&A3{+l>8IvQF8fFF9
zeQa`KG-jr!3`dwBHjn#<66(d^IboWG$V0<@x(d&wNU%^GA&{Ag{(%9vA?6c|x`AU<
znO&SvAUI`nX_<p?wk0S?@jaG+z9-y0X|(xNYfEQrtnGSdqP?9LULyd$Q7_;Re)SKu
z#NcpyboI7IPp{f=sxQVThg4U=UVR2w_%KH)w<z>%;g~Zphl0Ea%TRBOjSt7v%vcN!
z^+#`Smj?jcmsBT|FJOi#BFfK+xmb&?tttcQwWyD<fV@+EtIO>IIshlOVclY3nG_M$
zFO<OmM&xYj!z@IL@Y;$5N(H7nsX*Y_9p>Z&_%+0?V2vWNy>TeTWINUc7f!N%LEA|{
zhT$?B!67{;ip(g|X64W@@-|+mZ4fx8ugt_7Z@w9Cz4LCo^WOXM`s;6rW{t<Z6p4TN
z)o-mxeDb^Bdw=gW5wx7$1D{khrAt{^WL<1j8yQ*BOEBJomf(hvUiu#8Dg7714Zbi_
z%kCQ8y**M0IixQ~#S!|vJ#J;fr{;FL;3g)bT#e77Y$(8tf_HWG_~69eo<47Afmsu?
zpB&7$BV0q-Vw++1{io0JE#9VxDpQ$dfM$3-j{9rXZ<*CV2Q*z)Wh%vlR4j0*F1Q>y
zOE_hkeD7`}T0kALzdOwDlx{^YGR(=yC9Q|DIKI5&solUu(-1R>T&u#2G=mG!b7mG;
zXJv(vUsVqrX%?4xHh4yeO*d#Sa5ZVtC0Bh`%6yiNDrR1z_BO54AMa4oppCQ*8y#Sl
zXeFScnJT{OfwEB1oyxQJj_ZNG<3KGAL}lTz&3`Zrg4gI<ifEkRYEPQ-4&z$yRX*_}
zms6hi2}`cDS5=i#Cpo{DufH#fQr>5s&jqh4ZQUCPZZpzMpGrgV=iv!U2W7g|ckCec
z%By#MU0Gd;ufO>w);E~(RksnofukPHHSXzIeUJX<S$BNZ=2;%T$#=vpzO<tSF<%pp
z{jM%f4nmP<dhYe&?48Y0xKbYSRi5-?o1QXA37ph6gjXo*UeMCf8YAN)@yZ*oOPP2(
z#wXx0EjB}3SzL+74<5(k`wvCy(i9Yzyj(YwnJSjDSH8H*&wtMSrEy;tFWW!+QWa3O
zaymWI6z2|5_<HHP{{eUT+dq|7!G(HPaQ5=mah()(IsfTJ^*$St?3}0mO88mbxF|i+
zbWr8bTotF@Ip5-{wD~12R>9mP3fehceV6^5G(|gkuRA;kWs2Pj|3!1eQ<~KmhX?xN
z%9ZJ8>FuskP`N3qvsaGr2WL<+CBNz(cNWZ#o9fDKb^S|R<KdnJQ2m+_{zdpOSzyNj
zJUe|bYIP*s$37YY#B?d<6<>e39FHHdLk-M8ok^Xqy_KDum`z|tLIh8p4=q0Bpi>F&
zOv8vMO98=-m)^*Bkvt<HAl$>S2xlK}-l4%maWM(vfS*n&ESP2WvgdABM|8AFnUJ7|
zLV+^F`4twNlqWbM(qU#?YwK|OJ}nStArIW!#KcJ#GX%;_AV`!9Pt5}H5Vv4tnzbbe
z9&K%y@Sq@|1i=8&7(tQ*{^f>X#F#W_WC$FiYiA>c3Qez{>Q4R5wE3=-Md=y=tGL`1
z;ebDokN5nN&`HP^vrEC8cvW2aR>4RybgS6FL*Fsd#R!-?5N`pT?v22dg*K5ZOvdF@
zJu)miZ3+Y#A3}^Wu(+@o5AHvRB`<81GNAH!x_xiYiUJHUJ600)^ylg7@9aJICJy>}
z=1R(^%@NcdXzv0Dk#&!7CnoRCr1JQC7!ir0tVkJTEkJD`h6X3phTW<^W;bjfD@C(M
z4Jx4?^AippJC@Xiti?z&GgK;ZbV$f%_o9{-)|wmzr)#cn<gxurOWQFwr#{3i$JVJ7
z*ofUN7F|oR*g3TV!HE3+o|u%n0Byo9KfQhI9M%^7gSD8P9*VoKPR1MWT#2hUr=!->
zp)xR%Ek#AlnC<P@rG||K2Krdo$3ky*i8!)Q-=i=V`0pR>c?N>D5z8wp7HoVtVIwnn
zY#CY8eY@a_kk4p8=7-%jG3xDM7P8Cjg*nsm;*xhCVOw9!jp$&EmgEeClnMGmr`kky
zv1t=)0GRnxUJX78v|Vl9CYsVBMWRtkM8xLmVQgR`DcJ38?~C?uza{3oWTwkulD17&
zdwZRg(z97wAdq8AdZ;#QifcEn#m|2J^Z41n`kVOSkNzrd-h5T`q$b>Z8voz_`@h8h
z@4x+1{K1RW9}E5)!iN*D<3Qm-mttW!3$@wS8nY&*yJ;8WoIF?e2BTT&pWr9`g#M82
z=rvcudlsF}FD$sNP-KY!?lNP_2s0y<8&W(uk#%cxNANh1g0mG1vn#Q>AVqESC>oBM
zqqRW-wLIrQyigmn1*tG!w6sY$bZB^?)LNN1>qE?*P@cR;2`P*FT_>gC840}b-rvgO
z+;^#)X-n1r?Og?j!KCQh9-##|dJEmOK&bCXBjW-Z^+738C=I}g@1cS8ZLZGaOfXg5
zJOW3V5%_>HpNYP}OQ6~MuK41}kMC&<=m81{W>Q(s>X1~IsVbCQc|HqMa`{ZXG4(_E
zXPt(;`V;~o17|EYMgg}%;9anI#{sE_IWsBGsk=Ngz<`M`@0p3***UZ+WhxR3j-+5(
z=v;c2Ha+Dj_ZM8&-?Y&1%|REg)PuJ0dLQ8DPjg_(?PL7NHu4+iov_rMe$dy_PU(qT
z+P4Vbq+R+!=A$sZht|SZ^unX9w5LZ{s~`6W>)@h_e(sdc#eA}?g?0mjQdWxVYkpak
zC*zWi;YE7Ff6A5ZdhQKQzCl@m=c~WxTA4IQH2#>*q*9Jh$3T@$!<e~YfK=_~bumf`
zu8xUPt*0v{ugrLz#Q5Zt@?f_2g=pArEHA9Y{Ihw{G!|SVQzZ`t9_nzpT*hUUcLDiN
zc{3~vU4>P@680Tm0P_-e4#Pj~FSrVB`7VR%qZE`iY%jS#nXX*<ro1;!SMmOQddLHO
zbM7y`e~CA6&i&c5-kH8A4eqU&+h_l>d{bRX$IskX`JM^>)w_#%8a(pVPxUg+x~<7w
zSX39Ed^|9!KFw@c3HKnNo{>|S)nrhKH8)$U8?m{v7W>=Vil?-czWWXRP8rxMpFitd
z$b^Rv<kX)Z39kH~b4~y5@5Vl4bf4J-2qcG_EzK{BxpBhfeyl8s2#WDfjJ3s;=@BK`
z829hb#-m54vAM~J3{&kiY=Z=I3EW~{Oug^3Wt<2qy$v9`11hqGW0lN;h&mFZi%e(t
z_xHI$Fop3*fC#RMq?jJYp@5#zVUz}+Q6Zwll~F8+03+XBU0o&uj3BTrCd_~ZuALn<
zD@WU#8_}r7XYus0-XWy0Lk~m@#9<T+oBkB8lVCs?Z7+v6F{SKj2Ejd0hsZJoMH|@B
zfZXe<S2jUmPknX<alL75y{m7aau2c<@8>EVKN0Q=*L)9R5>5;gDV4iJ!UOFzH8~Sk
zZod+Jo|%xqD7ef&dK3>IJ<wxw43AGl-_U3riJGzQ!vxZ}O{Q6y^Rna>p9DOKE$x{z
z6S}Hnz+4b$X-7_MoS&QZEZ6YRkQL7Ty<OqWhE0VK5O3VN8DnGPf>E^qTf(ciN|W+|
zqb%Tvw*XZq^NbD*NARbZKCxNCEgc%X@yvwJmQbe@F*os-KzAxBigwBat$Hmu3PiDc
zz;XSXad5mBZMEhY9O?04UznrpDNi~ki)YD?3^CvHR*UMT8xjN;mo`1J%*Z2tcboMW
zC=Vf_5oRJnfIO-AnC9(lvXjoK%0{2bk=UWZuDEh-G)Bg1(Lc};BjbZ{_4;&NyLm<B
z8E`*gD>)AEW4k#PBp*pgVas7~d~FR=F3e=q1`V98aHxCLZFfhome8tlqV)9l4S7U$
zTY`~I6Qx)+HZhynFTCs#6J^IL?EKkkuGwwKSrIIJrd{Y)5M5BCEEHo0oJM9Lq-e9y
zxWWBIiiGIEiImzGCo0G(2UxbmffTBp%@?t?ek{LlC4y0Q1T|)R4n-u{k!qh=4*>|&
zHa^k6<gG}gO^!VW-Ce(VJ^rWv_#fi${>{IOTen_~zW!mA^*BEN^y~P4{U85V{KJ3#
zhj{Yvad65ki`QByX9w#d)UTOEQF)u8A*u^`Av(h#iogKgp9w7L2I*5Ua`{vya0H@_
zLXE<S()axNQ=7+;F0&YECIKFzgfR;%{Mp!!<)szT*fpDCE$g}|B?CrrNc)KPsf7#^
zX7~1CAoAO4PuhgRA0!b$Mpu<xkIZVgjZ{$RDdp#$b1cko>Ej+`&?~&6@C-dkB?%bA
zz|NT>b#^=<Y|a>Fe<$f80HOdWvJ=g3li}hE%}DwIEP2jEOB^1d>EIA}4V|a283(<&
z$Dz}}6*^Qrd65>g8SoJXw1C%!l<p!>Qda7MD`n9?d)m$wlqU-DC=>h#e&97yC;(n>
zfEi^qLajI{VA8hf4IkbKrl3!*sUXetS!RlJe&t0GU?_&)Q6!g%y45?vvVFKF{L~&w
zgY=R|Ibp|fnCgo!0<!B~x%%e<pO?i!_<`o<VTp|1>#Fp?w~Ior@{=z!iTXrs(i^vL
z-->~Oe$lE`&rGbWt{Wc41(*4v0>L}?L+U4Z0&{%9$$~d{p6OrQC+vX-?=qg_6Q}x4
z7q>ZiIdAZ{qG!kx)T4akJaknlT^}pEC?fhM2N)TMXp7Ow(YUTUvhcd2qf<19fzXp!
zlLGPh{^NM^;8Covtb!1(pA`xD6S|F3H~cDG$^3PpoYgg9d+rVMx?hRo_-;R*(@~e=
z>hJ%YGY&YdOe`-vXB!odgLYhY_3vNi$~($b%2gl#a#=4wD|P?RJMzl!OB$E+uFt2?
z0DcNI&J_A*SK$rO_1}1aJ4;jHC3UXFE5e0*NZ`Wz>gr_VOL)@5?Cqula^yug1MSR6
zFqp=g*cXSAA$GUpg@lZD@w181;kbG0Ms$r!K6f_+XR~iO9m(b;AG@=g%;T~vaPga}
zo+Z{xqMx~@|L6bucOMQ83`VB}@wOIDX=e0Z1mcv1qD>;m+hS%BBO};ukE)}znIJII
zf-+G&<{T{k5Zm&8)VtdVA81x8lxv6$`+<PtKBO=@Hw^pOL}qk!M3sMG^9ckCm<Zn&
zV!&uLvj->?G>{k8>3cdgGa}4NblRLF7a!)%7Y!DsIL`d}Q;s5Me}5s3^08@^m=|VC
z$R-G?EQsUKI@S+Bl$ge*)2jp)*4UZGj8lyZGU5O{VU(R9-5TKFl8Pnb6t>l`butpn
zs7wk*#g+Qu-H7t;nBbFeL8z(cjBf<R%hA+4L(BzHhwu(%sY$+7%ERRJOkBHlH~L4%
zr2%o~#c3=*eG>QX-}7k5;K)dH_YFv(PLnQ@w!q;$G9MFr6AzxUEwj$&WjaG#0g`Y8
zZ$@O<UztN=Ak<88qddTHy-lx>hQ*S%ZrxJbk9tN4W{n-b2{E0!f-!iruZ{z*P!2OO
zF*DkTeayWUN9BOgGXs)}1Pn-cB4(f4%_5v4)E#>#qqVh_c=}ufesd)@ceYf~7ZOA}
z1dr~jy*aE+3bM+G(n{G7P{2P9Z(3YfiBCU$7~gz*p!zF)l`ZoUF3BrP#ZO#MoaNjV
zDLYcwg(`|KaH)E%r<s@<jW^$ZEv{XkQ01gBNhzpxw|g-$2d8lYGO%0~Gh&v2H0Tek
zCzzjG6}*qse}!uu!Y$>+&R#8Tjouk+Ps#=ytGlx=`h3#)X{@Z|W&%=;1k0NGbickI
zQXlG7xtjG(0;_~oc1Sv=KS+>bZ~YWPg{Sl#!NjIOD0jM+bpTX~^CsA}=2SG~g=om}
z5euYWc-z;tmA%+p+xPki1Y#D*v!K`}kSZ_0LBXQe0eA2z@hx}xL3rn!8Fqwv<Biwj
zKmMox7(e;hU&W9VW@d{vq@eub7r%-B{eSyE<5$1>wbJFx955F}mvs%4SM39gB)EEZ
zPWXgyfw>e22Kg^*gU&{a>3XWW`-2if2+3gmz<UIXwibIvJjrWdFc+)SlMqU+FbR*k
zy1La4FYLjK7Qw1h0toc?SbcGSHwR*l<e8n>Twjmng{4^Gyppxe*xlL{tT6pj8?dNX
z?P8M|ln)dYO00UQxG6xXT~M?U%v{l`F7E647CPh2AXI4g=ZYpJGa|K&Z-H3|UX%&b
zGTI)&5~)!xXDE{%;OV0lDuJ22>6`BpzGtIvx`%cn(6M2L71c_appR!dnFo$^N?N6i
z3?^|nCW<cWdNu?%l$|qrTE*i#JLOwO@hB_w12}Qdzzp@I-+Qfr!jDZyR4<j8Z-I#}
zRZ`q1%+l+=3NMa)Ju6Q7<@@3t@7$lKQ^gfb4Oh={qs-6_IXI`chNr7p%24H7-~{YU
zH>)-%zBu01$7R5*JZF35w_JJWw2C6(j*vc0Dn$469^Q)Y>FE)jzbU06m5Fb@`C2qM
zl?iBI5$gS+9Caw>m_>Q@z2KfO$>%gb$~G&{T#cuSn^)4IEfW`gt#Gu|jk<s{X9Fba
zzt9>hwNftNQ{CO2(c9DQb#s`Ga_}bynGTN)dq={tsVVV~0o9|~_zL}GKp@YRcoK8Z
zW@BGUG5v*nX){h-ckAg23_TKl*4`I=y({ff@9SNLRS$o{6<ol*ERV~MI3>T!?>(TR
zpJMsXVRdo7uMVYt<^AP!arN(gJ@X1s-|-iE=87zbE2)^yDDbTR2)@Og<)v>)_y5n>
zp9guC9QmD?XI9>K)|GW%U0q$(7rGnW=xzW5pwR#Z7;px2k;5S=Y0acHla>-i+HgtH
zxROYd?e1S|CfUkt*CdmzY|U2284hQM;*!G=Ih+F+U~r!e^xf56_g#5c9#yID=j+dZ
zzn8zx#=tY<_uh-c!^6WP!aX8hJg=@nR}<E8es;`Go}|H5T)=U~hk{CPBc+LB1SV4L
zROvhPH2ood{1M+PwDQK7KtCE`Jk;}sbmV(WD>4!`V`gG3jvhG_ryqGJrVh<1e{fej
zfkq{;F#!$ELisn>9w-{~nsxI&XwvQ!*7v=y|JU+Iq&F<gDDWuUoIdd5AN@G~^iTdQ
z-hO8_&YYc$XP$Z3FUGxd{zBZixf0i|b8OtYlpQ)^(JTdONQ3I=urgAAOF`JzJu#N2
zEqMutMla`zC7O77^3+k!C^7rW5s)Yt%vM4RvmoqEXHfS$7)ro^48u7og<~x_Zjoak
z_wC#3`<tTu!o@3b{l;xCNz<vUtui2B@Gx-p$-?W4*~x)8q|N|EBWIsjqF$3R(dUz9
zPz={J*sfE-ia#|ggFxleh#-1$K(Q0hVY$N8*+Og4hvLGKqN6e<c=MjT`PdPIFP$TU
zLUkZ?)bPf}L!vc$;!(f6FpM(A44#*<6v&5gz{N1jC)HJe7!%F2W8?A3PrMW_{mQS!
z@u#0*c|-%_YJC5ze-{7Zcm6-IOY}eX_@i;`)I-6m0xMF$d65v=CQl2q^1(^clO6@6
zu(LXoD+AD<vvnGT<$)nVxjEMI^*3IPix)5W9h~Dwk65YZ6sLFJJs;Pv--y{gyW_9?
z;xERj(+>sTnP7VeU!>$1NXi5scmW&3gHx5LcLp8yrn5ZFY@xqgiGr<rUesKW^2N-p
z6*4JY?6+o1kUMGyVi^^zQn2^#!NC2Pn0^qWa+9;WV(<RRQ~*Uk+YxdWUhx>9L<e<1
zqeh-?Twjb!m)PqsC0<6Ol89zHFY!x*?1~h@m3t3l!=Z3VhS_?y_8?XlgqL=rL5p__
zX7)_0BRCn4J#kw3-jBQUS7d<Q7LS>Ul?<xR*apbQBv+O<)V_E(DPKTN<GkNAa~vj1
z@z5=b<Kc%+h^J$cNIndy)-BOs#bIHd7sT(WEUf5_i?$K9O-|fGba)?wz5VMdcFH8P
z`5PNj@`cmoP$Ah$Hm-KY?7o!&dbo}}j)-T2D&v6Kn~bnE87B*ye!=?gEhZ*7))fUo
z#=B_cod{mi_i0c<gJOyT0$?<7N*CXapiF(cwp=C*^T(ceGM<0&6Y=<y&&HnJb8++P
zjkx&k#rXcWz8l~9=C}NA1S>hLlCh-;g@22VR;V$C++i{Rie@V_zE8#0z?FW!c~mO>
zL6Zg?m6?am2kIX1mGUVW?}uaV&;hS_T)KAA@8U=;QaRcFvLqupPbK2r3npzIG~0{@
zWK3|%SH1@!;c)*Txq%nzKj}=pl}~`;Y5qwz`<W3CA&?5)T<K@=vohP1g<tnIl3SlD
zBoE48O15X^RDbxTqLvCYZH6)fsr1tSGEjk$)35R<L*#<Ao9^3@7y5&w5&UccMLsDH
z9?)9G1bJnDp8XB0w|r$D1$1E7u#~bDDWD57u5wmd-GYJI@JzXIny2cQ4;|hgb8~ym
zhs&3*#Kre6*_gAU1O4JF%CC8)dgGS?!EBIUDW@u;%Z<NlU2UNQ@}#Wgh>^TEHm3t>
zAg7=9WSYyX{;^wtodKLZxVXp981#_iMJmJu!pl;M$*aZ_(v`x9^8`?Xgg7Yd#vxDs
zNMIiEr}(-x9-vcRJ$scxZQjpHRI*V3J=IPyWEq#g@C#pvhfbgJNlyRzU;mppFCBs;
zWKVvqWcWY>)h8ugm;7Y^izC~l1E{gD<0mqblKM$d(bfKQLc-jm8bsIV*ogX2Uhp54
z{IcaU4F{h-D0#7=<o1W5%)|-r@Qk|e;mBh5gK9^)!l<^@@X!XA{LuyId-xU(@0VU=
zGV>De4ldg`ArbA&?`+W*DKFX%yaHdF=LPNZa_!Vd3o1^p|4xumz(F0QJQ%;8Ew2Ln
zs$U?i+C7m<t^lWNgCF5|H*Gnf{MvwD;mf{wdw4s2LHn7!_%G<hcX={iJAUO+uJU)E
z*g(a0xE}lr4!P=EiQneM16^yYHQR;o?QiapW#>VB$CjVdDUF}_g@$a)Zr2#K8V5(E
zD~zgIc5P}*Sy_zx_oTaMs6BpUHePt~srcfr{9+va#0x%N9Q|gC{+mNxG)6GF30|@?
zFEaBE%{v05IO6s=<`*}Ox9=O+_4oevU;ih3q>B%B&CcwNxqS!Xz}%4-ADfD$yNmJG
zn{QdM-#t6*qbWIEWOn!7IDY(OoIQIs9y$9^JpAy{IDBY#OpeR4AJl+9HW;HCzy~$>
z_t~mowlUv181t493Y;?soOZLdwV^>H6-H)&awZ%mps3~`n2OS7EHAU2gtSU0;f5dX
zyu^HeJWhT$q^MrHe8We5F0HWj1spIo3+;N<b!3<idGU2nW-OW+4bvnAL9Vv_aBL$A
z4oVCsM~!H}8&|rKVFu=$LE%GmY@y)8U~FmP7%$F>XHPjtd65raPX-U^i3>mAOO7x1
zZFq6w6SJ^%SafETFK5?K)Yv;Nf1Xu>JhdeozJ*nEY)KL9lj|4ZM~@tdlV=}|*+WNU
z-m)?wCHTt4c=g9W5)M|I_QmA%bgU~bowqn?hp@v>I!oxt+2N)Eh@g@|1Ph$-uiY6@
z9~fr0Z{JjXEylDIEk61NUKD;#^k6H~;lqc#{LSe%<iji=iWmIMMxEP3>Zn(wNzF@m
zn7LM%XhwNPnIH|FxIe}v8A3sVo?MksV&Ox%Vbx<{fz=Ed8fuX1@<T~wR)2JYtvWd~
zm4+T=O9KSCq>NlC1N$10<+%fVzu}<T&Xnkv;%G%2^Hg<+BB*+UEv%GD<W(+|?E6-(
z*q*m%KVLdN7W)qDjlp3>@8`HT$&BPDXXtbL2d6<%JLHezP1!gS?)dRzarnp~^AA42
zGqz!I<nHL$q<F<@5XWS02FEwAt)T>tOCG0%cT7BFt06CNvx<>h%y^NVugbF$gffgm
zjxo$>P`q2<8Tg?=59&B(_j!RBLqfT$oP%+HW0wqvtysLP`o2EzcLG+Ia~l(1#m0!h
z)%!EZj9;#{{AU}W+meK-Oj1Z?#Bqc$-;H_x`4{68pZHWf{>0N7oG0Rq*WQkAfBjqW
zjjw$@zV-EQ#@lakj{TCVT(sHH5dzv6%~ki7Q@1H(IpsRLbD*v*Pz2)ei&&a2)vG&W
z;p5^{oMvZiRCp3W`7(p-2^w9IBO)etlqUuY5k&hHQ5Xj3B+TqW5&ak+0`ue!v(S=L
z+5iZUef}th^%%UBju89ViALezUL#vx<#INe)QSz^i7@@_<OSf}JY;Nc98vY0v(s%u
z0!N;;N|{Qkc%*u*nO@I6%g_Ahsy}aIk@Enm{>CRcjSbtFz@S-?T=Md(w+#_b^#(wZ
z^VF4o12)di@*%}Z`}1K#9DDSMO}tE{oZ*dq8plOJ_}c)&LpX9Aa!!86(FYN|k{K7}
zQCP-Jf2R>t9R2vY6c4+K(}cC3ZOp5@%%anm(TDMDrM<!D*zL5!nf&3Nd=PqKRow`$
zal*^ICm;BVqb_o{fMl+5&YNpF`bEN!Nyba}laf=*XQt=$rw#e>1#*sP$9U(vDzxEi
zH*dtd@4lz8VP3RDIWh@}S*GFz<v_N$B2yH@LG(GohwKfj4;`P9-m+)bC%MfXoQng8
z56Y-N8b^=gj>VB<(oK%gM;?hIN91?xcpN<{fB7B7|1i4CVHu$ZW8dEWlA%4pPK~|N
zak%c*l_PCg@nYo*<J1Q+=#R-s4xm9l%J1#a&)toi*KfwPD_4E!%eAXl{q7XUa8nLX
z!ePkj-epy}sSDRZQ}%+6Jhk{z*S`!ZZ9zNT_AnURgcAk1Y~|~4xk_GF;t*c$ebDl2
z$7|<Z>F_I8(2KLu1MPT#pSOFpfHd22weu^S)vr|SO3}%;aqVol-%d+&Zv^^Lz50D&
zT;s1?No0@gOSaTrvdtj_q+?uSYq-ZjA>gamgUF5)h+Uh~?c{c8I2o5+x^MSH9Fxo=
z>$^`L_cmkDdj+)7dD&{2hlhv;Tm#!_?}1KSA!xuXX#BgrUFdt|>;K)B2Xd5Umh;9Z
zY_aoo%Wr)3Yw^$j`EUE{m;3gQ$JxhZI6U&OKU%eS?_nuKlj=O!hQdpz^J+ZT;`-I=
z-iE=61)Rvj$vs?`mKWRsu&<TbPvY~k8plymk96S_4h7mX@2UYSm#Zl2QnlrzM=7t0
z%k~j5bbP?4^o$DclsXp5a`oD+xO{y<15s}G7$2MPihw&n4N&Sut(5QIGb#N-RZ7Da
z^HDreHW!zaFGqT+Y~BSR#b9W3Af|RtdR7<%fS0f_lvY-jtcbBM6#D6?*%HMz845-R
zl4k>_A;PIbgX#>u%Ajzz{!kWph#Em{`BbIm-ohZ9%vgFqF@;vGGK01(K6>dx@!S-^
zA-^GojY@ca?&F_~&;PZ*5f4BA0($^dzK!_N*ZwSi?|1(hn<L`f*)wtA*s)kr*_YSY
zLIf?$x?oJ?s)P*|LCrq7<7-4I2O<U&j4%cio4{5cmRA?zt+!r}t5>f`R`$f){{1$J
z_zeB)uf3+aS&nC(c_v=`*vB*w%&9Ji^Zf@Ilo$voR4@j^fYVcKG^pctgHs*DJ9xGz
zBb1Ne5zdQAON%I}D``}SXWaYbA<<}DibJ}=k%5?;9f{E?8EK;{F*3OpQ?mmxBZUNm
z!hM0#KnL<j1ILfqriBH*03RA)_QW9>YO^!@;?B**c<m=|#pO%aeH^6J9&e+`J=fxq
zjME2F+;_2!XmQo2&J4(4ICkn-Jocfpeph2`41TUCM|BqdNR2rCU?V#Q;k<JBYP|jC
zdsdS6?VF4P2llI7?y=nS*?(SmXHeI_js1}k-b3NzT^I}Sk)w^rt$=ghJxAB_&INUv
zbtpPjrwCon=FrvgAu!3Wc)~0>TY^eyuq;USc)y}wvZI<8&m>tJ{c1-K{LU48xh+Ft
zWr=s7R0paTPNKrtWW@!@efQ0l896#p)xHfI+U)qmxRiin@yOXn;`oW<@z~>!Ng+6{
zu)cWx^>^Z{fBqNo;~)G;ym`wn`?DpEw#eRj44(&mGAz9+AzCpmEZ=g<mp}5NsZrjN
zZC~u#&-x)AP7RK}2Amw}J3cWMd-h53*vm)$cE!TVeB52QBN}rx4c;L~R@lS`P6%S#
z8(X~i3bkK`mLHuPb#3KB6-nPfev}8r+%(HUishoAFO7-pD-49&bq8S(RoaY4J=2^r
zEx;Y^oc7Gi)3gDX9sdT8S3;>Hl?%E6F3?*m6Vw~=G1f37a*HQMQuZ0fr0_J@F%oi%
z6EnZlGZTJ~>c;h3Ryx-uw<vD3Uq0SMUgn2zLlP^HOi-OZeLRjGJ7fiyuPk4@a9R2C
zE=;pp>`aq}1n7__MSy<sgnpGPJ!<AbX2eZCCJz5>1L6^uMikHPk<Am*!S{4&%oRXO
z@zrGL2b$d1l3zQ#C)*Ad-PkB@^1vhSJxkS5RI=R0Zh!7e4*b-Ee;bflD_l9$%{HiL
zbtne1KdJudQ!jnWlSg;&-j3H^eJ$R3>%4UY+648^SAmn)S+A_{Lw`DzMe!_0GDw)n
z$bMo#hWoH}ipj}IPtGuAvC>d{rcCBvb48vR%Us7~M?;oBeINP;GUhr|eA3gzi3juX
zM_z42X^i)KG?I0$j43J0iHk`RR$P%MR^JE%5Bd$gjO#-A;^+E6K4BnP$@7C`qzxUp
zXvNJkgSYFqQ(B-Mk9!+4dQ2I5f4bi1c3NQDvs_0;1o+{I(~ej7b^U43#!>J=>4*oQ
ze=(Nzrm5#%oRto!F#GAB7*3nP+ra4^*5HvGc(Pmfp0u)Wf^)Ek;^=C_vT2*d^<hNP
zmC+NWLE4BIRsS}I9?H9)oA=b|ZAn*;IB|S`JpIfQ@%*Pg9*=zSv(Y!q1g7Lved2xT
zbc`v&s&>ID8gj%mXO^+b{Fgzoq`7piy|C*)`fES`Pu8VSta)1qd!$qGSzcI)OBb)i
zD?j>i+`Y39W8=fJ&^hI0CMGAQr5qmC^S+p#oRKoMM;6llIIw>%4r&lMa_ErULEZ0H
zyggF7CM2L+8gQB2<n$P}F089_kO3n<mPBDYg>*yKm6=)gC=UC>a{X4)>6mqKVUTxL
z1BNwl^YSw@w#-JdN1vr}8ujG#cpN*rFV37f8FO>HrB9CtA4Zc5gQ-z(54o?wub7-~
z%QL&2gIzfbiSihhVM2abi-W3IPRvo?pw4iYl*)r%3RUgpUiH-#E7u-W`!+<QjEWIv
zSY;&LmjSXiuQIL1$aud$L^d`dMNCTM=KWQ5zznVoK&(Efta3iuRP;|xPx#1B_Hf?4
zbIXdmw@yLkgH7o*n=!1$gCHI~ax@-#^wHSE=@RO&X;{}Uyz7;TU5dk(6DP!DwsA2-
zpB=CS80W!=0Tl+c8WqotU&9@PLfz?AhqShFPe$3iJ5?G7@3aVk2a{{puDXLg#te~^
zL{88^uyb{Y+1!){niBW!z!{Wv-)-Ef6xiTIZWYHXOX}ot9whZ|Kj9>A<cG$}SFews
zI1tB9?3VI35F-<;6l}>T9Tt!I(4Gc2)wePf|M@<}m}Gp$Wx&|PKvC0?T+d3WKj33u
zmlvft-&vFl@(u;Ef$~kc*xttgqx`Y8q%<<yOao`6+PiOe96WqbZ9qZy*#g8-rztm-
zJ3A$@bZeLMy?bX#Md_2VaM<U!tEyB_N~<!GhWt2ufbVAT@-y#1@a1!^yu*Mo!Hdz$
zOZQ|HElE~$dm6@pI~vIa?-j5rf?TtG1Y-ham_4Oc|M1%@Fv87uDc0Be#Dn|lH!vbL
z<L<51xG97F=8bthFXzCkfptR$LT-bB8In`sNBOoPA(<_Q+);iqJ97AFJonrS@u4T4
zijybVd(BJM8}arV=i|HI`d+;Hqo2fu^B3a!)oV6TkV@pqd!VJ9K2SSS>Ab>-@^6Jh
zj>jwxN`reJr!~|Kgcvf+m>L-bXQkm6dE^@r^h2DW#TGmi50o8dJr5l`6#MyZhx$w=
zIAj`#-c1=;DD9Y^sz~t^TB;4&n1@b1<5b(GyQ3WITBX21!PD&nT=oZ4LU5{m6GYp{
zGEm=XEW7u0I`RUI$^-v(&9q4iI8fHuK4PV#$%6F6D?phgUaek~awmkK<dSVvV4RAB
z;!hvSR#mpW`s?vatBRX^F-;m+O^<m&ndukAJBjAOgZtFavu(96Zd|+Ji3Q#Tai7-E
z*C`{(b>A&Bv{7)HCA@Dbvlvd~sWN#rj<^k8m)m|IOyC}43*)IcVj7gb^5a@gKPp!$
zmk>|}MD=%M*TGKo=C+`rJOLa^z?h2ioyu$CMLxa41#Y)@u8EH;NIH00U=eG8Vp4t^
zwyrbv4N~G+b)diV$80rbkB*PWI9rgXr626s6LZZW8}y4D-ODZlCYg>LJ{%{GVFc!i
z0b>>`>PO@b9XjkugcGMu#fg(A<J3c^<J75#;>3v)4%<IB7yG1h%*`F}_i$KcA5k8I
z8fd*DC1UA|-M@*Ko9c5o7-mKN=)(MB+`WC*+qrociPP$?UAq?7<*un8xoUScuCl^#
z<w{(>bj4o*$6da3+5T5AUsc@ep2Pvio!hs)O_w8y`8G~&&CQQ_S{8JZhC%8$1SIdX
z9su*Uae(Dx2iVEKoepT<11U>gyoCe5yiZB##Z%Ds`=Ia+-iEyL>xEui?X(>n4OHCn
z>rMMpfzDElCR*jwo8~XY(EzdRT6oG8aO*6A=1O?eQ}Cp6=+g9kM51leO4U2)D&s*2
zGz7<>?uTVO3_Z{oqA_ezW90-6ol{|>gAp^*W9Rno^@`lyvuCWQl1)xPB4^6aM7hHw
zLsQ{vi|$@1>I_6{T-SeF3`5_){_p>3$Cp8zQ#3eAmD5{Z{mHBGum7ii74N)vJ03Z^
zCmwn9sFc}Bb*_VP@}aZpypO6g84*$0Niz5_^N7Nbvsx%PYztY58#g#A@wQ)#yME(F
zEG{oAt@2XiN6}d1Wz2gh8k?dcU)W{fW$Ow@s?k_E{e_(zo10i#d?ZUsk1R1d|5V`F
z3oa!@B^n+ZiMfLZ;?(KW>ZFcpfM0fJ!4?~izlBy_XkpvMrt)X&3bTZKkcf9cz#~1t
zPPUcDGxd!!LdG;u<OjNZyrc)3P{(#aJb2()<6C#`#JmPFI$_4!;jxg?I;uF+GN^{*
z?mWw$Du@QgeNrY5$@pRx`TF&no{eLc66KLS)smy;L@M^T56F<4lu~lg_4Mao`3uR2
z#7c^)KjwFDCAP%tQ3=tIMDzJ)UyPsstG^j%p8r_H_>>K$SN{CZ;(z?z-;RO)2XXe?
z!!ak`tf=mmWkgD?P+4-|$GH)7mdf+j_1rY-0tP}>b#fL)291)iGI8Vj74iFq<?etw
zeu;JPzQh}EykR`gKmUBZ__2@1w3KpYoEYdZY}s~`O^?}I>L$0~fS=VB1_opfWs{Sf
z=GAdif5;KvOyT7>e+xpiP^WCC@o~g5I8oYohvV_5&cxO(DgPVSCHPA*wR<4;9hi#g
z*>M>a<9_j#cQJS|m;rN21~e;1moG_%ZqAE8@|)hBv;6mW#ce6hoPfshkEn%X%7dm1
z&;5mIDZgHw;3OyUdPOoeGQKN5^x;S2*%zO&QF;BwdEw%_3dmu87;kxzcPt*r03H&f
zF~IxPF7{|(J?<9(?<pQz22o=9k~ZI)xO(k&+`hZ60r?>t{JaaZsl3tOxo6u5TaLIQ
zzZeXu5UHYU#EK>IW44!>a6G)XSMosJ(+@JM?(asZ?Ca}&;#0)Z!iENFz8}Gh!)bIe
zJLi*`Bs`q@g>s#<)}_dzND9BgX-CvK4Trizh2q0`=N^4LKK^r`P=9ksG!3f0m%N35
z4^dvfaxE4#fHRP@y%9O!>(X2P0N`j0jfrp4D>&j8$Mz%kMWbZVX8hqP{wSOOQnquv
z$EfWBGMkD3Boa^0(2R{GHle@`s83;QBQMTF2g>o>f&EsJIAns^Q{<C(!>&o;y>?Z7
zp~}NHOoklLr!kQ<!4VAN0u)Tou0oqKE&B9mKgh2#bfJw4Lt8OCm^E7`MH7lWbwe8^
zDs6?2`Vp7@l|I?ZbVHXFWAPw)l4+o)Rw}s?FSkg-VdxWog|8gXxLgH=juF4$x_i&G
z`r4^@`_1!l`_{bTXsDH;MEhqv@kv2)lpu}h`!-0Qdg{?QefpI9gm25pWTl+5{<Du}
zxK3rxp!7w9nFc;xe8_;V<}HOHRF^8J6@23JPejt|O8Y4&{@`DVyE8X`xY-VrBzeKr
zh>f<usG&iZCfNMsr>+%;d{Ly?>WJf+zL|+F`UrSP8J)JFvkauFr}A$*$I&0Qtqn!C
zq^Zy5!?1arYJT?z#I(^t)g}I{9I$fAR#%MbJdA>aGnQkVI83?-srwm0<(Fk;l7sOj
z>kUU;`y;aIm;D8Qm5o&k5v~58m-YFGZjBW()V<oEaY=2xj;rUmdHd5A@KYa<t2bE>
zl#?;dVaijeO-n~!bPtbQM`k6=%xm!y`G+=SFAu9gN6HCH8y2oGa8gp;30J2M$gly+
zL(jFsd&7FuwxQwx#NnCpz-xeKJIb*Go_el$K*ekOZ41|>ji<sYT|wQq(*hN*!lAcd
zr}PDF9I4qi^t5mRz4^7F{5#=IJ(0Tb?36FD@OCmW)A8lhf_x=H;^c<DRpYI0S%K*1
zSyzcl3`uT=HCBxe>Z&;7(l@8omL}BC4Qd=)UtLr_o3VHQtn~6T@$}EV7?1z8Ux|M9
zJ}Evc6Av`L%ev6GmvqttCOzbzu#F$h6=>hL;e+@E^!-2o^Z(5jolniw(K*w&*avv&
z;-&cB_rGIf;_Rb)<H@H^#?0>NSe4Q?GJ<ltrGa@&3}6t<Mo1^$9f%kN)d1z#KWu+p
z*{*_-;7&)%0v$KoEtb?lpx{8u)Wnn(XO^((AlRP5j4mBv9wRCmr8qB2QDg>y7oNaN
zVd)6i-Y~OgCJr7tsE%wp=9w*!%NIS9fs}Vu0}l!&iUlX6pvW#Pvebzp*ccR&VZ>Pd
zo;pSpEme#v+wT<Y8Zs@M_`u6uY_V8g70x~>fPC*uaoLHqe{MpC&4HLZwAWj{uHCq5
zdf1P7S_aR7xr1^0_T6~<?RVqmjoVVrq!2xj;K*1Q+BNLK)#50+_^==^kiGHho1$Ud
zN-F%m&#_xFsxT(DRF)?ndpf@G>whE8zW7NQ6Eo)DPrml&@%#VccZ4@&Oq_|i<1(~V
zcT3{?)&rKBWk}!}V<Ht6uKZ@lLx-4Q=@>BC*5!8zBtIyJ%PR|U^VT($=@z|!m$)%9
z-+1Hoc<ZgV+`(c@Jp0^7VrG_?o=2z=^B={^ot&5s&Lu0QIY6*&$TQZGGw7h+*)mrr
zzHs)wWkCZ9t5vLwARoL>kaD>pLu@b}K6gAew(iJ?=#QzryJB`P-*_085y01^hnz1b
zaqt4&A~T%QNHKI-fnW|{XlP939*~iY(ImxDwjnDI^mlAW+dH>AHuQjjjBNAr^?ezt
zYbz{2hH%O_of(K{pHE|AVE93tfA7c2m)UI2v|o_nvKVi@As$FkKXf4C(X$ir%rmFs
z#EE0#`~8@mowRY!O!y6E#_lZWeqHjbwijcufB!Mn#c=X(Sw@cdN*Umj>$Dk2SS9ib
z1jeys*aN8w!^#`m2OYoBn7OCz@)h6y{$UN!gDN{8=aT^`W8&`ZWznFvDuW4LuS+I5
z)|c&C2rq*+Mg;>Z3bVH^D9=(RP@JKWx*8jui1D#W8QgPm;NW2?BYYEt<D)lxQqPL&
z#UHa&8s4?x#4FDRZsZrShg6Sz;b(AEMzYGXDWl7MgBmdTVYt%nQ?V9x>JaFyvai=q
zdW%yD_U4z4!$4dO08*IoCq2drJls%u>Hk<cM3LINH}_qmV4>`=$NJ9gc^MTLx+uu3
zFmfmZyiVTeUpPdO6lvNaeH-lyhkq)`8XA<Q2HI2(8fAeya5cQ8-GJY5UFI$EpZ65l
z)`xO|0Y}qftEK7z#Zz&?kvz-%97i}AqezFVe2DLNMOuD2c!59jF2%Wnn;CTB=f&7P
z>a*b$ZT-FTm%I|kM;{qVQn{1tnE#m`URYL?_9Gv8%Ekm+YrpojZ}_-f3>B}apaqK(
z(wHC0*EB%C{+eP*zRWe#ReS_Yw|Iv90FaRBQ{l9dDaaXUm09UQQFw_9Du#37r{dZS
z?oAwd8@g%o+6(ZVIAU8TE<k)9;U!}e@{y~FS)cG~U5?2ZctD@a3PEl!lTt6m1tZ(f
z3gaCC97X_gYaKx{iqYUa1~FLDXdsOb9}xx0%6z1ZD9!XOjN!CXZ)b%6qE(pO7a(VA
z>c1J!F$kcQex8XVXy!S2oIJw_@VKC{+go3S3mG6S{X9lf#U&j$=qJ=PbDS1GESFqG
zL)LZ5G|uA$X_`Fo<C=7+aQLZ?2mJ0{)pg>aZc@H8E&kL)?!FLD<i|bH!39p&ni3X{
z4}vPww*2(mz83stxB$xcK0Hmb-t;_I{`FiApd9a`p?R!0mA;*>u0R`qLDkbvVL%(#
zPT@+W-=@v><W;a!x`Iv^*X9;1;40b*FLmp8j>xa^N51;?h{3Q+j-HqovKcXa`YHG7
zFA2l3uz(RUD&t{P?Quf73a{pl>3(WtNXErL3~B6Nm43CkxfT=C<8k85$#~+$=i<4)
z`PZU<_oOrPHm427ra{ru^n1$F`I_1UXl#mQXy5MyKaE#?|Ha?=wJkaUkh0Cm=lnA}
zyte5x?caFowYYI>CC)v*KR)#ILoqD{hFSM}QjYi_8eh83y2?hwpi_-x%YzD?cOK<P
zftV%WodRCWLNTCdbV@0V6fS2yeA2>z6m@k#DEq1!HDG2wvLVk)X}Bo`iqnA>bkEE+
z9e_bn?4u-yRUxCfjfxk-AQJ&Vk-+%j)T+gWd$!m(c9t0u%7H@5GJbW)d0I`_P(V>;
zW|bp1`L2NXDvNGrBmIJ*Dw@+=CZ<PX_ujD>98n2(ZRmbHjvSxUbtZ0Wu)B5VmXxi{
z*u7`Bl)gC`DuZ5`V8sICpuca>iWD>A;;M})KIlcAkPZVbk3hvBLeZ9ny15=hD%XVg
zv%a_-r%#@XFZ}Xv#A6@*w5*`Lin|r>eCONoFMscMy(Q}5M^4+ASQ779nc!ty)rQJk
zs}t2&$Sv>PU|It%ai~`Ye+B~-e3S!>iMQT<BQ9QiPe$Q*oO$?R)z5mo@%rmBnBMY?
z?I%9*$$0dM$752)5d$0bim~XyPJayE%x<I1v*O^rsp{yF73i}up#cTuf$vr@o^x`O
zjRy=H@u`$dPKbg=R?GNk+0baj-a}*Y;>RD4eY$S#Qu@AomQ!X)*)C_pL`oma;jAw3
z4hA258ylY#k49w3uF2rO9dEz2u7O2#OZ=3D%0c}P#PIl-%FI?Q$@@JG<SX~%-Vz26
z9XYF2!|~zgACIS>`;d)^<&~Qn2p3{$@pjynA$IlZ?YMeH9f#t}CXSb1dL};l(Puq-
zd+G88$^LcM!9DSX0|S_X=cVeQVexZxR%Pa6b{vCy$888%&TS#76mhR=w8Qx3kN{p9
zr^8?-le(sFz?f!f4q315TKDfu)*oO<Ab+fM@O}zk_HOzQ24uB$Z)=jf&-N|hM6p0V
zP%5Zzw@2kqIRIpfc1e4l9G{AD@!H!KI0jV?*|0&Qx@F>n9>GeEuJ=~hPNlrm|Dfcp
zE6>r%vDiJgCuaBUR=ZTaD$c^<g5`t>6DCe*zoe%(s;&|qbUIGxYp#xrH-2*Tb@UIO
zX|%uEgvu!Hn#XKeqyH;q9OWIpdwUaooAa09qdt(<M_V!ukQWW&lV{d-kL<%+W?5}u
z=+Awm(p7r|j7Pw;+o<BvKYC`JA~>9tiB#lGz5>YBhMjIu@~Th8FJBr@{H`w&0g#w?
z#zEetTobK;Iw>WKL@1l!vq3q;SHEQdtF8_oJ`nG{doeCwyr%w+l`W#@7{$tjh<6_F
z0{YyK?w*PlUwk%BK6EUuUArD%`SM?=4c!rbXf8tte!<QxpJ|oj4?1~92{w66iOeVX
zMO>FF<GXyFdnYiCbl}FLaRSN;DrCWV5}wI23hA0FVGYUz`IV81(vG8#q8J{xBvZ6)
zgYu!AtN^ewQR5hK*kZ&%39JI6RF931%c;K*KHlAPn?zO>ms9KTLvHaiNKR?9@Tj(r
z`bwLnKIAAT6M`Hs%`OB!{>h1YywioE@3CBk10%Wy(u9N0v?cRce;jFyp@BaxaqGEZ
z5_r7gqWmovo%G2#Xoe5T<gO!mQyQlOlGjb02=FdH;#52Ue;#S866aQUrRQ3Rns#V?
z1j~=~Df4Z4qFyOa0e*#}@~$iX-8}Por)!1xLfgOM0-f}&WW6|ugRA_@0mwod6heaL
z>8I1v4g<e=)<8SF?GIE~g?Gx((3tlrPtD`|;sW@!pL_3(uj@t04Ar+5&bBduAI*lN
z`Z){-Hy|8?ipP0n9N%BY!;p{<DgL0`kP;26Esshc9MM?JcO;nj9Z~xqmH)8r2c>JQ
zuPsad@5RW(XdF6mB+fnebbJhB;^=-w*d=2^x{eGAbsX_PV+XXQsgFzO+%@$7EYRi+
z(0=awJAdtC8ke$jMp^T6yCjOii^l)XJ8wvFUymoAIvgK*=5$O<jmOm+H{#lLzL3pe
zu1-yHnN_1h^L|=gea}GUdof*Hv$8FqV0_PENeUn{3%ho)G{atIW=7d+qmGk7fX-h`
z$GLKv%EO^tE^_RrltyMEcv+Jx15EOUhMLQ}_wH>-v7pc@7oCh8vs*Bh4uN+L_#g;h
z(`MCULdtKR+`#cVbW{xdIrGkOkQhWLxC>G^nUVLhJ2kE_W_Ks1#$s}seWSZ#d`i~j
z)L_ic3NK$WR))83&wKWqL2YJc+Q-|XJYBqa$(=n4)Xenmn4X-C2^kba;?J!cD6Q`V
zXY<1s_(OSi$>`wauMKtD_r-%<qGw#;t9O>-_>qU=^MB>9#$zwOEMsC$O4?Rj{mHBG
z`@j3oWh}1vgJ1iP9f<{abYJCEf6%wJgCWt>g`Tp(G#y`tWygpT$96Uv6Q|kmA-HdU
z`<wCFYp=#(X~NGu{gjo<*I$2CGWCwj{h7~vCLTWfh~jfZAIc&M6AA#MKwQ67${-yq
zTYcC9$I643ucb&Kt8~r`Abik@!vZ)~7P@%ZnwPTclpwY{@m&f0sW-lytYJv<xEV)I
zOviH{eJl<ipNak<mc^G{*Ah6DhuLDb!rb?k6y6(R#+;S7v9T!?en3XyvdVHb-hBOD
z45_Cesj^i(brM5;F}+6yl5j9fZ{=}SIWO~0K{mjNDYiMCi|0S~Y)tQFyWhO%yQi|U
z@Ai($zNGSs-=bl7@IgH67fhL<AMxteg$w6pfUHUx;2SG5x=yN|l&1_GUWDf5RSZ{N
zc(yV^U7`F+R6X#joG5afio#YG+69zp$Tp8D@0wgl#^?rfy9sT5SosXwfLOS@BFFK&
zGW5hVW*^y7#>WZQ*Q7isA7$ILN!2UOL>>(GHYO}%NR}cfbCE;`;eLRvd_EQ$xsyLH
z;VMXd2FpOnL(;QV0?Iel4oWFOiCJA=jj7qGICbWsIC|okzsh~{_N};l<+9%wL8iLf
zoSeOBX@fh<W+dqea=F~!P!{!bV4!^vG54ureB~^u!z7>iZivc6IoUgzxx4Q(0hwGr
z3Gm8Ym)k{Gfu<gLE<f_f_L?c%&o~=4#8bY6HDnq;;hqO+^2Lw7n|w$^9zf3MW?ulf
zFZEsGtZUutDS!P8Z&EQVz!#Zhdo3RloSvPCg9rC}3*WcD{oS~9;fmS;?HE}}#TH|N
zvYRd@qlAxs|G<HL@lmy-qel;VMV&36^Ye>J#}JB~sea6dq}vpUuB<Ny8Pb@b42nZI
z&&tDo{E^V9^^06^+oU&JxDv>Z9Kp%+bR@C7sN9sj>c`LG1APvCAzO%itSSAXWQ)nB
z+%}Y_z%j1!0ZzWH#Dp1Pp7;`e<FEmd@fc@vf{2v`DmwdZZ%b0%(8y$qSE5v2sP_0G
z!s)YV$Ar-z_%u2CRp|=pa@=KNC*@F0Q#c4if1r;h4rx;U6ro`TMsP}!0NId3Mv`w0
zBNfm5SGxRz<V?jl0e<%(Dvu5DCYe~=@V2g$6I@y6C6fSQ1?_tvWnp{U^1ZGS!-m9+
z!(SV><Le(@<0@W3-P6`+V-?<x(@PJ~rlp|LwA1bc?R33)0TTNTd{7>M<);b9zxoKE
zla9J>06$!BeigUE$`!QJwetWf3<vb$29&Dte}Df7lwWUJp39G`Q|YlOnW4=CfOwW)
zMYK++tBnYp+61HMu%ZvipJ{81iy?{fh}!6|+Ww&M@wwpv$@j2ig01wNipu8}FfuSW
z`X6k_=+IcVAzkSHnz!-pJA5EcKlONgl5I}sPDYFk$Ck!IJ{17V9?+Hrv;dq7;@_k_
z%g`>_PlsM!0p`(v@~h8pd1gTkof%YEz={B!FH4i}U3fcg&o9N<b93?N6DMPGW<2gJ
zNV%HlEPoNqw;OWh&rZ$8dy?7znpZl-+&V$$n*(pZkHfkr<=0z4bWi@?Vk5??icDaq
z^U-q-WORTGK2q|O#(6e8a3j_qSG*W-=mb|**#@%^D^e=y^lD}s{()yqiY#<ajPX&L
zDH#!ye(~E4T#7#(A~W6ag4K=@Rn(v)fF-8odrQK(LavId0ba%gGp|D<vM5JU(sngQ
zg$$K3b>^y5@Z7o0ezzR3IR%D+gIWD|-@70IVyS;`@ReqaiJ8e6m8H*HZqC2+uHQ-V
z3JSayFQ%twJ<H@36Dd&ZQl$Gu^TPGJad6+U_{`^iInF)*l8lK1s;j=Z^Va$JAAjd}
zyp87g$zw5xF(F<_brxSF*~|boT|E3n&={mxfpVjhzw1SRZDWF2e+DGNIrem6aX!BM
z<uAt%fB1uV=+w#h=tp1l%<(I){4m~o@7>rxw=X{b`OnAcGY@-E15a*6$khoeg}O3(
z%nTNW1j-)+6OhKlmJLc?bma(1j0YPNDq9*8sn{|zk7ABN_+U%QT^|<v{W$sXY&?AK
zP|O{kjL`|n)-DZ*YOHL1;r$37$12*`<IV@;Sg{)9yB>0Ey|^ny`t{c@#rg9a?i?fr
zYUC&+7!$i}ARxP}AaPFq<~<G0a`$9VFvE|2bLz~oc;+LI$ALq8mFJ3-cNsLBY{_Cq
zmeWr9WLQqx*j-osvO0zw@d7?P<^@AelNuY}EuIajynHQNO1|1`ZtGi9JxMN^Jx9|J
zk-Q6m$v~TAC4y}XC@#$Mt*L#`9#Oztud*fjcMYqKRBjn6UilapSE+{MmSpVOm7CtK
z!%RQC#z=o4K66?biXRe_#;=?dX|M9hF9!y7kXCx6G{Fp&g*r#@c%RC%dBnzFb<GM4
z*Hz@q?VLZb$PgCYLnus=Q51Ne1ch}cv?tG;lHquaS%bKC<9b|aPQCH~uDT*W{<vBx
zkuzd@!3$(P+d_Nrz#w9^MVY!SlfV-*O2?^JAYltxrlZ}kvV~&DkU=YBFrv-T#<Hx$
z6*`v>eE8=~aQ2Znj3s1!Q;z$Ln`yZ>Z@?o&<V$cVEAS82cf;W=TZi)5c;qlL0n&B)
zVZ}koBn{9+H?;Ew>WZHU@O`5!Ry;ZKMK;GKhGX~MsW^OC<IJOv#DDk?f2{uQf(%u)
zWA$5nbAoRHh{Nt1Xxn-)9{SXy#}35{FMK%W=Jv#e_uh-I|HZe|Zt0sTAGyOD)fah-
zaK;blfU+8sz?E{5JW76%7%a3LVpC3C0C|!QzBxT&qkH5W<==bd8-BSBNG3)GF$P!}
zVCG+Aqx=XjBZ79wn8~|@82LjQn~a-2STcsIK6g~{Sb;$Hz>^a`qLA@i{IPr>FM6f~
z9I?%6Sd|eT*g%jU@15~pA^j@`1MQYoT*gqgDQ>Dgd8JqP@FZ6xGA%N1J~riK^w6#O
z1F!s!lsvdkd4Nb<BQyBnp}r&UntCC<uJo(m_83on=$>bzk}ElOJvfX+!mDdOCn7|+
z<C=O+Aa2HU+y;uzrdiMqOT9$bh9+ML=9(e?1xMj6Ji@dMSLI)Mb>nsOEWg54upLiB
zPb>XTgZIS+RS&*4^wNOq#ZmMqlmzcfQ_xP=>)(cUSoL9T=;d#3xP0_WT_+U2cHULb
z1$Dm-mu`y^u5G%24nFm7$aB6xjlrOM`3>tjBKaIuUpFLw3<6KlDV|p)L<1`o-u5Ix
zLdRp(ade!1V$cQ$FJ?;tR6N?>nvDr2$aWvx7Y{x8Xng!jzYrtOJ{Emb<0`WH$u(B*
z)jp{W<sl9!$8!2@iXaZ;O{auE9k%hM1%3bMKmNS>o*ZQI#aEU{nd#%}(ev^4+i%2$
zi#OufiK%$_k;Cfnr~M)$^Lflt-d|UzBn8S}mK5RSLnk>brI=+_caUNPh6J4t4W2K=
zGKjJrp^i~y(9H&pP{51(?&z#sK$SXMDm@)Whm>*Qap%fxpYEBl@$8Hk#~w+(sEdNj
zwg!}SX0KTRpu?mCBpxp|vp<y6dS<6*L^t1v7>@b7Y@OkBD84?d^ir1CUb1JO6#SWq
z7#zVslF}ljZ*6@!mRA?W6FOCRqrurPi1$m$8x>umS-J%;qB4^{e`h(y$0iisEHA4J
z&~)|MHI-vXdF(UqcqtpBf)jVx;&ev}E!!I~w|rGukDWa3qxbj*00MN6<z&&maQ$}d
znc5eh{Nxwnkq^Ht?el1C@g?F*SK?p%&Tq%<J2&FUvBPoj#L-wpF_ALo6Ms~vo|Usq
zWPM=Zu|Jgj06KN*uH6tBxF`dD41y@|fA-}+jaPp3qd5D>Bk|D}UWg^h(|5o3ow#)Q
zVjMYoB!2!2pNrEEKkPw`S|%S1N(MiluP>S6_&{&JlhMh_5Jn665syKU6)H~9@HQmz
zcX<iJZ`o;(DYiM;*ic=rtw}L?AY)`#8h)o9o{j@YrebQ(P)tk@s;*cWQ-iV$kJ_=S
zzySa^Wz6w@404T9%TcQf^K0?;Ti4>!B^eOw%m}J;kPepLi5QM?&dOIAIC+TUKR4DS
zA^w0IgS6yZ2F2{2$vApq&L?#7UIBFOl2OwqV~3SB-doteZ@26B&YhcHkr?s!E#_iO
zvWtB1F}gJwDSVJ_c_|Mw;GF`D4^@Zgro$C|DqCJvtNrj?^(&fbHz?5?RyOFcX<sNQ
ze92nM%)qG1#`)@M)7<wBi~a{TB(7b#ZvOlGAgUWWZs=AER9?u3yu?rBBwx0s-f}x7
zGpC&A&IxhMq_QfC(TEY`FH>7Pqk}^s#9&n)VcB3pK(y0O$QrmW|2{p>&P;o@mRTWY
zRM%Ago40SounhKbwcY#b$QLCWtctR2iuQ*Fn2J@-h#=!uK!gMEU_^(cYstI)<wsv;
zMV!7!^a34Yf(9?TQKU-Qp*?|>{tL7iQ25iUyR3xqtbb3w2od?E(w7E={gpQgK5c>g
z0q|5F&}#)5Tn&(hV&o)D#?7+O&eNEXr}E77OaNIqg*?)V7SgA3CV!X7$@Q<FCJ&&K
zw;oN8(h@cMlj36zZN68%OiYf%zPVZT3kT!Wsgv=C|Mm~#z4I5n4OJ{mrP(v<Nk=Mh
z&du_0`r%XY^wUpz??3NkeeZiekSx(xr0fut_!tuqjt}L}{ImWTL$V$k+7;LJESUu|
zJ3TuMV-70_Ip%PR5q&p?L>dG1`x^hiYsYbq0~&86iwfgJBeoKiEP&H17092*w;a>R
zgM4dyACqDC)Q_=Z%XsURghp1VN5)MbD57zX4@4p(R`!Ji#h=wf#?|se9;g>q+8HB#
zP=)$U2MTW*BkC_LcPKHmZJ=r$Zs|`ndu2uZqi*SQ;0y8!{^TX8h@GiYu2Lo}SC%Q&
z7kPE?8@uZTJkXGUUtJTIlKbBNdNw|yG<l0fd9^QonFIGgrX?)pGw)MwiJFh(MVXLe
z@D(0h=FM{#R|5ByUxs!25%9yidai3bJ#m2Qe}Hy)8%SGtfOg((=#6JE%A_6i4(&uy
z{Jb>rx?o!5Z<0vsKY2BIB))V!S9yA&;%<XP)WOxxycaJ-B#r)Up7Or%Iv!qn%>InD
z`D}1-#=+a58{|uot?+LL=pwmjmu#>a!HNTCdydF|RQ}^iH=*{7LBU!(hJm`h=3NTX
z=>`Esu9S|xj(3>Y>B2!lOdO0bCaOL1t_3?$R@YXfaBRhlbcYk?PRFyK{bcNW`S}>!
zzXw5)z9$_*`pbs;KnSGogvNq~Q(2}?;ggTWs~th%T`XPb`>+1#e<Zz>%FUNE8Q2))
zDJaKye*gR5ji3DFa?H&Q#u*tC`w#4~GNQ6d`I?e0B3ALmTcyhm4MNnB&>_-boGTqS
z%4#+mKF9$<?EPiyiJLdH%h9=c+m0CFM*M)m02~U`;G2z!dlg@|oOPZvLF~KbB{GyC
zX16u?3O{+%k<mfX5oVs8{h!LF1yQkBE*_H-%z!<p3cq;aJ+Dfj3|NU#Iruov9=XY>
zF%1?hL92t7a=N@SFN5W_EJeO<tNI-hPMMkP6Ng-0EJP4deDB<rvcJ3@W5dFM!37dg
zfbz?}|9;Uwvv*GnNEsCqEfl;ox3s_>ehH7tCG$ytIivU3laDD5c}mfhux-h5y*GbH
zZa#KP(fsHOFUP}AekNk#bj12_JdmOBFaG&Ii_6z8#=&EAapa_o38mS*F9NqXwijid
zY;q7t!Ab*;L6JcYu<|9KK8S;I1Alo&NxO0LMttpSUyZlld^65HayFiN@<UP#=Hpx6
z`c_=OaWzhyIuS2_?lW=n^eORe!(u`{)Q1&DS=D@7gDpPrpRFx?OTzgu5K3mba$1x3
z=Gy3BOVhmKuSst1+qgkF=i3z+7%MA_vGHJ4*1U`b84kyePsZ-Ku^1ZLijlE=&t+Ja
zJ}>NZ)G=c!`?lYC>!J(_zU{Er1}Aykl3}+vzaH1GEXU=`s#{eJ?P78wVoHX|#Ke@_
z2V3k`RNrg^VugbCL3wzu0KOp?ypV|7k|K}sq2?Ft{6jFM*Gr*5;TP_y(KvkckdIKD
z79DJbV@uqv8@GI9Auo%gWKez#CzK2FfVUX*L+Y5^e)^ax6c4?jN%ctH%BxQb_Q1fH
z+R(V&$gm7vIqoH4Zim#(y2^wBF@I;l@073_l(U)rspu=;T8Z$WtSnKcd;*UQjn(Bv
z8y^Su9gIiMJ{E@$9#I~{-g<OFmf2fxzoj~o5+zxHb``;V^ZNx%E^v`tAa9=al$()q
zcJ#1#bNoo`**)XQglktW$J_6`74KZS5O+l5eH1+z79%n^SWQLQ>r;E<MS1!H1{TY>
zuJ9@Q8uAwp6-VCgzj)5UPy>R?sWg-mdLdYV^5=Aohp&F-2~G(acS^%wDeINS*twFg
z?yc03C-tN>^l6lz?Io2*Vr*aQF4GVzab!CKtRy-__wX2HP54|cmo53wr7;yh<Ow|P
zTa~`>QTI7O3k~Cn_-edNG?|ADkHtsRBEFHwxMZ8HN=J_!@dVEw|M7o_8<OW$wH;)R
z@A~+-Ly<~O;PJR6$HyZdef+U_=+r}2qF;OMwRrQbH>G%s4$Gg)z<?%tw<x>(IL;Jh
zm%3xjV0D+T@Q<qB&G%cf?!6K&IWd=%IpsphC;q5pj?;{Id45>^`S1uUe5B8pk%>$>
zQ<fw<8q5dO?obYVs*y@Ue}*i>3uKu(_2dakJ=-!Vv;LS|Va#F`hRKoJx9{j)ZJheo
zpN(pb-=x8i_Gya3X_}Q2<z(w91_dV>F;PREx^CsXB0*cypUX`_n=7~Jx_$}C6Rw>w
z|2Cj7hjMj1f#OxW0?MECrbXAHyZjStH&39+t^*zX0@4s{AN=LIp@m_GG;;XWJ;1Y9
z=34&TXWBMhK$q^U^KLr>y6N5zg}dUn<5fCbJ6`z}?3A`QZab`?{OU@4+D0$F!bAPI
z?KXbI1={i24*y;pz0U>Q<JUatxw*R5Dqa)Lpe4xY`ceBz=;CkV<JxT#lDG2rFX-xy
zbeWcGrO!TI{;hVEL77(?8S4Ku#*NBKm{c!6B^o(8XIlNi1bt1D9$7J7XQY3|wZf@<
zkevbbS*U<R8n>8)m{7ksOy8n-=rxP0^YK7q`Q+qi965R*9{K3gaqibX8)K(WM4$R6
z#>xjA&LbLV@9O>NZUsp%eK*kMcS4(IK<B=R4-m&!xeD+5zyFv2F-g_Qq=4{+UOH0}
zG86i}?|nC3dF9>My{A9UJ$fPz9iFox!B?Jn5jD2>egLy4B9%dd&JSLx0Ay$-pkY28
z85NivHE~f0QsJkfl9PHeom4h^2}vi0Y%t(A1E=Czpp_AS2wEyEISYhv=f2ua5B_wd
z3;@-Mfx}9Ou5_a82UTtAUmdMFVwwhfzFF$e55@7~niO@>#;gqr5jMckh>VzF22?u#
z<>0#t7#Vx_?~aMdu~=Q@q^WriWOPV(@9=fZ+rls57?Gl8qb&=}EYqqKW`>3#2C9j1
z4Prx{NrO%l=4B~cTZ+e?I+nqBVfN7{&c*5mMvn$u(K#X4CuMT!)~%S^voD^1{^dCN
zD5p7{l5k1Dq|kFT?By%(#s0&4{qoL|tlUkuH#NqDSYvu|*=f*-)G@+K@fKdBzd}TR
z@-yQS&deO=?El~g-;WCyE=XBC7-!Es5(^pxzx$nUJM5u{ABxX@{&R8o*kKzJDcC6h
zD2<#%fHA?+`lbe03_xbBS*B0ga*5ujdMqt_wx5BP6{URP7h{60Q5X|TQtDZi694+v
zVrphMj-K2f$4(xI$?2hZ(7!4gmL=B@6gCl$o_jRLR6pn6x)9&`<`3ePAD)kUEB$fe
z#GcqQJL^3CjfRyC8HcNJ`4TH6l4nt`DvKH6KX_=LkB+@^`MPMpkjVEiXw-g@T+ddv
zY`7_*ia#X56YDI?iYCQO{_>-J?%fyh@R?)r_!Ez+<2dH{C>^YrT)c28Zd||V*?!s!
z<>#y29^8n?Y^v&x??s^CvvS53Ald0|6C8xV8S$fIQ<A+&$@{pCh*8}S42&xG0ny8m
zuY6QcGR;8%^Z61pbtTEL46urXY|x%C9{A85@_<ppzHYWS-MDc*?#kF@MdH(+e97OA
zIC1QRnsk5s@P|K+Kl#%?jVo8Ldi9d;LQr-@1KF|yB%YAh11VRjY~A;a&C!F0;`GTA
z@x+sld;8vi`iRRHFUEJj_r3W3D?f~PuUv{1wcS2d5VO6bQvSU1DZXGwGl@Y{qz%=3
zU)4#HUn#;c-m|0fV?d&O@Yac2;A1D1#%V=A)sT*#l=Qa0mGVyfjKc*#P~%7bZ6&4J
zT&3Y2D7v@N-#E{N3y>3lINh{aMjHdl!)0so%T)?$n!dmUHG`*|9`Wmxfk9pO@QnIF
z3HHm}Ov=cQ!4f%C4Clei{}a<PCJyb7BS#LoUS50c^|*NPie;K?=#v;CXe+#)1_`ts
z@|_qTjFZQY$E?~WWuZ-6y?R~Pc^3oYSaqs;9g-2r_YVe!Ih142bnp%WU)$zHIljc7
zryH@F&N!HT9Wqe>e#*>v1us<+l?gA5BMx9tKXs3l)O)cmd87Yg<(*YR+GT8V;vjw7
zMy{62_{RXEk8{7KD{1*SCRe{~Uwn1mDHHfpcGrdIp*}ex8XO(@q+S$C-BAweF((eG
z7wQQ9<(SAg2|qH2<ZD8fcPA)b;EI+?)rL}9ivGe4J^1zdzyH}!zOVdxfixAr0KZZg
z3y5P33afnTp0G9!$CZDzGoGOtAStp7{|KZ0>J&5lGpylBgR`Aq8!Df6T%i28_NFQ5
zjbCX#2ntW(?Zsby+3pFmxBUvQ<;VQMKl=xua1>OUUjKGpy?`Iic^Q8`x8oKaJNW_a
zJUh4=C|q<=@AL0I?g+%UYjB|GlBDp^KWWtRhvi0<!Ib)k87a#%x=xBuLyFJD7rJM*
z)1(c3*zYNqjO&vAOhD)=H>CE<<nxHeu0i2jUtf)tdy5*$Iazfe*fI0avya7xf8!To
z&ofWN!0ZeSFSgduv1#+*RDP|t^FhYM2SGcQ%g@!n34Q<lfAjaY_<}7js?rhKQ;vaP
zapA58`0vM$e)N_)+<qxCr@aiV<`pAjoG*={(HIk6K|si8d}5lO$&k(qVlhxq_>zHS
zpaLo$Gb(1IJmsTjIzqznqw`B6!YdVy&y2K278qW{B5)+~z5eBx?Q#dHhin*(K*>OG
zrZK^l{F#YM8c`w!lm<T(qB1BLWH=Fu8Q(Qk<i>qgRAe0S7LBY>KE$`Ma7T(teu!+(
zp54mte%zK)&tO7+3>>__p(4kGzsjxhp(ryeq&RCE%tmRz6`dFrEO}39fa3!_(0T33
zwYbfT?@~^obNB9vIDYzAEX!zN3l}HCOplGnmg@4_dza$m(c|&i&-`+nedaSVCQgc{
zV;p%C|MvHPH!fX#J7)LI#GEYJC6sT;z}ACdDODOAC4Bipo%k(h!GVu-9O9`BE7v0o
zrz&%R;z}4ZzMS6i`Wvss)hkz|IPQrfQugkM7Z|5(A3OW#x%kw}FU8!UIU5t+exW~v
z9tG9k&XBwyGg<#>Ok}4`y>*QVj0o;oy;+dV5?937@LOHsor*g$IAn~iEZSf=d~A0-
z`;o`vkw;I(&}e@w-@6-2D|aP*o09X1ICk`?l;+j=;rD+m<LA}5aQ;rH+GBEZSM1x1
zp*WE|P(^LZD7bm!Zd|)QFQp*Ar!cI>Jvus}!FkYIobKFNmGaGM7;Lm~t9&L=aw#KX
zs9%b^EOeELuh4I>A5yfl)nt12sK0W3>Y=0lF38x#m<>x-M(*C8m*RBOr`A}pSKG79
z3qKuE_VMslGA<pGl?mRFz{mTh)y8-qf;u0cn6=}`+kru~?GX%$G1tT5-6h$ei(ZBB
zb#YNL$jL0(c4<2ReFj@O7{F;eIA~yyMWMXQI~;0r$4{M#PrvkXeBxt27l-#BilM$i
zDgKw@tAGC0`07``s=8P*jaE+U+SEDy0{Jit%XSVr_C3?narD5!_|TJ2#<R~q8%K{H
zb^BkMpN~KO<3EmXedjyz*7<kh+Fi9B<v%t-+tYxrj(Ht9hjx{LiH%&@WM*37jeJ)d
zpwssj3@hTe?#2F8DE~?(N)5&!e}eoYXA$TCwWV0Z5nlM2e)8f^czHSANsK2GciNnv
zx=`(~;sJhF4Ycwioc85#d1jd;5A-_<qn||i_NtA-OIbj<W>wN#aO4NREwx1@GGgTE
zzB)Eoll`jdM~)eC42jMJlH%833iX^O#D_R%_s+y_4T$5DW1e+CfBwC=bz5yx_(uFK
zgHdls9U2(Ypep)QPrOGoE+cjKOn!TZ`oh58*g#GORlZ%qKOtjfB#jBh9ajISg7`E+
z3#Xjl*`nQWFC59cEEn4lhE!h4Nng!71St9RJM>3h5!JX!AH^XH`6d2*w~IcFl^XgZ
zo^4RDVo0AaKaMbF717(3#4`%ZU%i)tKMFqb=;~|gG-*tujCugF&c#vJN5-kL&`xU%
zY)~|ya_JebG8}=#$<KCK;+i-b`Ekl_UcQ&BG;M#!#@~Ml701S_QzlgUf;4y<Jn*T5
zp$qN&+Ubp3>GLW5%XMgLo{>${)5t;LYAb$ie;bp+O*rEca!nreA@CC*Zt|?5pt<%!
zwY@g}ipz5wDy-sGe8BpQ`(xt3E5aSrg07$Ld{yiYUI3i5^P)EMX&waI;x%$Z<iyiF
zw19j>amH!#DI6WBv5|hz`E~PzKg2J8pch9sO%MP4m+|%EYvBNzTr=F>{Pf^y(xj{s
zDnHsQ?GS(3srbdoOcUxf#)Z$ug!-!~wO_V5^=o{2zyz4a!wvP9Op?pneWOA>DQE1?
zfc$BpilaOTV@J=R`ZP{^dhlSw+jV)rk&ib|&kV-FM~}n{zxj)C;xnI!@k57bzOkio
zl9ME}4FQxkp?mkU0c6$HJn4bH|M!3Mx3}2qNhipRDfJ0s=z!+$+=*9T{c(K%dvC}1
z_<b7_oF;<-z!~m?QbhXurJ$<gR&-Z*ZJBajc4E7pvqNRTV(3mr;qQv2Q^Z$4IPlcW
z30H7BE^(R$=RKuTR#2XxmnlnPxbk(+;NS+WFe=M~MaAu~g*>uC5fXw>@bN=IN<Z?G
zm+-Q>!9F&a<Yi@5C_+fyY-{4BQv`*Xt?>y?C>oIBH=lb^HE1wX&KH#7$>_+a1cce#
zdonVZInOxvWz|9+GqcQ~VzeDRaL5D7HNLue>AH-(W$}Jj%*vQJcw~PpEH4Dd<FcQB
z@APbJF0aMgKe-Y|=61)Ie(Bfa*-v~a2KJm$c}GIpZ~W=M`B(Ao+pov?^jPdWJQu6H
zbR4|Y&c1bC#HJ2a2jY3wWw#O2alls#6O4h2x+gwB7%w04GT$3-z9D1ca?I`9@9jZ1
zuV0reeb;fHc<RY`{-ZC%z5{z>Lo$IvL;a?_G*>k?Rwwdg5C&dm_(ZFW4?1sT%CmU9
z{Hrhw22OHf(64PyOG^t@xbwZ2wV0S1mJuPv^XW(9lnk-)sgbz1u@raaZ~1)$>TK_x
zeNvj2;+?nNi+A6-9Jg<>^1&9SG1V*Qu8)h?oM^-`!ntz7dkVY*0WY9&Y4KiMzj;?i
z1n&Vb*r-s+BE?}94daCGD2$3{@)S=7WqdH(zItz2%9H5WzdH`8evchL5(f_M3y!VD
zFuHK@g1^$rw>Ox%MUjDqQW`h~f-kiDW#~|wlT0C>KK7QS6Up=gj(X+OhP%cjYhzNn
zcFSm<Qu#S5lG)i2#R(Y@%W>=44H+(Xyc)q_1Z+`)UfLBM9AZG*fp^fr)hmK-Z(+Ij
z@wAHg$m37Qi2O{v_~OUn@Z6!87#WM3m#@dSzW&Yl=GVUw@4WrCXKFe9EsY8FDUE#D
z;H1gYC(uc%?7QT5Xl^c^7T%X%{&YO~!gH~kuOO@bmu1lY;lKHV_}bV0B7XSFkK)4B
zD{)T){;2pqs(djZ)<mz33GoVrIE}Hctn=g<b~*bu^=J8)^Mya(H;zbh;K&4-$ci5K
z0%oF`J4Y#?(@&Sa3|1UzoA-$ysBnOL{LAP{UgZq6{hLfp3QE7qY{SK6%oPsW5~yjv
zti)nakRrDMvBD(dSw3vv;)Z^aeeoVE<i|=3MZ#DGFa5lgEBaTZ^B#J|HQg?QoNj$Y
zsoK@l?4%6m>EOF2yhnFeMyn`Pe9009gz5_=n|@?OGRzF!u=p}0yu%s?j_^Gp)hpxI
z!orH$Jl~#Kv%yF|3Ki5d!wQECa9{%l1&4C1(1-Aj0jtzj@;6i-whM`Vl`H3qs5{0Y
zwRghRPHE@fR^;|4zS3X$Uj3QM52(f(UdW$G6DbtPLDnNYpp1Fj{Fq1dEu`hjDh^Uo
zX$UtTxp#`Jd%}$yF9l?nLgi5kpX2DcWGwLj?Ry|;$@qB>5GUm-JF_;9x>g!!m#6(8
zru?B5Ku+Dppv683nMXI@#M>?ZHlV@*2_NyQoNljrPS-9wb<|xotT^qwz|r>qpt#V4
zBb++UJadNxIEYucDor<S=2ftr-gX{csQ83ccpHDmUp&eC_t#My+Wy<p3-GTp09=0>
zo;I}cP+bn~<b#*~8lE<>nqcs6=S4vWUjr4^&a0hfn}6f{-1v1s0r;4v4mCgqkOq3(
zk4cX5i~58nF3=%e^n<)aJtE#usHYiMz=R@B%5PGA>WH3saT>h@#B`(9Ye|dja@8+w
z#yYOjWIoE*`76DiBv~>j6h~t^?>BA=k+SwS?dfBq@zKBWSK{$s{bI~Mbjldn$+Et>
zt~gRS05WW#b^Egcg%v<+yKV~l{;U7|SGRm(hzO=|%y@IuA_*}j_%6UVzVU91jc&wa
zj~|Z{Ck}hok3nN_fTJ&YX96VxArNzL$}_LZu(|iQCDd`F5Tt^xDlk1I0Lc_?v(htd
z;&ia%k<NyK1{OM55B?sc6v2^<GX1<ywn@ixIpfK&{5o$lld7`gnBhV|ERdp?!GV3I
z7zsX>PjP9m9wb=WWSCL{56pVT-`}~AQboswk&uH=F83px<U_%1Osuf7u%z&NqK_>n
ztcVOsq0)6=(0ll~qCoNm<|Pf5ycfb&vza{;vQSwqS&ym7nV8$XFYe9XiywXGT`6k2
z;@5uV*W;xx{zgn3eN6dLuMgrc{@owOo3H;QhR23u|B-{S-Y0=m-EVGk^rEa@6(<`J
zbqHVKY4XeigqhsEm`_bS#}VC}rt{WYZ^o5Nm*c?xIlpvxUG)F(``=d?Bc6KZsrbkX
zABnwE{5C~XQ^mr=kMh!=--}qkZv%`M=WRSlG4+;-GAL9p_@Ur>^+ARib;>L=2F=3!
zeB6>T!TT!6AO_gkM~}tnGsk0kb}B~4hpcEXtt`Z?+c&Kg&B`ESMd8M^+iC+#qJ!C5
zDLt$vNwM)vp$d_*v1w(Cy7dWWQhIpF@%+0N;``tKv6LF<g+H)Mawd7@{gyG+9Y+we
zl>!_nD1*Dwm>5>uVmsln<A-B%nlEyr;P3MGpsUxf#`WvheEJGFNe8&T#77kT5w<hA
zd10iX<e`vC>GDZrGJ@E5*`E^u14ENBJTk33`F6&r{1wLbLh;@o1iZ}($Q>!>3&95-
z*)E2`fMS?0#k!>^e`e4Jo2!fzg(K%EWz-8w>Z4D7C_ee}XXDxDUyS|x561m_8*%r>
zt@zRRei+~R#y8`_yYGp2X>`+`Q1d$mg>c~z5>GIw-Z!NXauUX=BS+)czx2!T$)ErE
zIDF=z7@m;PEFoTBU5z(>_@nrjzx#Xfm9PAHymR5bSlhZEle4o@V8>%wdEG-XRCzek
z7#RoPBXtGM^i|Zem2D~FIU6b2?6eJBrptDq0RumIG2uaf<1pdCWprV0WW3WCb)e;E
ze{eB$=?@&vvkRH0bL7GI@SXBd4(?0Q6oS0A+O^}W{Lt-*0M#k-OgVXGqRGz~6PZ-x
zsQ*P_q8%s>Wu^TxiNLBIr@7I;AsSrAQ2r(JY!_nX3Yp2R0_q3#?Bh<^PO5aAUPQmM
zE?&#DahqLRr%#pPu68oQ%R{2qOM9$vsGoVTu@(pC_Q%n~2R(MN&FTF4i@`@gDLaGn
zrsQEwGQ&~F^d)>u5_tkAMnP^xqJM6T2|3y&TVHc#5PINQ<KP?oG%x+ApQ=adj%&t+
z$0Zk84{}6Ny_PbV2=Md8@qFJX%g9Rp=T)g1|1MZbmveoSw*yI$v67B-)M3h5lYaxI
zPhbTK96dl>m!*l9IMNlg!#t=1@-kIT{sm48eT{zypTdacxQ@R)UFW_*PYQVH_KDyL
z@|lh>BawzSgkgts+i*FPj!ygrOuxYCdh*U^ctLwh`T&w-p8j|H=T9jdUKc_OKM@PR
zeHw3ryeq9SDUDaL@KQYdfDV2`($~SMH09d<?Xs4CI}F$B|3SEbqD7ub4`Bo=ZX4VZ
znzX%mfwo^kZy3-^L&Zt|Mn)@)AK&(*_aDB^wSzxPQ!;|HVK1KG-%#V`i9h2VJ6hCx
z#gzQV?INV(+YKw8Uk1><qM0mrV~C%w;6q*Y1e|mNCccShLjtD+<jfW)rg{2V1@^>{
z#zF;SD-7?C_}FiJHlF(%zZ&}=JF5)UXDe`>56rSPf&R(GBe&*u2k7Nz2Z}eknzkmq
zqV@gbzyDE3;kZHu2xeS!);z34y!GbW@fTmaC?#n#&OLf4PMtdJ{m(=g9vTxOb?&m4
z*fR?9`lEp{=;Odor-kA}W(-o<;1!cJ4A`m!suse^*r1~1;|?1n4e8BqrsdIvQDmp}
zV6W#yFa9Zio`EJ_hbQ_&c|#abj6BeZHkQvFN&l`eS%B5)V93}h4e42-a0#718PK`V
zQ8i1{47T#~*F05zX4UD)5D-=|P<q)qwR(?Jp;pBk@l{G$HuU0$ORK^0R@}OER|B3D
zSQ!(dtQq3&d*AWCefwlg%*FEU<@o+L-}Fn&zxt(LiZA@tzZv__Jgq#jPVdJz{^*b5
z^;cij06rQAj~|K+889nSfYX>jc}hhPHnjyS%SS`yIbDT!8=<m+BJt(?`SWr0^5xhw
zyE{fDSFEi3_>~{IJRg!V@v)D8RK~>KZN>yuZJCh=+c9|YkS%Vk05I6IMGA$Mf!@ZP
zc*#+q_@T6}Na2O2C<DA0TwlP3_r&Man8Qcr{BbO=0;%l7qr)~%Zi&AbAe`dF_MUs3
z2J?U|NaIreCv8mRF}5kc)CsbR0tOx9W8*P3H7&;U$K}gc<J;f<o_KIsivM!(5l9Ak
zE8*e=M*^dCQ)(+MDzLZJi0^E*I&t!-42lDKWI2<SmwR#Z)(y+R@`~`OJak|fs-y+n
zp>YC)l)C&%uw+wLIw+L2;jvMLOVJzTIws{$hQ#oc+V-^a9uW`NCUW1~H~2!c`LVDN
zYvLhCLvo-2Gb`kSqrG6%yT(NWCr_n9%CWsN0%a6XXDEzMKKW#P;fud0Bl06LF*W1k
zWG|k7FTVH9@5C!V_+ea^Y>0_A)@hT}4Kjp;GkDeu%?OFNDN0_q*4ATOM)sNGC*yzp
zzx-dtv!DG;Ozz*U?g-&jRN=dScQO8l|JQ#P|NamEUA+3{8?hmN95{9~2E~tC3-htS
zs-TPtABSss#?k5LtUNPO^Z|LwO`T+u5liN2tF#TC{fEDrZ%%f=pTvo9R~z!5?%lTS
zM{uSG2zLW>blPJnKqgSv?2~c}7B15fj*@Iemw1le&;)MbLLQ*YbjueSs7oJLDrKH`
zo{;Ih%IK6KL4RxmLxw2wiy@(h*pkC?W#Wdq8=su?_<(G)LdFWhD29aMv+tMVdwHKl
z1WBgYlB9ZsZj-N4E01+%_QgO>O<P=AiFwH)tB*d3PtG4FT$N*cqxwz^iWwO(tWJ;z
ztB^Nu-HJPRWlX4P&=12q>d)IFSzU0OgB^_$sq(?2e6b%si+`Oz?uCeUCek4v`F9-|
zxaKqX%6aVR)Qt;BhOOr&y`J$C39gIuk_?KjGzc}8$iFRTB=6)`cp9)A;o5mLV1LpJ
zlFr1L@iNS1r>rf=IC&+!?#ltB<vy_!UO(IO1-P&Bkf3cSX9|Iy(_m}HSmSU&bM>vk
zyt90=Dw<pJ=$kqQz4H-$#DzxP=hz`x(A8n`Px?WcZL<I$*QEp%{$~8#;zXMxYNvkm
zm(SaIljX~2g_tHkakZ>eoP?rtC%?J}dh4Uo;L5M!mOlW!Ks$axr)~u8aQxbG=)9X`
z#Lg6%rVX7ZrK`TF;#59B;RU#65P23YU4QG14KCuOpRSJQh@ERMRGx$=mVdL|*t-kz
zpwB=)N#g#x39?xEnzF+9J;-F0JSXJF<;H}!PIcmgb5h%EG?1mDs{hU_Ug}0Azv8WL
zQg7^2U*xtS8S_M;^z>0(chMiYAG;sx8xhiEqED*F$A0}|@#6pd*W$p_Pee$sp&75Q
ztVul;UIsYlgfj^{Lh>`AT{rN;yl+wxWPUes`u_gk`;h8G3KjcOHAt|;&36~@yL;<S
zyz}<?_~tjR+n6|WW-cCj=!ie)#_SDSA$$fsOVRx(;;dGrf<Ym&v5Hg5J{2wmFK7HD
z9VP&>s2&6ilDxf0VOUZ;yIdsWhJVpjGnD*P!6gbxlR{4Cfq<wRwu0cYHp1vM{J|Jv
zxFJ$e>U<ie40C*r!c$QaC<_EP8$o(`#sS>EW+v`{n{0f>Jp!Pz;?MFpc|isppO?qw
zmoLpR`-b9m4^CM5f`RNe#`y5Z;G{5f#VCQ74E*ewyL#oW^4f^8$>EsWzc2Po?~T>F
ztMTLSzZolw8}a$ieK!8eZ~XN*`SgpTN_ds;SN{FKk2hX>GsY&SeEvSh#1hN?8>~vG
zoGM0jP{m)$UxmZZja&a&FEGeKzD1E_W|S>X*RNg6Z+Q$3#nnrfz0w6wp8oLD@u`<S
z6|?(h9XD%%l^rP=tYT?U_e<Ei&rTz}U5u|;!v_oo2HD~*{9{!jPZkniML&*J1j=IL
zR`L5))}di3HBxk}l#9PZGQ{Xe?#|D9#vdaBgND7@9KVX3qd;*&6I&+uupqOk<1#MT
zE(0I}%;2I(@%47T)_mp4wfNx=Ux^DBF2>zE^U6;O(Te!20gPE%j(9}@W5vsgoH`M<
z$|3O6Gn3v+&M}ZSG<D@&1=FnFqmg3qfvt6tD^(9GW}Iq9+4^^7hwJTtqJ2c=7#ZbV
z4-5&8t(*`IdS+Xb90nxiUXb$5H%&Mh3SO{f44I^Qsq<7sX=htrRq=Zfbbi#HcuU(w
zuKkj^>WFP&ANtUT;+KBuOY!W7pHpW!9G5Q2sQl4S;=3{=zWU{_#G;hGiLtSmk|IiH
zy6J6q7!Z>GMwW5_H>X`UeBHB8^*o`vc=+h?_*;_k7k}XkF@2CPfQniOrrwuV<6r&{
zzY~A_|ND3Gqo4dZ)`jcTxkqAPd^9dyyBas|+>T`#OS!dW(|rhuDHHt}9XR~;swxU8
zh7SFhz<r2Z_5&!})&C@^7!&P5S$ElA=#OxG9bb-T8w|?JVfHV4xs3?L2U@Jqhz8`I
znI~gXB4{aN!fjOPDM#A?-X#Np10Fb*G{15@=u_-iCF59KPBX%v@9prFdSseTl(xx)
z6UGKBYbg6$^z-Vc>EUhoiMRd^AF)Lea^jUM24nHpr|XG_DvqLXWb&#HW#9|R7!~Mr
zJ_1$r%a-@n#--&IyE@XFw$Dcxc^7I!ejJl%K8m8L@o_zqUU^9FF%pqa_aEjeGHe3P
zAN}|_>zR0}1JMePT}#AOeQmOZ*6~xJ=s4#LvUFV;RhZJHF_G=G@+XyX3F_(~-6J=a
zl?FHM4O-v_bxFTTj^4s1zb2ObOpq^1L!O1Z1(vg(Fz460TD}y{wScg4K!w4(Ha~HW
zqyOIAE6=(XF2H<jg;Tav(yN?FQ>K;UKcjc9t}6XtNH9L~j*P$OB0sk|obseSW9Tw2
zfR8qX?BTe_0mKDg(OLOZ_B7}-4*Csa5HjX|QFL^lYa#*q6GsBq#0!qzJb+$5T*d9Z
zuR7xyn(G;;csKz6b{J64bp^WV@W*AHXP@L-$EPo<%5J<Hz;n9~psu7NpPf+hX>N_G
z2~T<`m!eP?O?ah+zV36EUhbW?gX8_+w1o55T$3)+J4XKNIpxqfHDL)7DdX`~f3$=F
z@rhN42|Y}z`q}El3Iztm7`lfn3Ff-k62(e}I?dRkvm;-gg^W|B9)MKD^eJEZcw~+(
z#S`kA$B}U*-~)C{lHOatA8QI<ABlMJm!FFl{ueSPo_k7}kBe@pd~B_^!2*TC+mwP5
z?F1jlGdaTGdc`rxO*`rPt-pJ2i-CqggGOj;UUm;#Oz7m_J%1s7_=7tdWbVi5(+A?@
z$)hnoKH)w3G)TT)%?sQ#Qcfpgu;D{gHPhqHK>6~8R|X1%k^Ias!sWP9VOcRL6pH0^
zmf<3tXWi3)sl4(7s&LMePQ3Xed37c(d0m{mQHqJ<JOxTnN7r_Y2?4m8L<Z;9Jerwu
z&by%lhU3q(2YnTUXAf@5lVx?UYA75dg*Yf~tX8n?U{gw^PmNMs3<?YlRxX4sMx_W%
z&rG@FeCMrq{lX&jQ7=<tQ_;WK9~a-b5UWe8@ys*N#^=BE>+zwFy~M^kHWbGn{Odo6
zx8J!KyZ0T46Q@td>efaqNzvV0S4XtT0G8#*vQWsZ6PHU@@}ev-ztb_u-(i$Tbx!$i
z-n<caWRMO?K3JZ+{N8(U{`@<BSKx(@z7U`I)F)$l_e?5_LXS+aZ{0^DuBxG|rBMMr
z;NyLW+qZAqaDY!<<)NV~JxU(4^vodB@SzXG7NrP2p(LOX5|6Dn7#_@=LqB!L)|y-5
zBi}gS#3Bp{R_HLU8I(N3CE4Ztcnk@Dt=-BMM?>Z;E-&d%O;5$l^t2SQG10f><1IPS
z>iqe4<GuGT_z2f!DMW08V-*2oVrptK_V3#pduFC{g=LwK8e)j{xy-yzgOS86zw1OY
z!Igd597D{QMRgD#1(q#vy9SkBM(Mx+FQ2kaDvx;`7|0i^hlj=$f7pt--}{hqzAD9g
z;qJUA5f&E~?FS$Gh0kBmZtx=t(1#Jkz|8g*#Du!Q;9FW=Hc$5-n2TpV{A_&bOTQe?
zJ@=8AmqGR8AO0kM{DU9ItFOEo-}}~gymB=mC2(4NgNNkPFS<7*lNfp%v<C)o>PdgR
zI8EoewXQ*5^o>ilj?W#8kA38Y_?5r<SL4tV=c0d-lf6?0?k}#yAN=<3#aI65kK+fg
z{3w=WlpQ^ND)t^d6mPuqmW&C$C$lWN*)zW({^Xv29Lj*pFULJ?*Y7o{eJ1~NwTzi=
z%bWU7_o0frCON~epm5?>ZKN~k!%#sQP4d${eUE8TUGqXdc@ZBT*x*q*UeG17k0w>$
zi>y}vnjc1!3`-7?MNTyu;53kl=@^up`dtMX4j2!-w?iFp&kDt$jR01-*zVNyVUjhk
z2sLeyBaiuxgTMbGrGG{3XK`i8FWE1xEXpsxsLYl_R<@Mb#sn`PdvMg1RR$ar148V6
z4}MbrY)Jy#15qZ)py^N?d_7w&e$Zia>@2I3b%2KXt~BtARkG@vk<q%R7ejvGe7i0j
zFEe$%I3Y~=lg#IEk$rZ?&AKykfxR4;pv0FBT7i@U+)j{byWuFg&}L(TatVk-87q#k
z^JD16OI*?eJB0x|`R^3g$xmV2+qh}`ihn#qPeV=z+BgfU9T6EUPOE2+f0lRTJ9*wj
zYq*@yz6W?_Jg3eFCI6h*RXOI4f^1Xt)5w4EkUYtc7w1`-F1jkbLsv^qfO-aqOWwqV
zQm&Rmo)o{=C3Nuv?XdDI{PkQ==_*Z|jy8=Iui_T)47B}vfv_f;Aj?$HN#{7i5BGIn
z{+TxO&1ZhQdA7roPgRcOJ+2L1JPNZ%*AEZl!n+l(eJ$X58?MH+p9^~9>8JBdkn`iO
z?)9V5$=i@vm6+04#yb;Jl8XuT1!Hof>c_@3j*Ln+hiE_QPc_I{-)7Q4rS~|n+2o1d
zgih*vjV)6g5sNXQXI}Q+daxF&8q?-wFf1zH=f3z%eB^)jt1<WDGZE4e>7O2~O3%Sy
zv4$%-Ftyy~)n)5}@@t1Rsf+<2%w^Y)-9P+KAN4yEq;Y?zD+32#Jf;JCPsYS+Ke?@e
z{=T<3ojP?wgVT82onH_FW+Agdkr54*)eUC0*yB&<=MGta;DT>7Xd4rXPX}Gw7f?hX
z(#C|+IhCimNUdKwZ4$}3g9l7>di$P>e+5l+E9wX+PZXS_O{0eqfbrWwC;2vcI!+2Z
zgF19}m`r%5^XP(SEfh*e%S=TY5{V0cH5Z3DeUmoXo)yCY>7zY0II<eC;bUK!iI(#0
zuQzLe97ahVkMYq-DSivyQnMn36Kv`(qyPRu+_-$5ii|T4KN6q*%opM#pZZMn&&<ZX
z8#m(jfB#>_g^SnY@R1X7^2{T#EW>443c==v1djz3tmUShIiP~MDXW}2PBjXELDKc#
z_!EU^6twHtugBuu1@Cj-C7!<b&O85pkz@b>|MW>jK~!<!-FKDmp?K-jFU2#@emMFD
z`>l+*0ZEoPR*^4Gv(1JVjPoN)ylAb)vczdk^Paut3)4vsMjr}-j1c&yfxvXY4V3{S
znWEHp3H-i=u5|pXGFp~Ir^})EewkYX3Q7<8=4=Gz(f}oywz7i~gp#B(@=>=j)d4Gc
z3{D=z<ZQf1=HLUz)?xth{z0Akf()>NV8ves1*g<ZvDaM6+m$Pqd;%DQABy_242#v(
zWf@m#RMAF}VP)xZjf^qNCc^}th?ia|7?6k$ie}+sC1TgEG2t6io1B!gI4OUOq){ne
zgR1j&uO9Jk#45*8%4p#fEwR<@j{%l%UGNgH_qSs_=n7Aq7wtmjL3!rmkW>l^?6Kp=
z<GJUbk57I2<#_t(XJn||j&FbSJMo%~iOVu{uU@(umt<VQQ^2S5jH=&|Y_lq{et$jI
zkbkdEsIyhMsa1dIO|<bbo<YgWh-e&De{%ZhiTJhO{Oj@D&%YeIj~~u{lTLVfE&lEA
z{9b(PD_`?5u~%;2h$9c3kTG#E-V?8{i`R>)8(zp|H3el9X?I7dtNS$756bqq+T4r`
zz5ywr-f}0!<IY_fUXn{>6n<b3k%n;kH_B)F^@l70^a;erAFotc)%F}mx1xc8!$(q!
z?}eKc;XS+e*iDU3HhpcLa>Tt)ToNzprD?V(frr&J_wlOl4fX3AynlcJsXFs&0qg9d
zg?9sBG{*|lPA~#^!9U+)$wLs>0*6e{p6>JEMH&2Fh4O0BR@@WLb(N3RsoHLcLE%+a
z<!OULJn<l^yeKg<bbNzhOLfh5LkxBLZTLr7kURI=?qisQB8)MIG!NWvb48qfTEeQx
zysyVG((bF2CdU<+tvXR!`e1;1p&ike=(l|Am(NOzFcC*T#LGTUQmQM1b@h)4ic4N{
z#Pd6J#8bZb)p*EX7qagJwi_MDN<KSWMEScS@<JZW6J$&to&FjJAIQs%K%gM-W)yxh
zE|9dV`W>F*I2qetWy{CB0^M<;Yi#3J3ON3b(M@8gbae%g4IJ`7U+XLt?0Rjk0H}J<
zb)H?q=D}%H$MjFEOuz?Z2;e^3MH>9HQLg@8h>Q!Kkt0@P*%nLtr7Y#3dncFS!rvmW
z;^%$$8LzgYpV08Sr-b;m<7S-XA1>v(WTuDzO_~DyDt+&@7b^Ysypg{y4*Y3%1k<i+
z0=Uw2({<D6$-e|u9Kt9=7jL$!g2L0r3HY8i-IO0Oab6YHy7t0$|HO;yOnx^`2%9*W
z2G{jzf{E8eb=M*<t{Q&5Ix#LcB@%Iy^25NG)VMq~JQ$ObyD|0A{nEAg;3^YCUVT>i
z{oNx5Il->%I+RDhCy1uUCPJfmj0o2#<9)w$rnQy0rg7+&y7Ff}`$T-?Z~b!2ef+tI
zeY;JcBB;LTH&wTK?3AR#_spvRURIgQ*}n<vl-<wi=lg&74<2`19G(~wbaxC=%#NLZ
z``vi$wd(-9KtsPVJa|7&YLGs0{Fs-+QIbYwC2=B(vU7S1mO2>E=o=?wn}ntV)Bq`l
zR6ca}#33D|(Giqz(Sh?<FL8sz-LQL5HOi*(I1j~dBI%x=e1s-w>vXmXtZ;`l>4^ny
z^7eq#3=kcxx+w#on+9}EJS3<9y!s=bELRQ&Nh3FjD`(V`b~g%&W!y9<r9q%H2(u!=
zfJ7E93+eR7V1k0cpua3-h?#ZI&TGKmxX0I-W$X#%kt4_B=RWyzeEj9l#qjLDxO3%t
z{IlQwy|{YqR-8ETaGXAK)=D)>`~&e;Y<0&3&)@*G!t?4d+$onIbVHX00u){1(~+YD
zk^hy;m*U#ht2U01i5t8;EX8$FO4(zNJr?I4J?pYCi^(<ylx=2**$TuS={!nU_>~vf
zX=u=faxpPAE+r3T38fsPmVt?P9C8H+xna9gKbp1fF_Jv1%r+t^Hf%5B1Q(14^V{st
z?IhmvBEDARu1=DHkF6Gd;ahZKbnWWP7gqhv5DiEub7GR?!!Ik6x`$`b%Qh(PDFa)l
zSed{W;hQ0SeVcLX=5?>=ps-nD^h3Eg!3#eRzM`G&TF4z+nrui&9(*W*@KQ&kQZ_l7
zwm&fHF?28{#;765l?I<?Z#X=;bNiN!2@V6GE~%?(!@NsC@5l>Y@HuC5x6~hq?<yZJ
zl*2C+auq6O_U?@n4?i5w%8+>eBQL}WDfx?c?#9=@{MGo$4_=8)b!>a4XMHLUhT)(7
z>7U4;TT&IN4U(lqN-e;3BZgSPwmgcAY%5Y=Xsz&6v<@hM4tcL+`&WPU*W%}Y^;hHg
z>C>j&%v*mDfAG(KJHGrUe;nWZ_P65d?dvgj>~Ku)-xJr>0pDJnS6-4mm4WwjkYV^h
z+oPP`sv}3eVMrV~awKMUbL?}t%vWSE-@bd-GwIYT@1{_<*$1QyFi=^crS0e*KM=6W
zMIVaMhp~#0f)d2NWkZfHkt1hJ&ahI3K`|xAwlIznW<|m{DXw_o<4e_+*$TI*dj=`4
z7{F^Xz!q23U#JdfUmU5-RwEyGP5+@db$Edd2=Re16f68_2R3MtZDq%*DSWn~s^>H;
zMFxFXV^FxCmY@4I4VuW8l2f%!pOXCGnc7ZF86V`xc!rCAs)x)Yab(n#59%$Mz+FQJ
zLW2)!#rw*q@l3iKS`rn%-8d+s4u=L^9fx23DUSjh&Uc{7L)=ujvP3nu742N1(`j0H
zngl`nIq#vjQSvC4JnMZt{2ec&S>B+*3CXLFI#;D_o-@9G?PuTSksI@?d8!mS!9rbU
z|L?LYZSfjVcPk*h2el^uOvjb_LQeh7CZ($*GEwMBip+W(Q9IbBHnFu!c_@H34@LBM
z?0&#`yjGb~;YuF6j}ULTcaEZ&e%OYE@IVj!II>K?$%_77^;@*{t5>hZHH}di%~d`-
z-BU)|62RmKGz+QAfXo*VKg-Qil^Z<do%y%@yU@hZUDa7rD+c`1pK=nnDWByy^Q$~e
zg2TAyA<OZzLMMM+Gg;CmpY*er{5t*_w_F!?&a)Fo!QcR@92n5x%(m1-Doz+bVRF=O
z^V|Xan_D2m^g|lm6%RVJ>Mwm#j;n3C&YQT|?};X2HGX-ilnMCJBGD8Xmw20aRDR<k
zllLevDyCI$Q<Ae8_470A;!&Oh5^nJGuGpIT%yr${qynMBkZpPn*$y$*w=2dxfhMQ+
z!Uye`c;o%Qn^JAg{oJW|@o)W79Qy3Xr2NeZhZ3#k7E+-#xD3XfjN{jL=0_9m$0oeF
zrtkY9d8O+={D)77e^era1AFF~wHfd}cPjJUx6j90Z(WN%X=6tY?T=#`OeZy%(cl>j
z$0o;p{GO{t4T}bdl7b@VmlkzR3fXp&vjUKXz<9e7{_>+CE3w16e|lD0mj$NcBdX~v
zy)B5rNr4o`NewCdq$!7A69RgtDFy*`&!CJ0N_lo%o>gf8QJp7$ZqTq8SNOQ%nlKa!
zo{UddrQrtpOJR4%fWOkUY4d<ZJSnu`!=J&qj3e;flQLLeJms1%9k0aV;*vVht=O}7
zE<XCPPsB@~`-PZ3a3ro@xEjCxJHHp#uHTLmCr`&CkDSw>&&fS2abH5JBwfBLR}>@^
z@?2%YhjO4`r9uqdbjs8>WdexHDN0u^U-3*k-<*(esDtJsO-oinW@lzP9XK7dXY7?X
z`R3PN6BjRkGPBw*+2P(Vwd#3tax!PKHQ3;2e#IMrh65g&p+_Q>%`aign0T-$x#S%S
z(Zg(%kCG*k>xI>aRK#pdH2lZaF|ZC3F51A>edG}yC7+lX^|n9wqs9)s7*{qLn!V{c
zGsetwGjmL*;P)KZN6yC&7gdj(N3ZrFUa{?qS$~WPc$l+hsvm`!U#w!mdll0Ljq+v-
zQ-43(R2Xn|l@Y<q^?Z37V~uN;p_zr2@=n>^zM6K3qa0&0+<a84c&sd4e;iG$G2S27
z@;6S@U+h10Bu+i@NId)Ei}A#Vo{HVGd*asB>+!We`Eq>Y&;C5-Z`_K9j-7~4z4TH%
z_xwj<QAW)F^}qhV#ih#^V_|7td}rm2lYBV3b5-SFHc|C1o+hMGNtxM;k8i&!n)AKK
zM<?Q?Preji`juagC!T&LM#j`fslKkCzYzcKU;TgM>)-gAkB7Z_>w1jMNOq<sVpU~l
z+Y=o)Ww*QtH~kj;#E3xIpv)*YY|lG<<Z$ek-1<0eDeQL_R9D280Qux2)ivA8=o?AP
zDlIDqyqADMGK!&q5+wdn_qo!e`f<HU5h6_Yl#|Jpdl=Wt)D!(HvYqV`nph^cLN9;5
zRkJ1qh?km?9UBQuEQm(r#qS%a&R1An(3STI@S{y2<CX>52ePg<<Qa9^lU%(3n|)aJ
zRdfZ&3dHE&{@G{dv->X5MO_u-J)s#tm4&9`Hj;FtrJU&!3QtAPpu|O3;U@n!;NPLU
zaRuFU+o9rg!r^iIxfkR`j9&bdAF#i0(60OuQvO<aP!@P@r|*3(C>n6ZZ$Ni>)46t`
zaCP#<33~B6tPKisoEA>{dLWfqBKXlb8?1@m{u$`}6<S&4*qy6J^yBn}j3;>j0Ppb(
z=hw((Fz=Gz7Gp1UfJ~`P&}VVDg15{0O5082-!kcZp)ZAnyy&|zUVP-Qp7Yy8xs{H*
z_V3&0e&+J!%W?D8ZI27Z56VfO3V@4ejwgm@#tx4kO<b<+`+_!~fi`Y{XYO&_tL50=
z(8Q_sRPhQ5R~u)A<JEX?OY=ccH~`|eab~*&x_$)}x7^PPy>Z)REMu;q;?}kD1S$>C
zPGhL}`b+u9_~j~oFMo3!On>Sf=;V<Pd2N9EJPCBm$H3yyc6ncdcPKC>W+X;aia4V_
zeik{GPBS7o;v=l=df;S1wm6~vVEp#660duXrnZ41zYz^K9Fa7PZiEb~f4-|e|E3b1
ze(qp=;%|O2j(y=%5vPuzz(uT~v#89pUFkwqzFM^de5YrwRYm}QG&gx|LHQUT*Wdd4
z=OiO3x-64(gVN7zDU%o%FI<cZ?_Q1ddz-Om_e>l<G#9&PW;KZA^7G{Mgtq~Jjf_dl
z07gS)YX?dJbl{IAS&Bi;P|-;Jo`pIQ^7ghTImq*1!ynU?D1^GlkU&{Qk;@9s%6I3M
zm|Yg%=_Xyq!Jm6G5$8NoAPGebENps4khsq@`k^y(T-^)Iz{UaK!0C=?+rj{D%V;20
z39j=dKKKd2&ncisU~guA;EOs=r@=qduE>~x*MzYx3MB%Af_DOV-(X^LIzIgD3-Rfn
z|6&|^=#jW_;Y$4ezx<!#>eU-wwRq(0IVm&4ar@4^@=)D~*UaiWAB0c*_Dq*}1K*$<
zzzCpxyt}|2YGyX+fLTdFIbpSgZwqkx(4wqnI&zAVd<OE$-~k?F&)YF1SLV;A@{vDE
z7w=Nw>P1qHH=V$kklcB1xKAl!Re%PPMyj`vm`_RzD(GkMLGhMB!Hg@iX~nTA6J>^W
z%7Z~w1|R-32K+$EWAO*PV9!PhPs*5J#Q>Kp23%1R88AI~qC8So@W`15gW|pGF_pNY
zpZYDMCzZf^5Fm>9n)aG_oiAoXx3HNW6j5<c$!XKM;<TnVj1j>KkPSv;Pyz6*(P+rd
z`GUv1R-NqIzt3L>M%m?^4ji+Dyb#W|0FJ|b`0Uv@_t=wh?ujSj$!DHX-HgTgx89Dg
z{pnZY%YXDA;)3o6)CoTK)YI`Rzy53SbD#cX+`e)({@ee@e;cpA`MQt1U6CQhSFhRf
zh9O5MSuea>F)(qtMaTQ&)kZxFBszx%hU2kwkHt%$cqyJx9y5FP#OA$?c>TNIi*J7O
z8_MIgIDg^2xOwN6IwRG^SSl&&TjI5Ok3utGyvQ6gk0=Znf%G@E?QGB9Juy8!6N56I
z<SF?;GLX;4@SzWt3=9uT5dx2li&3TH{e<x`851K?LWEC@cU`*<<XAay9V4USEh`<1
zGSuc5_~_!24FHVFoS=~rp!i;;L@`s_s+9y}+y;Ui{UdvEQ39+`NhUCqy{(Y?Z{#iq
zO=P*Le}M-wKPw#Z&<}sv7LX1L3IBfPztURC<)(S=#B&@xX)E0;Ic1=|d=Je@AJ5re
zy1&<RdgX7s{q(w|wF~&Qq0$!Yl+NXB{Mzx`xbVv|e{jBOfT@3SGKKp;uS(!Fjwk0B
z%`<&3@zNicap4DE8HaS#A$)3R1aAw7hrh!bz^{uJAbk6^;$%GeG&m|5SHik@n|zul
z!0~lWJoyap#qvkaoQZw=_ItO3w~NZ)9vkHyFO3B<vU$n>zB6l9Xw~N|Xk1~k$}957
zJpH1s$c%E=wVMYQPyXPPo(~1mr(<9vPxNn0j`9_B+7~Mm9v5gsO4BZz>8H;@XJHjl
zSATF;@hThuZ`MyfyZoeWo`Lc!;JK7n?j`-cSERf)P_z`ZpX<Kxa>XybyMArx#rM;J
zxP=eco?qc=<LoUXP;~7CZU440u@i2f_qpP?(-j4(0k;(Z|8niPH~_zL_?5l}wBr<T
z$Fnt7RU^tj4ufJy(J?3{<!@s`*J+I>Ga4_ZF(@R15*kl_+0Yh#JGEs`P(ruxpc%1!
zYE*stkn|0XMaOty_rSu+THKI&a_Z@Q@u|Q5`8e^#Pe<P)CzxGS!@<~A8I+D;KJ^s(
zc$03ZTguihSC!S!z8BP!`NZ{q`@cJ>u`CfVkkWZEQ^xYS=#QJ%Z^WgGS7K#p!!PO`
z*uOXS?cL`%R*=LT@Bg-P$qTBS6RrVG%;6<YwnOoZ;^&z+4YWYf#mwz6FAB<ubt*Ko
z4OARfUNhq>l_5`eQhHb`T2bWOV1<LS+a<~tV{%aB>NL6`O$tW(*<jFv@f2=7ca4dB
zb__<<28f4quHc1P4P;o|Cmmev&!8e_oGsT)lhmhQHYl>N;^6WU54tFulrH(x@u7UO
zGJ!F{zIV?$N+BN{8i}*#o`_Gr{PXeflTXL3OV{Gx|IwfLq@lfg=i>C~Gch?i6}N8P
z5q>%|sb?DTsK5OBw0I?6TM=PqUan>@nGt6{KT04k33Kug+h&;kq2qSv-55L2N4a^S
zh;q4uSBLK9V*R0frz0FArD?#Th;LS~7n{bVtR-V(GGb?ZRO{#%d)Z~sO71XBXe}sa
z)H})_xKNzfy4E$w7+7+~wpJye3I0NFE`jHO;Mo}03w4lnoOR`zuu;G@x8r5K_`XTx
z4$*Rh4CQtw+KI?BT{&aMV3aZe0vjp9Z{=3}DH(-GXn=NjmoLKROM@5_+~a3O+y<4L
zIuLkfM^ThFZHiS$@=#`a*0oQ8&coT*p+wWpY44{_o)pig;>N9;aq-f{SXo|;OP4OG
zPnd|upLjB!fAOR7_>)h?(Gw?QXmm7Qd-b*W(?9y-_=`XNa=iD(Td}sX;12NY>4)QE
zANy!L`uI61$#>)1-}!dD{q8$?46Wn{LxP!oQQ}oMuk?uKRBqK?)wv>rUh$EGc4jDf
z|6^|7fjD{WL>xJEB&H?PVnDq2?z`&0F2uDP*W%XQ+cv(3$JO6VO~wFy6*4I$W@=_e
z%FS+-hxa08Vp__@gyJDXC=FSU>fhK}B<Is&#2X9<J{;t>AztIyPPBe6jZfOZeYvd%
ze9uO7_G^$pQQ=)EW=VC=y9tX6%?AP3(g?udV+CnxS&o$?wM84Z$R%xEMgi^JGp$yP
zQlYV8sWPeh%(vXKmbA0#;}r+tfWPoA+Z-}Mf7kRkg0|w0EEteY)k7-!ji11+6#qtU
zp~3i|!?I2q;t<;yJI*plK;xBYyZto5;37S84`dp$Z2AaYGe5?l%Da_s6So(7!${l4
z4ft8*sC;nkc!2%dWyycvm%r;o*X$bbZ-BUkSC56aoe%ebw{-}3rhclNRc`!~Z%zE>
zFCTPOc;NyvuT0AwA%v%^zAodHt333(Ej;h@>h$~xp;F}4L0vrbgQrfNiWgsaA<jH}
zCU)=M?fx#eUgrBq^yBLqe|VRU@rR?o`DPH`+2Or1+Kl^Z$v*82`J)s?KlwrjlSS|=
zCx>!nA!W1AD<kC1>f!vIyK()-4fo0379|I7=});QKK;G7CA!~lRxuTxIFpxUOzZ@N
ze-Kodd7R8s4)0Gxc>7w=`%F5#(m%`F3%zLyK8Vik@$;Uj_<s<tzYN-W6%PD*Y4}<4
z*eRV7`Jxc_O@L*%g^M(8{Dq^}zkS~ZVt4XJ4#b-@DDKCA<dW@7d~bpip4jq~#>B3e
zlFUyrA*%MUOK~1BwlikX|GLg3XTpir%r>V{^`~Rf(}oo$%|LxSy3|J8U0jYEn_F?>
zseSRWU;k{J{Kc1J*Q2LJi&CoXY^<R_s!tK=MJLULYg1xf^+z1a(7?8`?oe)gjF0Po
z^0$vGes;zvE;Ru00WDavEFv#nycny?Yxd)t0sHsuC#je2H#S$5hOJTAkeDfEmY5EJ
z6Q6Rs%>yZY><=DQLnIFoE}~ILy)q%|jG4UtKDJ_LAd^ACK#Tz~AmsrC0$S8<>KB2?
zr*5I?+*B;M1K_VR@#jXV9-Q)VTMeBuGsxn{kMTC0LGi=)3I?s2+R4wX^Irks&ekb#
zL?>K6nTh>Kf>BlZWT%(m?fmI%9nT>uhZF$9h=T)2#Wxz5ZBzs+RU2zkFmUP+*QM0(
zrQ;(<Pe|GOL_G1#N8;Ai8}ZF=e%C7#7!gO09*@b1DJgdI>X=df`KZ%+9?7`GDMBdM
zD6Oevr@LpG)J<DC;e7?Zhd@V7U7(mcT(kknXLY<99K0`^N~ZorGtZU<x#UCgG<lP8
z{BpL+Zv#hz3rCzzPGL+;+OWVFrGsZ+K)z~5dun>pJVOcLs(&z}8yHd!N?}QvWN`4w
zNzDFYXov>V0yc`2ra@#N9HSSyA)oCi{<3A04#yx?+oh;GWZU1*K;hRt(#PyOIG7#t
z_7vr1<Iy;fi<Av`M;$N|M?H8ZLw?YTLTh74wENu#rJ-CX+I)1+d+=QsNjG5`53Z0`
z`f+NLl4cyyv|EPTp+kpcu#Lvmn>S=o+=|y_AaGLFN1l5jKK}7f#JNWwk9`LYstEn@
z-S2)czVhX-#A~noI94=3bDG4hE0<wl@IgIJubP>ij*+o38KN>mIjKw?Xbzx#HX_y-
zphYV)`}Px0aPSDotzoQW$?#=MoRW^naAs0sY-B7(bx-T{ak{*?zqlYn;$DnSjK`^m
zPRB!!$cUB_(Jwh+m0^5xIwmG()klrT?%5e1Q_J^Wz^{LpA@b8WW<RxGOqRmWN82#w
zX)lY*OTm}Z@7!Hdd&0QmU6*;eJCdm-)&II=1?H<=vO=bI>+Kq}6ITiurO|5{K|V0%
z<xd-+kkn77{j<8EbUBgHv^DBZ{1P@fez_LNm|?<*x@~+*M#0m5F0dST5t#c9m#cmh
zvV9fJK-ZW^xh0k9B>uK`p7}tFvjh(pJs6+k(KfTZ$T@yRpZkW!&qjUY&z1fYXyXCe
z>46HzFX`>(UGX~QlPe&s;<ld)><<q92vfarrEgL|Cw`@Ize0RHS6)?a^G#R6%8&M*
zb(znVuIi+47yli{BoW8hx^Z~Yl}wN&<k7~vyzE~<6s1jiz>)P`aSERi5koj|Z6F-^
zpnvB5qsJe6JdSw`nsi-qGT{v=^f*2=nsn@PUs)3;D@%C<IO7R?uyO1*Kn6{|rh#Yi
zrJ)5#!<l}M_uK~6)^~fmEj012#4Yu0^sh>2-2{hGQ@nQhC<FJxOB-hL)8%Ty>9>2!
zUC+s%?Lb)DpJ&QlrD%s4KVduID?iJUfUu13C;Jt2!xC4e1uE<>gNnkn<nCvIop6`m
zPs7n0)|;lD+xR<abVbrD1pl9w*H43X+BOd7OMBCmd}w*Q(2Z?S-1jL@tWb<93GYoz
z5A^$`>{;o|(<8%v52BCpT=JI(W6<|CldX!!*Z$cS#R-O^Y7;|p95KtJ)ZK-<adUM&
z4t?lAJolwf#qq!L=@@+C5d>bH1$rT4jYumx1yIgC{hL7^6o)cpzmfNKC0d3xoYa4Y
zW}n~pzyG_Zw+N)M;1CMGHz6WXLU^I~!n+rJ!p@)yHX|i&N&_#AcbeHfpJ$()3%_)l
z>_J2!SX85DkVTQgz+ig-Mg@RkNasO(US4IZP2UzT8>++Q7%R00zt}1zi>)u-m+lpc
zREAufhABXHlu)o;s5^z1!`mGtovHp4SH|UC31)1!;nY2i+R=&EggJ?mb>kKgujpvw
z0^H%ri|s5GluC3djWt;0;D|BdG%X+w1Eq}(S+=<yWYdDH)F#e4JzHBgQ9k>uB%L^U
zCLTHan3SBQ_`wfe@sYo5gPM{-!M76VlsP?z&f4EwxVsR17r`&bi7#FuLfBPb0EUET
zNtM~geT)}YB-UJRI%^CH>YW*K_)K}|WOEBfu6R)Y97jr-;HhP=)hVG&l3#wv1?7cR
z3SO#aRg9zmhQuSB6bTyV^i+(Ek6G3@d1rR_tog*ClYEl|HO?<8i_GeHQ5q>33OWN7
z22{=jclnQkkKCF(5}RfX`ZgvoB+^JA9DZ?%75Na6LE7~zdR%TPqp6g_H+Y8Z<;guZ
z797qy68Qpb8=wirHs3kQ`tVhJ&N88h;F)wVIpBd&p5=&N&hG`N2&BOYjr?iL><Ho|
zuY4JjuST=w;JZKkkqx+yfBcj2$xnSc9y#-<p2s9NOYyz${UCny$}3(Sn^YWTWUib)
zAD7>LC$3$-6f5_ZrNH;c=<rAkvV};vXH-w4wMSa~x_#%4`i&*k;iOlcC=?SDSFc>r
zV1HW$gXc_o$SDkdvWiAM_+@r*tGpPh7}gkuR^r5uJyIN=e)idT`niw9?B0E`uCxnF
zt1<-EVtEw>o=FDPlRFfQUh!CUhb*kg@WFAy*771BPvpBWtl~8%8p*hz?p9Q%d{N$Y
zDju>L#+S$?!FG9AK+dQlEa0H-sTWoSvzy4iz!N*-2{<UXzl<vStd#SN+$s;kXpbTs
zcIin#8Cp>7s-T|RZLSS%8Nu04(JF5CbGpwU)10(d+LSd`Kggr<>86L~Y#Vrii#Vk1
zl~bV7^PFk34*^Qq!><=AJy7`Cenn5^SI-4$$Zdy8U)M_0e(vCvU*Sd$t54?{879mz
zn>3wt@&_m*emqy);#1{UP-*e4u#AUnWjc?4!qN$M9P$!RncOJ)loxRvyJ&01*UYC&
zZ|0HZOvrcz-85M)ctT!GFb-+#XX_{F>3?ZQ^lct{Y1<l`Y{a+#!qMOXAKy12UN4kS
zUFpvN#wK3i=HLU4KAxSK^>!<8rUAnw09&S%MtumtWGEkmWmOWxg6-QjILxP}Tyo@4
ze0Te7WRd3&g5nF~RPl#j`|B@_sh#V2C%<xp!RL&FOlNo-D$i<HZKynZ{ejB2;y`h>
zyUcqhX!{pk_?3U*1U?7{&<?9KJK+HE$Irfs;^)S{7tc=ZX{UI&w*ULXyQ}D+!U$k1
zu2(4JKd5|$#HxIcqA#Wf`(kEjASTs+jjL}TQ6i3g?Ptu8F5z||ZJ#lS&nJya7v?2w
zX+JVfq*p0FzO}=NPPbN<W6zmA@zfVS9;d$a*%<rqW8TuMK|3C>vMoKMz~yM}O{o4U
z3Ub?cW}K#cTotAB+~I|vvXATgt-pWP#ssSpI65vAOE9CvuWMkueDO-$yuRc?ZDwj*
z1NNvAOL?1`h~Z(BAPFO#2UwLOM#HK)1#i8OQ9);n0I`jL2I+wre+E+}Nac7?9mTHT
zI|mG?45ECA7(*jhCJ;akxOR-tSuht;^mjXD{n0q|r{W2}uBr@ZD`vPe#Y4_a(QPSR
z@<@i^DagUDQfGQ!T`x_f7Pzw}5EK#?E7r~o$g7G&Hijw#;7-;Jjc{F24CD`uDzxK}
zr?+@ufHAAA_~v7kNm-}ffqxnk8fv9v&q$Fv;MImV-#jnH>{@XAX}u?dmaw$CsCwa@
zhejcGy(CY}2bGPEozsjM6g?xP4E@0>uTIEt;G_-7QGVUKQO0Qukk-)Xle}b5@NP{t
zJiv8V{75GFQc4RvVH*jv?0ozTBZ8F|%dGN7;bm~*)Fk%!&&}=kYDFqu$zK~1YP@WT
zvZ4=eXpm#zLW$4!O{DONe!st<w9uGK-_QYR)DKE{ZqFh-4;@Hji43jmDnIJS3W2HM
zRH+T+kssq)R(+s|!iU=80zE99V=yoqPx`iOz@Cf;9bLr>#oR`l@^G1&uv}?K9;adh
z-6*f=2mK7NJabBvw>$|8qQQ3uROsY~%C6=aOZV2}<<EXTKK=6N;`p&s%4?V8>4uHe
zZ+z?9ruXb4kHlk-oKXjQKfeEsZ^X?@m*S0Ae-c-&UJgFOhta^w*o+490|yVp?mc^A
zSOYvfURYX;ch0{Pd!_6%YskyYtT4Rs#v68&9m-t?x<SS;;-~|+3-N%XN1+2*#Hb)1
zwKzUG5f6*D6Azt^K^Z{Hs_&b(=jE1Safx?-=)_qSVrGbXgcs_+sFV8~u_u{Du7w+A
zXh8iD`O_ZYHSGzyRQckuuF!|VLt6otbvEJBGxRB`_>a*+pN2!`mDc^8?v)0^hIZ!q
z5l+Ihk5OI8FXbwK-3uer+~Hwgzzi^cn89BKyS;Vd36fcvfAh>WqY1Nlg10=3PT|x=
z8wJjIvVTwf#M7m7&AiP&;bb+2y0!8sf5sou!dC!%?oR|<aV6)-Jy)K~wIONAYx}+(
zuN~I$*E8j9hy66Dcx^h|=9+RK11SrMpRnZ5`vE-WP17gvY#a*9F)qPWDU55UG)>+>
zm!{-fJ1yX9RCuOIo&w}$Sr*VgS$>Tt(-W^8Q2C_m=1}3j%8?au=v(P4*}e+(_Omg=
zv!Z4q-}$%c$BA2jkK^?~`8ht(<!C=_jilfA4hr#|l}sie7^~Q3NIwUUphXE9S?`Mz
zCr(;-!H{6}VtHi+f{YVr^TG0}d+<Pj`Qec5Q1OA1@s6L;w996CT3{X#r-j4$3GmOl
zOP)eM@rat&OpDxcZ<!#l1(|L;jcw@SPRMx0gY9W{!j1bN+y(9MUV3_Q?1T%yPTN9u
z^b8iR+u{pPr;YnBfjGrepq=No^eu2d*xYkNt<n^@vW-a^i?=;tRQTg|18Go9NiN4#
zXPo{tBAFUdKR>Mg&&C2Y$g!$DqVa~~&<Ca04M<NO5DpGULMOYUac=tfL_GER7vqs%
z`$EjT__UCrSL+|$LhTK@JLM=Mb<f4+Qe499(LkrH&9!j&Ue5Tq{)2z;m^oa=1Yjjl
zjhdM2tLt(7>h-wz-n=R%VrH6|KfapF)+EkpR);T!(C089(rKc|anvWL<?!My10c#9
zd#+KiFeGw|hP2*>K#T-lw8oI&Ej1J#UaG{%sQu^|6E-rWFpw%0T<&rf&@-$wBKc6j
z&Va1Hw#7x_WE=x^3LEKsPxkV6Y@#-C)1&*$HF0-4fwm)ojRE37kbwKdNsT4K49f^~
zOjaZ)1NZn>dS4r-@XjEsa#L4UOmyYrB+SzCQK@y#f!{z8l(O}J@(ub($eXwBSdnMt
ziI+jyy20wkI$u-PA9OG{`(z;9Gtg6y$OT(EcnO%g!(TPv@-jfeH|TZ)hKF+08ztq1
zY6c3IReo#*%8pz7D(1SpT(Zpe<=Om1KjVd0zGE;U+K_=hE9O2{R*E;P5k6sTY|OJw
z2M->M{rmS@DWU!u5TR{oWXQ8p90~w#pM<5dafIy(3cg&vXv~U>8X@ISy_gJ0OksEa
z41_MP=-L!5%viBii4~aI!b07EpDTRBc<@X%29=cH!cUla=saLRDiSWc^01eAUA#nQ
zQVwvgFKEHPiLVq5zI4<l#(CmP+2x)A)^$!%yy_%Hx=#k$;4Uw@`uz3P)fk<cj>kXr
zOnmVde>om{=&aiHMBKTx5U;)ZW_<md-;NtMZh58QsSiC7AA0<;m{4c-%J;q#cdlNI
z^KZWy7v6g}uHCrm?|aZEP=6Q|yim`32=J4-<mG+BnL*@4B>2EyX;vp}$Rg|FMYad=
zzUe4Y=5d}W(KBD^wk)cg7|FDq>FL=R<DCk%cTUh)Rv)*ziXq6mLE@|M=kgwoP8``I
zAGHnFI!?)jl$9HBBE!h1^R^)YvlU4+&`zi`=oMj3D{-+OUao7fLZqY6R*^Sh6)oiL
z07v5ov`*-<BV)!5PNzXOvfc}_?~rn&c1Of~ZsGbV(8lAo>VMK?x(pXSKk7l(j-RgN
zgWO?Qm#n5ASE9jD{`}yXVfrZ*s}(Q%q>}$~JQIct6R*I$YaHPP?fYJ+u)2~Kl5FS+
zdeb_*0m5)aLvQ~0!vmh_2eOVbU8U*Z&uk5pk%^MZtKwAIf%5C5hnM6{96gD94io3n
zzi}NPKa%Q-cN=fI#y>rMtvGF5T!~LPD_s-NiIZO4_}MRJJK!f*TXjn_lppc^>;(J>
z@YH=h*0VwP_2~S0h8EG3MqN%gz;lcVCQ@i?@RPp2=rkAs*Zw$m?3lM^VQ|nlF+M;$
zWi9BIzbgYSgutW}z8&CwVLa>K?b?sY({1z#crL%nzg*Ihcs<VU5D#d}QbC3NWl-_j
zxb^G%E^L>#%DbIrr~G=;0OeOWfOfpzdw4)B`|H2%x5cv~P&m`eeOWtRFMZq6x8mTs
zBB2|C<iLzSqchWeY<t=zxo72`_b8@@*}f#*Utz<FH!(CAQ!-w8&#+(dm{b}Nf>G&g
zBf>JM;klnxbhcD)a)RG-%&YB;9UO?KKKEig{u{p-`#<qq#MrQ9@PYIVj0q9jl!bN-
zak>}YDo$Nnoba#9Xy<=?6qI2uH`jl_cP12oB91Wu^O>P_14rQ#5jU>ijPvi@)$?Xd
zPt^yuc+vX46gy5lky6@x)M`i!U^bX<2VhKWN-1G~BHK@xF-9q<ucKyVX;6@#XJ}C7
z#XRz*3-p($b(O5EbD?pl@k;?@>k|Vb3L**ujT+06jsaTeEUYN^NjVZuGtXsD9J#WQ
z9T{xESTRGuDStO$TuB9Kb!i(f_*hUII%tp?3m^e;5)yafr&EIm@I2}14zBf#!KH@|
zEj1HV112;;J14%ZNg<^JW~Se>Ji6zcc~&4WDA=}v457TEsG+<vP@_EZ#ajk=zrd+_
z@>&0)j}F)6Rj$;rS01>+YsDiED@Yw)B}aTKe?Pzz;z>CxzA=0J0AsP?IZ8BgW*l8s
zJ51aW|54D?_)vCL@04F?y#0kOaKs@!FQJQFF*!9ULt@m5sgIuBvoH4Vn^T$C@*=uP
zOa=|FA=u048btWE0LQL!tS0X)Fax#Bi?1>c*ssfgV%kh-(g*$Cf+SoROUSgB=!Flw
z-rj<1?p=?{lbKiefjod$49`JLg@wnEMwv|_@u(9yWUP!f@T9!ZMjH8QWC=Q1wKA!y
zMW{xU!9zUwh=Jw&1zb_okO@{XP^9la;60NEF(PANOY|Z${o?t_GiT%HKJ{{(ee6lK
zwFyPt6?g9}#XE1m7jM3GUTJp4=)_nYI&>h8A3YQ^Q)6-a+Lc&cxEr@@3|xu}7vGD^
zmoLSgyLaQ#rOT$5^1}=2GnIbvSNQ@S^i?juT+a3ar67*$6W)k#*_UJ+@hYWoS(aUP
z)t7jN?32!Yig_k^;-2=eE0YQE7QVx;<Qr{5&*0J(#U7&pJ`xxH(|3|L40l^rx}3F#
zrlhMxPgyx0{|1Mp=Hxn*7fI<KNXs2%;<o&#Q6LS*q{DF`Tycnue;4%ZD|`T_?x2(W
zw#5USi}Oot+;s6}d9qD|tIFs6jXUGnnCm#zp?HL&zrip28QQh{Yy{_OA@4x38pn1v
zWR5-!+@v8K{FIk=LR%<sze(NGVF*m0X@e%EYH0b(pdGK$vYoVzr=alT$RF1R=q#vo
z(CmJo2`hho$-+<I<my}cSnAhL9e<!`D$g!yc*xlw`lt`z;nAeadw}Ppqd}k`q7XlQ
z^s8|GK$KzI0URY;#I5$?hz<QI6N$HI2WQC>kr>yYg|_NAByfB#nWjNv)NSbSQdi%~
zai6q^7<*QG%B{rCUwLPm9Gi*dF-buCrZ3vJf6n=Fq--54%Zsl`TQ|Mizw;p|!*DJ;
ze-$REd#)~P8>lcr(LtPcymtKdeH&-qXBzXj2MS-gHWVJ9eg8r6%D*?f3r!y-AWb_R
zuAN64+G&6`{%UV+zux-}Uhe&u{uPg3J^5ad^IBn~(d|yOwQ(0NApHpQlWyTL_ZcQ>
zl3uEDM4r3U7xYWF7?qwgIoKcL8UjZYZ$z2$mF!9B&7;Iu-SJ^hzU?$2y?d;GsAEiE
zXh`XZ)#Zh_tLv`a5l_7Qk$C1ee<=>X^rEbkF*==iAiadsq9`h|2z2x=KPR14mgH2%
zvsdG<rkapJPD^sf{;l6Sr)Vf``2sc-L?`H(KNSLn@Yc=SQWj*5udK_E7>+%=`I7FC
z8r6~%xCt2woTS4!`lB%>MS|r(mR*>AWIr*-D`H5nJ;$e)cxF304Tz@!1I&^m9?YIW
zd(O^j@W8mx6@|ll$P^1@h$48~O*1o1sm4SD?*I%Bj>I5)>t$4sj}<0$BFr3mJCjhN
z<hiV3keS!elS`Y*pTSyJGnjwEJP0H{Iwf$|;0BErCK8cA&W%{ll!f>?bDRnvytd-g
zC?|IOsUsmB9V;sWtOmG~Q(S<Kj1J8&No&AeSI3)5EUOc_bj*P4Syq%J@ydBL^s-l9
zlqqlWOu^?qKZch1GuZi^2hUJLZGyT4mlr{#c_uG}SG1LK>cXkc*z&_W5v&|w@bE4S
z+mz4&j8Pf8j=`gUaANeJyVFL{FIg<}8U2ze>RuW7ogAO2A;+=8V|vDrK-uVgt$1wQ
za=-_P*yqaWLl`snWE?LoYB0bMlZ^Tte#s*;lppY7FZ%=YcXf5mX*o>5a@fcQwADTJ
zMBO0Q7%cD~9$RsOmnxkNEP=mUpbjCS1~+2B^W-1Ar4DRRD#XeM(Vbs2z&6#9iWzMr
zpTU>%C%!8G=2rO9Ib0WH!EFFJY!VZaGsoOR@0Qx_fDE$TYD<IW<FOMb<FO~7jzdRJ
z#I7NBee#Y|e=IMp$)Mmn7`fuId~XHHV`zv~MD;7`Ec+!#ba?Z3ZpF<T*W%)(OY!#G
zZ^w-rHya+OqK~X&7(*4D$@U?l<iWcJ$+2WXQkfoZTkuWSEK;WB$}f>!SFSJQSIN{a
z>Z_o1u+)zt5!<-Ti^`B#LQMT3cknRV0dcDilK)DoxWXe(d;4l6zQS{5>pTJXg}Tx|
z;YZtJWr%n#J91$}plWc+-;*MO?30oQ0O7{0n^IhxRK`!34NoiHdh$PfAVCvX|0zr0
zAT9Ej@|f~2Sp=_>;7>lpOGrKltMWmvQudh;KxT=6QkTc=@_h~t2N)lpjOm#fr=Oo+
z2)=?}d)Vne>QDjl%59eLIP=1Rk1>X_crwi8a0Jr=FWpCFv6^x<KvexDF2$2qD)9<X
zx{8ZG_$qzM00wC0K{W6<z8tCA`6SOUWfPPn(};KOCmfagyc&qq0CEFB0rkP^o1wXI
zLv-jVpe=3(MQ+mh7d+-oH;u=B_$wf<<Xe`JYIeF-{6cB8s@vpu<1ai!HILLd;0m~T
zri~UK6KSGll7>gx65$yqBe)Iaji6{CgUlmc${AO1Fs3E4%oiE<_^p)4N~Q4xhCFNK
z(*}oV<j@GlCi*HS)3Wc!p}Rmx%9q=#p`kS|NrQSKXew}?WTt=Df0G9Piby_y{S>h)
zOKlm+xS6aQC-0#l<8pPL#+A=IT?;C1#Rq!h6nv0>LgCN;!a9a<a^Fsu*|ce^GA1nv
z_~F^elEb^asQ6uwN8<^v3lBKTHCYm~1DQt!DKF9y-_Hp~C6Dyu7yP<ICUnD}N}Hj^
zmw0&3&!ex#a|{YrCr0HlwyRHu1;&I72aTB=DLbKYhG|n?z8;f)Gc`IEQ-h;1_&_)|
z8e>Ab@7n5e++JRct*QPv_j6Cj^S|-MIQ5y2N9>*=cW>Q&z_^O6H*|J-1s>$r;fel|
zb_&y_oqIcnanXZr;VI?n`@4VVwBn|aG22UnLVjs%?vNF>ytEu|zQs{&_hNElF!t>m
zQztMG3kwS|yJuE|z*vlqjkv+_C0`BK>J0J)Qg;f%fg+fL3A4)#>e&d%*U!o;w<qMR
zfjgTd*`1L)B>m;-PB~x<4&=57W|C2+Y2=)F&laqHuS&2=k>4@!x%A9#)8JJ`f1J*W
zgcu?m#K8fS9i=}zIuzR!qU@kD4%h5FvJ$*>tt&3$;Lr9TDGG{1r<CPSJ}?vK^mH(C
zD67!ncL}Ue<q>(z&aKEOLrGh0c$(XK*qW3E20~G@{J0tf5ajAq>4@tz8q);*5>TAU
z6|I9Znd$VXk1C5n5hz0qIMf84A14r**Qyh*e7J)i@Hyf5!C&a*xV_uAZ=1HnM*~wv
z)B_fSpB25yu6S7d9h+1GnwD(&M6fZ*X@2YtLjomtdTQ3kdGZVn%8Crb2g-|La`*0?
zSX^9=#rb8=q;u>hN{Uxm#E(r3H<ViO-yKIohnb^)Nb?uCrR3!xBAFxrUMODiss@x)
zZnIn|4+-u2ouD@yyzsT^)OZ!A6HnoF&lLwHmA5?X$8#r*Aj8$p#3&RJ)8CW<Ae{G4
z9((+8zkGP=)I)LX_zBhRxQ&LaZ~8I5b+O?W$uYpF=bNhUJK{G-1|!Qpx^a{5B1r!3
zEh!Bj$jhU7*?Y_irdNQ3&}1|;{47v$t9(`vnlc0Mjs8N6a2f3^AJs-$Jm=bFQOCco
zv}f?wDlYfwpLNi|&vQ>9@T+UxmyC4M2uGEp{GhY)kw;z|@b<U{Xq#E5-8_h$zX_Cr
z87}ZRFO|av7SE)00^K)p#YOjlv<)1qkr;f)4CnhRaVKsYGT#KQgoDRlthYfxACr9-
zTV<1WRwiiA<397B_hkT%7DuLVHJ0Gq{)8MF6RN++{)}>4IVXg&Hene~D)AWpweVHk
z;=g$#2hb-Pe9e2say*=$Rq1tK@#WLR7RZlxd3vGpH?Edn(Sj==yc0{842PChoF;7n
z_r@((&lO(&?%$gv(9-4y*i+e6JJWbk5>oi;TD+Bq`;@|0;ehjQx~9y7ung~|32?cZ
zXB=_1I`J|t89AKT4$Et{`@)+%rr&AHLP6rs`)+?rf6{FKWb%MG@WJtr4JMAb-wqu<
z^xRh&Omp+x4lBT~?azJD3SEVx(p22$qBvb#9|V=Yoln~p4aKK+7|_kL$+QP5UR?{?
z`RxSl@N)R=l&+Vq;tg?&?`goKKTvpp@+;_l2HNSk@5RHljZ+cWP_f-lP?#N`8^xoI
z@Gc%N7YwWYaUE1dY2h)9@i4#(+0v=DHvIDS_|Q;HOaGaX8%B3RpTe+Jzlbh#S7YhO
zp0POn^rP|dU;cbN`SYKQnA@$+SL5pDz1WgYd0!6sU`+8gb>&~<iw#%xVd%2<(<4ST
zx@o;)fZCosppdw(jQ9P6zyGNE)I>@gib`S>KNJD{F(xiuysH6ZU7bBgaE=QZ`;<4l
zGQpRKM@EKhRPZHVwn&f`MUFV&K&fR$uPl8CslQZmH>{iv@?xzRi^4?nq2R2Jm;p+M
zLBI?!gE__rjT^d9WSEU79tH_7R&nGh%4ZJRQZgDIaL>vGs}}ywfoRJrmZD64R1(4<
zy6Hd&%Pj?p!czv>6_oIRi~Lx+C!Cc9<Dg8+moS#{Q+dVZj5afUTzLT)#fl?!*}Kn*
z3x)$TB5UGl8W8!yJhP<C*7DJkn%SizgeMqG*%35?qI5WB_}o#78CFW0{*(sCOSpcn
zax%amD7vPn<N{eya$cC75P!Mn_}CTE4c}}u%CRzl!HMi#yMEpKOko^@5nI(#4mjB;
zx2Jhq8U_~zpPtP_*PHTp8fJufuLl`w<Vg6CRU3{<L)mR~$xy%u!;nEPxMz=R8fx%a
zJfM!7Qo%1?swF)J5P1?nxovzlbSNVmZtk2AB>hw0*|@fY8&@L}J3-RkO;`8MN5FIW
z)xCMq;OyWg4o=S<BS_BSJ#h;kfS>D^MqzK|0Uuny$s728;^fIV^YFud{{$lrgY4SX
z>#A3H%IXd)JI$6mwLi<X<rL!@KIIHHa=RtJp&?d4p(j_~s3Q(($TX?MWG815I@f}x
zDBbY97~oIw2v_ATo&tEK<45=!e-N9;Hm<kx4M<!$Q011LoD8VEaKtJ8$)}TV;VF8!
z;!4Kd9vdKT;dT55Cy?#68)n5`VK~x26Ce@p<WWExsPN2`?pXx}m-`LXgC`aw%XMZy
zeJ*{4qF7FZ&HW7D&;Y+M8K3)9pt<5l8lh|G#5enYR;=l_pvyD7^iwK7vcXnZ#*d0K
zH#g_^o+c(H)GoJtZ0B2Vz2&|UpuE*T0PeTtkYR{W*2<@8<CMK|ekOuwa9zQthQHib
zoZ_F+wcB`IUEsPedCh01&Aa?hlm--y!c}<z-8AoqZSnlm#BIFulJ(mQRqn*!m9ql-
z%WaF-IKck;;}M^fg{s#~4-n5+|8mXoh47Nm1jSb*;!}5oXWmV)!h7@2)$1lh=a6_i
zuH(-ouO8FownNr;H!mRRU{#AOks-#Ul${Ji9(CummN6m<lAm2V3wrNMMmzL2>D$od
zPr^<(Nar!Cfp#3iiqEu*c343>-=d@9Rajl|FIUh`-wQuWen2n24@y_@cS=`)U&ZZB
zvps&+JAi+sE8w}8{{@As(zMe76;?nvb!8voZjasq;Uz9k_d`m<%hY~JTUVY(BsN1b
zCe&lbE}S|N{KutZ^PxMWvQK^VeJ0Y?hwjqYIy^HLC(fOYm%jA*_{bMujyNb|BIq+E
z8@z-q{fBnpaYuQ1g16{bXU;8Uz_4&%+R#sPu+GHG-_kp2T3Z~vq_~bJ==+EN>0>tZ
z<!6QxAe6$J8E}Lr+d{5gyXje5ZzYi<J__9MD2g5D!4G<MVqhSR43_DwK!}QLka=D_
zKn`S2$He~pQgTy~LorlgQT`RqtTFEZu#JRD<9!M8_Mp@>e7`3khSJczO+{Buq2Wbb
z_R!~kdoe)OMi^y7A>Ab<l$mwHP)M`$K{#r_Hh+CjNJmf{uiLpOg$G5zPm03|1aWi4
zAdMl)43Ae?Ns-c<FX}RQr^Aro2wp2#DhFwkM~Nr59f?n3Bg^^TG9yw@+*obcM3EEk
z;Rl()ZPP9~Yz#r5l}D%~EFF<m^NI)Ua;g&yHSkxt;4|BOa?4M|M02Dp<Y2_{vNGGz
zYzU|hSwVs?%qA`_E!nU{1~4S@#ct{sn)C-h8=mv02lLjeHM;UqFwYPw4*Y{x)VEJg
za_5>eNC>;>q=Di5EWRjC$~9%D%;Je~VRX8ltBw@kyb-@Tg9%EO==6Ie;*A+*o;4tR
z9|V*E`~^JsLY6ZJf~qUi)QVg8_!qtoUefcNPaS{Kv=z^Cg~Rdqcl0JNUFtaB;t8iI
z9gt$j3J|h-_bx9W-gSM~HxPtY^d~OxQU-7F!pTpy=kn2^4XUpg97GAstp?1f@Qn%c
zUx-T1!D(Z}*CY=xzSeW=)=>O&`5LI_Zn|z*n|Z~b!cl%*AUyLZW1=@d%849#GNcs_
zwCe$Ahqa;n_$4Mdorbh6f1vWVoHlJ0=*<ftUKtb4yZjV~C)&!r)fI2|^DHdR2}J`=
zx9%r~cI7JJ$gun4T=kW)ulz@36yOF$j}SENFpYdVD`c{^i?}~gIA0uQyx@H|z7S78
zhEc)c35O0Hc0Yde=B;?`wbv{!0Q8qJ0iE!O_MW_I_|*Vq1j^5{$Su!uE@u<gj@x*3
zc+`eGt*Gs%dp{=6n%gc2t!j$P`*3wY`DE_flNPk;;J2wA|5m!9A5d)PNIT5(Y{y!D
z<zM*Q@ls9_6aI;-$-jx#<{7|~qPYzKSvU043B(7FjSXXJ{2J1`G~kyma|AL{#c8t4
z&_ZC~zWba(;L0{mTIk{)*<}(7TseW7{SNUNBbYoO9%Hy=rvZ-<^3PoK+ho_x6{vFY
zlf;_tyzw)DPmVbW-Ew#D+k7lP>IWbUN86!1?X(5Ov$_Y`;XvCT*G>ae{I(zW75{^v
z;_sBM7kcCLLOV?_-rlS3d|yd=(|!=mgcT0rRJ^MHO4|-AppWb1+a+V0Ty$_N0Ds=2
zAXR?oE)56f+{V<8asrgMGfMuzH>}8`(#`k+HYYuCge;;nJ}Di1YAlYOJ|3U`;^*S$
ze(`fLcARfN@$R8y`o3ga^36mMIoGNyGOj}CE}zlId8}d;FD#y-i?KTuW6>j`%@g_e
z{o{Z1q>qLLPco_o8fK{^oFa62dDWdcGrA~)DEuJz0HT3Ix|UZY#>Pi=<*a;;x=cd?
z#Nb2uXMnQM>uQCyQHaUjN;wLp{!;N);dN!^9wnPCD1c|CrKt0vI|eESBWNKHwhy78
zyBsR$fbzhKB7?zQ40^c=0a;dpU?!H9!mRv@sVI)*ku&;qCK|L%m%z+Nc&K(7nv!dj
z<=OF-4;3V~WgI=D_+sFoaC2o)<>g)!Gb?T2)sGb`g)@`SI}NoB00qG>0P2qyp((34
zL-uw&`A=o8K@_?=^(`9UubuGaKK|yJKOT%-X&pyt;iY($+p&@l9VOj75I^9F3@{SG
zZr7QvDD`~Yu+GchROeQP5%1a9fDnFM)d47>Kb3zBfx4$$O}LyDP4Nw$cEaVL51L&c
zjl-{aLHW9TKnZD-JpDN)X&jCCzUzS7hW^bPVkHvMkEbFu@q0UzUOYh3leE-5F6%Du
z+q11nbK)he4ec;rn{fYfFF%)2*UAGZzl1JM>YaGW|CBQZR^C_f=_jHKUaxUV7&9SK
z2;rX%an&i9$OF_i)YSHp&+-t1t1s~0d8eVRutiFSb<3Er94Kz_B(EmX(7p$H9$hwE
zn}^`(mbtD#`#zx?FZoe`cj8VSQcqP^l^^l#L~|JvX;3vdh0ZfR4c<xzABz-+fg>LU
z3b0q#2Pm9r>^5~;`7}`U)IE?fnsVq$888OKXusUfqz$VC9OF&ht~V>6ThtrMz7+d`
zQC1Pg{S6RSfQQsq3=gr&t1&=nIewEtj?>ZZt*puby&H=Qi;|l(5_xY58RB4r6_)4c
z?*>~rC4KI%p%tS6Iu0B-;4M$cA|Fp>rLkl=%aJ9*uVlsy>wyk0d{s*SC@t3}d}rwL
zwBi)2Go;cbPn{2sb(LqvYiPy4ColZ*<68F>r-PleEx?}~Di+Z*O45;c&9%X4sA~i5
zyb9{slq;>%+ux76?_kc~CPSx_<0<98>tA(7+;U)YR&>Bk7|^Dt{0cg8gpU5g_swsE
z#lLmkDSi57xw+4_ER{_(BA1yzrO&v8bH%?LP*)&v?hJ(WhPTrKHFg^mwUZ`E5p?se
zcx{~JPZ>J>!46Pq+c-WbF8<v%l~8fpe%$YjW4j;mGTjH40VqGN-85Zj)6g3Lhd_A0
z29#f3%O&)t<+)rNSA|y`c)F8+xoyza4Tz^d8`eh8Rq6c)=vQ8ZyOj~&PP&wyM>V!$
zV4whjB6sIVpH&`gS%oSzG~TxxmhL?vW8%oMgYm-0Ux<&t{E0Yv?sN=IPbiXbpg-PU
zmrG-UDWuF^h=reVi*n(~5&sf=)8=%{8e7-OF)b6y>7Ep(@BjRtJ!50SgS|Ywoy5v?
zzYU3dD||0vHSXMEHgZD<lr^gdjq=Ra6VmcwvGED^v=2!!V3hzxISDhf<q%AFDtbgY
z$(I>(B?6#9Qx*meoQ)ZjQ7P99E<6*50c>J?!c5?WP5hyQ6%&+BkT6@H?=diwz1}fE
zP?A#tPD2T0)mv<sNiSo<UkXko8bgBFw_&zlQN6-V5pcG~NP}gk`1sKo-DB&A2TQhb
zz`s0hlobc|8sln)0Y2j(n=*53F3K`Ia-~r>x|o-$JpB9&9h6fQ?YR8uN(XLbPB<-;
z^2^i`5Iq>B{w%z9go96l<W8OXQMk(EP87w@nx=S0uyNDz%1Z`=m3@(rh8Z0%^@&Wt
zk2I#JJ7yut&o}_xi#&&umhha#>>3f-!4fy+!^VUaUCJri^v8WR03ZvKe!lgUGN!?o
z*OIqNmwf4V!1!~kU*G%Bw15m8{Ny!hXfjj0E+j6%;Z0hi5X*Q~9_p~dkqHd+8oY3p
zjf&Gi-IwgN=`Pm};~9A`d=<C+94DzV4b%s6g#5A{#_<%Y)D&93X}EC@z0?J5u=ccr
z2RX%XK-pC^`{BsH`VR6$=JTcNT&gww#N^u0SU9<qGZc&QYi>;^b*cP2Wm5vK+o7Jf
z`&T^j1u|{6E5eGeseoiX6`vE55+3oxw;PtBE)R~h#HsZ1bD5hf&|t_r;+At@-seN&
z^|PSj6&SDUTle}S#2@OE8&QAnfkbsbJvkkZoOvXkfBpq2^TSs5SLl-@|Ejpyr9O6@
znJfAQ)ua0wA?4UYk3YiA*S-0;->N_Ex9BZO3o<B}NTFRbA>j!DU03DDcSTlKmn1{^
zei=ZY0sYg{(|#Wc;CN#9CGID(@9r&gUehTsYlEv?^B|Y}tN6w1g7$O9+1|!0ZH8@+
zTk*L1PEnfNcp{(^r}A&7E5AzL>t9f~yK&m_pv{k+^aX{bt`4m@N>u&{dQuRfP?NEU
zQTVsVE^rL})U}1)KkZ!Fv~L5a_m6>S#+gXmN7<O@go7i)fHuE+?}^K`?EqxTW3O=d
zuL-L@6CiEzw5|n(r~K==@YFrfo3@R&9Uo{Up&k@ZZI!=PngSkak3!HS=?QPH0BM~|
zlUFxK0W|2+p8NoIqThHr{=Re|vE*ISOuyO_sRf0rT;;>H?%On!UoUk1D^8b2pwjM?
z7f|69r!50DCiVJvK>mtgXMI~`cR%0o!a=$=wMYb~=}^~7>#a<TMMPs&e?a4ouF|BG
zH+}9d_21Ho)i0y;&rFWT-n}#N@WZF#si&TZ$DTYFb0?3*$hh=ge_)#Tfz^+z&yQVl
zCPertQ}w5!XR~=Hw|2nbCyy)}4-ng=!;fv^egEV?dm3~`3UgRN;0%5gZ_0G%&b&{;
zSy)(#6&V)wMPvq=d|`w40ys%2Vq#)UZdB>;mougLo3q;ZWj-**C*D9TGg&ZG6~h4G
z_Xwn17?haVOCQ<EQn{Yla>3D`e0YWx3ugUuCLKjAm-t{rZrMm<gvw+qmX#lKh7Kvq
zk}rPCPq@WO8#=jt!%s0bJRY-?yJb{RVf~`*zW0fzg3N%*UgiE+y~h`%({NZ^TK1I}
z2<tlt&_ai`DSp|~g9;Q02?abtNa-LEvJ8Zt>NKT@m2&0*7sVWZ%AyNif;1*lFrlAH
zG`}4o@p6Ya=IVmIvaCjhK+#yZSD-xEwgg{u3!0patD(Vs*_rJw7@Rfh2p{>79LB_b
z^D1YbYr9dcB!QcYWkT0f44l2gz+>9@NrDX+8pO*IW*wP!#RCUGc;=CiiXiy73b%M(
z;dRBYItK7KZ$aU=gE+{s+m5R`7+i%Flww-`z3~g!f`T#6>_eq*;{n=f2uo$YYj6}^
zuV|50|D8@k&+Yti-7+VQb<LGTDUG@(M%G8tMH=&(d>X@GQ?gB)&66xLzB?^s6(p7`
zU3~8!<07}N5uLMwDCu1JBa!bLnJhU`QUmwxE1*)kjg*pJW#0}de+2|nzJkn?2<b>)
zWiLYl$8&`xj=Z)3zk<3?+!a3^Q1u6-e;a?r<C@_amHNc7Je_e&r|PEqiZonldwHdP
zFeZ?>A+IEEsh=B-GpEnQ7e4<B@teQ-H~ig?<)ziQJ--k)WenV2n2$S)3vuPfO`jyk
zSIJj(^}z*l5AJhi!ek}x$&k3m!4bNjpT8?Sd=M;+2}Fl-()ZH8ur-sf=u-}#_@_Ps
z0|G}mz3sDE*+oVO#~FvhZPc{u0hY96r*mzf6OY6_K)iCmPH_N>bx&BOaeDrEBp>^B
zf7`grm0uh1b005wUtV%GZa6==iqnmoX}h?|YdbFPdFrI~FJRw>hLlN8J_F!2P-!~Y
zJ9tG)b1fXs$DU1y{OFgHw1!tLAZ{05LgGk#<;uUCKI0VZbf0pQ@o?Z_T=qnY@MK>{
z+-{tLcD_J8_r|Go<=^|<3-53D8L#jc;Lmf`N%rBofz(Xr-g-MNZD{B3v<<ZVw}bL`
zJ#NR-3!Q9qZD(0{^ZJ+L>g55hc+guGpx3|hE5E$HznnXzX#?TKAHqxK@h^veZ(5#>
zn{ZvW8520g0jr~MuLy2da@%lKS^|kIpuIB@%B~)^>v7MjGTTaROsM~#9v_W~(c#!V
zGa1K@9E@{kAC7a6JRJKE?v3GzffyW>@u4^)6O+*=y*3FFMdXds*uZ4a?aP;>^WTxq
zJP?!Whxxu#zN$br53I{^-0YV6eRyMC`v34-=eBI&s1t`U&sHgrlUz6%iS147P3Hwm
zPH|zT9mSD?bIc|RJ1<Z&E=*5P#q8{~lzMQ$)XWq`11iKer$f>4Ky2kWml27d+gKPZ
z*#9r9NXjdUJ_;nW`Q%N5=cQ`CWx+~Bo_Mrkn!)3DDX1ta!vetUIA6e3KFR}r0VrWU
zD%Q#y#zZQk&_G&mSAt>pw^UhUv1fX}Dr{Ww23+ptMPBTM7j!}hUc~&}1%KaxcOsz4
zUpQV~wop=LmhdSvd`VH_w|h<FKs1^cX?&?VTn_V1{vO4MuehXSVCE*1rm#{*;x<Kf
zIKZ>x>Cehur_)S)@XrdB@lht_VZ6jcF`=H|8%EiX>Akmtp(j4@a*yf>$5FMssEJV_
z$u}<Q2R``y1>+R$%Gawpf}#W3__HE}pBq~$4e)`q$VVxc&_-P95oIGGaV01czlw}Y
zK2na%7n*Vq$?yW`Et(Pt%k%}%Lw$my9O319p>S4SMNivN2j)NcR7UwaZN()JGV9rG
zh&mIHe+MslAuEdOR3s9RW+!@+r<oQyX-7bHf+=>11Ku_09G5iN#_?<d1WG1IocWwa
zNGV99WeXo^X~)!`X;Nl{d9p%rjYa^TDkC(3rerk3^Xhy}ck;Vj8;Yi_4ss>ouSsK%
zhW3gJKo@m{D?`q0O-V?{GuL(+pz`Nh_dxr;4T(4FwbD3dbDxs~U4VyK2i>$q3s8JR
z)+Jn>dS(Bt-;Wcyvr{wi<l|4qFZ{w6<BPxarRY=LtSzp^yXP;&Pu_SV-hAhsxN!M$
zT)ccWZp*k>kdeWb%@r9A91~7I%U8o8n|CILG!C#jwJCp&Va|79kgtZ1^u5hQk`Jm7
z5xgJ4B+;`U{&4J{n~TfJ@AcPT_ckfN7pC!{$^tK{F5CuOR?4C}!>HA@6So2LvE>Tt
zp13=u;Th+6{QJV+_HZ2X&0klzx={5-nznOVhh!K~_;vz#<?*eua5QlFX&V%d$}8`E
z@pH*SC#mRe(77zd55L1ED7nHBZzufnY%W;_{7F{;PM$06PLQ~Y2Z;m6bGCc@yLnc4
z-R}hWyA3q=0M9iB;8!vQ?sgp(9li*NOF0S(U&hO4T;Xk(zZVKm;V9^pMWAq&|4#Y!
zhP9#c*eUFNfJWP?gG|$=G4JJ&8HN4S{Pf2!ZYLV*p0EmU$7_cJ<<}1Dh4fF?#*2SD
zk9Iv46yB0|pj{tTpY>dEn+G|?(N($HtJ;3XPC&ePA59rJkG$2dBJt$D8BaZvxARw>
zSYOu|o$mvyuU8a4Fgr0a5+mw&$7BS|%Ah!UXfBQ&J{%_>IvNL$?T^{LQ!yp~>6zIW
zl)ldA0vHdHDzyVn483(DUVZJy@!o~^)c)_szWw{+=+R@|a?04x$&K?13*M4XJLHgz
zv9XafCLoMid(R@O@K_!&1tpCcA+|Mv1b(9^FD~S)``!8ZSXx<^a{WL9K!5DrJEn~H
z<V(*|oO!8^m+MwmSJYTjS;e|xu;5*b&UTaJoS2;qL#ZHN6nU7%KtbI=IKW<795ik=
ziU?4HdkTgIGBMjrG<sst;(T8!oNyEXj0u!qGS68&7U>yq*>aPL@#Y2<C|>l5=7C|;
zu)2KDr!B#U72SK4LJBl17T_|yx?>h?RfDLvjZopxCVyt>DXfJD;_UPTeTsn+MSKqy
z2xEom;(-cg2&@q6mPc<>Xi$+)TiLKs@@GE;k9&T7=X9`)FJ|ZLQ>KOPG!W9rGW#(Q
z6vy9-5RY6ZO&Sz@zsKS<La6*)27tUxpmJq(L%gRRsK2DE1x=a#r{oUCsL%!3ga_0Y
z<%AbGOPTNRkgn>OnOZV6F5~M0bp#H`hu@Ch<Xd>TFUVUc;J)$K)e5hm>ZFrb;|kh-
z#N8=A?Eyy^gBxY7JWUTJ(0`>T4YXExx~zZKm4G+}j0+VH;7YlPLmTSSk%K8=tdan<
zO{MUC#UmWx3cc{ZjT=xjep4nJzIxzxnDSEl#E-V(S!acj4}ab&se*ulxRpL}!^h0K
zpc@7#6*At?T2S}dTDoA~%8Pd4XGd$}OZ-(1%1URReNN_4SD^S>&uu*A50Amsh6<~6
zhE~4iDjZ(YhKhf^X)E9MHF>IM{&|+1(SJ;gjmMEghvSjcXX4a}hm_AiEZkj)x6i*D
zZ@+Usu3W#N{(i+r`0_4=-#c*I(IC{UdVrKaR;HbYNI~vO!KSoy`}9v#or3I`RN}iP
zQ!_Jh?(Erk^2sMXzWne9KZtkVeb0R6>-4pij(!E%gojmL@bFW`Cth1ViZ>O%lMZ~k
zueu;jC!PFFxy!1@^zGou{;6rxN|Q9?vv2)tP8jN)HK1LO6~7JkQSkdi3Y@v`nv3JN
zZhtCt%HTBmQ%0248K!6QqM<F@MyBaN=e~PwlB!JNhtd`=yZL3@o$3a_zs--juQ)u{
zWD-#IR_WWg3JOQ2-HzL3Y|&W#U*YIYlc2CxJy)LXxb5^6-iG(%-0s)rLmT%_xayf}
zyF8V?jTa~!g|8Pn@evB7Vu$oy(0z*+1@C{ReUM)_eZ|=ce>;x~FMpofY1{V&nZCl(
zzhtEig|8QSaq(P^=XQUQanr9e_EOd+UoLH(43-H=h2|rFy5U*|eok&HmK^z5FVG|M
zX=6hAw0ANnJtvcn$cPwHU&>3`<0Hc{GdUKM<D+rx#GyDWV`A_A+1R~jCid;!8~YCI
zw-LY=RcTQgKkmhy+c(wcem5>$x)4{cUW%pVCF$D3v2Wl0IC0{n((V;q=-3$G?EJ}Z
zlQHSv_h0^BUfj~XKt!W_iP9pzxu|q74domKe!$Axz18)YUtEZLtPI@aBWU+!p^U_g
zteuI8@tB#JG7%hIy129y3k!=4lat}Dh+?j8c}~Se#oFMgNwx*$ay-oK&>~=~O%CFv
zoTNd~777Mfz6)T3MSd6`KGLd*15Rp^J-F=s&uvoqoej!o1zJ~T+3&5gD#IyBTV9n|
zlV!&C9?!<h*g*f$SgI=nsLF?uAUxh`B|2a>iob<JnX28Hr?pHC8{BzO5&ExH2@9#h
zQF<|kC@=mNC|!Z%O_tech1a~)wIB@wW+S01>ByI#VGaC~px$S88s%6#l|KUrX)p{>
zYFVj)PbS{DUDA9Om<0Wp7C9^O>WH$8DN9~yQS|DlDTC#p0n*k01XjXw086|m5L`)D
z<peL`(BnA5TXhUBj12NDfEvf~y`Gvt@U)Eyu8BA4N=T)y^3ad#D&z8)SD@{ev=oju
zz69c=gEz?!%9!PBaQ39fRl0h{L0=mFdRE+RapdRK4h0lmWQltW_!>m3jD*o9v!1%0
zVbveum2deM-9VRDStep-e8p*a>y8lQ0`R5mW0<eI?I3%?V7`DK$o!I)ChG=omln^W
zHuSck%H4>q0sN-3!Ku*Z+J#ETm2%Pl;OJM9XW;0@E4~sJD84{n`4@ixu5}Nz%R|4M
z_nAH+aqvuAOZ+M8;K_99sywVN3Pt84udF9NE}1W0j|}H>`PlGCOifJ2`aSh?^zW<p
zyj9a*+jo61DWdWzEe;&~X07La`EqRoB{YpZD3C6XkDBt%+vxaM?4R2oCr_R<jaRQ;
zjZ2p<c?OJC3|6Q8jU4^8<+ONP(B@+Se!c$5)1;8FlIyA`hv~V>!E?49oa)av{I~Iv
z&PB5$G_Q`DG|PwkHa$g$^YxwXTWH@;e_|=W3NJiBMdy;<{A9U$)441P&isj|XS^DG
z84p?Bo(GVn0W!>ORTW$HLOX8bfd@d<S^04VXggI01(|=hyzQ{U2lU1Rh+7W472fu5
z#{*yWHSKuyoS>^JsLIf$zg@0^_B~Me+JIN)m1Qp;khWaGPU#EUVRg^7Hw`qG-v`oX
zV*=PIp6(kVtunWvaJ0kv`n%-;imo;t?fYK9ujsBmw&K^dAk!D_F3xtEf|8-OKli=y
z>Y4aON7YmM?^4g0VCv<q-27~P2ww-_+r?dMrgvb{+s1_cn502(B5F-mkrjPCd2u)j
zMg{)#rJk%{JlI?l&igSok_N_+!*elra5iRkPsF4Q40QCVsR@l=Z1WfHU6hrNAuq<w
zo7dv%^(%4X#&z$m;T_oByJzF%sfWBmF*Y{tR~y#ut-5XdU0{tNegB1w2}T&o=NERl
z`nO@6{yFELTW0z(5Kxvk)}+7-;mX>o5yRM#kx`%UVGC(=)PpSwDW~~B6kC7D98lT)
zL&t-{S)ZoIQ8}K0mSfh;FXqbG4<@Qp%UQu(;(^XuLI*#_04pMkOMGj9m$q4o;cLld
zgfKHcsEXUAaFiBiU@<20m{{^+&;%E=eXNr3K7?eqgP)a%a(ro;K{^L>gguqAf~@Ei
z9>#WN=4oCw9vlHhOmS2els$6+A)>r`ro5CDJXW~e*-|m8DGOM3@_<iKG6GgD<R@pk
zW%O*xP*5(8V`ItuB#oRCHG--i*Maiz%sW91ev7;E>nMl{D>(s(C9?E`M>a+~o|=2Y
zDGPv+<m3*^M(H?EDw(2uZmbP&NlZNFr2CRpKPwmFlwab@xS1U|6R!gBFpdJeoWA3l
z06~}b3^NX0QG5aX3flKTbvA7Wi5ag&o31$XiZpn)fpj~7@ihQl$R9Ea2s;_+u3M)(
znSOy+$&?<MLS|`C$eXCNA3%H*{wh~(t*TX-0>Z7pHuUi9H2P03>aH6O5binwC$#HI
z65f$oURjc&%c^5;!>l&dPH*HL0C$Ds--J~Z!W>7u&|{V-Y4jgg`!)F6*$6T$jixdr
z*lNl4P|8qw!)JdvUf>i;hrhS3iKoN^6?t}`!3~hD=`6Y7S_T6=clx3=^TZX9)|m*Z
zTmW7eS$_7tbMeb_(B(dgNa`0b^62|4yH1_JeM(TKZ9_e^ujK$!p&w-2D+}TyCqRvI
zVwlDizE{Nf!8dA_mzKPg&M~#s_PrHKv>A<_anK8m<!AaUtjW2<1N+;{L{<46!@((g
z#?gX$!k^@Y&-cbzNZXKp+cC+PcGTQ=Zv?8Bq=);wW?KEW;#GdMKcI~tf8%KA@6c9w
z`SpK074<<`8h-KL<cHt3youA`7UI0Npn}|<8(Ol@d4E3vunNhY>1_A{ib}Js9C@?-
zzQW|H-1Xe+U$|<O1?c#<!rN)9k8G#MU%$R@_kreN6Q4eo=#ENUuHfLY^J7oPtLKcL
ze&m<qGktKE9(Y^4Nt#HcaFRBVV_p~9^w$;0vLsZQ3p(+c_R&2-6`uHb=1D(#X!F(3
z#L+K5j@HGW_2j!I-3QUt8&24EI=W>8Dr`Ic9(oOR&AS?R9KXSt_mrPli6hHF-^Ey!
zx(oHeSjoNVqk}UvY4pP%kaCjYZMwxk$DtJv2iY<`3i6n$KX6n35aCHS#X&Db7j}Hb
z@r1mM2`1y&b}IQ%`aOHbWe|+T^z?{y;n5f!OMQ6PpwbL5*)E-04qbUuhRB-q`o+ch
zxI4e-3G^u$Ag4|p_i^1m=twxxi<g#{y@Jem-S<!bkIzZ=a)#KS`Q|DnK{);}sS3~4
z9}d$X3ZO`Nz}KJyu#cV(vbix)H1|rGmN>kOY<!J(zC=P^Awy$1_-Zz%DXg!tIw4QB
z37HvUFasSJld-_C6o1bOY9O~i^s3?+0C~S7w_)+#0v|*Q4N?sL9AC&0u^2>D0PjCw
z5P%b91wUq2Q(0GYc83Sf_2!fZ>YqkVIs5@M!q)R-5}Gs>oQ<Axse*Nnzr{iP)W3z0
z#)m+<@RW^<`0I|>{8oI$DPBs9@N#&!vl3pp?8HcmkkJVdPdIE?sKdK2f02(rBL!QT
zY8#jP9nlH&^F^hm^T@ojQ*r}o9Qo9>;^7K1ER7?eI=+H>Ha_UiXq`CixYe1<)Adf8
z8t4*=mP}X5fPC@EbaLP!Ue*hBB0T0{;%viqe3^zT&!t=grcF_ymM~la@!NRHRoU$n
z+>XaDCnu!=^nP9dJXidJ!Uqs$x`Zvi0g>_BX)<p5bB{5>l@5%yXyZO(w(37|>c-D0
zp9yW;bzk%pbe|QjTOKPeSJ3BH%iO+*qE3IXxVYe157XDwL($ZRHVv6qmO0yCH>_(U
z%OUs<OXpvn<^05dLw3<mI8mS_72)j2y~GPo6b-zY2ZIZBV3bX{mA_=jWoU5M4Y+-$
zI3?rtYy(7TNRLZ8nn(?_X>8v^D-ORJ@Y=w0VM?PVoqZKpWza+Z>j6G?t^{o-g^Zg^
zUMr627N4Mp`p(HIRwbx&X2^2289IGxlgAatB+5kpm8(`DG-`gJqcJe5oaE=eRQE-L
z^KRVF3OmK^;DZL=wftJ~JNGR=bvgEjK7ca0Fuqp(H;?+|nsu2zTuEPbY9Gp*h&oVh
zq~ZYWIKA;rQ_HVt@06hh6|WtJU*X_t*ZA)gzwUEOWB0k@<EjimZ`zJ)o`H%_+IF1E
z-_VL%{*|B8=^382)3@XF`jxzuzxl0ej?duFxNT_T+Wy=j&j1eSOj^?q{J5eY2PpT>
zZ~AZNUAw$($b7qH;@TTeRCMrFyc!4VxfcpY8&~DC6CJ=$i_7(=rQ0cO!TZxGOkR%T
zXvsIWS$3eo>3-1tUkiuQQ3rX#Q5qBOm*nVgfb6fhZ}7Czldt7OP<-vgY2w-7Xv&b2
z2HY!-{;GV$hjMgQzc(RTh=UHnP=vmW(#Q9`=u1b3`=kT!is`B0n3^8-E}}6R6cc(L
z9^Iurm=$6^sEj^M{XWpxyAn5V-n6m6v9m`GABc&`are`#maMF<DQqF`-sN)!8`1Yq
z{&5-;016h$M1B>TV@~q~4bFYHlA&_J#If-)&nB_bz_EUu^u){2@Pa*D%=D8Pv;53B
zVLf3?Y;NZM__6}I&vV5eNcmr~^02&gFIJbggij1nu|`$EsmXzuQsGBOM$BAT>r;iA
z!N(4xL|5|g*P`#qHA=pM9F}oY0C+!u6@`s;8Z8*1DQ7-W1d>kOD>AdZKEhNvQ;7%;
zXQY=9<qQIyOBxWad~1u~<hcg(>>Q|cqSz3k=6QCh)WH&7I0QBZ<W+fC{uG@u;AA@p
zz~A^0myT!?0mV;va2Oi$PrP}b&w6rxT=Q88{9B!nc_H3$C9MU&xwaiZ9HQA!60~t}
z&GIMUr_cuP`=R{ZxO|sIA#eY-(tY5)95MqvmJx*^Z_wfL>p!D+8!$k<@?7#&e!XQi
zdmF#Df8i>>tTV;o4?ognhym8~vMRovG=jQs`?YbGE6BVtAk)9%a;<VW4{;tIP$mH7
z7$x4@jvOWHsoG07hwb-GRTv6?(bIb+a@J%wk?B%LPCw8#Cir4G{R=Pd7LDMkI&isk
z1$9%zE>Pd1#r=cIl@kyuG3lE!yZ)OpnmSeFb~IUs+i;VY+AH!%WP*yXrl)yMC|Oq`
zh~&=lmj2y%?+1t3=|M?cewQ`zRGxJY5Qb~h0&tHbjC?Cz8`|*|$rlEvPF$`IgNNc9
zlMuS{F-a@zgz+Osyvh?^@n_&!@?COfS))DDE-6bx61=5dGD$-o$a0m{dvukzXTC+B
zSA!IHQ++-(RyjMRQ`T8F=t+;Ht?R!XC~sa9-k+|UMsdhX>@ubb<KJ=9H30^&g1`8R
zgWdQUx={@k2PiZxbmmKi*R>5<#)PW-Ucj#%ukL%VZ6Hm#0^)UC<JX&Cg%?z4#p7CZ
zX4`px{=Ip$p<Tvyx=tGW<T~->9J@&a^tP)ulwZ+Tak=6Of6)Lm**FYGSh+6LeO@a(
z>C9(g*p8zKGe8#i_~Z14m+29-`I3<S1x+-AV^v%Nj4#ty-qjzs>8ZThP<d87U^^W@
z8|;(@zqSJ^kDX}gg*Ht<_g=X8r(uJ)!CgQ+#*{QB(9Mz`*3T*~X_J#3+6uPQmb9lH
zCfCIE_(FX*zm<lXKvnDNYcgsy#-U$Zw~-$<5y!{+2-$un<vBEISwgTWz5aoWfx(!a
z8TWRlnc49;a%5lZ-Zv#fLJ9jeg@?N4Bcsb+D7ba=hQAX832b%p2TE0Tb{p~P!s7Bu
ztgbweuC3(%r+@U^7J~x}@dw{@kK)CkvAm+lQt(+`9v+beG(0GaW-NB^osH?~$+&ap
zPF%ZsL)OcJ27Wq!l#(nsnX}c2(>bu7@}!pgA~uO(m4M|E@WCvMi6t2(>$0K{eg-a#
zn#u9W*t2_2?AyIJW~OFhWJm^u1`_abBq5c-5r%v*m3-Ll#L=>xA~fi@+2L~H5c_^P
zvh~)jyZ)m0!s3eJ(-0V#QqhD)zLVj~RU_H}ABKZSqVmAS;03>2x%%VEpys0v^$hv>
zPJ%iZ@JiW7x%1X3&jf3blsx#OaLUs;5nneLF5ZrWVI!B{a-c)VuzVpLA%xG&;-w7Y
zfV?lIE9)WOl|b-?L_j%-gX2D*32Q%RIB9sU4ili^P|k`&I$s)4^i&jdO*%5L;^V++
zKYnyu`IbLW_grjqb)MnR`<13PI1gR%MxLPuhd(q?E-N7!EB`w9^K>Jw?Q&PAT&^<U
zzQO>{TFM{SiQ9mytp^;cILv4?PaUYV9o)j%hQi;@!-TamN?N$qfLCc!9=OLJ1@HgQ
z+n>Pdwq*I8P@H+5?|knY^1Xa7r>e}X94fQ2TvaZYZH$|4+Y)MF7itia=~e@lpheS|
zcB@cJG!lXqZD<K0BMV_+At5l`?V$l<SCy;EHRj-)UcUL=c|P-a&*A@DEB@b(Z{P3S
z_o|Ut=foEgE7n@EV%)Lzjvd?Mg^Lvuq8%fYVA=tMJmD2=b#S=FK^Wt#%CjtPF7Q?C
zTx90c5#E7fL5ZWi4Ae76EL$$8U8Fjl%*-d-4gCWz7rb(eLGFsS?Fk@(3xeYqByM{E
zdY7HLV;l^4DM#d&eG{KLWF8F;kp-<Bv_o~+fYaJ^=h5cXry~vLGMlyq^l!mDtsFyM
z(4GhL;HXP{3MdJK^axsJm4i0K+18j_RGYqr3)Pa=?LrJ1>*ju7V}vwt^brGm!LKk}
z{c-`u;kW75r4K9`ObfMXOs~|)Pm#a}UG#%49|IIFAM<$5<<GZ4rZd-Pn`JtFmoK<9
zY?BWh(3hozXsdy&_<4SAGT(Fj6~LLtWdQI=Yuif5>W!9P5#d>;-6#NWjm(d4dWT2J
zAg$HSx)V=SC$A5!{FWc*xn-!_@;ljtvnmKfN1FME!MZ0t!?@%(kWIvEzezA%;-H~O
zNt-z5H_&qgbn>1Dr&$($TPDC|23lF8aOVLuqj(KYYw7IwY0m@eIa;<HzGcR5@;D!E
z0<(q{GU|eLS<Z!#oNJZZo_zp^bJp1x6whda*hmM*_zpewAH4dDYfSVxuWPsoSfA{Q
zZj#R~h$fy+XbsEc%CX(2CkI+npOw|sxpMKc*6?d{<=XWN<+azZm$i*0Q684P{VmBo
zEl)R}Xsx_i9zT8*TTZ#jx3;zx4;W5ln6P7jgOzwl5I6YI^WXk!KNUC3C{%vmfwRE#
zg-e(oJ$j@?_^~Qa#=*j*42hL;?b?;HabZ1<1$=P-VcC4Ttwk<(a#;Y^?Gj?(2-oSU
z8CCXa++{>A>D;m6VlXo^9Vb80A$b&YDx-x;m>1vj(o$KRUo3MobMZ1DhQ#vXN_?=6
zZ(y)rE5CNFhM;2ghkTB@nwrcnT2n5zAn~LEBVuz?`aXPAZr{FFZr#2kV`4XGkio)$
z1ti~qV81ivrwv&Q!~%p4pa#oD#7W?3=s4O?f9_22K0<6SLLUh^;~Nw?spN~!bRK$4
zUKHsNq@f2-6xcdN*YeAfC-H=VNBvWg2QPicHyxA*7Adso59DGoJ6$f~bE3-yPOTs6
zUo?>jca4*l>BxZ~<#ash0C0vJ91)7RX~-AI^h^WVFf@a__{m^^)4~M?T}2{%%I=6Q
zav&@6L`P~TW!V8XCbEt?eG;5!wpk07Zy=e$jSS?4EV<xDhFt2BF5;A-C&v)Rw#G*A
zR$Rzzz;X?({0`LC>T6Ml0|O;#0#{|&wfvSn@S{kzp9H3-Lk&}|RtKlIIAQC=(_Mgo
zxZ58t#08K}oiZkAC-i0!s)T%6s&N7!%lk;pGiVEe%L_QohJ>gFJWxZ`^1|mn11|CD
zk|gPXoY=d5#@L1^#-{vQtN`od3q{BdOb6E6v|auQ#`ocQeG^()z$jU5Uep5+d>I(x
z2QUJaVMZj5IWgptchC#3AIl?0*G<lSIVYtq<RgYotpMlWqPf1nkF3_1fOe2q$n0RC
zJGx~kAh1poz-I@*=Q_3gNUQ0j;mFf<287;?GYqa1`k`-F9yG`?ZKiupMn<3q^vrLd
zTKPkA&{=Wg`}zTv6=6BcYULS>qrM$}AvAd#`h0D?qw4U3^hQshO?MgGM(077SJoXo
z;9Td7Ur`O^12~>EDZ_Oc^HGJ+qxwMGmVo(b^K`>)k@3WZtm+)^`$*2HadKjhG!>=E
zI%J!-)yXhYe%04;;{eNZ8O>)N$L%Xr@g2^xI-23I!7>5TO@j~G#+N7~o<QJvo>rE9
zL-vUrn~H%5FC^8jWjQ<r;HEffjA!a2$F+3QpVwP0INr5I_Jeq$p(4<Ixr{+ov&ox_
zip2Fz|9PW}`L7$}%!T=sc09fhhg`zJqKh|%1#9fFvy-y7{Zwo0{j#()Q!ZUyFUza5
zTE{PxPk-i>a_z=OnV3G+T7O%H!$f)b=s|h<<gwPv58~hpzWH<Y>QxyRk7W})EQiOO
z@|tV?nK^b1DUbirU;Dz@;7&7t6qY|J!lIhJ%e+x8eZ@3cmRA;K(Qs1JR4nHCpdAZb
z+LQsAlZ^_gg)Up12y+?;UoT}L!QFU{gA9X1g^rUvgp0@3oKvp6IZ=cs5({$+u^)`D
zPh(Wf&CZv(>A5mHGaHL?^7Mz)lu_hOb0T|x?@)z)Qnog?%TBz^c~Eww>-J85P!2=l
zP#2+iCnAgp?tscj&>!udi#o(oFJrO&C|^XU<FE)-yjG?N#80Q7{xo4|WyRA4KHSFB
zi8L&P<0(QI8eDv`z5uaN<H46-gr+$#N_YyQ`Yd!BCaz3B#+VQfmMC=5iRB{?opZ8c
zv6cEo86-3(EXtO2$s_4<`9ubuqygGN_#!YJVA}8+wg?KD$z8rca;pX`M(9Wj`lL|_
zBJP0$P@Z%!<{(NL$Vvw{GE$u7b&D7~$4x^%kw%+F)`_=^w5n4_hYX}=9gl+G311H4
zAt3*n*8r2?IBBilLdr`#<%z{=ZR^mn=4bs4(T4?XdiD`9s$BTat<853#+QpWV-^ck
zM+cyfuu)IhUYTFeX|73f>saxEXX@`>!0?Eb`eWW<%u+eo=TqP83uz=~ehF?1Wa9=;
zAH?ScaORsd_WNaU8fY7RmhJj7&b%Qf*#g|>O)CP5F^~<I*JT4T89gJ~K#M?ThD<-e
zS&r!yt)$UMqOT0{>gr<%pc(ior}_=!#9+S26~am%?GpH`Q~I0va$c(Wg@Y}1u|9y~
z$fAB-$FwyDr_+(s^{?;(pFG3mgHB0MRNrYG4ItJ4<N&$FB+Fu5-1j2A@>wVI0bUTa
z^2r}?nVi<rq)$Ba2FMT2Rz6^zfL4xOi*I3|E3To(4dnq{c*FUIw4O}o@th1n8+Uk^
zJMpAgjsbit+jZAM@K#*<q7N2+<c87`vQLbbWuV^4le@Km^$2`~lHP4#fX=+hk}@-%
zeaG2!be+>rq(!*Gs4!luzGb0PTaJ$Yc^Rf9-r@k{0j6o=3DCorVSJ}mQRf1Ax!x?l
z<4xxDGd#$$2hJz!6{2BT2A41B>Rgn}T^@Z%Ia+Hd46m;zIMS^DsB#2-v<KHI9B~YX
zoM+HZB%&oQdirBrYcemVF^6Kb#7z~}c={vEe2B5Z6MmldPfbsTF>&_nxa>TABEw*>
z%*>335pm)2a=9pX>B>r3Sk^jvO4rRZZqOfU&AeTBXJUJ2C(3a7%BAwkD=)`86nF02
zj+_6y(l9GSf?YwJ@HF=C|Mjn(@nvZW7~f(*F-jK31b4v?wY(&gXHsmU%r9tBF9C5{
zi3*yh0&($HLFL#IM8R@96YpM3<jdFFJ5do7+zVV9feO#NOjL9>D94)ewP{|K9@oNK
zOKB~fWt>b*2Yzy#7fvK!dGHh>&)S`gJ-nR$!FHynPq)g$2an^a3GZ|4sf>q*T8OKh
zQaOwW-UZ>y)GR2JiFD)crz%zYV61SFigKwu(IwH5+Q5(jt1@H<Nn?V#z*taIs~jvw
zWn76-Sy8Nrk8#S<Dc`q<Hi|mH0p!CXiD=jH+<}m59G#I^pkN&6qMt~|(GD0|;0L&n
zll0U*4h#m;sf;OU6zKqY0p@Ar#`U3-&rkX-ZI_-h(8=*WM1FynUxPdy3Y{x&TK4GG
zO+xTxUrK#}6rXtx)1^aJ<R@rp0xwaC%YHk+iype*YGgrUHwuETCYc_<8^tn^S>wQq
z<b!v~`xMsOK4TGMK6ucp=H)tOA?)$0!~qR%2;e~mJT+4&T6$P=OJ&so>$*aU4-Wlj
zIwj`iW5y~AxTtI42lGv3^kde@Lv)nQ(3U0Tr(TR@DkpfrgNy_yGH8ghH$c>Ky+9>y
zy4FXwxhW4}rvujMIj}rr(3Jz3zbT~wMWY;ygWxk7{tOQQl+%MYBmFuuKIDQ6K5&Ug
z_NYbIP<fO_I_+i1CyOa3WoKIlUuA`-FPFI6A0Q0am;lU^Bo6obvm7}0fAh3yeTQO@
zIZq)gK0%xrsQ8%AD+;h|ppzNolAQgV4EzVKU7$b}lnS?fo4($c4UEE#@>}}W2b^vi
zgJrk00C8xIH{U23hE`6W4<H0O=9L!T$c75fPr4EkN1QgZq0C%He~U1*H~>#+>Jd@D
znNC|<UPAB)$Q2mKHC~xI&CsSHyKO(qPaJiZW22=FnfB!jZMcKPrMyme^IDfyr%^uO
za+&u$9AN!{DteIK@jhRl=YjQeemIZ4F7GT)C4CN@Z|au$Th@?nRUX$FWac`ywo$zc
zWO{g8NG$$v|7_<PLz{Qx$9hZokS(lCiBZHV23B0GJ(L*X;7vLk65cXu**vAC53!D&
znVBjxv%CT@BMnc>-qu#x+hrYoQkIsd%e5QpVNlF1NZ|O9*28T3KPp?>k}+{ore`or
zPRs24Y`J{tV!3kVa@pM4ltIDMilcZ<fz$V9!kCyZWB-f)>Cc_*AMD44oqb@O)-)-L
z2IVmrcwZtGnivyG#|pvfiAT#(;H&|^q?{AG3a^D1h2(@MgkiAo1S1v()QBL~fPxtC
zwZftSm7wmUpejg}J$9;ePFvC+etD+Kl3_6^$CC+kU=e?GbgYWlj<2P1niEea_*x$)
zF~u3_YP>UOl^2EtPhT`-!-$&Xq$jy4UKp3*f-JUDQ7)F{sESsA7q1aVN4Yer=#<cK
z*NvwVe!7sIG|E)i{vcyQ#s}bwMv5NG`cWtrNc1qJg~8F{={WkMLsDn(XU7d8$~QY5
zkR3ZO!mjrKk;Y_5x}sc^dGRKs=x2zwf^i<$nP-I>G%Dmon<m1%_Kjyi5w-(b+CHC%
z>f$!b1V2gU?|9_Qqp)S;TYrP)bQ908E{HLoQX+2JR+jTNzh(3x^lOs_wD2bCMPQOv
zem_jRhCxL=gw~mEy#aKmF7d5<TZW(!UzWQoA1>>?iyI;YF2c}}UTGyIC+k|@>TkZd
z1L=uVv>eY&6hl8~)jxCG)v|`G{jg~Rco6{&{Ttmw{}i4Otkijs21^74Ne4m3flj~d
zUk-yF$PKSyl0fn|IFx*fNleJ9@<Lbxno`8WPuAH0;IF|pjI3x|a1}q$O<)Pkv$>cL
z(g0L24|Mo#p9l~#<y9&ajPNiF=z%JeEdUwKHUV$hL-bW%=%OG=P7mH_%N&2C1Jun)
z+<<B2Bz-48<)!{9v!&1coEE<WnxK=Tzjz!3h0~S4anQjYi@yP~gGj4<28Sd6Z1jMn
zi%{Ye8C)xH9qB;#N*JL5uW%tcc#X?Aki6#aAnB8)g|@C*Fb+TYl7EzpVd(lJU^y+H
zv$O2h8GOT|%OA<Yd;HfjGT+1vO>PV+U$C@b$kSgKf&(CRka@~?SRUiec^cq$ywxw}
zu<F3KG>#8sDc;iJ90s6qICC^)ICq+uj!;)uA%6roA4Ad&mjfJii7tTgshjf*Z;;mU
z_U7OQIy~hc&d+%!f2MWm9Zd@@jde8*zS9PDa@^z!e^AeIRjlcOHm|O)gd#PZbEoxk
z9ys$lk93i3WMvTpNd!m>Ss+BOUx|-?jUmB;o+o6i3t5xL&40;fjgBFKL4h&BH-rwf
zCg!-;qr?3&Jv&}*ytH06E-jU*8Oa^vU5I_HmG|QWR9@6xS(_>g3%st7?>5iLsNi6f
z-QC@o<I;#&P+k}<GFbkDzwt9?d@q85#ghjN2^8TvjZ$39<HC;eQaYa^bdPn%Cv&i}
z7>$>rRUn~baC282o~bF$rRBt?d}@%BKb1Ay!T6wwRxfD~%E^EVV`5y3?=iLwDgNw8
z`D$?<cm8E$uqfvX)I5E_fFO?pg)u~Ua=}Ew1?NzC(yg8;<0CD?d9iwYYD|@?icn=y
z8Szm*-usXtfZ=l(#t$8gCmEq|IE)AM2Tz!<Li4e=G$_EaxRFyDAG+)Ao~y2?SOuvs
z%C2q|va-Wqe9+M{9+RwmI$anO^4(G085KkVnev+ufv#Qw^CUpaL<m0Rg+59qPvSu{
zphvv(%fW3hZ8qBTU|M+T1FcL$%IJL4jPeQP6M{~zbsD0pVU;@lC~L|G+AwgQZosg9
zaw0Qno7VDsA<6FmTXDdE7L@QQ2V5ru0Pj2uj`i_|;88oqB3^%KfXY=)ZW{335o4Kh
zGj4g|nJ|EEgmE<MKv(&M&vUNB=kRjD9{S2D8czdbM0RC#aroX1hH~C%PhQ3kV*t8*
z+NSo0q4uw~E;1MO3E@0m@V#h}R)J&ac92{VQkvzpasijg5OQjG2&8w&AzbM#H)MvK
zNF1b%k^@al9Rl+@|B%zbd3eCM7~i!XfaIwexcqX|9Z9Xs6p{Jkgo$`~f}pMR5IfK4
zM>UKOMaZcAEofunuYxK#Zy@vU#<}GIeB_k*D?sQ-;V1fkLgd%T3m_hO3y=p*p!HkB
z0O`&@@+B<^J6eYWZO()wRk~PX`zhg?Fhm-PB#@8gxU2)CyyE+d@Q$vdF&Jl>7P`4K
zydmCEJZ)Zm0H6Hl&!+~<evv$bY2dLlv@&pA<>sO*Z~@9sb=Ka^<quIYN5P70Q%9i6
zl=EC-$I-=+7PM8TN(c1kQ~IdqNQ3zXax2cf0C9NDSF;S?IQxLdLSHw>4Q;$HC)3QI
z_?$yKV<I7S?f8<`&Jbnm<qsI&rnmCJ4W5iwXl3VmCF>jL=iP=4ZCVZFwMKq!d(Hfv
zXP<wdCwwF+nz+Qfp8#>);#wv;7a?_O<NRb82ub=0(T)VD4fRWu_bJ$s!!99i@K5QQ
z$aS3UM}9gHpF)_)cav9FR$^Vev$YwgJDq8*eeK#(xo~+!Yv2-Ya$0RX@k~a;R9RZ$
zn{hH!W~h612XGqep>(Bwcw)g*5kB(F+A((xjQx#&>kVmmCSe@QIH{VT)4327AQzgv
z12Nbii=r|RsFY_~;bVLNSr9&eB+EzxmJH&f8d|s@XP=M1sX##!H5eI$hf)|5**Szz
zvJDBgHSsj!aPPS6?HnjPUiNm6%Kq+wDk!%v#fRmPeZ*wV@mMkvlrK6{*aKBClQ~Aj
z-jOUgk<Bg0ao+CIe4ISRQw|K&XD7UyAcF(LMfrj|j&D(*uP$cL#Nvj|0#4<Z5rW~6
z*9s;m@LAEk3>ymqq3HH!r<9)t5S`-u#1jGyioA=50i;ILU(^M4gM*ZM3t5`9NHebb
zIB2PR!Z|QW&otUP2TSs}<3p#ZqQ(;kU0nQn5IR+#V9mX{#Cw7v-0BiKG+<dRn5NAK
zfS-PolLt8HkdLeHI0P@gyhwF_uKh}2x_;gre+Mm1@~Ce3s}q>z4CpF<D>>4fE-y4x
zhlspD*EBpY7%bE6(}zfx-`Cf?_&j;bx=ABe7)?TxalpaS(sDeJ<0;<e=F_shy%TLc
z=nv|*JeB05k%NVO1EcaC<swZk$`CH=FaG5Q`h|`{h=(@gCPh4Y7B28|EIyGFZ5^eT
zB=HSs$csGnpRq`SYLEwUX_glP2$QP1gs9u<T1e4dxeP7cpxi>C+5oCfmK!9);)fvR
z)UY`F^nD#%KJyLqfV<{Ndd|D4ON1P+G=?-LhRf>o7{_vcN(x%P1KAZ((^8HhWk!IS
z48DAdmKSG(c_PP}M?j@RF69AuujVA4h{IE9BY#qz=ksOB8$1;!FEnas4B7t_s7WC<
z!W}L1{3zajzR5#Qc%%_f9r3GtxNMlmd|8i)58ei^2!RN@KH7OUWGdXw$95eMb~{<8
z4q94h$fqv@eDa4nV48SG*XJ?x@q;*W<vGtp4a+9Yb3?cwisS)Ay5<^oUikRw?BmoI
zd8AC@aj-y5LrmNNBp=}kr!Hy4l3!exstdZ>4j~V`{7h#{UK$0VKYrvJ<f*s7X~>E)
zE9@72rBx8RtjU}5S_cHLpvARdJ%KizaPnrKV3<TqxTxzPS;%HR(E;hLv#z7SGEIY{
zoGl-&ty94DnY?nOIX&_v-supynr;~ZHLP5sttwTZm2Kp0db1hN8ZdCRXf3VvO=BYI
zK*4J{0C9jSv?z7cfQXx5;3YF?PCLbz;3J>B62SU9*Wic2$4(+14zML^V|}e0YK_cm
z0JuFF9+xi9mTT8H;>pD9+;~~rSSXjSY?N0&bzS3xcHrGaKK#r{tDBO$pIfJ7oaFAL
zXJtD+4Y8$l^|Lbe@BI(nly;{QmXAcDusA79Ct*7fooQUm45lcX92fg|Yh6*SBsF(b
z<O!!I@j|wEuwDjE1!uv}$x0{c1PqSYY6P8RbGLMyy9LNnCcHNxo2Z;=5r3kU|K9dt
z*?O{5HXrZGppZczW8&zTkJMrD>%u1MbyE6cK_OH*mPwv6NIvgu9Lb1?GHHQD*XO;E
zX`V)CVwstp*90>MQnb9_DMNy%AX>}t8UfcSi4d^B$vX`4LiZ>)dMc7<`zO8y7GiNH
zHWKLwS&yb895|JoEoE^BgGB0CJrD<JcJvStprV8UKpT*ME<~9$lh3%Ig$B;(+`=I{
z^r#br(&_AkV_-o_8&L`HN5dx|H{xNKz&SkdGvAs|1@k9;f@xbYoq1U};?N14Wk*>e
z-{f(5hail)3_=rq3v}fL;BlS?%Q1d9oic)BvQVJOGWa4bopF(O;|w7qqQ+r-=2`lX
zd7%xoIuUN2e!{1I>~eut^QK?LHbjjXjZchu@bTql`LR8*jvL~P>o7j(Psk$+@}}m<
zYw(w#9HIy=9}c~<F9ksbbm^Wd0niM#B2ZqCm7enFSn1>fQRYfSFyjq`jt24^=BG46
zeATPDi!Ngs9?mx|H@*Bixt&a9S>-$r(2xeH=@VD;s*Kr|`J!iDkZ?I?_Vcts4W79B
z0Wgp&s3piSf&@|j)N{qzW#7v%K43ZM({G!Ww*}<1c%~$mXTTqAG=e{}uCx(s<jD6z
zZNMX0wdn$ood`LRbo}z^*aF%Lh`CZ}(u7nn;o`oKVK`EK*ix6ciIiDo-U=Yg`hn_<
zlbQ>8hMpNVZO1cQx5Pt+k_p3UejvHzL(=f!0MiDqobh<szq)+VWxv8Dopb1*rO)(q
zrgb`f!S1APPiuLDrpgH3_}7nchNRDa1uf2f-4F(V=*YwUgR}P+78dFvYbx-uNLLzN
zAx^;4+{ZhU$Oh0p?^dS@p<~u~tOZ;a@ToJjDU-P~)(R{ZB#Hdo@1s;fpL9_~q#{2Y
zGOG@zgAXN)^PkX>hLCU_bb58;BxzZn<xOD6p9htZH2sH8wN9+3!Ln0tFqy_;pX(Wr
zY3WSwy0ZgF>&rIIt|mnN<MS8)3TIOC5VItw8B8+(7E=}aw9d1|gCiDl5~XnZ%*Eho
z)@a&B9L01I3{E#5FK}|uNAd_7`g1&eh}}QkdlR{aVBMyAlP+A(r`SORv-)NnoV0f!
zBVw#9E>4u$`LVLHmWIWvuU;!_7gt0-rnN9%xZWyHo;;Mn@VLw`&XlE<*>dU9hQ`r+
z>;$@Z=U&;|+>X96_V51pUq7Sp3k$PhagbjO1pU$R*%pNA@ltlox{8a4cMRBvmFL+C
zU4aJ1n4mBW2DS}FCzRto1QzzZg#Z%;je$&!D;(dS;LFy$P(3E&WB0H;ez+M%#O~%n
z+21)VyD}=axAvq4a`SF9PaEP+zbfo(%r9_H#d{S>kYS)sO~p!=IMPd#1(V4ndC8Dq
z8`hc94rN5_Z10!d9kw^AFv=so7OgOM6sU)I(jb|*?5VUijVO<LtFDDUUL}veh)ISA
zNlc0=2X)7Vz@6BgB=`eW@p4=cqmvL0`Kn7r6O}yVHK1WJVLgP4Nh>6QsDoC2S;pZ7
z31Hp?$tOQS%Bpe%(6zK}S_`c%_Q8=aeooZgA}KET0Uck`WV*rf6so^!rwj<XI%uKe
z>B9E=vJ>x--T~-X6bFqEO80m&pC?aG&Vvh#{^R1KN!WtHyDl7${((%|C-{=zwBc&l
zbwj`9!U`|O1Hi>(QVtQ+3F>-b&zMe`lEFod!ax8%z4RH!gL2hH2jveL3J0Yj!_gxv
z&u9lxS6{e<t`5>4ZgXHb)<Gv@1XL#Xkyd8Q7kJ`5QP0stAPvG+f2Ywd2GWM|tL+7V
zr*CjGCd0#Hw*}+f-o*Q~#5tKyp;`p}3P-vfaqCAP2&i(5NckPQmESsp(}(HO2f){)
z0dBxAFOnL3>!PpYYI+cdysq8{AQH%Mq|mo3d4Tim_=qc?I>&XO^pKPE{!MDY{OJ$q
z<RhBvt-)!AHtsO`5VnIm%Ew>$83%$ULJC`_h8GB)O5gGTaX^m0*>@N)ZPIn)qm=~=
z$Iai#A>DGiIVtBcgVRHz(2<X2)U*gzT+XqHqYn^f{t61kV=k2gSai+J&Xl>iIjwUz
zT0Xa1^B!e<C>9>XB13axEVLyWROnIl9O$dLC;qHcdf`Je=$9e8lVLXm#PwfARGD_k
z8-5MPU$BgjC5V1)VFKhqhc2yeARXs+X(8s=8g^Rnj>s3V8~|ei2eiC);HYQkivtg?
zPj49L9!~!?K8g<L<Ij^@?Y2RBAFNx;<4}j`z-0IhOMtV$cY2#G!Qhdj8}@;sk<Lwg
z*57ez6F1{)J*E*?=lQr6GcHuWRcNfywH8OVE-hLEV^oaCI}+;~bMXm?qoZeWvwna7
zr0gB=e)LIM+ZZnwE=-iwwZ*c$ycDky!~=o@-iJOaV}Jdxy&7AYXXj?(MGX`R%SuS{
zj!9ZI9fJ#e{N=uD=%=xWMAKF>Mgs#i8eVZ+ESTi;R6_kTo^)WK#6nAsy8#S%jE`qW
zJRLYKPakfVySE<4(}=@8ErgCrY-2(`))-GTG-c!+*OSALg<d>@iBF8Huq<#mRuwlU
z<6>-#ryuE*Yv4J;_3(&yIu2xk?W=OQW1?DCy3~$x_+k|us+J`>Ui1Q7mNV09Yc~-1
z0gh=X>5&m4grVLeFZjU+P4byFvy7jV7ZXPg#GnqsB=#4rNeLSHmKXTo!3jU$OI$on
zsk$VcX_Aj{xZue<qkX!Tz7L~hwm7@Q(eDhE9#?tD8<$u^CvSikYwDX3=M7}PuQQ6E
z<L#iuMVN_JBEYt?ND5cuq(pgYo*iZOEnd3MyDVu;hAimhBAqcKJP`cC1x<~|BvNxT
zFLLO!EKa%TGX5|qNRK;ZBF_U7@{R_RP1GaCrQ+ysaM*t2Tq98ra>)ZP(P~H#X(3e$
z{-h^8^J-}$i0&0<Kgy1PT7x00p#w7I#W0jt^<z?M>M-+p9s*mLS~-AuLXv*cbu><I
zWj`lF&N%4Q)X}~f&h~S9OJ^S(?Pwel9Sj+@A3+Cq@DO*Q6z31Vim5nw{FAAc3v~HC
zA9mdOo)6GfUW<_5a#^11V&K<2q8TX-<r|S-kcCTnNa(9Tr)N4K^25+>WLbvkGunj{
zpznARQuqE`X*qzyZMvbQwV@v6R2)Cj6-koR$n-5Q5PUTcLz~`$>GDo_UZV`HjE<jv
zSc46a-sT(htxi#%3<Wr?;yOL-YCks_xXRlaDy&_&lgd_)_4W0LfAGNvWp97KK9$Wq
z;0WMJMee3hS)EaBx5#QRKhp-xyTcIV0_w<p6D`n-ke`OZnEfux%=~OwnGbzC`*(nd
zEGy&?hOk0cH~3lqgj;7GyK{}X0X;a)JvAK|$f;!mJU;>E1=w3}{rhv6adGVuJ-)%$
zq!uF2${%`%9PnK|D|hiFec}|Te32S;lGmt|m+4e|rinl3_<bI%gTW#DUAe352HN;@
z;(N3K71z=_tfGZu$wN1#uMWOFA)-MaMTlr(zR_Qf-dZ2(TN9JQPMnsx`EjlPC1+aw
zMsipl#IBzc<s%WiPoZHWL!*?%rE$60vbMTfW@qN|;lRnWJk9Cf|F`+hM82@WLVW0W
zSj|lAg#u!aAqIt%lPXaF7WDC<3@OKA2t4WRC5%&`#LJy9@Ar!iW21yZ%Tah<(&n2K
z3{BpHc=~8t2E<-@D);!|ZrR_FWppYFR)zrALf)<5{SM|fj0v`AkRM-|MoxUof^x{9
zAU-#@DE()p1CLg7%MvFnQDIs}1&$B%?d|dc%dyHrC+CB3^`&a8YhDhge3X)6qI+Bv
zXd;8^frgTjX$200J7PaYqWGoivY~^H+LW*VxKQHg7z~E!H1O03NI${LU-0Ju2A6d9
z>qHRpJ0Q-)m5>ZJCR#9`2hCs*D@xP0gJ~4B_^3ZQ<2(=ma4qd9X^=wi3o|b7N=0fo
z%QIN!kZgrhPB*c&(9#a4DI7`4M>MVr4`fLD`NTMgF_AR1D|{#M#DKoo_N4^U>3<=M
zCr1@;;DQtev-PidfZRi8E(noV>cdC{&p790#3;Ms*`K3iFa`~Hl}fneZAw(b&N15(
zo~$SNgEWA;p&n{f(9tFk%=E0k;D;tsNo)H>8y<$y*bavC?DVX|rk76#frH-c3bb?%
zBa=Hu8BbhaUaJG(^p=YD0`EB{`&NX`<u|l>8%z_N@GG9;Xy1w-$mw|dx+Pckhyj&l
z5A;eO-t*=WAnHGOT<29z{K~JSmTP&fA9%SalJku?VM|Kh0Yo_acS8|0unqDHNy!I}
zbXe_Yf+_mYNjAi_fjlig<UG&aruY5ek68A_h$SMxh81ugzhMv-zT+B(jz0BvdnUe<
zg>2_T(txc(mzD{kON5Z(GCdEb?|goe$9XdUISjD&@SM&Z$KvPJS6?Zg{`Bkd0o_0N
zgFh^f9zTvdtI$HjJQ;kN7yaIZ!E(c|;l9po{;iK|%WMEQOrJC{=j(b9>!+w|<pXV`
zN*=#KKP(%+3+FY}u1!nmZZ27e(cpX&LYDlN54q2WJwDsMl4Wxk-wwHboa6S93qRM(
zl$|;dHkjUN4l6}z-8#eRI5<cHpWLG%ub(LsaLB&%_I0*K2T;`s7^F40p|AjY^Q3<1
z0<R(?f;d%|X<5b&L>O`56@-pj4nAG$r>1zP`K&CiOqJPrAtzL?;^95#eO~51RNaVS
zd1b6zzC2MbUY?J4DfkfH^73+-l?B51;VIvR3m3}RfBYZ(${EK=9oILtV?iP7BrH;C
zeV{ZtHO51{*PwJXWAL(QXG;<~XUEJ&J&A2c41_!*KJTnjFe!r3Aj5=(|B*TcFFo_&
ziwPMD`#T5a@q;Jj!M!IkD0a)i-m`MBFPW$L=6Jq6t_r~(k<)OHuZ|c`7dTcGqbFaw
z9+NS_R;($RRr7O8;*<Qn<Fd1J82;hWNxU0@(ZGw-Jl5eW-n?)fPcO#U{(+U2#so%-
z8)wjPbs=j|*Ni%ddQ^GaG$GT&HR!-(QbN$xj_w4)m47ff)V%2&L0R(+zS<56(AA(I
z)9ts8L@1INoCiM6bb!m{crXUa@zK!AXa}|P7zII7!{V+X_?du)`Pi}}3{15Y{MJyV
zYjptn`t<4g^gzTx$;F)x1*t2>NltRq+w(GT-30=6{*o>xAaxAu(ANd(TC@R>PVVTA
zVGs+WO28OM$aosLnV$%QE_iz7j(h|}ofbHBAXv$=Jc))Q<AhQ#2>Gkx25CSW28Q#`
zbmLmQ3L$ak#qZ}?p$%7nZWJw{!D77*8`D`2haC@G6#;4J1WUz{p2kFTcM`&yhNEr)
zrw!-=+v3SHX)}}7PaxaCRDR1NPGaO40sd1-%deb)wd+IDK`FVFFF0Fb;Ubr)o0?ys
zE4>&7aw^Vs0i>7dKpQ1&I)Jjdz6`A_u<?X1IRkedWGcVn8t7H=16HLWKSKlkGxyBb
zxbwiep?6!Bme2S$-O!iqJRR=K?1OQm<iHttCsk1q)lZ17@}ZBum9gZPgTFI=lg|Em
zc>^Ij{t`?BpZ!6aTqhTdej?wzQE9!Lr{^zjdS1G8v3%k4UnpPxnJ<;QckY(o`mNt8
zw{PExrxWmTvo!mQ_!*m;BMp7MBSomlBNd4-uB)34%x{{EcJy9+0))7c$JVPj4x6};
zT`d^IAJhqh`lyigl-J01j<*9m2LZ+b*30^O%tOO9!#C}~0pJ}T@I<*-_jPLsr#syG
zZMdZY+H~alItw`87A&Km2Htf6m<9*RZ5hz;S`OZj%u3$NMY^Rti)-^U1O~iuAv=S>
z09}1ZsdlsAXa(X`d_WrxsPuFsx-?~sT?kq?=k}<xva~u=7MI7%+=9X~EG}PJD|?4K
zWpi_1YjnORaaLB>X3LeU7bO5=?nrBPwhPLb*gq&wo;;3iPh)@gKmIFc+}Y=4i(?UC
z1jH5zbp*B#?Hx$q(Lr=v6yR}67#6XWh>yjwke9A%;Aj|Xg*&KHn3tU~B;o`Em5N2w
z?EIn@_%pGX-`?KMG+it?rRkB3h|R}aQjmqdl;#P7T%2P&GaKjTa$S|wK_0U*spc0}
zL^CU+Kpmx^&$uYmp{YC>D^oHsPL57hIeTS$OGXNNz~cB;RjLdJ`ehDiby_u17!fjd
zvcIZxNERKKizgi;8Z5%;1B7ESPpj3|PFe`)uoW*jZ5fpouBHKz64C6G_))jQ5oJKc
zM1XwiLv<Vmj}(r4;ArFlsYiBnPZS2rh84Purz1{YEe`0)7TW2zlWBY(oDP^DT%X=J
zJN?uY^2&ueI>i$-MWCllUL1R?juL8~o<wN#ylVxJmgxz>Q^9qFkL%-@@THydl;Nqg
z0Y2op-4fb72s=jqg)s9~Icojmq8!z=dLcV9l#eWQ7^EUQ91j37CNjl&hJNznRFbCd
z3|%<&w=NM6P60X-qe>LCY#{l&@&Mcs0s8sa0WPNje3y3zrXR?RoYM_xdpw3V&%WG<
zX>eEZ&jIx29R&kCf<y~_RFm}T`b5%^W@j8UfZGQ!9zXbmYx$lB=#w_(r6iMg{ivM-
zo~2y<*${2zBY_ycB&X~~uMn>P2v=y^(e0Fwav?Q6zVqv{&9sPCekVZ-56(KYI@^aW
z^o<xW5H6k?5Ll07OSzV5u*4zRjgYvRptM1qoCDC{Jof;mw{MzH3gl_}Ps$s4$+emW
zmk@n!XrT+aBv$$Ty2l~b{jYmk9ONknf(GbsF^9@;7jdQ=$gW|amC@cv3VryKJo;Z4
z2BHOz!<e{w^=kS2=RaTGeDlrn;fEiV?|%2Y<=(w}aa28pgAX2Vm`d=V+|Dz6=Wnm6
zf70y?ZJ4^rHAOno90ngys{yWX9sB=JP*tpcwhMaudKm^nb>OYnK)>n(QNNYW(658w
ziS;1m?B3fmP0QN?c%0?n49Ea1-*QIDYWZ3@$gnOg9zYC|2Dl~-fG6io;h~i?MN3N9
zD%tNt=q;bT((l|TXyrMaYlR#~K$a=n$TR^8hg_VJ;~fjF-Ju^JXRFg`nO~YHtLxL{
z+DohDwKuMn*IvI?#-~op!-x0FgNOIZ;o*LKV`6D}v79{Pb%NXh#M&ob!oGJWW&WT4
z=D&7EJE@w&VNArfA-R(iwkvT4Js)p7j&w9a*)S>?w3-`pTNBH7432z)!8RiZ$v8gd
zro!+Yhqd(!GA8EB=GIQRb?c54ep+^R_R7QiPs;A*p~R1unJKmyjm0-3j*b}M6H%BM
zSwlQEAuno)7pU37gfTHSF(*6~n(sQWJ&PwF(O_p2l5<jaWV|4gQyXTcXUm)n5sW4b
z8c*`kD6&F0_KJ=ZcPVti;6-f8M@>a%P#yU6k?0Y%UPy<E7K{mxxWewFAz{D+WXFP@
z)Rb%a8HNdi!2`ObK~^_-hx0*j#u7+M*U=>r^_FrBrXe2&pySZsh6bko`gb94mgBI&
z>F^_~rL}LGNVDIh#XiBwZel==x9lXX@M^zsm4`ITFpqtM)A}@{aOA;Jt6qo+88Qgc
z_8G2zQ(ufU%SsSG?GP@(`leAJ(eVpDsR3yczaa;CwX8{;It}t5U!}+XV2xW0bi`!8
zho;|7)(?3#Es|^Aq$wB_K}UXBzpb3UF=3s$I?PEPy!pVptqkBy-+|)7HQA9t4Rm_o
zx9!;Gll+P|#)Q)i#Gyq-t9J{~1Yi7%zP62JId8ib9FMdJB@idSXPU<>zh5trWmVU*
zan2VY-a^Q%^Mz%Gi#8OW@KH|DMt-#~vv_KaiT+lQJ~&^?ZFK@jw=C;{AMu)w5`@if
zz2O5aGiU+FgR#rc$$-P1jPQi-WHzh=+0F|v&N?G2t{1_ZW$1(HNV99fyatCy(E{LF
zo|YHr#{(=2agEHrjG;V*3#@!$2TxY}JZ)PVX9#=^lXtraYH6LP@uWC?pp&AJ))*8{
zBi{BS)-YTbHr6-F<;$1L?95Dg@Zdpt^7N_J0=rQbK42HQO4PK{Ewh#FvV~v6%<<9g
z7z3!)<V`;%@8rw6@j8QeFobMY4he}F25Nvh9|h;7p>|b0<Y#^f#sf*0I@_4A{)2uk
z<|w0FUMsTC^K^oA1N8xz7MJUR%mZhd7W%lB)~>|^Ef2aF`aD1zw+zcSWZbP#!92Li
z^CDw{a5(*FS@Il!PgyAsV(Sq684L*8nnxdu!wbDK8ez&X#9RHqWmLNP&=tmt*4EQ!
zT3e5mwF|T5wKuPpFMauqa{c9XRV=R!9GC5#O|8{gqwmBquygZ#dm>*;VD}FAJ9lmu
zK7=>+fBE13x8enCO%aT6Wv5PG?{t%4-0`kQO|hU2x>ybo)?k$(vA@4p4i07Ei=K0w
zaoqVNUF&u=0P@Gn`o?7q=xb79s@%SPzx?P&KPex)f3rM!vR$?}`OsX+m%68?V}Ncm
z*p(emGG^!I;|SLYg^`5}wlqyn%_=;tta%?ojjcriTRfN~!sw7p77C%K3<-=0J~qRf
z=o}N56GCo1i3*M?uf_wNNBm$B!@vUC#>xzc(>RKWCI%5)LQtmYG~g>OJp4tcsyR9h
zeeyA)>614GXmogRk#Rz#WqRZjC~?*>N`)lEL|VE4mC*pbCkumRq@EckFAqEhvIlc)
zE1;o88b(yZ12}HqlNIUs4xdM#?}>Eu`IId42S@q2XpojXNgLnzP+DLHX)cQ=Jy%GA
z)B1P-T!*LaQZLIDg~BRd#et1}qi{SCsQNmOAP#_5VUdP(1xd*}pA|<OS9wB@NXYih
zNK*22E$A{K>W_;VhZp2w2Yi{H_lJ(5jyzJ|&>}z@Mjpl|q_33t<KI9$=u8SU>3wiL
zxIFd&_bY>G4K7>u#U_&xvIxT)^<Cq}MgD<pp~@x{iAmwb^iX~+RCG`<2Q)f3FZU~)
z!TJE6V*t}|BZpM_`=CFJ;K#MH9q;#L;{a&lj$}q64#WaUKs0Ztz$bjYE+EZc5jF(A
z=3}2!Ngc=qtQUYDIB+p1fF!ArMVj4l{kS42#sMy4#H&upJq*yKSmwJdqYeM$AC_w!
z&V!E5{K=d7jOGR3*1w_U2kf`9`nZ<X;ZZP1se0<CB9gZ+3+U&eK+f|S3uyBUxzz_x
zO8y2i+bm;4d~FyQrHdic65kpVhC$xd0rV?QabnKn!(RJ)d&1=t2-4Ch&asu>%u;>0
zlqxT6#kCPfINGj`@0b(hlx@W0<U7HM)*j5Q%(;2ZRGTAGew93^3&(vqM%Mp7pjMB7
zy0XuFz`l51LeCq*0Q%d3i8(`Q#u*3?>PRK)^2ju|Ny~E{n8ti|hQ18qFT%C_#`!wK
zyaq&iBg46Y2i>j#O$CZ@IQ>kj{(`TOW8#{9@-)eS<ItF{r8h)+ZL69bxUL^|4uS~t
zr$lDhEVJ5qa!iH++ndJ9%KBt^<x>~S7rywZa_P!)nT)Su^C9Ykvb(oiwzjv*`yad?
zC%A50xL7teE|&GRjk377sI~uIxqIhknfNPz;cLIf9bR<JyQg#_?zYld*oVzF4&laS
zBuveend!MQD>peY8)+!Uxwb6O0gji=mzg?Cj)l#*7XKGyOsub7QmH4)lSez{_DA>2
zom&sf_S3z>hvHb^#}QW=+@X*1$B3AnTP)MFY>!${;Z@MdWzj7wuQln#qZgh$@U8}L
zp)=^Yo2R_PlA#l*Hsu)1dEP!B<Kle@YCAhB3wSP$R5BeVx60%s1J0d^MIwB;Fw6^)
z{cJlCK`2Muod_U~v|MNsre0FEjgh$H&L9?JSen)E7BQWRfx^5NX_AkC3dRk<aeGlG
z83v%w37~_%43a^$`53bF2`|(`kad~$PM)LW(XU!O;h6X<KGLF%lE(Sq6YN~gKkzjy
zNV?8h=0LUppZt8s#*TPQwi+v@vEu?4Z5rVkSHB9wr%u3tbs=UM5ra2U5~N-DtxgU@
zAF2AWyp|cdCYN|Bad@bKLX7@OIvt33>zr5o=r^u%ICQL_D4JTI0T1E!Z;+l63?073
zCjNP-Ws@Ir1|;B}mM|Ra<aHOg4(GN9`uS4tgF3DGBrX@NeP{XHe6dm9PmA>aqD0`n
z12E3bhfZjP%Xx;r*YtnVsvmt%VfcW4I(~$S%7>miRP<w9ymVUii2fe)TZVbc?DV`w
zBtH8#aueq~aE>#NqO$^_hrxOdfpJmYSqB;K>VmmD>bItW2ZqZnH=NxNxX-sbg#0S2
z4Fmo7Xh@nq-Kadk;T)%~3~f1D`ZkU;5boFgb5QxCUIi^hpT3pV*F8WeX%AwEx1|M0
z^R*4%xKXYTAB!W{MwP}`!ngj$(_iVUXvfV(PM={-GBcZB4(3}^900>{vC}iNbqx^A
z{6x9>W7bb5SogTx_{jD8!iGeA+&0!gA#NnJH8)&~9jGMn_i@mtp%``_2b96(h_XbP
zxoHf)Bv&DAIO&IXDgby%7tx?O4=k^hX|PVCz<BF`Zy5k#)3kUX@u?qk7+@Owa1n#H
z^IItuJp_oDsJxbuVNw&?FhJZmRbh46m)m+xnyQ-_YMBS|3Pb~iufHgw6lN?kAIH2u
zrTKqqEWS4}JyXiu{6tw?;by+p(tNP)_(1*lB<6X>{DIaETnBh6u`s_>=I1%sh}R09
z#RG+jU;T^U__e%HR3$@s7CkDF77<b-G>Uf*CZ==&N;iSAp)`Dq4_=Pu8xcH3n3Az^
zddA&-xo0vAWIRlc%LtJnaJYX~?%#b<KKkHxxqa)NjEc>2bjX>v$_p3ARcS!8P#d2v
zv-2xuVQHhx&8<rAvgj6s=WErtS?M;ZPEX+(=qhsURMecSm&yRt<)Yv`Y2hVtp0rSI
zj2JGgc~^lB801+WVg(y7oU0Mo8^+y>Jc=tDKC7AwoOI9x5{3$m2ES55Y8VTgf<^0U
zVDY7K^9vt~6s%+Nt1TaQ`lU>i2VZ|#5~7L6g`rV(&XNatMWI2)Go0m%coGN_NBx0p
zJ7$+p#+#%efF?lc(P!l6B-#yf%XPdT9tCZh!$b7qC+l>o3!X&Co#7&X(=g#Qv<m0K
z4I$ai$nEE6=K<XfZcm(~ckyu9sDUFrzqP`6sj|r*-&uv6)T@QI-ZC#pt6zVh$pXn`
zQbH4jCP~i9-u`YR(1+<~!7s>OqJ*J*(AZMWhtMy2c>|-^8XV$XmW0#;KhtU~^t0=D
zhIohMbYPk+f0aKdZw(s{2nl|Oevr5OPdC4Gko2+GV17}U@h>b~Nw0tKF;*+MFSP~&
ze9S{E=9zDzuc@89sK<pN$Q+^nP6l%l^LF-8%Nxk8G|cs;r4Mj1NW++XkcI4&83sk=
zaXyw4vLx3KMXY6$gsh&3*77PXlEd%n3qPdrqjcZtZ2>;hvq*CtQZB%9Z~>G#yy{mp
z8DuyFCS8@$`WbHU;rP$A)(rzV`%W8VUenSiow|<tRyc7^@9-Iq--r|L(udo*g%%$v
zsNwn%^mzzIzY6-q(|ZR%y8X&aJm-x}6Z;^Yz7g{I54oh7Kwdxw2GIuGo?V+(y4J6G
z;OGPHpMd$?-$NJSi1X!O3<(T~*pjX5f80Q29O^H|spPSk=BXggW8c>he&TMn{J41;
zv@#yTn6UAYw639?6I^2*^y8>XG3;VXIqQvZlquVvvL-Fz%BLTE+{Zyn)2Gk6_nHa^
zGSj%8<;_-@QWw{8;*eL{M&YOf_yZToFN8sxSFlW=?Gt_9bffr<f+W*jPw?RKTAOv6
zdIQ1ch+tT92O&j2;rIh_4zcOl@bhtG;9Som=rY*D*Ssra5REd8^0C(C0RpEo;ijk4
zsF<0XC`-$8@!g5Tll`)@zZ36F=PL~Ga`Tj}>FV>RV{w4d)2CbI(Zh%3{)78v>_7Qi
z|6SgH7BK|5b%{)=hkRj3gT{}8Ls6xuG2*3P4H!CHu9lPk7)wTja&P}ogG!S>--S37
z&Dm7=n@_jP<Hwuj$&;sLYim0mp|Hg$J~+pqnVFK6G^YuOQ=%qi0j-tArFG#JmDgl^
zST4qJW=bG*z=Gn1ag2pPQhSl6@<<;BIg<h(*ErbQkK^H<9iNotrR6d|I~R8(7&LJw
zK|J^|Ldiu&m8xG2z>^qFHZlhI6J@xQ__DC_rk=z{{l#}KR3=6dFQTVTs<)^otUX2P
zDEgz0nTXhmg&wh=9S-U7+=VhA<B>+JrqkdQ92{jxo*YO++@w*yNbex!Atk%uAP@Q3
zNz(;?<t5(2fQA}UL%~l6_GCjGFwSxF8Js6{3RG;;dB6{Jh~G@JzF?ta;sg5S0L&j{
z7OJ8H2C^G^U=UU^B=LlY8f%#wFn_v~8M>%`LI}SSM6J^gT6zQNZJ7sgvF8{}IRl<v
zJQis38NUxS@Y#kqzpxOCNXxU#7~{1@!6BpHMy59|CxE2aZ|(01*1_P#xY15;WdbeU
z0M5?f^p<W^cz~<^C>gm(S3vCm4(I#`yAQQ^z&QTA(2a!yL6Ib#1WPXI#K6UZTy>rR
zamoek18glpw`g<yMNT10_{^^MkFFhB*8+2)@#GP0E=Sqyg09KG4Fkk+To-ts<3x^r
z*rqv;QSyu*=zvrC03YOGIe?<shbeQ*SL@Dto5yJmLyt2Mj&`alxy{Ts4B{1?Bq=?j
zLtwe~ohJ^kzAd!rEkK7O4Luz1L(5O{C_1OdFEouF!~B_F5G5+&mf6ZR-u!KP%80<Q
zU!gBc5&cNc;J1@uTI!A@(9YA&(DJy=0rv;CHVupk0Y38{yC=Aw6Gs*LK;)+k>C1e7
zBA+gFeH*Y|`3|G{s={eRboHxlIs(w!X(}HWRMJn}7^qA(4AQ4OB2m26ABY9MWJX;P
z#;^H~3ft*7Z~(FxQ!PlF@I_Ta-7os^eK2mm#@wa3ByXx_y(8Dxqj1QsxU5^nB?#Go
z=?ut7opR1FxXi{oK8kmcw_KkN50q(#q3hR%zN`$_#2%O}=Y7{*9}?SS5n$QqknzN&
zE_Q(RsdHVTg3A<!l-B`WjCP8r65%rrCS*`^EUXOdxzn<CVY<BX>5Jv!)j1gxTV?lP
zx9lIVt#mxzNoFl|z>%@w_<Ta_OyXqML#^XCF(&>^sg?u56Kp6f1*np^Xg$lyiL&8G
z2`Z1$<EyH=s)eFjsHi)|OVt<?7!cdr2eMu^%iBMBzudX|KzN=AXdxqkTo8HOGOh*m
znB0^M367j%G0*!Db8~r8)AaPbI`K>tju*PKQahc{HY{m$QG&X-d>TnK3SvwgY0-bU
ze-L{nap2ZgR?D1>37UorV@wnwR0Bkn9*rVh!l3ZQTE+x56bC}7v=}k$E#pGSLXDTa
za$%zURWForLS=;KK$af5#^NC4<wY4of{6>A<15|zbDM-tlxa{{%9{A@qL^h3rx^Iq
z-FQNkKL%w*LSq+`T2R(x+O|%lkuHAANZBdVlZn%Sc-IE2Px8Ps`4}v##kITxeT0h%
zOiqzOp0VI<=nbT|6ApbW>$sB%n0J6v6qH5w>Ds&oysDcdqwHCJgXMyA-2mxjnlz1Y
zV1l05FulRDaD;)Nkq+(~102wcd!I9P@leJDcjM;f`SSbJfFD}&jyu#;2pUEiah}{q
z*E@LRN4BxL&HVe&$;xzO;jBOKTv>WxJ)Pc`zm)@Ie!0kSSU)v2VStyoctXIq1z*cP
zqzCNrlQJv`nxMmvF(kb-w*Z-zeb|eAaDy_0%q)WqEd$8`=+TrJ@EGo>fvSfgWFj9w
zojO5H>Np7_k9o*#O=?iK@+vqFZl2^qfg4MydpkCf0j|}Fcr0vdn0kOG7knz$VByi^
z1CLymIr|G~cAlf!e7gBnEaVo&%~FM|IwnX{G{Ewnhc3TSfV>u<bDsn7t+U(RvRYo#
zwa||XK?YTZ$_G!NpXUY0MU+wJz@Xn%=y;IRwbgm@z#+q6njx8u55AQ(ppRes`ZZGp
z(?vohZgp?<1loFae&Cqn@QtIN*#%z0=~vv)%)X4gic=itJn#cw4GeGXjx|zM6SW{u
zjuBAHf{(#4ImK>=sW4t_Krjzk$1GFQx8U-S5l!oO;37S78BTnx)ih4CY~5Ug7kMZX
z{t`#paDND2j&X%S#z%&yv>2UgZb~8BE_kS$8s;zRhrdc6G{~*K^J}4P(@}9C>*+kY
zh$?PBX<H!L#{sQ8`{x1N&XYN<53T%uUMV{rX?BSZ$&qF{rvvmqB}BLdXidYkPdNIV
z`WS`;N5f)l&&&xob5fSpCd!RhSIX6wmdnomqp}$<U+<T#&1YqIH~Sf{Mson#(V^ri
z%cZ4KE?z1{T|vgg|KZ;_i$SD}u{5Qwm~wDX0CeQ~5=E&B6OQSmc78s>#p01iCUbN1
z(NVWHcgoiGUU~5FX}S5)-Exa9O;2{p;W2|;IcuQEzKSOjbMtdrgwK`<84XiYvVgRh
z#^4|qUb@ET7+SVaB|Zl``b33+iONzy`iF+P=&+WJ_a>yQU%bK{YC+Fa2_E;%Xdv^r
zQ&Ul1$S9)(A=Qn^MRN=%u`KvG>5B3apU`9wFHEau^9$n1$94patVs>FQ>C3Kd~lRc
zFs8<t$kIVeSuiFrZpbe#f>MLV8w|`Sqx7L@wcOyD1Y-dt9#oG7V6BA#og$>c4}gm`
zpTPBqkM4s60X%hWF#zk^PO_euLqFj{K3P>+PIGz-q>TdThWIK@KLdOpTo&_0{bo-c
z%D3eWJjJfg(5G+XKrS|6U?G%uxJ+j~(6cR{^B_Izrjv<y+NmwusC<V&3?Zw6`9hdv
zcTwqD!;c<4j=t#^60I-1<Q>LD&DZ6?sdUPKeqIF`l$9uIqyy^$Z$irPVwLoPz73eZ
zO|uS83q9560P6_YA3#IjhgLuHf#O0AfE&~c2_k_b=<#cr^Ec(SZP&_8{P5k;s87;^
zuDt-)_6*YO%i#Oa$_MC=+<*x-{bYNQ(9uH8950~*Ui`RWQOC1&blLl0zPwmUX4N}&
zPyXP;56Ub*X?88G+tPIx*P<rfFyL2J-~;Aq>C;#CbsjdZ1>(Nh67zk9HorbtRtt{z
z<pbyjv}sUVXTp)vq(f_ZX!_w+mXIL}l(9}9qLoi><cA!1tQ$bx0pf^Xpv`+IPXY8U
zLz_k%XsI2xK11|9ABgliQCS|KR;r<{{(x2o!gkK9?N{z2UZBVPAs=Agb-yzmV(kNj
zNlPPwYc)3>YvMUDD1idVq1t)rmAk^+9AF-f#hlh3Z1oOO(YAfJ)8_?wHM-*ZH@65|
z2Kq-CPP%J&mb)wKCjr+XH>Psix9YPI;QT^=<SQH+#ZOlr(>ZP+?lRanO&jmyfEL%b
z5hN|$aJb_$^z%87w*_NKZ^NUY#pk@=<v-kSP3Q8aF6KYam;g)@OS|CV1##BXJe|vS
zV_Vb2^ii2zd{#Ct&z1{UX4O#-%i)Qf)~ow_X-ohQ9+YzTjtq)@PKGLFPNbJEr%^HX
z_x{#@Te_eyi&7>T6+nuq3{^knUKdSKRw|W0P|}b^HIKfg)u~XBM__yi=8=qu2lpSB
z4?nt7-g)O%xpil=tgWw>`K6UICd1&2#WYVMWYTcgKAu#ZV&QUEo`u3#Y(-*0NSFZ<
z`x>QVR45fgN>otiTxl3I4A6XG@1*SQ?v<UL?XtTggGM!Rs>xuQ3YAPc`;m+d7W(k=
z5_x__n?)i|H8goBt+h6x3+ry0!8BPkKSWs|0Y57PNwunS3L%|Or1P_=(6caXC`VTw
zCNmzF@??Vg<$V(kK;ak+@gyd*MPjrMbwEesrP|ortoor(K_3@LDdIs8lZE`Ov4kuN
zKcpue;pd^_0cx$sM1%{RgxOJ^1azS29e?JL;4nO`-vY8+Kb{mIOZPl|D}OMd)`3VG
z@zME-Pf<|D_w^%gT!!&s%Sko+DY$I2ND;3ckX|Pzpc-8K!q?J|f;J5!*Lrn)zQAe#
zH-@3%@#Du)r#zY9LIzJxSSf=(j^A>|on#P^wexTEX!6gzTOi)y+H~k!I4{k3`=q&_
z4R{(}LrZVh#tqP=OM{xUEVKCaPn}qn>ncDZYG~kFs-z&eqZJSekq!ux>CT_9L1Bq#
z^)w9AWjtKW*~4i-D<jHQ!<lcEDFR9(9rBdZ=AE<=Qh*d}OULbilrij_9&-`-RXW;{
z@k1DB(?>zvQ8oXdts>fiL?O<38>}BdDB8G&M?uK$m8pfc4+dQ$Kdu$H1_uyBie<Gr
z0d0PUR*!)%7f&W?7#Q$~#`*+#<w0LN02;esm22~L8eko0o0i4^E@*1pJ}{6|WdR|h
zi383H7|08@Zp@*((=xp$vjO5RD|$oQ^5%L+@*};gi=^$;GcFh5`BZ{;@}mBONPI~;
z08}9<)*~XrfXF++T34{}rtiwdyO>%(93QbSJvif+xf>m`?5P9!01*rrF0b|HnwK<`
zQ@>JxdT=`WQ{KEb=e3+u{&+gV_<{#W86EgZA8V5U>aaT2Jn+%)`+Y2*T(VZkGPMx;
zHa^gff6E7;`6T=fLfVEctEKH^59ir}>A|~P9bZk0-ww}jUycDD(%SUQzstwI*F8vt
zuchllt7qh2wHn0b(>}Syi6=|EE62;%!cQNUsoA5lv^G{QT%9Qk%ai5k*+Hyf`H1t`
zGZ`T2yIY&P<-xt9^7IKez|)vmUgqE>$^UzQ>t9F3P%;*w0)DZE1{Tqo1~BT>GE}Vk
z>{&V1LY+rlSl8Tf7gO2X+?MrnzudZYzudfizdU%ftzI@?RyQt+cD@`-PQH_%4#?9G
zDk$>cT>&k~WlYd$TU%R^4}*_E$AXOoIE(QXTxqD!1%l4Yw=oX(cVqhz+nd;4ba=pf
z97m$j0$hce;jVxTnpiB<4$NawUe1c`N?LHuskWwP*e-Jx2FL!s<Q=kSOoNU;s)fO+
zNd|+0K@~4^a|cj5^CSg>Vr)X?lKzU7WAaxz4H{dily!V5T!WdqKyDaxny~0hJbj2x
zCHZ)Sp##sC#AADg20G;j7nUkzQhCBC0q3@@esnODuaF%z)6N4Y!$mmyOOM4118J^U
zH~9PV9j2XwzDYBrj;cqpk;^!aL@1KSc{~ro!y7;5P1<1|gs0!8w={j-(}_lZE<=|_
zXjl%(7B{M_fQMTPecGY=QaiVkjpYOJ?oiK|fM;`aGZdMtr<out`&TX!T4x}wo8XNf
zrU8hz`V#hgBBpKAN8t>8*)0sEi8r{bjDG$=OHa6;FMfv6S2TpE*G87>ATNpu1G$!$
znxw1{mvTG#DYA=q_$n#EV(U*_#9M-730!5b&=FZKU_H}Uo<z{0<_~F5DBXVJC5R?y
z2N9(+uZbvS)D-#P`pz^s_a*T3XB^PdkY-#9ZC|wPP>Qhrk{;!Pq@x)D$Qei{-x>sB
zTbcZj*63p(9tMM8pd&JB_`Ey~PH*W4avEsmkUpRZ`V69Il)lX}Xz{waKNpqIWql`2
zmfJ>j3(l{_Q)i=e1sC`lZy{)FJ^<22$pDPAJRsVsp>5~eJ{aHk;eRearI8+Kijr(P
zlQO@O|4B6>2<P=z?D5x=DNt2P$CNjC!8L9?PL<fwZT(ZVkRq{S;J+uY(j+z`nSVN<
z??<}iIS$a~Wa;t%vHlPe95-DtC~RoNdN9cbfSv$4c9M}zJw_<{(O{ca9oK10<jGfV
z55VE)_3tuH*~4^g+&G{KJ_qom@zlu~hHl>K1N2S{SMhxSPrB**U|c`{#2eS~CY|Nf
zeyZzw%kQxBbr`OcQ}tCEb6WOItv%H~yf1-aF*B?FIeStjXAaB!(rMYaI$JJXUDi64
zo8A~6(`rA7P`jL*9B7?-Ru1-_$pAi;9M*hiVJOR(_&3kE;M0J-7~JW}FsVZ&RyjGj
z1|%}97InFJ)P==8sSbOh#q)y)kIF|M-Y$3UJ}jF%hh=PXuB@zIDQ6QiQgEys$O`8R
z*vGN2m|XHt1gZcQ*!fWn-fxJ?pi&ffM~}yh3UeY;ylWsoUIr8n>zWs>pPd}3JUnT5
zmbz)+aO^Bk9eAf<PX5B&T)6qUc{zn;Jj}C~TxmRQSz1~y%WErTesLbVr|j?Tl+CAG
zWqW%kzFEPi#P@gi%HH;N+27e#1@D$qP8*RvoQF<5aa1keteB87fskXB<CMj}jFz<K
z^C=30nY$B%ffSu1?;7U>kOP?HFf=eG<n!{ECnxI6wD1FLK|_W+ntdQUn<4qY84qng
zZvMcBNhUnH)im%-BzgCOc#ww<252m&)xC9q^UeO33y%SgoSTJehaZNJ6=X@nUqI!_
zygFG18%2QiX(9Dgx{}b#(2;h$);q;LFUx>o4LHBdW0)R;8hMdMz1O-QF28VLwq~72
ziqhH2vaP-fJ=>fxCII4L7>Y3Ugab$3s-v0*P`w^qLU2U7>E(nBTGHiL_JHJvr^S9e
z4Ct#*ywmS_T)-flx==&$4_(y?b*e^D<^hNXgUgEvc!G{{)iMP00GO+|DNyAP^6sFI
zhc;zBPaE|&BtHRM$Waod@*>grK7@?g&z(Q6FCSdnrwyhzw0!5K**Sd_-Jn1B@IZ?L
zaz3p6y2>7e^#^tGBwtI@0=RZw1RNglR6fQAA#l&p03&3)2nKwDR-U1i(Z}PA3o2+8
z2Qx~Bc?{4D_#3?Es1{n@Pm(r}8@~-yy7Odp@$>pX^7hlZ^yl$T>w{@SUW5_|Ta5F<
z<&J^8vmV#rz$YExTn4@ZT4uyuUs{DjSZ2bA$n8KF;1G32U0!#2xi!S<hdd$}KJ>(r
z4~Xl39S7DEH>eNE%XL7m9fs0^7rEea(GM=xQ!<z+Ytp1_dEx7S2-+yJp>a433T~?A
zje;~xh8MvG^V;>ndIN2G%V!{%<Q<PSh5kHQ3GH>dt1Fyw=EjG+%>`)r0LKmexMc%f
zQm0qSAt4h}AJ17ySF;NKe!ln-78<^IYJ{VGw2qja<#2-WGCQyJ`|N3%nmsPFODARR
z(o}inHLcgjcv_ZyllPywA^+^zf%>tG5&Cak|Ii=j=jY1U|LuSO-w9(Pl^kwx`dmht
zq+QAGIJF@)9vCFkQ(AzlbF!U?57FJceJ@_#n4DcKs~cC#g#4}Dqq2AStQ=~AeXxHN
zPYP%vzIct#;*6u)=;SP}!O^){m>=xt2~2$co)<`B0WPBf$3URa0n$k&QAZw=fiP8;
zSLVx=E0@dq#(G&@St%<kG7y%Q%F@C@SzMTp57#X&ES9Ckbo0W+(X=WmhQ<8yQk+Fi
z{Tv@1M)~;g-R=(CrgqBCmW+$1Ph?bVm!}W!m(52H%hr>}WoL7<>}cV3xStk3je+fS
zqB1_?rLbv^rK8;W9tP+CLl=Wi*u0ptbtwmbx;l|Gj`|5u+t3~{_!T7_PJhrDj0b4W
zY*XTC)DX|S69_w=^1BPU@#qFG1|2|Jwg-bHP)(i{At2nw`;%ijQs@auOP(PsbWax)
zJz0@`c+#PUGrhV>A9)0E>TV#PR$k;OX!8k_T;9d3vgEpavTm{tJ6Q=;sS5C5z&anx
zYJqs_rT!_o#N8f%>Eq_Wz?d*jizoFJ`Yb%HF#!aXRP4&;ej~kH9!0x65{J(pkPGmv
zPsR%nsY(udXoHW0>YoEqF69*xYo7S6(@)?d&%im1S7{W_i>I<QdB89a5M`-qIS*@z
zKFl+Ql$|^lJ1Cm~`ZjJb&in>bD%U}%{@3~+hJl>;6)4VM2D^GDJ@ZOo)!rE7kKgw7
zwp=Z=^oBm49bo(@IjxL-x}h)EK-hgAniu5*v@|V%v+Pzr(B~oD@s@TJ?vucaWVPUY
zai+E3;tZ{fKD}WSjq~WIwY+^Ar{j#n@AJI~I==8K4C|dV$bDWX1GrWPJD{a+ap7t>
zaBWzAUemhw3y9a`fE1-4A(6}U4vu+#c6OpH%+E&HxV$;V&9_Rgczn}B>AprpL_j{i
zfiuW6GE9dI?>N9&@3!3N9qDy<1JIVwvY{soZLFhen&srp*)EUpYMOCC$|2<Xpv)L6
zIY;Cgvm3iDbh3x@2HLQBjcaKw1IYZcTrIS+`f0)p=ChXHd4vnHK3{@q0LwP7p``&@
z9gMc`yoIUGIO<?nhgOcm&+DV|DL!y2E4veDAC82bnL8^p^Uq{VJPTvuQ=h)7j(b}6
z_77x4aD$w-JJvdNS8MAtt)-XB;^G30LT*|sFO;#r`#1mfGmdu}lcUqqr^%cS$YLUO
zqXQ=lDX1VGcl|jL2@B@#oqOd+KYFV?-QJUBGhJ3Tu9RiDhnxFld+)d$9P#vkrwbaS
z3NwV*9>v%BCa3FXx~0c)U06ST`ZVsQvM}G>*^Lgx(*>S3OwZ=_s@b_26=t?9EG?7;
z6_Bs=5no+iEE^Xtlp8NyFBjymL<PuT&_ZfP#t(8Pw4j&b)HS72nS8bF7U?3$@StHG
z2dGSFWe{vVDG%=4Ecb5TED!J9Deu4YPTArp+y@Widm7t2GKh|3l&D^&WL%*;3sFvm
z;K(huZ)L+%-&yB8)rf_<+9U3c%B3NZbx#^R;Y1HE@S+PG>IQzeV+#|%IpJlwJgK=O
zrY_JVE+!{Az_Oe-_%tYz2l-W_h?yR6xbGvw@_>}9a74TOfIMs4kL^=e;^v21DM2el
zdXD3Z$N9zpZC0Ab1D5S@>pC3$L3GkJ@~-Ky;T*>$n65?vg?Z|bb=HCN?m%+L3$%Xv
zbq2)59p&kaY<PJ0XLmQp5QYQ>h3`z|U*?C)v7cpO@(JCeGgYN)I^c47t6)H<P7JRB
zsEvjT-sRasTP{0r#v4K{wO@-1AP%bXj)GBmmxX-bC60p@=XAjQna!{)=;)6Ir<=bo
z3kaH~A4XjNG5#tJX!!;=3P7oS2J-o-WCMvot9&yy0C?=kv#(p|F#^nEFmDT%X&5D|
zl>s>3%EPy;Xd3}>;~Z{b6u!j~HqOwO0hq^O!zi4gm2Vh@bDXsEcv?Jui|f+^qvRM!
z16p3oa~jZ3x9>jPrnk_x%_wMjQWoXOcng-FdNHrp1VNj({Wib8o_(DLyfr-VE1v7v
zuptrX=b)-zx^GXh7Ghz=ZU<f4l^wf1kkjTJPXt)=giAe>roEntR{7cxkoKwPVD#4c
zu$lWPGmdzqH7?em$f&dcVbiwj2-hFi2HqpZc?|+PI<#_-n>we4fsvZ6K|#5I(53PY
z%jooRT^TIf^liGsT{^mVxIXk{14)<p;wXD77cker@5yl-zs<J|lh)_!(rZ3Va;AYu
zB=J$k1dwUCPKNU`o_gVlg0JTo6S4hn27^Ltf`!wvusT*Q-&ib{udk@@9+z9U9+rFe
zo<^Nqyf~-z-BDr5{K_lW%e8A)q#3;-KfN&axBln<4H*->D^Oxlo}ICFK$Hk&o>$*%
zZoY~gAFvaGZ9nh4^Im!To%gj^UMLgOvR0;Ky-dxOokJ~#wTPCDpvAM~s!qnRcBI6(
z@|xr8lRO_(AsHZa;71Q1l@CAspzQDOOJN4-ybAGh`N9`IFC$^CT)4Pa);3nk(#oQY
zh`F*Xe`VPQ!+d;14|$YH-I!ohqJE<i^h56sns9?sCQS#9BgbJJ^fc~?+@Ye!sw=*A
z!Fw4e<@8`rEK}v)yYH6|-g>9};QMctZ-4K{<)gdXYP{umNt+ke<IBWH`2u)=bcsET
z$839ILgFro3^389L6IFN=b5_5j}8NZ&Mb{Y8|{ah6)1bwMFK(NVuI7;(Y;8_Q;*zs
z6?u`z2$y*#J;8x75&2R#nMc-p(gBVGtxiq{GGE05m4lQ#0R}<4|I~bOciefzZ_3H^
zXmz>Nr_({$x`;Eb)enFl9Qip<+@Q`>(QSQXnbJUOW#V%2`~012VZ-m!11YwalmGCu
zJ_y?dmCC{?X*MP-gA>f+`zq=i7!o{{<3<2{C-pIOTts{2z|fZ_iluAY-gH3dTcep{
z3uqFmCtvjee77smm+LsX;kWgo)anu<t%hqR8Sa5)S`J}!i1I4#@;R?;&nmz27%VGX
zjYDsqHofI_82rFj1rE{Z4<oIg?|_FNp`l|8$2`Da_+1|OTtw)11LII&ImCmfI{LN2
zG=^3Wpp|9!Ulc5>)yL4+k+^YAH=Ktz&TZ2N@Ro1a=7Td>?s<41_?0w}Azb=NZ@!iX
zuuStAOpl)Kqb>C5ftJQ;mJgVx<vUOAi)i{|2C(1eXFum2#(gITdeh=kR?elPWVK*C
zbwmG!9-TaI7+N`TV@@>o11CREJ7S&-(tyeqU~%k#Scn}Cb7f|l16BC81zW1QNq8u^
zc~dg;uVv7+!tbM64*C0OMT}!zH)1Ys7WoOPi>gQ*zuxR7K{jP?`*4)4mMfhI<?x-J
ze#$VNWdT_q?5;?CsFS27&s;k|lUS!vTKZ{Tt1&15%f|t33!qIS*-34C0d2f3YfT;i
zD#giJT^7OHmzU7VMyBbYw0w(cWdLm$XmtYGuxTv=d|P*YS)#1Ya%ivNwS(o81+Fi`
zc$$E%XS{R`KgI-G|EFii<<7#GSlO5?H(p!QntHm7O`K{i|De3})+4R!N_q8F`u12^
zUtcP(z4lVMdgWrYE9>l??cFl=H~+@3p7A0ECmYf5TxeJug~c+>3(;sNPfp!P9axq&
zk_Ai#bbem6;pC*dckh=U{pcrUXaAWN_Zt$nAOm8yoQ=<xqf?(~#C9LaQ0I&zSA{>?
zKQ7Y~vgDN)Pb28KY{$5B`%Y|Qx+)9k({H{}-uTR$<&Dq2Sw17Ty0)qX(M;{AEb5dU
z6(z@jj&exHQ!Su)8T(8|%PH?Kut1l~7qw4h^h~QD>O8!RBF7ddp0IHCb4sA_YV5co
z$OX=L1D%07!EwYJ#)Ta9H1@0<AMTf<gZ*;n*6s2Kzx$o?Tfgz0_zL*5Bh`}@bi4?D
zaClG-Wk5WW-p8t!V=c-tAWl!@Gl-=x3uRv3hVF>^Q++ZQGngZ+T(R=vT?@tOFfm!^
zk30EUH#yM~PM^Cf`Mm+s@G)d~O2UigJVDBQXgAd#6LU;FH9vk-6~yAg5XkC<PKk)l
zPags&6jAE29TW!R@S%-6B8tZ#m!E@)iNfuj__#|`CtvFUq(M?A2I8a(kdGW<$)h?$
zmXIM+{m5yqx4^?>e#Qx7M}Kj&u-Xh7`d^$KF8vTq56Ou)IP0GvGmxh?KFd$n^9K()
z(+`7Q_;~*X2Q7K2FAVy*$_<WvjO@=PU%n;}pntquq;@^21|mAf@o@C>P%1Adq~#;@
zP(UgujbK10&rXD3>VHnSQL+^2p8$u9hCbv-#z1CJh+X*(L>QWMXhWN-J0R2m25Ha*
zP31vujmOv+Vyd)4`Hml2M4ME6(jhuhYDN{59vVUyc=+%u4DbK}3p9+v3DfEbR5`Rg
zL8sHE0P(<7AE|A*&x1A})8j&(oa?JiYoVnzfP>Fr`wsWp0Q^un8b_M_KJ;;;Xy66@
zIJA8jg@?Z{yNz3VpA#C>w>p4t>&&hXrWpl&dSp63rx^zr-=_7UO*i!E&<UEh4O<_u
zjP#z*XBd)}`N;X3zt!FSsHJJ?auJsG3L4gM%z5(DnBbbr_Gz9bsBTGb^n-7}tN8dN
zC)q5-mL}FA9MYnxqAu8yHm(I+N5sH;ljww@!oexvh?7S=-I$O-;SP_E%fbF(=)mj_
z9oK)-FgQ|~)Q$3R1ikwL`Q>#+1OZ(eXFcGuE<&#I$m?V3f?(GZ>5-AZxHNEZ)LY7h
zFV`T$ZGeD?tNgjnV6E7VdArD*-w3zRZ^H-*z&X<KJjX7``Y`|VeJK58q{l|QrdeQa
zWA4CurtD1V%QsDnZ^67yABMK938ovS9mY8eqC;c?;>t1(31YoIq0Y;@!_zW)wfvMJ
zF#r!i@V;Iamd4BC>QuRUV_j<pt@Fm8m4{CrmD_iol<gh$Yb`~zTrDrXvQR$r=F2iJ
zUX@|79=m+*-Me4L{@uU&r_Zpg7iC$eagYOn1xXx{${o%a4C=6S7B0A4Tya%VDuwq5
zZhmyD{P?YRMK@EHmamkFsTC<SS5C)g$^q{TXc=^>#WipC^ZvmJk16;p_r3<N7S^nc
zkF==X-rkBMTrniR`E%baU-`;c%IClE`SQx^ua%8U7t8X>QY<35o6j+?7z?~t!AVWL
z{QQj5oA!3g-qu#U)4(ZAJcc>e06p3hy&PLUrr8dq$zWdcW|rpV7RutX!ZUO6PQ~2x
zY*x68s_|(}78(pOgG9;YXuW7b1~SeD1;&KKI9m2JzPx?!=DqUcAAD5a`^ksp{de9g
zPoF%F_a(x>Qj7C~b{JjE9a<PtKc}jjc<Ee&@K^)+=-F}jF_`7@&OlCvvWD`V59tR!
z7L_rW^Ft-+=P3?d2hHzraMv0`f(0kr%wplm;#jmy0O6Ex8WXkXf>B>Q?qqjqR9T11
zL6`J$7G{3{9X<x2f;J|?XcW%)I5C2WO8SvSrwcz#JlX?rdZ%mXBzaLs^1JNwIJHQ(
zQ~r!l@>i{AzJvU;T_S6KkRu+z>1x1Cl%q@-3E@L828>#ab>fkqW&%;>7*t#>9`uwx
zYC!!3JY)DQwk`@KNzwN)RJlmfrqWTALD!ZEorWg*_^bT{m54_GUjf23PRRl%2zgDs
z55_v3Jb@v8xfj7G*?n9qqm^v{*Wvl;(py=yFEC2}C_SMwzMtOZ_k2DrZ9gx7^z-Oj
z=+gn`(T;*vNArQR14hX+eV@kdYiQ|OnSj#_ZF);*Xz4A}(2w__&8JUG*y+Z%^oG9t
z7xDCIfln&abofrk|0FVryA3SMyvD~ITGI@zY{LEYQDB}vT`S8;eNn?SU0$f@H0w#+
z&T+tetxLLk{i-)YA@?<t`B;~^TqKJoj$#MTyjFFr;v>#AloKB37H7+X95-deAe!8^
z&1E58QWc(4bEHM+z`Jh9$y1Z4cl6U-t-M4LG~5)5_eA45%#Bs)D;|u9*alk_iE9sK
zQy4wtS04JcWd}m;p#8J{phtG(9O>py9)Nh_JC`@Hs!^e|SaS&ABi+#EO_*yncf?tP
z;J~rIz|l59UxqN%PdN!4jxqpkUc-?20!}k6ehm&Jvd(G2Jf8o}YcL+@%Ll-X!ULIq
zme*;fw>wXkeIW53!>ybM1`wwC)osJ^ZZn4KtgMJx_2cOo858s4>d#|kPDaJrg@v+n
zu%$5}qxR&m?Cc#CEnU?IOPQN5WpQb;T)DhdUViy<xpDo54C<xw<jIpV_SgQ^ubr{S
zZ)s_<<kuoqNHr3VKfq&9uw{Zlh%v!}nkN=4MDj5kgQ%44t?lyu`!~yb@7+>I*P?vs
zVi})WmIm|krGc#SLm3B0CkHYTa!b!Ki+nATF(&w!+=*~|abnWb7)&p}e53s8ul{QJ
z#y7rEuHU#Wnaf%vjh9^+3K$%G&6h&(6k~r^#)I%D+)+Q+EBi{{dHSU6%9!8|G1Yt)
z?`a&xQwpBMEG{pW8?U@vF28iGOfAfo)5)<iGdD*|=gTQ77bjG4y3@@3e3^`QEHJP*
zy=o@5ip=qNQ+37GlG%lsGC8e@MGLuj=i*dG1*Tx_^w=A0@!HusE|2a%DL?w&zbOCW
zyWc4vy!XB=KaS7U#G=+djZP-MGsQKrogAv(WC(FLgUOB$;GIzK@drT9WJp{CNOg^l
zv?ni*&&c^}$1J$<{q*7BAlm{%!dqZs(v3S7QdoM#piu_IW0D0QK<=qUEg@BxAVhTP
z2<x{QFa{5(8VSc=G}aIZAsvEVF=RlFUQFyJXQ3Q<)(H-P$62<6vxOXdxTLW$5r+ZP
z0ScbGE0JCu`BM{N)S!j<sT=fQlJw*scltwkWR~r~M6U7U`lP0y^%E#$=eI-ZghTn!
zrRv8xL<fNUr)1>j#Va4RW4hDf7^5dg=z}rAg^KonCRH@CbAiGkN3N@fuKs-YDfGj4
zTJ_NYSwK<{u%8}~sw$6&b#nT7w)JZuJjhqj(gLH>`p}mTME|Z3xN4-ea{KX7az=sE
z`ZVY1XMlbbO)F=V9_H)Q0WI#23w_>Jw>At|4}-(djiNQaO?wf4OXn~&jFpxa=*LHa
zY5KIEgr|j;{yaT^QF8n8UPN;qZJ(!42ekU&h{t?h@iE_wsK>r+!>Bv}WIB&Nw6aL<
zlMJUj5nXFg(CTMraGLXXd6F)#XZYr0LC-Zae9gJcy<Su^Q7I&T<Wa8l`C>IYA{G|s
z%hK|EoYoZcY+N^)SCeO2*IRZRuujUS#?0%;Q9eBFM_y9LjY772bA&xN@sNdn=$i}v
zvzXht))CGHyfUfRCNlJO3gzRPsJQC^1A^;#<XCY?;CjHdB<do{UgtxlBR+Ldj!`G>
zFWJU+t}_H8a}<zY`n+E<UT;!QMmbrBq+GZ&pAM1_7{)&jQKlY^%Q^}`M|zC|=NS`#
zWm#Su9+fZfBAuPqm*aW`=)WzHZU>l;{s*p$LmR*0+mY-i6O&pmaLO2m6wS=?&V>5x
z+^N>Or)73=qCDk<(nGBq;1v&DC=Ch56~^*dSzehdmoKl!DX6<U`(^Ck`qgiq<&mlR
z!VtMK;CM0t5qIzTFdM)kWs(=486;ZFu<ZgCI@H65Ps%%Qe-JNhOiV6G!Br_SCne^j
z<Z@Y9UXgN#G6c5E!Qq|^3I<M|(v-)n%2(Q+7S+5X!S^J7@fW^T{-uBE&zCpfcrzC6
z93{(l9_~MUTpm1lQ10EmAI1bPW`{AcyCVbTK#TAF@=X3S`NYvXPEJt8aG!t|urFP>
zSgv2cUcUOxpDov3d$pWRPn5kQ(Vv}!F~gIQ$I1hrhM1a}jTgjOl;hZ*w7#}lHa6DF
z5=Y0Xf;TR0#OY7t?B!R*OzCH8aWMzFWRQCqn^G7zUhd!gsC@hPey{xQZ~s<#`<=H7
zM}y8Ui9ZIK1}cj`Ed&qLPzQSlG2zh7_&u&qit^+s8aSVrh)vOn_^LQxBcF*EZTY(R
zL>OI9R5pBG#HF4wC@{7N8{Da5QLg_Oa?vkpN5%=#kj#$krzA0F1y4ef4tI1Q;uROa
z3WKvVp2;?{58r7iFE4<hopkl2R&^9umgAoUV&avfzQTB@1Dd>SOgK)Ma>t-2Uvy9d
z@brcY1W(RLJE<?>Mg8$7J@W|<P7PM|BN$PICn%q?;v$z@)TZLe74;B0Nype@8wNx^
zBIi*#cgeVzoSq#^KV2kbJREVCOBboQ>lMbNE;vnjLVwp&<k10PV+Oy8jCzmih##TU
z7t&8k(?rw7Pv$X5j&M4I+syS05FSO>LYoHk;|B9tW+<5o3KvPw$?oS*yp_{}>CU6E
zj229P9*j!!MTa!E<BR09diCKv9>8fWjb-)29uok*@rE|;^p?McQF2;(<4AAEWgl8v
zLo2)WN744-<9J#;;5g9I_T>Ph<a`qNFT!VC%<DR|YtszIjmo3tv5Zk*nItoJXEbq6
z1I*u{o|ozTJPw@hu$}2VKQU+LT$p$MS-)_-WnN_dZJ~+`z>wfzmt`3d9P~0X%MCjC
zncwpcyYozAVkY=vR{>j<xqj=Kj*DY8RVJSN$5Ug`a8nN#287BTBo)vjFHIcC!u2-e
z1BiUHY_a7^X|Z18+MsK~_;|Z*n;LiB6Rvequjl%WICTa!baBm6<qUYBk9t$NRCnkU
zP<1w4M5UNzp%d?vu!hX`$CyJ#8e?wDg!HrRsz5=uY2q^;-g96%Zj%noSNjMU$@(~B
z!ak0<&gm^3VA?+3(DHfCB%JlB`d1hUVG!ciJfX2OE+PB&FfQe%T*8tgoR$W08Y;#F
zhTXyfH^uX4*vaV=El<^tC$#3&x^rj$tjN<ELq^2%Sh;v<rmU{9&eVGGfc58DnN=4*
zk;QR%C`0;x_SgQr@*~6i&|BQy6*(0f?;wOFIWutFrN{Wl4$9-cX$^*F<>oC;ae6o2
z{GOUxQbo*bkWFb}JXhuymZhxX%7=FY_6~N%lOK@VkpZx^wW)lMv~b@lhx@x4XlL<4
z(|`HT{kd}e+SM>57(9pEp_jG9ma45yEr@sa;$00E**xVq+~3o{<y{Qk!H}gjsmfHj
z*cQcr<t{usa@JN>%coy^y?pCi-zwkylfPURuU{20i*j}BvD5PS?yd6fyYG~@-}-TR
zc>jSG+R{~dut-l!{8=pA7G*@NuE<c*!kzav))Zb`T2!6P#1nyyix<nr^&4e^7r|p@
zU@)sOo-xvupX7Y-lefyR|KtBx`HkQF_2@iTRNg5KG>*XK7}~?#gR;A|6Hg}GIe2g4
zD7HmCt4~WN%d89sUWO+R^jutAQhd5jl<FL+_XiIjmfN>)hXKY{${E14A@58OZaXX;
zJKq7&1fsC-P5|`Rq{C-|fnIT?B`#^YZr#KyGCDpH{P1!hr0e1^gxrWzwm=0h^(9Q?
zADNB-WLR$zJx7*#BoL0lEf*Kp<^t<_!XJQGQ#N;(v|x|Emye~X&rS7fczH=KUg*=H
zCOdHP5|lXQNv0-!xqR0}NXZ1xf}b+;(W41AEYkm!eiC&pe#WPa4EO`Ymj;atiSaZT
zgiznl@1CG5dg14tu6*pC{B1oYgen9WG<YcAqZtK#Il)&Ul4>$^b}c|}XK3mAX+SH(
z;pgb%KlzA9dMz7ZnfCiKahBCOryJU^(_1>Ho4<`4`Z&fT(DL}gOP|D<2QUqQZ)bTe
z7}utq7X~b=ABN7(IKccZ&d~DrVH96Wj}J|Z*-C5Q(9#2KdY=|BeV?u`!|5-QKMGn}
zpil4gex9S?MKrD4Rwm$h%S&48`aA%*;l*AD=4ts%=d?a#y7AAa1F*r{ZO<8oRwp~-
zTY97$$2IXp>zwX|;#|a$uKxkRP8@%7YE0MExqL@L*L@bmC#Sr(nEbfp@wFnK5{J>E
z`ImPePdNVlfE^jES)_CQBM*!SuD9q$Jk*e^C=>HDN(h^2{9!~_BS6=Kxc(|F)DS7l
z3oTK!PT|YjTC4H>E?qyksmhJVd?KOi6#N7buRJJEzJsCb1Uir(ZjeV{!^dUWwdiD<
zdU6E4A<8G&QTN385%7^S)^+Gc8tXIGpKeP7vYZ#-G@|6&kPd<I_~vQlI?OtkF!F6o
zI4^*3%LjPAv0hH^$43F_tsim{5oOP<W09nM^4i?$2sn?-C-Y2P%1eCG;X)p)@-Zeb
zBIui}K^A39%r8u7&H1cMOl#dR^{gDIE#?+58OCK)h)-+cm9>fT$}1c3sRbSe>>upM
z0|B<^Z*9p4VSN5C|A&7`>mTm;v(1G?K6f-z88<9e<l@r2GL<#XLb=|d=Q{~(g*sG6
z`rw1x<@-N)yF7fjqk*)bjHJlnsTM(V8iXs7rp{YUY3|svxIH??9hR+~P33zKABE$c
z2^Q)XF07Pa`o&+6g`Z!%<}N*pE!Jua^ULboQvQs^(s+4xbX@j#cFOkF(|A9koSn#G
z8ZYY`%jKn)ub1W3<v3~T!NUh-cYCL-udJ1?e&ws>=f3&#W#f&{l*##ddX;KgHLgXN
zmR#lZ=%C!b^<nww!w)5QM>=sP|9EtGzN(EO#AcGX8!R2^61=}btMSFa`GuwO#+#oh
zS6+Ed79U%uQrELnHMR!x*?}5BE&Jrb-SXSN{!hw3`p^G<Ing5U;+0GB^oPft2Qr%W
zWH23U@5Y^FfIG{)e0{{CP$K6v!7a*&;RAb%GTIhoL@X{YM}D_&-zhh5-V6ic!J~)q
zG=gJxnXmvR6i?Fa<Hn&o?BwIl*m(oUp3(v7Esk)X#t^hhv`oS{29O6uIO)ljAupN$
z^fWjO?Rh|}hV`yLpI8OJYd(DBBQ2iPG4d*KaR3+T;$h5iCnt<c8Jb+2xVw}FSjM5<
zl`+UkNp*h6i;@QKEL+bcpO^L6a>u(l-1R~(jxA0c-?}irC|U3mUr^tek$y4)R1OS_
zgZMrPNB3c9AIA2-$7*Y~^QAGOUMd}JM0ia5PNd^1{h%HV#Ik;r%J`@NeM{Tt15#$^
zLlbmdSp+S<^#RiC3<G|_i_#1&j&=xo{=;>=Er9RKF@SIR2#=D}=G&(s>@>?W&U7tJ
z3m%)KyF9chFeu+S03WVv>kb?r#SgB<w|arMT)RHK!L)s7@%CH(HqCg##se)qVLPCu
zZRrf7aII{hFUxU!T$@J=qhy;N==1dH`*KIg`=nqw=jk*GS|0kBq3uI`xubM$<+o{u
zmZp=_U2sw+X`K#znK571G0@5AbQ#dg2geK6`TYdn5$_goEx?B3K28`CoEo{nO*`(+
z@?_&E>qbOgXyce$b)||0I8P&<vCUNZ@ey9;<$Ow^Ihn&%IHyI@Tr>DuR+UBeu+MXR
z{gf8nT64*l$XLLkx37oXVC7o?hg$U=9dQb#t_gCKhvHLm%E0E5DP4Qm@xjZt<MLUj
za6BzU=#5dqZWEP4{9JRg?r9FsKXwL!l)bG>C#yBWN1g#)oX+)*T@zgI;mxNqEKE}$
z0I#2?0bJV`hM~<fY5KIp!3@&l_r(xAmEJlTM;+ll4hCfrPnIol)PIH}DVC(M7E0;?
z^mWMbl046+F+PasUu0}^nwjG86&aSITUeY?Kbuhh;>Pre)|2C9VR<$_WVay0nfEhy
z_q8t9dU$zNYs+PB>}$P!uv>O^cFXRb`t-?IxqNx0jQvM{{a@AE@iZ6s2TGuK2dHqN
zg*u=zX8#|1@8ju&22(CZc<i?zVQ1yTkM5Np{_vf0@BXGLU{-Rp@IF3O`eJ$g^*72-
z{nQuB+dp}yeDL8%Wpi_@>}uhCpghmUWXQ<U*+XIZ+*Rd>wF?*4%h$jDm9o6Fpa#<7
ztPa3aEym|&moy;Qk2oC_bR@lZct>J$Q~XC-biuz;KJ&TP%QwIE&2sh4PnUzIPvSJC
zhYuc>V=d&DWRSf2@~6s6KlSA@EelynYt~?5kic;{S?9+`d?;@xwiv~uMWJ~qoV|^w
z>hwJIKsTO>a9R^j094bev9U5M)9%wUR2J4RX%5gx9@C(Wtw_q_RAJF^SM=fSkIL`<
z#y>6p<e&ay4dmmpvc6Jg7i737FWO^Y#?PLN2~MKQj;0QxPB6ivLJh=K)zj4**R|+W
z!)u`yCq<~<Kl<oa`QeX#RBqk6CChOuWMRM^9AfaON6}Mbg3gY|Ru~M5cnU!tQQmaQ
zHmHJ9Kj`ryNQnt;7;qdBtr&;vI0(8N7&MJ8XiU&*BVRQ7U|iq#o#(i-QU_gJ007k&
zWzBp#2M15w^6pLYV@z--1zsjTCcX68K2+O=Jh;Op$M&+H|M{N}S$ko{^R$2qlLoHZ
zjw1yzBp6p1<Z~KZ92`KitgWq-^|iIwjz|7mo7-h`Yqvan^t3#FqJAy|<)|9hSt5bT
z$qRt{`@3aJef`1R`)P#NJ8mwUC%ZveYnsa$<%m#1;H8sDf({s_&k3sq%Qih=I+JD+
zKf}C0$Jg;g=X?gT6$RDLzXj4_LdKb|g2UF0c$<fPr;UPE#(B`n0^0mqXybi7d@(2P
zF)%96mai|*&fxTx4lrFSqtAy9rW-}yhn62O?I^s#at)-p9&nanISvEn>*EbhAC-0<
zm<MQiT6`<F55}3M<pa)>+2Z>$S~xH7QMAz5k$xV}D7<kkO^dTL4&Ssbn5N}7wDg%Y
zoDMi2(;~xjdmoP@M*qocT|Sv0UZ=$808vxZ<7H)ezAP_eNN|)Ui|@QHGFiuab;#xk
zM63zU;;t{(cb*0x?z3Y;YaOld5BBp31!|>!$^<dWXKrCq*No}1w7gU<Uf3v?E^m~L
zjn#5tW38;NEyv9#t{v1x-aO)o^N|*Hyz|Ht66$9x@_GDZvpjveSsp%o6kEX`JmjOg
zQKYO3{c*RO8(VX8Y-49f12?Bs7ePcFvIxzwMmY54`lj^`D2kuZ)oU7P)aM{^0OT2R
zG8zysvd6ah+~tCdsKY^?`X%4S^?~vj%-2Gj-iNG<>euk7JPdtV(3;nA%0Rp?r{#4D
zB>5yN%i4+$xe-zx;dju_GiVewzD+|P8xuJ0k77*B#(U4S|CH9M^o8R%9(H+cwya-T
zmcJM`;n~9X;PHcU`}W=P=n1-F<SIX{kx$TFJ%;0AUwLJ%jQxlI&VN;F(=2@Mr=~lp
za8((I0q&I2_;EUcbY&sMU|d>UmSB#CdaT9kds@6c5D!l+_`2R?*|>1AeDfRMEPwGY
z{KfJ+zw>+LAOFAqNqPT+kIMG$UfGwm!y~oB6D^u$>F|X;3Og$+d|iv`Pk;JVEvlBo
zI#*RGox#htjv4909>9sP{EuXr?{j+7&bG$+i3a{$`P8Rwl&^mEOXX*O{u^cS+Ld^R
z;_>|l<>q@Imb<s_mvS;*E@+{D;p)q(oW(ND3)Wc48n9C{c}_GI*v`&YyqG*WRjqgd
zbwlT;E;Pd8<#A4p;xVQwO%~sT$}m^1zw~N+`MaDk0l}-n#h_=fsyr(5yFdOx`RBj&
z8|6FS{Uec_mc_LtO(@bwjm^T8iRehi40|j&#*?i}7!h-eoIJH$RyH=|uf)?68a&*s
zJLLyI_))og_paJVi&*7Gr=?@C_(cT<JSRfsgrNyagP5lfM2Ls4rY8s>0|yv)9+>Vt
z9&nLGFWiUV)epmp`Hg-WhK1@r2aE>~Mux*lm+6qCWDhL-K^np-Z1B{l9*V+&0S&|>
zU_tw2!joPa5!_*--t^C8D<h1%Mwc#KDu4N}{FQR)%4PLgl-Iv><mnSH+iz_?iKlSr
zJx_b;V!`&j<)y{)>DOiCzWh?W_j06h^5n@=8R56ecfa=oweJIs74|Hu9o3K6c8IRX
z!4T$bbH>804{yeMMHrIk>L(G5TZ8Kfmv=?7tOlZii~12A$nbN%;4f>d1IuV-1Fm}j
z-}pY9hX?wyGwZ5LBeO5VeaCpywsHGyn$r#acng+cfY!K{&v@t4g5#~Mmfk#eEtto?
z)9sH+e-RDs)zb9&7zf~5XIj8ve5bcyIf;Kh4R}5+>FLAxkfBHFQb{Vbap&bQyeT`O
zO*6FVe;hF1C|P~nC>=(@i_-o$(B_HEnC}|+B>A{sv_QPYwe#$YWVX7PuB{WmX)P^4
z9A|w9<F{oRlBG0g={q(ia^YLkd%%1;Gd)o_t!Zsl*JMt6)Vd(vLt`6|WOH|4g5xi?
zT5D0ywO~@$dbYbBvQ>FU>l|Ibxv_f0_q9+-4pz39T)4Piu3Wub*5zYJtgWqU?Xwo!
zyt#YK-FvPNyf?=MnHyqU)BT}a${6o&u4^5zxL6KkXmGQ3duKP^{bUXBk**zgblrIL
z=y7@S__3}%2ZaqBT$V77!bsuz$ym~LE{?E`F{S!de!js-{n=5DC`&-L3H8ZS9}bEc
zWl$L0hExRY#m#BfJK*DqNNp7Iw97nGru|Niq1JB&lAovPfHbFvZ~5`l2oex$B@N=1
z-_VC#`F%a@gC&V{8xugEuA|0FACF|yB-9&b6-S*C?%IpYgjX{6%Q-NJoe<p6oSk6{
z^@$RX!kE~&v`}uma=Bc*vJpEf?mv_varZ%a_-Ic?>Zy!X@y~MlQz=?Dm*o|{|2A1x
zSLe&bpZU4Z{u<hOU!J$V<ogmVf~(T<C^QBKce%H=x1-a>zC-D<w|}gTx>X*_@v${;
zi&$FJ!v58l%NM`!1r30s@{_mTE)RGQK?`s!YPLD?=#EZAR#;0s(&Hs3sUpyYBOJA;
zGP5X(guJM*IOBT}JmGluj3*r@vht^8*{{i3x>jC#<$9UWpyl)%Hizu84MY|@FMUsr
zPnF|6j*&fxEk${@b3SH^POSlpo}3<aBpt~+w|C`fPP23MC1)+@*z(0)=7ojDva+<I
ziDIS9%r8k#3=S>gxHF*z|F#x}PZ`t?AC|ixykFk>!S~At@4pjIK^B)dRV(*p&M(ZC
z*;%%c%}Cy&7N_|o?^i$da(VsJpDM4t`FeTz)329H7cZ5a?Y;6t84ut2&iBfXfAm&7
z0ebhHcjCo#CN~zmv>kqQE`^y);~f<F)MK2?#bOu(BKvJXwqob^fx4nOaB;U?y_M}d
z(->tk$aw2NJ31Xc=~>)4?7YsC4?qGPdWX}3J<HRdJaRzU1upWg;qa2jIK{$uoJ1y2
z>M0B`>PF6!8DS=KTpqK^-6*AnLpLsNY{gqB*KfR3*4Ed{%Ia!-iT(1`tL5UQ%khLm
zIfg>?C*HA{(Fk2xSymgrBt!Vsa_#z!vW}s+w50yPw!Sg77x_`gY>84oTU{?38yDh2
zvY`vm+S+P-<(-QQdt*7&07IBNSr`)l<p5kqu@DQ4Kq|RDh!7@5&4{{U)2A*55U7?)
zA=js%4FmnSWf|Y559L!~Kv(mOaKlTP=yTEUYWue`TNssQ8Rw;a5@>l_p7X*$KMst-
zwRJ-Js5<IHpU3_v`39%;<&Wa|q&UF%PJa9<9#H{^^UB2dWh~}+?Aj^Q;&;H><zv?Z
z{KmEEpQJwfVE$GnKzx+E7T40B2Ys2qC*`A#0Q0r{IN~3tPoMXb@QlK>;QY^nHr;u(
z^ewK%J8lPETsJiLabv~bk!aK(0z`2GleJ5J1b1H7Voo5M$VEN<3{Mv_rqg&TfpLN{
zp=&DJU5^hA<KUEio+2OQ8VR<A1$Ip^&tEE^{oHHi8{ha!`Nd!Uh4RK{KT~eJ{BqoU
zT3OZFLDvIx<^~_e&)(jCY%AH`-jTl-{?^uZT*sGmjktF0TDfxNN~}rXUDmZ^L)VWh
zmoJwaH?EaeUcFYXU%M#7XtB&`UBQjv@v*$Y$omc$0*t9plH#dsaoS(hq0muaCDbvM
zVO)ZGSsjbrO@{NpZIvXAjL6IKGj2J3-;ngfI0H$;c_l9WE-(8eLe}O%*s<`T=?QCT
zJl04VnTJv$YzIbud%bbo?@HJi$iBt5hD64nzji@0t^Uipj8_LZ93Zyn$G51oo{5{x
zr)6$_tgNmrsUJ<3r<+g8oqG>tMC{0LJuBQRXYCf-@RrB5CY>!;udZm#A*(`pP5jxP
z|J<*I$s(Xq=)f$BTsbVBa^lX#9Cx<awy~oHHZKV)S+wKo)Tgoi0*javm!g6-g4nk`
z&zJTlr=;9w`N>b-E_d(U)qvQJMNWKqSASXV(^!0BjV~^-Z70*>E&@gaon3NS80ErP
zi>ot@QW+5A<67iB+b@U5du3wkERJpC3)vSgZ<JY02vowDM*hBxpr;QX$2TA5X697E
zQ_^iZCYhaGPLkRSg9-UOMvRZiNzY^9lw(wy>xTh^6+bCQ`KIO1PN|W`r!}FR%Gla2
zTgq$e!Q=Aaqg&<HJMWeEe)5y@_K$y9-hJzb<-K>_iuW>(j`qsR+M*Vk>soxSl+|@!
z<era7j0NV!ix<mFnrL2n^_6m6?&^&j<$@-T#pTuV`2OSa&wu;(%CG;^-z?w#_IJvy
zo43oR3>H2(N8K}7&{3%8ylasAMbnVbBH0Ubm4`OQ8Qd4$C5dx4!&`JP7TA);@nZAp
zd~7dcpC^ybc&W$X8Q~`;IBqz3cyye;47@N7YX*Swh<xR&it`7oW3)l&?LNyyn3Nt!
z-9i#9(ku}lWXfm<9S~9_FXZDZFK3LgFfcCUyg1Lt%=mfHf@ODiuRMMFRQ+Hx_O$YJ
z4}%j?JOXFiUZMWOX&}#HOs<5XI?lU5C;4I<-(Y$8;7Qrp;>0ptoOID*GQcR|y$WQ;
z3)O0G^o#dxWJu)qTDte3C?;jel83VVcJ-UlfI-ktb^@bcc@Yj?Tyjf#t!F~v1-LC7
zZhh*?{&_S2;no?aAO6EPjy9*BV?vGgue1#~-%;Q^><s7e^ko}be4pmLG{XISM$tmk
z*4ZZo^Vk9B$?ydUXnE|6x9>2}mZ6^p7}w$*G;78?x|SCCj4^Pnp3ui{k2i6{GDd~_
ze!m>{?Eup_ZfN5zy}`JCTAL@}w3Y^dZWLdij<|6xP3yqhnbu&QzTZNh2WaJgQrCjx
z;Va9~Pk*M3c-4i4G2pb|sd3Y_hJ>F;RQ7<%5<UyA#f8~4By=6-x<($%Cz2F;z{_$Z
zwqMH~A8Ap3plhleFIeyG9LNA*uFka+GGpuYn>WgrzWmwpm9KqK?#tziKl8=%#%DgG
zj3=a9o>pn2Ve_fhDo;1dlgCfX<Ht|R<3~^8@RX;DKT({c@p%uDn^CM0Fe<po%epML
zKJm`#MBI$JA|vJc^{etPiFU25u5u%klb*)3#^IQFuLX2{Q~j`3iVvDb)n-Mxjwp}A
zK*qJ0jQ2Q5kR0&`UNN;u{EY(Sq^xXP#xw0@T>Q#&JBG|iOL=Z<gQJ~qy6^;DAF?wb
z|HaVJCNFc3brX(jkLtiOEse5+N_z8wO<dggQheYC4e`8=`Q69XIn9QI{^g*n*FF#-
zKjk3<BO*UkE5}xAZuCylqgh8ki_J4Lb7NYIF2$C&=u7I4j1`HEbyge?E8F41#VPfr
zr8v!$9Tah`{SW`g-8QFqnY30+Dw5k{qO$mt&!XP%SJ3$<C$MnlG(gYP;Wx|85AT%Q
zH}6VO@LKT9E-2&6@wGe_<?p}$VYzkZPT7>A<Ta&%Kn0|+a1@UfH+XMgF5k`Iu0MK6
zSA}Qdt}Bo8fT+l*OiGI}vn?y?krrc<<x{V|RKEP>FUozXyz=Jf%A^|k^zfuS_~2&w
z;K%R8Q<UYUHCa|KmDP<)GLVj>*WL0!gLG<cw%qvCtEytYHhoy8rOWEZYI(Y~6$5Qb
zW#+?nv$OLuNM_5d1|9EiAIlIq&_uAK#r(s2_scCA5Zh08G{GE|{k_9j*zuHOUj|Q6
zBVy65ZLE~F3mV*#zqqncCTCP;<sVNwv`Czm{Ka`5e{Pg1`D5cU3*)8pld^gHe)&)T
zlfPg7!GHD-%TL~ZM~2)^8a?U=Jc+@8;JuVX8AZ9xi8`b6)2{NVHriQEL0P7C@~t?)
z=>T&!FLA|(PE^Nn(wFLnEjm1{ArFQS?M3IyPOgY#Qg)tvgb@H)m5eT&2*9=-Cf+>$
z6eCuhoXHI;{a6R$t=|HRNbVl_v1xo9XPgU3_jCsEBUqw|OFcu^3hmPmfV3Q>9;moX
z1E9+%J2o(Q>Y(;h8}gONmt_oo`t{e#tFOE&!)s3CXJ3ZFZY=s^53F?LP7oh}S<yvd
zV~rz+7YC2e-+S-9^3g{hmdAX91tVRuc_b}mq8wo?VniIOy)`zscQ(uR)@E#T;@Dp%
zchq*7Vw~$Q${mwZcSng*<!{0OF9)>sWC%GU@IW7LSy@jBEe&ao<9l*6pVPq`zyZeh
zd4RtEB7oNpupGyMQR#+JwDwy((DI(g*QPscSvccc-WK{i9|xuf`t+?V17YjX0(jb@
z-_~~KY553~W+3RO7lo#GTPHx&$!%pAT3nlFpO!Q`r?oT=w=(+B&kNoAGFmy7)yiyX
z3{D@#Yx<UlFnOD<pKe~DP3y}Y#c#h4rg6IEwEV6Q(*W)#9bWlz&48~P?nAmT&ojp-
zG<-_;ljV5FKt4CoI5DT0-U5O{e5^B8mgjW6UDV=yLDx67k>;XJISy_l#cl%(2(2m5
z??_>;V;B_}6FWPHl6Y38#XCPU6)&iM^BZ3(U;gTs%Z*oFDvPVjF@GP*5P7<}UEY7^
z!`NZK(`R-Qa8n3_Ag&L*`^Gh#HIf`BUveXLcYimozn}fW=gXh?rC%tY`qZb&M>lWg
z%h=*!-sk3Teoug#PuzH#67EQAwa4*3<*l2y${+pFcgwr)JS>}=d77s#FIsr=9f9RC
zHl?*j7*MJom6@s=+)Ps#pEz~R4K2nZ#u}$=awClEj9Mi4sXrip6KE<N^&8vhB?BLM
znMYD*8>W<)+ZF2iN_Ydgl_lFuk)Rh*Co|K)WnW1k-T4`up8QE;zQlEUl25w>AZRNu
ze=O*M;6qzyRd3h<q_yC=t*whbO^Z(#$Z!L_RNgoyIO>=&BYm_6#f^=#ljubK@wCh@
zPRqFcl)|hL)t}>=aQj-f?rZ&<4+$oxOIcl;EvxGbWp;ihP7}pA8T*g^#(ynTVS=W?
zQmHyY>wnO3B1ILaKjZW-K9uEX(XsO6(PsHji}m;4yQxa$1rJ`5=8+odH2hc?5r@ae
z)RGD!>qv{DyjyUj0zH$VpoO~z8h7QfsM$NnL_wZdRP*!mvG`%Zj8U<^z8Zs`FH*8y
z>gm(RT2Sr9KF60|x>jynyH-B;na`Hxi&x6Q)2;GE3-G6G#X8WU``LK;{FlC3<~J_K
zqsh$&56g$QZkKzH9+ve>7d2S-%AE9g`BSfy3zsjI`+VC%Mgj)K%#;iYYE>Pa6B-Uw
zu7`K;miu?ae_QV6?Q;L_y*S2JQ&gFsWn0Rs42Siyab=@ixN@OfzIM4R$)K1~gYdz-
z`Nah_413sefBxLUk_?%Jcqd%ehpJD8#HkwPbWBFdR{4Y9_^tA{|MuT5-}~P8!jPb$
z&(vvRkG|>%Lt;-0ON=;PmZ4{N1C@?8=yTD&*~SSu2r6DbgmK)ZS>MQ0EwFg=g<_u0
z<QFaJ7xWz#@6fZrCqM4QBQ>{Oh4g6C=wu>fqTq4vp(aqa&hhkw$u#aF@${=v4(`6b
zhR~@R6md|1CUf*?sgafCZ*dt%{~VCgOUcH2pi(yc<iTZm6PFiAPPU*f;3EykQv~iD
zEzU25j-(?Xz7N6HIkh>bfyE2mO5=$cpQ-1)4Cb2nCWtO5$X8}%O!ORG8{-fiLnBR6
zDi?+cr!X8I<xBU}Ig?4Wi|C_FK`)TnxwSDTMyE9t0fRI-*9GAgkZ<Uxxy}rId85Gj
zboJZuIn4kLFuqL#oNnJ>I{Y@R#RJsysC?USKaV!eU|e4gJbga<eYt(;)3tJezU(%=
zPir_YFJM$YgpF@`+x+@?LqD&KQ`Z?cI1gWcEIayj@HnsGGQJ2#(Sfs0eck%>q~nY?
zJ@J;-bZvUycN);AGZ>HGmAeDz`#hE%b4LZ!wek$+2Zw9rIo{$7eVpT?XbpW?eH{aI
zy&#-hj8cb~_bM$5Lt&dT5bw$y?-+1XCLHsnDH8CRlR1K$C+57bxuW$0TbsCkNJ^c9
zSnuSiLp&{xJNVL#rxSa+&ND|d*Y57}J{w2S&y*{dSIg_Ky;8pR^{<pa^Jo8bxp?_v
z?9jOT;C^}h<Vo4w+Aeo)-7njldu4Z<8%&Z@U#i9+;06)f!_Q<q;2&#FXUi~)yyIs_
zA@A$o{MqtLzx>PP`b#%tAl#N7hmyCIPs9)RRSvB$R2JU(U_BK#|Kb`s6?Jp-qub>V
ze*Zh=5B}gg<=uC0Nr${~NPW!A&z30}LtG<#z)Nob&H8tpb3MyC;u?_G_IO&O>p0ES
zwd(Ugz0$t1z-7TJ-CA&4BQyJp@(H{RNP+Oy1wZK$hi@R>rW;IaXWX!#*EgpFeR`b7
zB@WKO>yJKWAjpmI7lGwjzI_0^p`X7nAkzuZ2Sa9j{4BPh^PbGOsyKFwOw;eB+m!n9
z#51j<$I7KE>spucPHt}RqxE+8cH%Tt4)CEGre?KH)pC?RxTiG+$0d*bNB`bm(fT4M
z=v)}_s46esEK2f&ENOYFFU7zjjCmmy)bO*aq)Bz)qjLY=!}39V1^c5gBzQW(%g1yk
zIw1T1uyCS)ao0HCf1o4sWMW)}<6Vgh8>>=oAznOU%M&`$!Q!aQdOR)ZYyo0_Atygw
z)Z!nl?mxJz0di1gr^m~Z7U;atdSQJ-6U4=`#pA+zkK++!?2gg`{f*Cmsm!fk(pr5w
z?lkN^ep23k_Z<z)m$Vo^VBKC;);LjNT@`s$KKkfpdGzp6yvX!K{^N%a%OM}klQHq&
z?wzvv_>qi@Lp9n&+*w|fLBKa3R#w-_g^QQU#jBUgg&SAm&cx!1j0N^$YOwNMib<7=
zhRaSt<1>iWm|;N3oynl!Wo}N-dh4C?zx&_(&GK8n^PA=2<A)lwxzObC6z6|Y*N3WC
z?qcVQ)*5&!3lr%8YJc<OIsiq~Eb45k7EP{Lo_vHS9*wCZ##cbq7U~gd7q;$9#XvdK
zB*?M3$fXaZC7gFFNJl^J-q4Xyg9$V4W{4*49BJZV!bd0CJobx<hGTsX4!9sNKvP8n
zV<%AhaT4_xmbhex&eFq+Y05_bMK*p8KGlJA!-32GPkkq?9RR(@2axaBWP5e8@*-xO
z@zepvmE+U`p{OgyogzF{z_8`Z>rAX%ocQP+MndEpWe{PVh^Vi4B7|(E@u-<eO3N*k
zQDawQnEYZpnON9PH7JYdo--tHkr*TqcRFPPh_^fj@&{U6AK(K_gKwO{I6LIFvRnDB
zyq3>koWm`c_T$pQyKl63gK7RaU>TNkUfzJ~{o|n3uMeZr+B^W$;Dhh$WOz|NeK4)t
z&j9{BIf=J>9;lBTU&D2nemA}h&Sw;~w9vKk0OK46`suB2=*uxc_i?oS^z#6oQL+Hz
zExYZ%fbj<RG2@}NYw3VC{-QJkIOAKn?h|dgfph@;_*l+q7!7WFuSpU*9hrMufQ|(-
zPY#Y_0Sq;A<tiytX@0;<3vJ#}T-J44izD9M<TCK=B#&h$Kdu{LoXCJUInp&t*QHa<
z*Kv*7<1|dsNr%gqmdhJ&zFNNcm7gv@^VKhx&wTDPWm#*Lr%#`hAN}aZvFP94+>UQ$
z?e89z-A&%FNF#z5n7JPE9tH*l$I&0N!{Cr77kjFQGvSYPT^=i+e)IM63xDz#%Fq3Y
zpD*hdHp-nlcVl}L+n_Ke_@0-`rAtFRp<oL+Hv?y9^K`<=iRn1`=?6dfarvV^<h{#x
zq~G1Lx5v&APNvLV60`$Pby#9hk6d@+=9K;jdrWfeR*Qt;g<*sN;%j~f!!g<;pHAd^
zggNKargExx+L5O<L0f-@jEM@-w`v%nT^~tsKM6WH;&#TF$9O}Aldc^v=xRFiSO%cf
z(AjvL`*RB|kL8)Z4^B5Yp1$bGBQQDl#`;Hye3z9qKJE6b%*ZI1pV!)MRyvMrQ9Lf(
zoNx=gYA_%D2ICbYf`emNQ*c1h+~Q1`79HPO+1}kz#>(%1{Wt#Wk!zkz!}|!ih(Uc8
zA~aYy3Kf6qH0n*nsiv73UX-5HpxBR-gg*S>whW3}Dh&JK*`lNls5WOG9}6GuKIan&
z-rm<pRYCL40r<1Bwzd?<qOlOiD&o=H^i&>8J3lwC#W^PztjmzVh`68y`i9(vvT<=;
zE8j^C=p8MX4pniSU?by*!LNbN*~<IUar5z34BE*t$()`quYUHY%goA#c=LmAC(`r&
z`yZ6q`T4T9zgM<YQ4DY|c;9>P{rL98ojbS7gM0VOlSdB<N7Kg1JgOJ=rLD?HxOVkw
zx$*K#@gnV|%a_X9hRU=mBT0JCXi}Mw;lN^-Dqul}aYbuKM~m*qL@UR~B%Y~(j_DxM
z;ozj)|L|7%XTSa%<?sBT{*UG6y${Ro;cgfcF`%P`IBVMPc_h$sOav$Ks4%h$h`Coz
zs;G{*8^P0`d}jrNZ7%LiXfjc`(g1|9zPF-&!GHk69TZLxV$$Os0KVMK9UKN73tZY9
zIwqK$DAevUcIZY7fG{Q$$LB?N-W#AM;_ilAoYQ@XV^ZaA2PNu&Fykn?Weu}%j}Np+
zZ|WMKd`V<bc=GdrB_8!J+<C?XVErS%R*M=;(wtNxKP6DE_#mXx*vHCOD5;Q;t$b;5
z46kPz3tZ$dJn%{5otMarf9aK{LrHH*j;A0h3v$UTjfwH>HnhGHQr|46%NXw!NDlcR
z$7LiOpu7Or5n!4&{z<_+_$}XgGEDEfHkjVN(})`f@5kxb>f^A>+NJ@P-A_04<3O7)
z>3uyc$Gm-d17Y(0Bw!gW|0r-ezQgd^wa|~Z(B?BL?^a%$7tl|89!Tr*HZ3qpZp!HV
zmcE5PepKFVpE?ive4}Vb$p$_NEi~ux8%Q4|%Q*X^=)kq|TUrCSkK;4+aieGq$fQrZ
zo=t1nxK<zE;1s9k(R}LLjtOW<x4yWLDG_l)O4r7C$+%woVjjg;6fzdDb21$K9SN>~
z9Dc%$vdlBDH{==ftsGA!jt;e0-rg_Ic!F@41_m#=Ub(zpzVtJnFJJn~7t5Ew{xjwE
z&%9Y?v^cwa?_T-A4}VzRdF$<ROV{wNCr`zzSz6cjlOtWT_m9e<j1zWlgaIUjf$Lp<
z;1?r9*MZZx3#x0vsq~#z9$zSb@>hPjeEny?UKV8p-MxFi+`s#vJbw5%-kIYG21k$c
z4(Sxzofc-w(vplNId05yJz*<4H}dY?dr;o~$-6Rk-YVb!{twIh@82v>p6;tY$Kt6*
zKB<|CIMvU3gZALs22j`3SAM%9cN=g6mYoK*IiqdP1DcrY6i?-OA{5uMXh+sAx^CfW
z`vPi_=j0I^(7+q<_zhVhFL0e4`VROO2ekA+lQhK3TF8C5<p<yn3CL1ftfS=+YaT<N
z7ijx=tBb+7HqAlu1475pQ|mj`E5-yzyK;K!^vrl{ZJL!)F+K&ojN@_?^~g<Q+LC7F
zTXL+!*zqv8FkV(RrpwaWOc)j%q7+Xo{#XB#Up-@yl?`RfGAj|I0bKZ@mMTKO#N)+j
zDLXkeBPFL|zu!k6-6=P5H}6RS$r<CW^rVy?i_-|WLr!Ie<*9{J${drT8Z25c@mcTn
zbzbUNi_X5J#cXbMN+V)<X}PQ{uf)qn;5n^;rxOeo9>wt#;h1e5$GluUQI<4-$IhO`
z`yJ0TP<J+W%EJeb%I?-)S(p=_+_jfqD>ECH%h=2utFLmhyI1ai@IkqM|9*M+_;DPK
z^+*FV^w)s9dFy7mclS=bK+YmgwG;*fMh}CO6LmOM;l@ih%H?ZU%a!Zb;;Yun>uYiR
z>^NVko|;v$)w0pevK9zNC_0)tnFegE1vL<vaDMpRe-Ycrp5A{{K78lB^26`{puF|=
zkITI$cgx<<emT^nLEZA9H6~4ri9x$z0H{4<B32p6l@3Q=BYjGf>C&R~m*e9+oCvhG
zwpLcwR@8A8!te{brs~RKKRUkT18iYpZztc|pwn^(;ejS5zUjdpevCoB>dk_l!IAwc
z7uYcYg&s`SIoMbX(?^(0^4$rgt3J46!yTaQ9q!N^1#hS!hfzei00jex)0gupg$)X}
zK1~dCKpzj{HclI86hzn#U|i%F$q9(20-PQYL2Hvd$c?)z@`>}b1Rvty)qgy_Q5mU2
z8xr8)W8ohLyja-YME=rpAeXyg&MV}q%(e9s7oofu6Nler=$BDQ{v*GkatYda8@8NL
z5c)Je@FKZD%VTGN*7TIY(8_3e&Xau}n78G}x$XM={W#z}+q6;o;rjXkeVL<Z45ob%
zjiHt6`~i6FT4-^7oy>n8O^XNm{B0UAoK9Mo-qN;mtwZv5eqXlXMR_`G-}$sOhCc4R
z{6_HtE&X{kEqokJE6;ix;2o7;E7R$taBW(f-iBK^4{vC7X#qZuXUo7h9zGV;G2ba{
z8i(;ia`idMgx&xE|MW>jK~%Y(u<(b68<`ru0hLX3%pbX(XK`UJwpX)VgLeYCZg2r+
z{ewJs$v2*!$IY!%UGGlf2IbMgVL8*{eqnZ|yz<iJ@})0-p?vKdUoD^e;^)gtuf1Ak
zmlw*Dt*!FTd+(L^-hHq9<cB{g@BHW|<)Ie-V<%^2qD;w%;FL^VM=>UjWJqW|#Ptos
zDZgh`uTxSh%BG)-7iY`Ye)fyyPyOjXRlf4Iuj*PoSMJ`vTi$>B{qlrEX7=|)_bd#E
zl{J-(LtU2GzQR$rteY0{rk%=0nfUk}Cv@JubGQ6ZM$h-O=J@ejAC`v?x1%mGu&5{2
zE_v<olbU=wPdKg_QNNkaHm5Wu=#x>E=K}4*^)B}H>l!z>`6r0$b*+OC5q{3eEs&OZ
zB^|1eXM~Yo<(UR>9|Md7T6{vrapos{zA;hd$FHFBn5LEA%51^qYiSK_{JAvcZ47By
zr#$t^8?5Y|$C%L4EDp<=<>|x><BTJW_sYK3SNXmv-yq|gXX#@%@^$RBOXKC@6^$9K
z*T*G~lb<Gj>1RLlYgqn+jH@4oa>0a*!f+>(MNka1N?ASj3qDi_e#!^wwD4x}7u%TB
z@u&_A2^KhU<g5${R$*Qo#efkNkJA?F9r5L*r8pY%;-w3*P3eNd8|xR!`WgnrYHS^1
zk$NJFX-j!N)Z+Ed&0FQ>`yYnC^K`R3(;(*vSc#99uXi?gV?oX~C*I|V+qZ08moCWJ
zmi~9OQ2)t~-ztCb`@dhl`|WSX2}%D#gX%2}u6KX(lXCOJ58}NH9{sFofL^<PH6As7
z>8oEZKmFw|l`nkhi{&#v^;6}QH{K}gmoJy)ix<j@jGx(+l`=l7ff!Ff)PiHO@HLI7
zD=J*KFG-(>6V~dTeVLmN9+w|{=X>S<`49e4`R(8Q&GOs7@f+pafAoj(WQ2D_w)eNo
zfqaay_%uFstpXEP&rnR_>QB@+Ro*lm9cy`Mwp>v?U+1KU3+t-m6&Y=sd}RE@NkM8i
z7Pp-601tQ7)z5M(4*deWWG~Fkm-W@PGOGz{kC)BmpvUmbEnk|vWF#;_Gcl7m2n*t=
zSnZ3rxO<L#IsDja#S;?7f*gYw9?HYSWiUSwMOGSvGHG%G7&!mcIQ>2r-1>tyrB>NO
z1ugyG8IY!bv34@U)g%-Uv}rME1XHHPujpffX3~U4DB6-SN0_HVybptm3cyiz+Qxp!
zhgh;A&;aqO3&7;(2@pO5;iz*L@HsJ=4?sB_H*OS+O8+E02G?QVH;j_g%D0||QL+K!
z+jIb$exAUK=vyHDljbuDpfhb>j-A`nbbx7%!|&r>1km>LYk~ME+{b}w?4O@U7w+=_
zEiS`$&(~$2c9e5^A4cH-<NJA?2he*_*vf3fmi2MzKuc?9Xv6LoEe>eo=as>HAICom
zUc~3Q2Wazd>HB=p*_jS!eoJBQ<60Ms@n`w8n?(`QLS$1^o=WgttGHpgv{*QmkVP!B
zZ#<FUT0kDUcJMTU<6n7Vz@aKfdu%`AtIeFmH&tGJ`AYe@Z+*4=(x3W8tp$F%y!7hJ
zy39|OM^B!Vdwla}a|h+h_S5Zh>!aJ{w%jJiZtfqIM~|PzQLlV|i+45m*w%zWG1wK7
z*8{GU_`u@Qbh&=xN_p+|*UGgU*L97UjP2LA`Hq+J<Y;?c8se$K9NShF=fsmojW25r
zwYs_?e_dsGR`ygD3?xpb<WLz5jX3;9#z?%(y~k6M6J7h&pHv^3ZQ}ae)<1xz(MI;=
z(4R|hLl&UctDz3)6+u^NyLKjm<0_f>Q80u>MuUK*&*%8Z`3c=T>;BuP%f)|FfEH?f
z*csq;yu|~gjdJan#`XE4+`@G<EEh0*&>encvrSz%H7ZXbxG9c*beyM<u}s$zE5?p&
zLg6?pCrhR-qILLm9wR$D8(a3q{++-2=gv4?hJEk6;G7mY<eK#G74_mN_iKlawUg%v
zWs&w!i{5wM;g0^3^61f)_^0Bj#h4b&+!e@oC9=>|1YIJ-6JbnHAifX5S0Arly;81T
zxl%4&x+nu;RRdM}$#CHD7o8Y69F56IHyAN{d%G&+L7e7<H9oFNTF|7hvN&JnXJ>Lz
zEM3^Yx3j%lILeJJPMkJ1Gd-t?WLbJl#_?-!z4epw;fEiV4?p@)<#-r(6!s2|<M9yR
zd{~t}8<#fX<>@ONwRZJtdHK~>%A244T)BMZTHynD8r89-Xixd^{(|rBtF+RcPLPiZ
zv9#69imN^pKGj5(`wdmUng})@J}lq=!#^s2_wW3ja{taf8A6ZZyBADsEW$TcUQK7^
z`0OOxg9%S{%DWQL;ni_sAWJrNO{c&`6AQ3Z`EbtV%NymY44>7N#h5JE*2YAfk9X_c
zfXI^`ky~Wwps`JcJ|R5q$2T^3(E#i~6Tf(Sd%N71amQT_7XO@`&5QHU@}z?FJ+&K(
z9H@@?7Q$frlX|bjMHj?aP)3nP*)T$x#{8ZDIPQw@)F_`Ep=0)5fMPlf5a&X{<6eB^
zQ;tZhK7EnxSuqauPniMyZeT{M1)g8xK@}|e6K_6UqAr<S5}XIP7&{^eUde`^)1Sy6
z88)Jr0KmnV5nWtZGSjXh(UJC%3!gZ0ks8621%|u_ak*%}5RrV~0Q|g&b^WJ>Hf~u!
zpRNz4Yhe^mtBcc#TVJ5X<61e!0e$)B;a$gumJZj}9ngmRU^%1GTO8rO9(`!rsP&yc
z4jRjHd*Qdx&&#}{06Np+o9;aNQPAeqh5_f@=3~e-LeKNJJT1;X^Xe!tO$#r=jgo&J
zKnG1r*YX2SiyIXU?I^xB-a;$S;l7+v`L^;#$pFsdX?cN`t`DX)jpg7*$saYZv^)mK
zp@ql&+BmKS6~Ry7m^0&vgsu-f-6oq%O@I`B7TSIq&o(FSzH1_t;lLK%xYH;e&B2nx
zVj3fYr{eerJNb?T`_Nx`<#PGiZ+^M_$zT4Z^0jY#tt_vt#x?oQ{rhoN|K9##T)#N5
z<-wi%<)8ifKP$ieo4-@;-+mlAoix>eZt4Y1?*|mC2uiC<ce#3PseJt#UoOA=E5BU6
z{FN`qb?4omyj_0$?eCSHC!4ySOq2z_U8Qm@uPw(r6&EgEl#jbq)-PU>pqX;({SV_q
zxAEODU7vYJgu3ChLAG>rGxgiw`F{D}qes%=B<g5JYck%Wphh_ugd3P?G|(RD$9hCA
zo?h_2qssnqz$J~g!skXO>xW6!eTuU#p>4B%xYS907>dL5vik{s!~>W0%8%(<NV(7B
zhjBU2daiXIrZKK%8Ima4fldvduRBgNklxB3<pbtz`RP4SK!<$R5&2yS-kHFFAm4le
zduDb#wmHqtPnL<P)AINUFIt~!%{HcWYZ?(dyJuR%mvGZEByPMiRo-}WBW|YiPUl;1
zy;a8kdw=;C&#*{1!jip$7!^DcL04QnNx)K9n0Y!IhR&h^r;1Z`#x|u#51*9Vx9`Wd
zCT`y3qjVSu)8c1=lrP?TFC7I88R}>(x)}7l=)oh#br}<UPM#w&v5YA^d&#*A9!HZ&
z7Ro%+z?P73oN~bX4J@!{#e=D$`inabiwk^Ij;&a{B+iFlWc6#|$<bvzHId<6wmH9F
zI8I*T9A~yLA&V_BGYp)$nKHjLUoKv`ST0<>Sgxu9FUxIU>|Eqjrt2EmFO}K3MRodo
z>JU#GcxjqZK+{Uk?D%I<*QvXbNYjjTqD3`#Iw<Ee85K7__^|xWZ~u1rr~jXST;BT8
zkJJIC%GTB<3@Qs-!;Z@S(SCWnyO|F)#D_Qr81G78ykuw3iI#jd79(3VspRo8mMqIv
zKHq<#tS-x-k_3*P<IW4)me{Y!B+A{Ha8-c!GLm0xBa*Ob)jRb}`(i9aC)Q$pUiHtr
zEu4i-U*ZWB?M*r_cJm~QW9m4Qn>vbtg|0CuWeh1?tEqzekMd7LgO~R?J%dv)XxGCm
zr;0qN-Ucj!NjE;OC%t3alVp)EK)&cluGAem;Q(}wa)~y4NrXSzma#y6>n~C&ZBDSw
zq=Tp`4>{Cd8tbYL^-=N;8UasxU||6g^%8w4A;iFjuQF5U@Wm54*YZFJ4S7qGpm7}M
zvbc-}<LoF${K+THU_Q%p{5)uRfj<3ta-46pdySjMGFzU0KIcJ8=X{-43vGS?>2{-F
zlnldpctbzkap!N}czj&+k*a6w0Q-H}E%ejO^KtMZnTA#u>uVU5mdT^_H+_rCvO^cB
z=b+8+<G^_U_{csFtVinHmCN`RNMnw0*uKLr0?X>lHGeB(RC)`otn*+LAMkNJfN31>
z`vBp*J`L;I`lJzd{X_5i157t;OfX;Ob|~I;^cEYyxg(9GNc(;uFI;o8kZbF+XB@+N
zl61WDscQ@$wB^WE?&9-)#F64h`#dKo<>gm5$~V6CGv!bG$zO;k5*M#rF8jK6-hc2&
z>xoV2!CGaa9ExXQb|K!e{K5CWUw-|c{6_i1-~Ub_86ZD@!-;DgHy<e@*P$?uxCYni
z;-=rL%J*Bpz)4QOR6g_BH_OicZh8BMKPoqW@@_fY*)4Ok(_u&~hB2|Eb<4GK^~Ou(
z;^k|?t!p8`cI&Bf=c7C2L#;RN-npl>&{kYa*49?yWo>l*;Dh(eAO7JVl|TH0@06Rj
z?nQlbH2(DLY&;F2O?=WI2f&<2Nqp*^Z82d`a08Gn%QThB;I=S0-r`xO@Et2oi{v`#
z_8+vn<mdH%*q8vok$>C20Qae&l_Pb~KPujGTK>d!={TpG-nbg=pvoBd4Vbr|7jD$G
z*=a49*JTza>T$5ELvq+U&DOE_aNV@l)G{arhQyq3?7nB|I&o6Acb}B~gA=X6r<ImB
zb9eXnaAGM>pQ`U^?fB;B*2@>a^h&&P!1@hiV&dn%^r>HC;2iGD%2tOMmoo7vMC!13
zBuwYdCl*<-Xw>K|7!+(#(!kVWk+1w-yY{jSiTrq-`o4HsShM$?ftQOb7as6q7d|CF
zr-dgAPTq0g*fjRsbBFux?K|a;ENAuu;_ltPS2i_BceZxQq4MJN8?0UiIqyuw*C;iK
zaMycldq-}!+`s>@+|gjUb2sl4@J)#iKe|<JO7=bF^Wc#NzG@1?h{YXNJjbrR^vX-+
z)1P~zeCA7^E1#3Q`tr51c7?4=t8(jQT?WRb>(|TN+>#nhEyDIG84{Wj<ZCprm8#y|
zP_ssPsW%=m83=N>pA&tKRBs=>`+oV|-}>!vAH4TooD#vuY3L--F{yCuEAK@-lXW?!
znME}kgPi(Be-`o>FTD61$J=SpQRU3bGc#jlL6hhTlc?&PZA|{=!M=>D_=bpNbMg=U
zU|O;`ooZ6;IHB^e#c4{q#!LIk6TwV&!K;20P3!)yFq1LH9(Vlng);Vos*S`0VVv+J
z!ZcFSfMDXufkO)YjEe$&2;t&VLS!Q!qmwOE)Gdy>jXp-YoQgaI=*J`{<r5(o^y@kx
zJz(8WjcXm!eQ*JFAGGn`6<$R_$)K;229rd}2$qyVeTV^lgI+lFF<3X#2dDE4KXZ0o
zxB#askbZ)7usaVTP5<UKy~_bOt(6Zr-8}tt09;Gg)=wXg>qAR73fi=O8QSoRbZ>pY
zJpHyb7&po}y$@|V;P|LIYr*OGc8>S`ejcXp)Anilaf9*p`~0onx)xd*`#q8&AP+ms
z!-ZiUjL$PC{(P{!KET(e_2oGJaq<CZ`p)=1^z$1<OSq38RfbVGgJqb16uA7j^ZGk5
zN|y6(@vS_=C|<y09@=((9su+>aPsj;_Z@4lB|&CV`MN}Cy)qT=Ir7AOVV<u>vtY##
z;b}aJfAV2EK8^PYxWRUM!i~%PR>gr9TUw8mix+0f=f3b-`Pw(XR=)U^FUES{(bFg8
z?f2h{?`dso?Qs$>S+hk-JX#;jP0y7HUFY_-_<#85p%(Caq1U8-VzDP(xZxP@4eGkv
zmd7xq<;9H~m&zAD|Jm}!8?VR9)Q`1>xp(V!*_5w~YZwD7t1={b2XcL_T+=$`;+3mq
zVRfyHO>-LKWI5i`8fkM!>#41HpOm#0$KPX!p#uwfwt%tioNFsDb+fMG`VFui2_2=o
zZ=6x(YTZb8TY27gobtNq2pa;QLCL1AxM7z!`|`;|tplYYY0$nBgmXS^KkE0Zh>OlR
zB$&qitp(GuHX*HxcQRY(qg#4A=i|7+aq1LjTI0{-qh4LN>2+&?<dI*)tmnP8i6?V>
zRw0(@;~0ELG8*<`{k5_-S2i}5%IX@Yn@!2E$Q>cvh+bNnP}*YISl^J$2^oE}GA6$K
z`mfROSdtt;#v=<(?qMqky2jbg`j53&RWuC(8XE$*rMQ$_-?&)*#Lxet8h^IjzNHR%
z$j3198xwI?Tl7ht7Xd%@;N@J7MrC^vkJ)zBVV^$EFFrng^t5a~<voXOEvolo!HQ#H
z8mHr^Gx343xD%mHPDhWgNbc^%w-=s>m;HzjA3lxGf^Tp0Wy+ik_)y(}bmYi5?&MB#
zderhlSzcc*mtVSEu7B!Ax&GRXa{2n@_^8~%>Qb3sTQ18R>+v4M#-%Ie;-%|lb?K58
z<O|Y;$9~iscTXuR6(+-h%F-ls%H0i3F1sq%4&}Q4s64%^$>iaavd6KtGIW0M?eCV~
z{%8NJ{QmF#UU~n6_u@<A;w+~VDyRDR!I>;X-kv|zqI^t^tNMSYMK<4&AU|Ht<*tSY
zbnKIhPDBUgX@n*Ro&wGB27kT}!FLU4;kXly6js`DlUmmDEe+gsIJT~3-{7Si{P>mw
zomCPrM8}nJlwLR{7&;#l58IpKYvS_L<)cyBjYrYZpVSxVBfNYD?{OR#oOKz618FB*
zOk^r6Pp-JDj~<-*#Ux01d6}7)wqb3-ara{nLh{x>9a}V{Rs6u`M>G*hG^ZQXgiv#k
zSaHbH(9b0O5DGss$~+<;IZup0;-$NARZ=?f4K#p;{!iN@qo&u4Ds(a<zxWf7jx5#T
zKwP+*UV*9(UG0O5yq@;~Xzcp&J`RA_@ph6t4{rzb<L5z}-skzGyurhtqS2la7d~8u
zGo;_P0(;Qr7cySp+WKwrz<KdLjA|Q$<pM3fPv`mv;OX<VX)Syl-r&?)S`k1KkpZ-_
z!sP@feBOj%9tg$&1SC&CuNM00ZG048pWgLgTwBgQEn(BQVEia(Wthe?Xp@$&pWcRX
zPInsMuxZ*bu9Xjf>*s46J}`>sMbPTlLSHV>8WQ|Qn%W<nb0y4;OV$P~{Dla-r6{NH
z*}=dfn^TTB(PoU3AGzsv#?4Wc{gms6WF2t~D{r1^eSVN@42%mhTw9weFTHwIhQ!y)
zH~+-9%B!!x9xpNf;QK!;w{G4p`#XE$!3fbh2tA}1*T`AcJIWtmw*(&+<upX@w2x2e
zdN!>q?<6NZ>RK)uc(`e$3dkExa=P4=S3Y&M{M1i>wtVWfSIgY&WZBt#Qr>&#t@8BA
zLtXa{Wh~{k@k^IB%B9N}%9WR2E=yYcP0Y+o2)iSurQ>nge6Uj<-F;l{esHIJ^zH}c
z_6Hx7L#?%@wb<kv73?m+d3Qmaa;bF`PkpB1E_bX%qpH)Pq3SR+rEQrpNmrcakjr+!
zCFqhH>yw$8iCCY^%AcNQ_W;HUeA!RhWj-=2CDNilkY8f)!$DE~*rKh>5+97=$6fs3
zdLBZy`l4@E=zrkDk@0kt19Y^L5TO%bze4hl#${g5k13~{<~)Ec%2VmAJBq;;hK1^y
zwSAnJI?sDTGS=9Ya>RGho|Vg&*UHKYpJLEBiH|rgh?Z|su#4!l%*>s}mMC82ex~uk
z@yHWD`^8WF8oAR+@+ku^W#sw%SjdE9L5GHdC{}IQxXX@$@-JSvq>lYF<qKc>YT*R4
z_uu;{o(OP+S$w!e1GAG+ClK)=JB|%ioJoK$1G7zLbCWGehvH>R43AcF!N@h_Km+fH
zh3Lt%Sdeo3*sccU?w$;VgM;uRu1=3p!PYDGEONXR>*VOv>Qp>2VXMvT{9JtVVR?O}
ztX|OIyLKs_KwQ3YrL0}tC`&RRmR6U`+|puMy|`X3T)tc`$dKUM70XNOWoBYY4WI#a
zs&dpP5uCsfryT9>XpnD}tw)c`)`Q38soax?591pZgr7cqTpshJ<in53k2DGW!SDTk
z`ObH~T|T<`QJmk)9eIq2XIh~gseKNPC6f=<@Sa56v6rtx#X^)ijXKT?0U`PjgOPom
zoKmsKi&)%=QTy^pl_8%-8BZCQgqbi8%27P25?<}byD3cMCm1bq7%;~?B~Z7>mx5_g
z@)s?4Cir9Gr=4q`4}(H-d2+$G6fg#n8}E%s=UG{{+$G{}5qD%T2;yl3i|M!-fv(gm
z?I)bF>s&6xW6)H?0&AHHKo5)t)EfmkP(%ByyXvBUs%-H`nDhv@7r5-l0l}8EU1r$^
zpiF0Sz>~xPYBbB2k(MH4H2flp`h(Vao2Hdv7?SHS^2F26$9HaUrZw+*V0xT^a69ph
zO1IzQ0pjrX@ql$1m1eL^yFR>#w}qBpp^Qf*ckxWiIH1}Jzn@1NKQ9dQ>znjGO$%*0
z{wHlO`+a_gfqon?u1{xh_&gn2+JGv{Pi&pu(peU`3^TsC4l&+yQR$8h-q6Y#<yvTQ
z{j~GoMRcvaKD79@o{YCoT0i|fnLwM?r)%TL{3LSlJq~f!$MJTIeo~nR$6v$)xcvO$
z^8L-sH`DN$@424oI>j}Mi+!YYAdXnK@a5j6<pnL0Iqp<zkNEHu<;th=TxU3xg_n{~
z4t1sFiG&OZo=Wh&Bkq=8y|G$;PDaEp|LI>TpZok5R1Xv7z4tySKmPGsW%J3?_yz@A
zm~rrragCPttf_cub!KKZbl%<FE;}+PxZZKc7)QC{h8*)d*XOuK!=J#Gl-E9eqx|$2
zKUZ$NbS)Nwk00DCH$Qm4>}>PC#JCoNi{-+`dYmSC{pFX+%B4$Xa&}SW#E_X*-jii-
z>!3Wi^`Lz8&WGjgAN;7i{o@~NZSp`m=JzGo;?4SKVUZ^sD`8OV?Ch3J>C4T*F)b$1
ziZMu82&+fLFQd~q-vA^MYe9iZ2PPb4p4XxloiTvslyBUq)Eb2~418W7M{|n{bcCYs
zWGp3T*nP^B10zDtvI0*SKO^ZXnBJ}rpuqMWVEXg?^T6ru^TZirb}c_N{V_+7xcS1@
zSfL(Ci#k;kVL)KGvAu6bMxlDVXim%3YZuF_uU;!li}OLN;S%ddo@j9L(+c08kdevL
zw%yI0vi*3g>^$Bn6JP(SSAH!OLGj$;L?iNoGgfK5tg5(`)v_&q^G>%Eg?@5cmhaW_
z>DND7=9kxHjh~bs|LE;_8o)^f>>1?g1Pig~Tyjysy6{s>1L-uj812cJ;8-e*f8M#^
zD}224yrhmiH9Mz;;dnXZR0Zx7s8b%PKo}Q$;@Ovfph0(}M4mRV$mV?w_V7isD9<rX
z0{QuGEm&ohEoeeny|`NVPQ}`VHTi4h($y<67%rAI4aC*;wX%NULb<5&UAl6mT)uj(
zT$CZPwz{rG`Er?_oGoL|Cdw&qy32?;*x4_KGGO+1G(l)Wd2;_jxqtI^x%bhXa{u<d
zFe3Pv3~qNv4Y9?`%J<7VZ@*o>`@QdKBKlFebMLO|U_YNo=#Lk<j!urs;jxS>O&Y{`
z62Y-=xoCH1rzR5^EOeN7V_zMD$GD5^59I?kb1J*)fx9L7J_N=dhC})o7fi(NyySo`
z7!nvUN2fA4&W_`>i35cXq|YJ70(vSP^XE*RGWt>&ilyEu;>>ATvC=uF8zsa??bI(g
z1b`Ri=Vv*ZZzei!Tog13=T<703_p#BFcjqITNt7YDvX(st9-NHazluE%86@GK@~!p
z_@eh`#|(ZRF>-N;di5XSnAy-EI3gVQ#<kE6^pGzYMS~_SZ6Az}--e{+!N-B4UJS+q
zq}c&|UQa#_1JTENJgqbI;|Al+Z!m4&H&_>-%^z@hluSFv4Q&|c$IpWydGXspo7O_8
zC3HWnjh`3p17#$=r7;fZ`#`^543^W<w$M*&<1fOG3IlDpg&~^id>2m=lpet$y0*-U
z#XnK^6VhGoTx=%4c^o!<+rI$IZ=Lammfq0EwR(=CZ=p?Z^KYTe-!Mv7J3||8p^rb0
z4zSGDwa~|nqG@4NS|9hKdKjb)+ODG^jk$^&SNV=#o0?&GSg5jq;)EmKt7K6XH|=!2
zfksz7@#mH%aBOkn`$^nfJJx!GHT}}kT>08JzFhv5zw{U6e?!;Tx$?n>A8Ot6<8trr
zy*Tv=1433xIg2e$$8nvL=3%ICvutjT_c^Df=d-f)^ojC1RPI?mD7n;NBycKI?Cwy$
z(G!GUTpBN*e)HAx`7eAXKH7^xa_{!-a_9Cft-tomlCGr}*s6WyQn~T!tL5^`uawET
zB^53WiE=ij>-|L8y!WKM^Mjw1?|=Kd<$K@$cKPuAcU6xkWpQOeMnr!7nw!ep=;Pa7
z7(f^iyj#IX%=p?i!1XpE$DM25+kc}>v=?ot@=^{rpQ<HpZgMm)hwiMcYONR7?mC`>
zC*=VA20Cj)Kwk+j6Y=pPzb;I)bJ@Rwrlyn7#{;8qhA3MN1APj3tT)igFutX6dK+h+
zvLkMKUzY*<38K-j4W^JGhhY#~oFoSWniupj;1(BW<1{i13U;nszPM4=*H(*$MSS;)
zF`(5@nVFg`tIHeZ!rH~MG_z994#vvXy{+=}?xu{1y)yB&&%gX@SmP}8sCf*C_&Va?
z4ly4P!D=D3F6;zRIN{?>`-v&G=#|~=gL1MjYh0c5JKz2n<-z?&@raNo2eFqQV}fGp
zA43B9@x2M<&l85-y}fv`n8iMK4K7?>E7xvZi|;iot*oijO=<zlQdO18*P&%>jHwf!
z@=k_?vP3=8A`sj-dWoK%D$_@vsv{mD`h{vOUNvCWFRhgemp01AWjQKQVT=lnP~*Lc
z>o2`rUV8O4S!*ws>o;C57cWYe1q~{Uh|>uT+|zQjBO~bPt}Oh!<>8%s<?;Rd<<b3n
z<<Z@{G9d0G{_ex_<iQh78e4_ar+7z!y=j|IpT;q-@4okLdH4PIVh<&Qm2YNn=buG9
z1_a-cIFv!bHZ%rk{86XXsAJ(9lNFsue{_VfDkX#ONX*5l6+E_@p_W9e&ZtS3&+~If
zhQYkNB*SWDIVPd2s`FPr^;&uPwO8X?85cAutz%%RewI}~t8%llP^USmYJN@=mFh#2
z&cv+hn0<I^0}KiFhB8_5#Dm9bEYkC9+T0ORy7Xq^r*F_P(l}sWDvy)ZmtqEG+Q>nr
z;sr2v(&dB$Uk)z%X6ECkYO!5NhEP1^sACQ>IHKXIoRH7p^S}+dnyI|Z-(19Vu(l4c
zJluKErUzRkwG6;ATRD#RWw&`7jO(WnALZIK=h04DqoB`w9<+HIJ_%n7eLY6WG<*`-
zq>YjRuFt1Zx+es@NPb8ZnzYun(8u?~e;ogl0JN<$ya;bNFWf@gHm2#fm!Xy6y6$kJ
z^#P!59pJJ095g7Ww{q<O%L4joK#S`;gTte60Ac5I9&H<Mb?U=;a{4@dd_NChRJ`wu
z2iiQ0GxVa*>3JSFpO((yetxm`pv|M*xo#v4^HjW##~PvD{YHWvz_pUK1qKCMJUC!w
zLhFLOM&(I*$mF0Gj0fJw<ViiofY&~}*!=pZub034ul<$si@);AWnp=xeE88v<zIa7
z`|%?4-p*cZ5svo@Vc{AkH>vdq*C2ER*y=RDI3EYk<QDE-t(OktdKkdGOIW()OU;R<
z$m<uU%Nw8llnjbb%dnX*n@=8<JGX9@yEktt-_x?Xv{<fOy%JlmUVh_^GPAl-#wO~q
z^cWIayBzO5D<A#%z4Fh0<G0H1{m$={pS<;>va|g(*9psWWkv=UA0>+&B)X>Z#cT{9
zp33a*?!`M6d~KT>!akIRi7%T=<5P7>yWj>3+n5<Wexb{C3hNS%rN4aXLbTDdQ`!+D
ziuDfd$xY9^VHgr%)_=&1#;w9oVSrD40<EBNqnwGJ!&MlC3mpV4c9>!qTG^J>`VJe6
zYn_3<J_?eq`h$2H6TE=UnhfJPUi6niF*7?^7GzK?EzFgrMXibBiNm1`;FEHoe*8@9
z`N^}1GB-J27H5~s`qG86zHp%|j4f)xKUVf0?UjSa`{is;dHv--`_pG}Y}C|voNB;Z
z{46TDlTD|Y(ZFHgo@k)8M}^X(V|$Ve2^Q-6`yA1EEX!s^N^O+i_>F&FKHv+APh`1l
zZfWVOuB8I7$DF|cFL#a^XKZ=lg<>8xVhM9Z8y`)}FI}%jCuV|pDp?QIfjNDNJJl?p
zS*+4wm<Z@_bVMp`oUKM!^XiavO7?bPiOtMSYw*sMm9?d^B1<?v4bC<v4L(kg!idNt
zPUp3tzFuB?{nO?0mFsE{J|~a$osD=R18M8=mW+i5G9I3mXGaWtCZJ<Y2)x9+Bd3aH
zB0APU;Ak`%T-;&c6XeG--dHeiZA-?32j#8|kB5&QsUBoVXfj}e!%*P8fdgcUo_7gi
z>yPqhNabDl>@fPJck4eF)x0MWN2{sMV;|zetjfoCBc@`~Vmld+ZK*RRqwCkNm#=^I
zYvsm`m&)bKSHf_5`PEOA<x5x8ZYsBQ(F~+~<Omnu`rj(M`@4lBc^^G~RBmb_d;k6S
z<LJBZ|KWGa-4AZ5WBQodJTZ(XDjXF@R|m&CE=)QYr8)U9xo~`0zH1X72~zpf7|n9V
zQwhcKp{EadaM2gx$Hj$)Iut)Iuyg7Y$NNPw1k~dYNPHMz>i6XD56kh#g@h+To<IP?
zF;L^8hHeSqaULs9bGVJSAW7jHA9NZoKuZgrxSi$MCtq|jPwP7jt?8}5;{ai32ld*d
zS&n5HS{g%}-zYG?^>NHMqw2Md1ICTQ8%EW|9}C11chv;_d<=bjY^@V&lx{#<mh)s9
zMzuj7koTgpou|hr=*#KzS$ESp+{RmQyp`M1^yOqY<LQ!ydWeOVNZPT*+`)K<)%@{0
ztVTeaufy=z0bRbC?z{lw@QnkEZ*f3N4=~@hV7%cxIR@j;)5osGJFSfaqj00ZY511^
zJdmepDyB!)&!gY3@mm^)hv*s_=6!UDej&#^&sri*aH?~BCoNcYEgaXnVs-Fjg5&5}
zT=BjrH$GW6OlZ-6d~g&-!7)!IbdB2E)pba#`Gwhu^4ZV4QvNG{>0d7Y&41&smX-Cj
z^8N?!m+$`3x68Y4zf*R%w&T<$zP*Hk(^?;~%Oh``@}hE<VQOZkEbz?-twA_E<cHt;
zUc4vCla2ci9&0^SpRCB~k`!e<R#@t3-85ESdTpWn@}K_2@+-gcC(HWkO8M~*eo%h+
zU5<o(Qf74-;b{Gz`tp~{mwxVB<?5TCRq3bXcq)+w1vfz-y??j-=l_p?SpKvB>;Juc
z?+4#5n>$b9DE(_!FO|zTE|ra|8)fC<T5R!VJ$6m&p-UGpt4_ztAO69&%Rl)0|ET=#
z@BX3Wjm!9$*L8E6UJ&&ZJ13+shsDI2CHjrV5l8{^yN_~br^Lp_vJ8of(Kh@0GOV<&
z;H%0!6@r%cZ#iHGcJhEP+P(IHaFR=!`v-p8SB>w3>7WdnN(Zj(d**Lt0z*=j=Ic#A
zZ}S^Sf010k>8-PD)A!>d37x_z!rU0fpw`<naz4#zV&b&SFHDy8wS{tFO>1r$kPBK5
z&reU5ll}eJE_W!Ois7VAjssp=URHRnoE|<ayH9uG-DVu)XX0C5di~dy;zK4Zu2<r8
zpM1n5#aW9<2|B@$A<+PfH#(ELygJIbEY+QI>(+f)txsfJ@G^A16wShy(y>5BChu7w
zijKkr3x0Wdv20vejc-e=tuo-Tf_b`-bS$X3TfhPxJvb^BJpnBKtn7%!LX_=5=*O0(
zx%q{%6faycfO8P9uB*Jf2eGyu=PzHra#e%yT6y{9SIe6kfUmy#T71}La-1(|YcM{W
zDBCh<?tXNq+<O1E{Cnl`JyqtTt=J>Ex%pIv4%?SDwP4>W2O1EEGAMY-il;8&Vxa8B
zX;lwoL|{lfdi=O-(r~f}$7C(?0v1)u6Nz|7LU<nC@{+t-AnG!Wvgj^y454h7c<LZ0
z+c2Is@EB`BRwpksoybr*KIVgc7+-vGe73y)`fKGI-}pxPi+|}al|S<r{^j!JZ+yMH
z`sN$u(zUB)jt}keREm74Si1m>mFY4!J71O;SIVW0i{;gqUoJoWna`E4eCf;O7ryl;
z%Fli4=gJ#zz8U9H|M0E1%15`EAe4U`OEwb|D=&<*@QUwRFgeKCU;+n3dF9-`*$3SZ
zHHg&q(8S_X4jA-RG58TI@tA-_Z|+eLeT&X7mv>ZH)J96AryiXVfjbYJKFYP}AuHN4
zR1Rk?TOM4NKY@7bJI(R)V3aI_(<}qsa4q!tEXzEGR!%>y52nRg{(1VfV4jbIQ98FS
zaQcguP&}XXCC*^FQP8H5Q;yFbenfe|wXP4wwQ0z7-13Q^ryuxHJV48D-?)~}(8m$)
z)3-P-W`J?6Yw^RnIh{HIn>t}zCq98NH;{6SLK7VE*N2u5AV25XLgdrXj-qqiY36D9
z9By$yOKWGmp_S>l{ZTyjffm=tpGViid2#&IqpJ&&><rl!IbXnYo-;Vy`mHn0^KBUR
zax6^Qg2DO$z3}74offlj4bND)N1TN#FPyUDf}4wRGg9TDUlHb75=YANex+o`F|ZgE
zTKF=YFJD?OU;Fyclt26D|6KX>7rsy)Z$624B;J4by;uWqViVhvcnZORC3z=!GIY;t
zkJdpF#5IoNRPi~Yb|H3xOsWjWaT4av&VJ$2OF1|V9~U(*b*E$yz5eF4@`a!JTo?^J
z&A5B>mewNo%aPU^%Zm%;<yT%QZ+!j><)zo(D3f!`S~qaBu0EMK882J+o|fPH&EGBm
z@bCSn<@-PQPI<chB(9;kbUrOhS{u#G&*}Ov4WutOe%Y<U7UWfK@XBW$_UO^0xXBwg
zaoPD%%L+tak&7FDa?~w_h$fA&)+uTePL^EKI*eU0@q|e2$VYg&2^{qmvMevd0DXZV
z_()Sc``&Q>0*ve9ffg6x8gF@88X)o$L|TM8=+gpy87+V78=MZfe0HuApr4lcNS@Mg
zzIOA#Abc$h>1UHND7bMNb0+I8Zd7ZXeyqN?|7f$EZ0(h^ouhKPt+nTt@K5*4@#bOK
zeXvs=e)OQ+``~VQs<khtft{+)hY|jl{!8EbwT+GSc+|%JK?WF${ZNR}BS86XF^BUo
zWuYc3SA!R$VQ&}1;YbVUV=bl?j+bqv2!p0B4D)Ibfyjdk4ab^oY^-TPzY&({^2%~_
zuDRLSybvkx*srL8%}EMue_>BQcd3w@l}kk;Rh6ng$`D7J&B)DiUM_b67UHSL>e@<d
z<+*g}Qn`5Xk_P0}a{c;^a_!2Ea&hC5Ed2`_Fft;>rpwOLy>kD~qj0xB#EAH)Jh*+g
zJZ9U{{YT~D{fBXc?e^wodHnc+jH!oZfA^^>cuy6_OU9~)Jv9(VukIa$VX-CJt)1=o
zZp2d=Ae-Ay%l_eBIaS?oq6D`hj^kx(@yY<;iG&x_lr7839Td@GP*J6dYZ%2ui?KjQ
zVGkTHU2`e}pH4rNrBD0v*?mrc-~&7Vij0U~{Ih?yeCDUV5bud#!10pR<P=-TxSK(x
zroqS;x!K0`?ce{y^1=U?wEt?-b4l_%F$Yp>ory$3>jU6uQB|izSBv5F40~s1W`|4T
z>@;b{8<NKT0wizhV{naoN277MW^=V9ZDwa?TbAgm(^W?c;6Up{YOMh9|8sx*eIp;d
ziBoO#%y>h2c<2cCctTviDPiiq&8aMw=jLr|b=}%DKnzM~8yXq25tS?NZ=5)D*3MtL
z<d5QUJsnT`lYSmCTv=If2$g&dXlzR&R@EP!!v#;}$G1qSY`jfIhCR=K*SMT~5R)6`
z0O2?t+A!@8JatC39rTs}Zsp}=s<s{i1$3gFB~;P7U#2A??XU7Hp@ElPHI!CiN_dq&
z<&yO4c&)<K@mk>kiI@I$ywVRyd#TEL$iXlnaazH5;iVhupt{}{MkvB=hAYfKX+8)_
z|BSI7=D>DWK^b2O*h{}wc<ElM?}CHy$~Kdb_Ezq%g3|w>{NpUD{>Og_szLomILD~P
zmuYR#mfpaYd#V!3bfjO2SK=Rpn|i7`v~sIMiBr-c&e9Il(~*9y^eht~!(mUigwmdI
z0;PL1T#TRSa+TYwXog#vj}o_)d-|0CUhp5aJQVJ!$L<8oY7>s#1m+R0!5kcnT^BI$
z?7V=LAv`pBQ-TF7-jrY`1lbIF2rZmFG@Cg#Y45!IroHpQd)6hM>F)je_CRWFbyW<R
z<_Gu^2{bBw|MPT8d5dIvVm^|yFAQo<f`@@WiY>|E$8Aae_R=PK6N!p<#ILiv&5oTI
zvWu6`+41AY#XH3osi%);H4iWN-u9VeGZHo~*og}lZCE@_Y;Woofexjp!#0*T?dBKP
z>@WY_f3)BH@|QNd@WghWiB}O8Td8&xVRv7r_4N0|xe4(RumBzGF6ij60S;pi40!o%
ztgqSp+*4nfCeO^_>^h<T5tMjif<?ouKk9{YY1rj}kcZIa=W|^cIpm9{zGY2(A6wR9
z!SImJ&2bJxVc`{|?{6FX$E`G{9cY>(9J<q~Y(T5<&2UX{_sR^Ecxg_1D=2YGsD~qu
zsY-jATS3x*hep_QAwqSAfCxY3f&{sy9g3&sYs8NDxUI!i+m$f!Y)yi~y4uCM1Pxab
zswH4-$!|lv+t&JKZ2wc6hfdpn{*S)@GxiRWZdI2@0JBCSg3*hDW^iI}TppN}uXtW1
zQVb{8e6Fr;*t!(A9kf*s)$u?ad+TTzC`z4!xx_{0{CxmP(QIdexkreA>5oju*AW^H
z0t9hvBS@%m?f9yJ1PpcfU<~PTkdL&}``VRW-;=1$;P)hJ5Mf)BZ)H+v;C6!kKAHRa
zZJ=+^`g+tkNcYZjTVL7quMLmyKe4+v@7tYQ_w4@N`*u%)#3M1#PyJlPQ{gQ5Y6L>T
zswDgMwM7>sb~jh-#rB#V+S{_{5)yVc)@(-u#4fiOZF6lcFLv+vVe@s)Pi)Cg!UKB~
zS$W`WgbNZ<zO{+3LevBKtHA3=%D>29mmlSd&dY`(<%1u*HwUU1sE$le*z2!f(O~$t
z{p64S$bR&PKefxRziy*rWAU}*=~KUHYISj0gM=i}ZFS^b22t8ayFaRT_uE_c7yscu
z*(ZPhJG=hnH+K8_4ZC;awmrIY-=0XYc>3^(J(Y09b-t|dAe<bVIc_IUpRsf2PTTQg
z$NZBiH)<^@o^?sW<fDVF3Vh{KJ{b`Ftd2T?H&SJU34nGHtDIih6hbIQS?PqMtwx&#
zfr&zs41=25Nd$_?BepHMwQ_Z;`*cwh%_!$^%?=!89eUuT0#Yxbl^X!PGAwWqF7D;P
zQOB#(Nhsax;Y(F_t3#A)gfH=4rAnyV>yZBFj5j5eVaq%K_<a}c1j3f_l;A>RWll)i
zsSbj&TmZKnjP6jUonaHwUdJgJ#e{?WTH%#Yw*$D9DxsAdfW2NP{}Avx@xrPpv?$3K
z){DBsT$XvJ!QCo*;<kdsPn=ib0wrGEjLNjsA#s5+zJynWeHCQ9r7GQ9K`XyZQ>LMg
z^KzKL$0_~NziCbtP{&Is-Rpi1N|LXZi{U1Ds@(>zr&lV|?O!Q;aWQ#;CeW{8SlQY%
zIMgS;K>P(?C7i#3sSj>ly~mkXzc5Dpg)^|C2Xl<uA<e6&PM@&v{on(8{jE1_PmS{N
zlP9*qHg1KZjdIq4TboD~^9~Ql`hjpI=#b)Mj`xK@XWe0tIpDqN=|Iq6I}~p46l-Ga
z1_!(SJ&BoP6E-<DWD`=yPEFdGb7yR7YD#%)v*r0En}7V&w%0aoXkf@rojYr%&tI^~
z<0q|qa8wa%{`RjD9rkSVnLWAvz<&FSU)#_B;xFy)ty{La!}aq~Xryx+i9&oDi;jcC
z{Sqv?d}W(*XMuJ=JR6I>oQvdU#zhG}9LV-^q<)k~b_nE4*i6NZ_T|5zVUoXsxUe99
zUjP-)h96&r;_OQZ4yt(%%)HK=9yEeKO(0w>ga_?c1OEHN2K~yR20`;>z#X5|>u^xG
zg!KO|JfP9mYl$prG|YvaIRPPph2(fMj(!wh+BiGQmb^piJ7P9f`MV~@)1Hafdm;X>
zjW2RE3S?`wZQ40~f}$fGv|Cl1^KNbb<PYBR1K7Lye7qXDcLgC3VM_gUNIz&pL%|ai
zgbFctT))bD5Ilb3cMo@Td*@qTT5-l1fq)K`ck+`t&SLa&PGN|13jKa=Ath3-gM|_0
zN?Eo@v5LLUmyRuNhEnH~J6E<v+Q<~e=RdA%sEGkY%qM3AScS+p5hz^$j3yeSmA!bp
zY{@x{r%&eX!JUV8_x63e_3a(IDaQG(1PJboekk?m!2_Fn{MZ@pl?4e8{(uXQv9T9_
zO@hd}29tG(_Um?Ncf$@#m^jQ04ZAz`LXtagXz*!pojsW3i!D#=3*^QI6)9pnD<hi{
zI5<P%gOCqARi})MdfvRzRQXo(kew=vyi>WnAHlV4tbXwM@cj8R_K*McPwXdu_=ooW
zAAV?Wy!W0>9692L!0+F^D}hM@!QFfIMERVb<x2|!5MNeaSWj1<2GlMWCJ-L})BpG%
zB}{y5x31r``zoLNckXFWx?{IhUf+Cu&2HbgY4`4lmdc8OmUIpb4cp1%C+y_OQ+7;)
z+~~-t-xToZ(PJMt7`Ri9gWJ*G$R~A39+&`Jz-6#TsXS<lv_mfdN<(o~#pHudsA&A&
zs2wOz%BDhW1$a@WqFa#`&<QJJ5cIgyEpQtMEulUTw1PSv&*v-b2?xcU?)A9q_}EKT
z0%2RFy^fcV?&ujV-405B0)C7UWqi1`qE**_A7~W@(?K*^x#O1Zt<vN!t;sc@q;U|`
zang>;c+2=p+*Z)azwVZH>Hw(M1ArSU?X74W<OigGiIe6AUH)-**MJfy)71z|xNzX7
z(tkfF=>lm7>gH0t3QC+-bb*8HrF(`g0srq39D#b+RygT~eh|Gd?(erH{(kou6N$R?
z-;dKY19gAW&7TJ9eR+W42G8C2mp`F=aY%C~Tb#fysXM=+xxbfd(Ro)w{D8loC17p{
zrEpUI!nKrN2@9X0If?KrWcElHx%&EL`{9p1w9{hd*S0qORz_B4S-e!PJ-y6dB#+Bv
z*t*PiYbIsif-6Q2G#EMfD{eO8@*6*oC*F(IcJjyNC|sKI+NCr0-g|G_mDkVP>9fb}
z*oh-H$(e@fBYvKgvmMOWOS21nQt<PRr!SnhW9QD<@Z_|$b@nNj>@et%u+wgf51-l>
zAAMrK`s<(D=O2G;%PaGK<vic*;Xh#fca%iF3UG@P!i2xYp*ff`@6i0qPLjT!9$#37
zzt~neuw9!4Qa=x=`sX|=w3ugs5D2TfqTZ+^*(i4eh7cO$PZ6nbvQ82LzYN=iY&+RD
z^#2l06ZU&#+loa>fD{H~J3~)<9U?xdu((^otHLE@_=6zRSL!9*(p&;I+-SSx%@1Fz
z?({D|XdnPYcyf*sw?!+lPE}l+s;xr}g0E31crhPupr`ij{Is7{i!;i;W$;h`;EkUl
zMC?hJ@D+YV66vCnacAF}H#NC?cWG3<*R>?$B?%3W9zL=8xkb@Hx_>UIdc%c_O)<?h
z3?%fRp+u@P_iz;`w?=h$J5A0*_`48m>&B~*>>Fft+|A4EA>^z9EBk!<iqC^Y&vXY(
zPkYr+qL3L>NGG>R@lf4{=y5;tveLWc3PyKnQBwWff`o-fcI!q65clrhcj4i&7~Of^
zgy1!9>2ptKZB60WwzQ)T#8-wG2qeu(J(uv%_I$^>Y}a}YNwqz*9?n3(xN2Y`{maVN
zvh)Q>@+<4B{`CQIfsF9@9M|J=4J-$*{c|1}QUjt9@kd#aU)SXp+i>_wq=q9$#>g&J
zEC=e5KoQa|U%71m{J;2T_P_ks|0_Fx^@@#6PN@#M{NT*D-(0t^zx-N))U0hRt@@i7
zYZ9#3rbao_)Q0*8t+Tt=_Eh%wZrrlJ{-6HJu1jcGM1WM?FU~Le^|q83FN)92FWBS9
zPwXKtm#bW7AJ4`%D7B5AzJ3jc({|$I37Z%jv&Dt^>dVfm1|#mRj*iWX?9^eWCIRld
zrp>U;3Gl0HtFoaRyRI^zzhOrCVB`XWd+LW`q#R?=av_5i7?5j_G{YP}(Wb(Jp5|2b
z&iq}Fg98w>_YYh^iIb4Ht)O%-@gm<%1*CsHj&%31=#_gN8aM@{J^j*NkE3o+D8rO`
zniJBk#HmA>&k{;|-Heb9yefSOm=7vD{e)i=eIF#A<5tir&V&5Z9k*8DUIq1dOIme5
zAnBL!WW0%6hC?rz-1Pe{B}~T$MBVkKfA}_&K>W=#HCH1+lo>}&i1>#4cLR9!d;!02
z_!0<n5RE#$R`zr|$gl3+3bJf~G}q~tICZ*BGdQUz>llOU_&~ZfmAk&=p14`o=~hC+
z4UQBVUdJ#q;Z9gzXwx|EW@k!cTMYnK=8n0M1u5pjc)uXF(gwpSUZR={_!_~277Ll&
z;)Ec<TT$_->T^4Ja>hRV(GTqHci;AlVjewyta)<7A9d~R>bCBVPR&<c*25L{ieY18
z(-#HFf`5^~RdkX8^Q<sIDKQ*a-ey~m_$1QC7yVNwkJ~%%ykl>?`G%c36}KvKv*q~s
zgbhlN;Ec!W@~SO9owtpZHO;Sw{L+}y7cSV*(`T%2cwDqObIP`(PTSttwXZ(@+<y6&
ze{COs^pQPyaM#v$*NqsxJfz_C_GIA@H3xJD2D)uzghS%Jo^}<uU*it{##ah%W#nth
zgNF|_k82JWzenBShRz}=Rf3S=!U6wsix?}jjMN*v=8pIw+E}`0JMg;lFBooh{{(Ox
zufh$#^skT8w7E1@AO}{~_n}p=KWqSTrMt&n=}7>;Mw<Tr1ZY9lDd{JpN&2N~LfC!L
zS&b}&2?Pn;SbVMsN<#?8$R3bRl|i?JhVC}?70$z|ZoQpWH2elmgb3W#))AWdI>lEH
z@Y?>_pS<-me|Y2tFJH&!Kt4CY$PpWz4MJH!`ccD3Nm6D>!;6cn_E-|{?2|<?+0VpW
zuBem6xe2yAA<fW9V3=VDVGMmAp2%?!cyoQt*2Pe=($C&Q9&q8H^YVh2Ofk6Y?9W=;
zv|aWxO8Py-7Zfq?hqxX|3^+>7)(gcuC@EK?-&4cn8djK8@`4+TC^s~4N_|6&JqM=N
zSJ%W8^V0UB1Prz<E!p~p82v2?AQEskwl~zd;#L8*dfz98AhE-F1}SbG*pqs$fu%=+
z#Aw$Go9KRK6Mbg=ozEo1h^B<0rS)xlA_40m476`;Qra1trCi_=eIHOXc*UnArP(`R
zD6+;B={k)CDUr>L3fn=23x`KXhKG7omlE2#4%zoUe9!*P|MtJKKl_*e+`9S)VlWmS
zuNFSK^U!X5eZ%(Fb|lRyYt>Ch>h~Hz84h9W+ihrY$hsv2Fu-zVgopJWYVg>Ru)%vC
zJ5v6xhe9wI@tEMY^18XWZYxh$?CJeEo4x;3@jbA`$4_lnLe#)eubsPk!A_l?w%Mgw
zTUcGR4V5K_(s(U<Pr_6i+qIOvI4{9xY2@b@QJ=~u9R#3WytPtZsaI{4!g8U)6`xI2
z|KNk;Zcbt!q?1tOA9$1-t46^nh02LFVSk4Xz%AV)&QL(Qm;R}z8L0b58iH;eufzc|
zERbPd0^Eh~ds}6V37|#?RfR3nSJK2S?HMj{6F(vGGHjVA6#b!<I|^tOe+eainXXdR
z@si3uH>q?hbs*Egue;+_;H63^?dWCR66)@#EWZ*D_%3?_?nx6U%cE5q0PNI3<`KZ0
zZuKy&(wAY011M>g5blI0-ll4T;^MFBcc8<VB)xQZJajkXzCS$)nfCNA@k+N6r(S-j
zIu1H`ncoshoKj^TN<GbWoP=}->V81_mnz{Pzn9&h`!a2yUKS;!J@in4+O#K5GcLKx
zE!!63e7MsF!H1b)X93r}D{c=FVS*I5ziBR#a)$LKc;P!CxDu8#gzz^64B-R5L1`2g
zFI6`lQWY6=Zim|5*8D1d{Qd8}V?X|bAKBFOq}{oD+rIeXGx07Pn!XQNe{Y|SjgE=&
z9#tCQmk#-RS-w)Q{IHtKISJzPHlleOrJO0|HqCoVzZCglUSnIY$Y>516DKnhVY>vC
z!GU2L9vHL9;Yr(G*|CL3bGAIYXwSBG{Odh$VV%Bs$wrSH6OYv=dVJmQQfAuh#us1O
zum1Y4>^HyqTf2V!E1O?g(3~BNBqe}Z>|6u(#Pjs`b=c@gzYX^FcuM&K!xr>zgo&PR
zf5?}`$=SI%zmbv!+ruyT5+ue`X=ba}jtcV`JZAOhbxRc>UQvglOTJ-wwznz&_!>kx
zf=j!IIi9j3rL+_HBTpuz9J~$5J=#Ua1{(Pjc)XEURE`yXiIBIqmx1`91*DE!D!2P@
z_@kQbw+VGSkTkO0rz*h(DEffsApBJ{Qf2&@V_XCcGil&PP~S-e&!_V55P#2W3GhMO
zB)Y!JqVbkk>1Pr(4_i0FgjAOz=!BmXIRp>dcN}e?ETm&rouWuf2>8PP*Z=&-KZ`@R
z(NR%I8hqAb|5_G|U&BhmDx%82e^of#)+u3PS<K}VG2v_KtbLMJU}n35*`#wSE>_!}
zA(JrShd?U26eP?YhmiPsLF4nia_rAiA-cbqcRw?s0-$pf4m|vt2e7G3*QyF3vuUP)
zDG2cQT(DuF00%wXpd(c^Mk<Al&zFr|-gwxNupqNPi|=VfisnUU-iufjqrGVz;>ZUi
zNDTLy4fPzdUR55q5-h3G7T30IabwSxw{~p>W|=}EfhrT?qYV7|R|yk*V)XoFT7*a0
zMI}Z!bT@llsH~I+`b0PPW^1tI?8y&4{J#CG|MFki_dop5x-{64m&1oOFdb_5EoeNh
z!&e0!7IZ0A)VSJ~w9v^pSWjHN3ft4wWrtNCvrnG*3*B4;OF4QHm^-B_wW??8Q|T9t
zEpD;W;KJ7=>S=av*6!T9ZTBDDw-pHwZEB2D(^JOFnu8KpdVBhOfMs<<3Gn*zK_jy1
zaYY-X%w+Lu3QrkT4%8=gF6Gr5ozGDOgvag^j{ak!s_;P%w3mSRh?l;>p?7qT)id9O
zR`$}((;_I{zbi~D|569aFs=L(GX8|LH{7L{X>^^uR`io%!(QPew6a65j5l%0@bz#2
zVe0M~CKXUmZ!0MM%JBQbG?Um85;rD4{&3X4_e%`!-2mJ&?<o3cnv+)DuMUal?M>!X
zK<Xv5ax0-8w#3i)GktYh>6Y$w|AbcVxRvxvNP8KsrJEG}HC@0-Jm#o+8R6&ijoeEp
z!_}dTuQUTmHvw}QXS&sKTDi4?R=BOwTE{ExsCqf1N~qft4sve=|IqN6CLr>IxtZ3c
zdq9&`(o1`k2XjKqjdD?&pk2Tpu)tYiyPg2Q(Zkuv-Y#D(g%tA<Z#HFmXeAAL@E`Jz
zof+WpNL-h_{`zG*bM~~mfA!Uu_R+6D^0Sn4kDuD|!jf%o?1+i(^0S!mOL3SOCQmfQ
zz>#JE1~C_F==*GI&Mn|X1~UyW!ko+%?rUpn6p93iK^qwwalwUnIH&NKcdVW~)?B+`
z0|R|FbNraiD2;>T6B0st)Cbs}blBGCm+bF<`D^>-U;mAL`RQji_w=cFkVp@`3guI#
zRJbZi8M?X;+dzM}AG+ptDkWS(N{99K^;(zjF6q<UJ!s3zD>gU#)OR(&Q?*O@<LY<k
z2b33o_d{V-`c6CQ5k(oelbfj<@(-5Rqa2!Tfw2khmHtgQQd`l&07?lN*2&a8n{Mct
zo&?P4UWd9{QU)J$*{(})|4IS%I7?{wNhRHkD8p3V0hLbNDEO5I&m!FqxpyKSuu#l1
z4qIwhY*AC4sBI!VbZZ>#6n>|Qm`l9a97hPg?8TRjkO4Y`*CD)ixrZ)4)ok0p`lsLf
z83@l`1Vhi3r}!!XAbC(QoCHVMh&}A#V`@<SzW<%C6U)nMYW!SrHlPSQZCH(7ExBTl
zQCsYY@-GvfNmd8cY+V>|MqVun+B6oGns47x;)#!jPN`+%0gcY+VX_njnYMIGd+K!N
zjLR|P?snYe7XI>MKv9E+S>`MW!UV5V(%qbt0t>?hig-(6)zoclSOUbvV7sf~UK-7g
zt!N;aQ`yX~@7nUt3)_5t$ktVv&!8hkH+ZHpr;ONR2_x@!G*{^Z39ZpUrBVkN*buxp
z9L9SRLqh=KKpnpWs*@Re@4a{Jhd=(%Uc3C7GwU=!1~&+Au5I}iujR#6TVLB${l&n=
zAkAmpIMi-FkRix*c6RxvO;%Ut7Z>bzAAf8&Zr@OvRdz}Se7)<cx}9eTVH6N``&>Aj
zf7x+bE33=4zPYBsbKSrEtV*$;bU=gTYnQIr@#7~nIQ9EpJcJw`L5wTSL{9WrE^=m`
zxRgNB7~L#ZraU21xMyc{<rNlnjvWNoMACIH$^B3juR|6$?U6GZF`!kn+`mAs0Jn5U
zPuzskKcS@8%0D695ArXejJL$cz2R5Ig{__j7cy(+(!Xx6L+MuHrrt``{aV4k^a~Ec
zE1{(M5<b=z%F*~qH@ChzZ9v^G?Pb_952br6djj}%w}iTTNhAGAoYD-`%M_^N)a`Ys
z(@#~0@1g^gIAu8{q+5=Ssemk}RGGI71C;)$>h3_&0_uL~b(#q!ULDHtt?-g&>D~&j
zg!FHv(m$bYF6{~FUaEtjF+U0FVM@O&s>Z^Yo11uv8~Mi{zrfGMCH(wPI&&5C46FLd
zpesbBlP;^a%69qUC@Z?mBhAG+;dolwd_Hg<L;vstd9ISFWpZ-F-hTHjJ8|lSZEDWH
ze(hVk@zr(v=JT6&_vUSTaQB|gKAP2hx}iCW_p^d`VmnB_r{!NkNKlnuQd>pmm&N3k
z5}0uo2N9A<2RjwIB^Zp2joZ}Z5gQ*J_c{O8jhpt(*I(J)+jo3N2iLXoxa{!JBi1)M
z>RXdpUw*!K$Zmc0t^L)1_)q>g-NXBLgV&>w1hDQeC|}a#>C!xlhZ1&JzMS0{lAzMp
z+v9u@b0B%*iwWmlM-&&gR6d=Z)4aW4ywAao6zAoXUugT5ZKoG?<8>%I<&BO4@b!6F
zf8Hjfe7`>F3lw>>fJz-!Z2?tzIVem+=U;j&RYKAMO8*iMe{AWGne^4?#e}5S3NpSD
zzl02n<~Rr7IXG&SCbb{<>rk=N2q7Xb9DodE=ZoiR*s$R}^6Qm==<iS{d>8dAcQyl}
zGLI<Ap7`|Zv|O)L+dui~>wad!SHsnSGZUV$N;&f_XRk7dxjVNb4A{w&XT<;x+0OPe
zztM>U(g+<LUEMC#bHLcIH&x>eVM2*g4oC<9Qyzuok-B2lnLiwbEsa8mfhGr=6nG*C
z_E11si8v7k$O;4!-<edoDTmRCWlw>F0S88(EnvRa%;O}&&|<*ZwC6kP*5C2mg@_|V
z9X387p{nzRq`q}~Jhy01mp3I$JhPSE!zT9Fwh#bB0+AsEiCsUa#fpBcd`3PzD$i5R
zPTKvCH1L&#b0}>R8hY*M(P?|*^{e*NpZ?VMS#EA@+WmX?{cK2IZ@&iA9x>^v6RxS{
zV0$G@#EMLOwShi^9Ea!Ga>PmmZ;|jW#q!FEefrs__RY7~ZEH^*nlz9$s!4T=-BZjD
z<tjpaonr9HgXKJQx4yMzoD1pn!{Hb0?70he<j9P#z90efg@h|~Q^2LA<fpx;x2hw0
z-SRLU^%QN#o2e_xQh4a$;Q7SpAcL1TUX@cTaDV*}AHq!LE!$En@U#^=P{%L*fQ+N=
z-U?nt_xG`<TRpxsr@Idz*!{N&Ww_Mq@ihIZCh#*{=@)cTrCYO~huc9woCo2x%43OF
z4^!fmP!E&#efh22+z*AH^qgo?Wxh(N`=x&kZC`wK8bB+zS3!vn?eC%w;En>fj`Lk%
z>SmyB2THenelos<FLQn^EP;OA_-Z7`yp`cIOuChRt>9IDW!WVpeyeb=q6OezSLxm=
zE})KELg`-mA7pMq>$rfDUfrCkY}1kM^lSEm#LHa};Dr)zpg$J?1Q9-z|H<z2w1f%g
zQS=99rWiBt7$8Kj^8z!AZtxBn2lNpn!1Xf|ETq9~X)g4$6VG5;&CZ-XX79ZBwoOh=
z_(I&>TM{5{-nDzTXT?A-+v38qnCS(3`uM4>XpUriG~2;>P_e7K+s_v$BIhw8kHHWF
zT*V%@8B!*+-Prj+e%d>5XAxJ-G+SPzV66EH@af0Dv(Ggb-@9|yg&4N63{Q;N;N*l2
zO-xvi=KGz^UE_wukN)<T_N%}88@qY^x~;CQdVT@Vx72eYCS6tOamoTe4wtiops%;X
zMuvwJ2VYzy^uVO5T>A$HZDedzc`<wV;E_GJ_t4*+pxh2CPtR1wyxt9M{-O%A{AH}>
zV|N#LoNmIAKAK;6U~9t-tf=?c7KT%_4f2401)e49w{(0NBJ6(Lx;X)p`-Qb(cXNRU
zB;Ckss1o+kg=S%I;7UtByi)~KVeoTn#rTWAX3@S=bp>CD&WqOwfj$pQ(bu}vw)(}>
z4f54aZL~-A)hV2A>Q;2x6}f+z0H1|axka+7ypX5LOz&&^lOJ9B84N9zBRhP%CyWM5
zCB}_9k}P4c)v)}!FO&%I#>7DXh&pSZ8uOOjyZ6ABmX?X$8AupMm}KrC-hH+m_uEPl
z2h3-qMgW;fC$vc?^z&X}>=YLT$Dor15``anniG5lLs@j!%0DXz)P{20$dBhs<$zhO
zTyhhG4NPq8W8{%MPY*<7_Ro>K+}O0`UnQmnJ8V+I#7Lhi@6a<_WYBrMXmhK(Hox)0
zmUj->_VdHWFo9GJ<Ic@gzAsZuvkzpXh=dWdM^q)Up_CUi_lx?V;S3J+*v!ndUA=nQ
zzW2cg{#e}F>Y9E1)z|jfr=Q!ShmT`*qgvHKfZ%Nq4z}~vfHMkwg&|T_PY|Tm{JhDI
zcVxb9aHTGSz}MHlabaR(NAW8ae376&4r?$YVbrfTKqd##qI~^qh2q`P0F6+g6xh(n
zpj~?HHGA{zcWiicLiE}0H0(mpKwrOvofSXF!m0<Q?)6UIRR+H2ROzB^(sX@bQQ6AZ
zl^0J5I4B*Zd(=Di9o`9E{56QWpkG7`0pS>JFX#c;-qJlG{Yxlu@UMp{?RCF|l5S}R
z|5b6K@1tMoK>F3)4g%uIxIhbBXe6y*-q2m(bG#~!N|lgI2#`ATOLri{mTq+@{p+};
z`*%SroJQ%UpL9&Pry!PTPL+^$+O@ZBw+E@_*J1J<DxsuN!tdi=ZyyP*+z%@266*e~
z;2`%BTG6TNK(6e%s(`X{dzFVqPTZ+%ms#HX^99uHB@jNt)*)$^<p(7GLE&1tCE(Yp
z&6HuvJk|9y*KvUKFY)Sn6ZYo`$g)UCI;rY#a9Qt<zr>3zMzNqr9m0!{mU<auhbM8y
zG8u&eXcpYy$ORr<-MnEy9DE%BwU}QCD(2T0)reby1dCXE(pg)RmtDDf-rjiQs`d8w
z`DH8*?%eY;fGdmK;>Z{H=eD-KY4Zz<ezOsmwrptLB3&F<W$xq^VgG^-t3}#Lv*PcP
zKMR<|$$3Byb#oC72e@J45H#2!z*Xu95xi~1?Tep&{IOlTcEg@LS@hecx_e&O$mE!f
zP9L$csVQGLT#=yj?H6C$&;R1D?W<2dv-$a_E~r#FSJn@SAMqr~ge}UDEiUd)6{yU*
zyUfPNM}524VP3=)bIO^Cp^*{I-Q&Kk>CwYkzx{IV>4NgYSzE5TMOb0pg>vv$lsiHM
zQMp+PGZ1Nt9}lHA0S^i<fg5%d<wsefS3&U%0tr~X?d)%xNuxw9p^lr7rgSUA<Cg6+
z^#qTrQut-~gu1;1Qt4rRj;I_PIS#+34(%8Z5&C#n1x33@!0<Un3Za!VE&V+`Hq_T=
z{oTBL-Dy4QD?P&PR@hDi5@<<r(4X@6FL;W@`=j#VU-tADyt??%KmF<JKLa@?Gtwh(
zRWz%)G?1>YZgm9q07_OB9ug<3z`h!w4!O3zWy{N}_UQ2w2^07I_9X@aaCjqu)n^oE
zCY%J)k=h06aZtH9L#H@VNV1VQxmRx4U{Fz53M}0V*_BgjKr@3e7=<HrB0PYOJ+SrX
zjU1dPAZLn+L*)`oK}f?PNy{JeuDapwrr+K)KG0!DhC6LysMC78+k8NJG`D1TpRC&4
z`g2>{IwWDjY>Pom^f({F*C4(|AW+0Pg{bU^E8@xgg(5AanY2?j2qY#3ywhfmPTJKg
zuh|d2_dR?4>Q&Xlil1%x@{6zR_U$|N^vP3y=6`BpO3XS3vtAgF3a+lL*@}dY_%bCi
zNBQQD>uUXsN;RnRWny4x(9?bE?rpnv{hF<B>}n8F8mSH`RjmW%MPu$(1Lfcw5;5uA
zQ=B5@c^n=du-7hMus{0aKe1ycPYb_W9%{_J12!=^X+xgIXA&kBd_N|ykdpT(k4n&0
zgHr6(L_(Pndz7DpmsizB+#$-8YL`}-D>ob~5PAbjGk9*-58aBtpoUWuG13@{FbNS)
zC?Lb69`u{$grt?`QZMlmN;(HY(saB+FQI``nhTu$aYtWmj3W;CN14^b0@zzI8wu&2
z@g$8>WnN2_kmkByX$DGjnWz2vYeECRq7&nh%;D;Q`JgyTz&RAO!W`vR0`>?Wcun;x
zNdMHyPg$0w-pW1A3F(F^@e?m~;GnYmeIRL;DxoxI-}5%4KeuJP`;*?FlkE-Nnbj6>
zP}<A5ff6t2C!{^yz(v*ZfO>po*u<^dlNM0dfuxywiPK5}CCw7*=}Txu<GUg8_Qwlg
z_c;wFz7i%F>tU$(frpFn#?K!J&lln}2Q=viurkXHPCb6LI9p{jMu|%z$_Ovs!xd&0
zq{M@0PG<gwrkvEk%-(wEik-V~&g14f^9SN}9!r?m5&wfALQ5ig1drvl4KeL=wj{o2
zm8<bZi`S!h)5_T~Pk-d!=O@{=);IiiMb0F0D<Ti!&OM#?TMgTlf6})px=Zs*_N0pc
z=8bQ~cP%OoB~s(=&~TegAD^-5lgDjrd|Y%7+tWw0_T^`v+i!mHE4z35uAgCazDPjU
zFrfpUD*Z5x^JM~H&gD{<&aT5YHZf$q1AYEgLW#0&2}mPj;}Ryuth;+a!p^GQyYo=#
zdukgh7rywkbx14``rdPHstV>G9xv*g7^D8PFDGDgcNsiQMZcv@IGu`~gD1Wj&YhTx
z(xU+A)Lqk@ZL$QXmo!mTc--AwglWWGxz|+*b$bcvr%(}=ig558hOv?Az@rTL<tqf&
zAM<`3JeUj7e37GK#os}yQvw>-`}Rt}?G`V^c?lOLB-r-w6%Rfg9vGU6KV*bTIU&Wy
z_(>U5s_kF=<M)4tz`&UZ`3JL1hf+sE^6Qe+GBPse=Ob7#?(XhW2klZ~yVc1%U3j>8
z>yF*{_NFbWGay8;noR<bfVcex=TJ(A3lg!-iEc%wt~#_5=3fNl7D~uDI8d;zWCdK$
z0!{BIbfEC5Xb7Bk02SlKL?Jjas)4^MY=xV4(&aZ;sa!zUkY|U)Y#-X)w)W@SHqw2_
zj*a!%v59^g9psRg+3Lo&-Fv)X4;FT8X{XKBp0(SS<o!M6i<sEj#GtvW4#8In1~$(K
z{)&(CAQ|K-K2I^I`cUNQV|x>V9o05AI%sdcdDTAr;ScTNg^RYiwP{~{@uhwJ^*0)H
z9{BZfyjI9L1O_Hv{_a*C@`eYWWO<F7j?Vc7RuM=OpLj_F>12hai&co>Ar}T8JbLI~
zCsx<C9EY}rAVKmazxv?MY672kIlPT9+S|(w#@k+&{StmIUp{R={i8pUKyk&|ctw3j
z?Ty={_MZD4$CNw`ns&J+!H8%3X(y~q`e$DxQ))uEer7~5Q$Ms_iqgwY7&LsEK4#pY
z8#<Fow7C%6!VRdJQ3FbQ(4<btV|Zk{ODNkSaU@W$IoJS+QxBgwiC@x7sM`}-(M)=;
zqLX@24DRN6t=e7=CM6BbSR<{^Dtq*;64G6!aH#ZfYvvV|xTrdvgmh25G?(~kF4LSk
zkYUrURoFU|VX|DjY%06^39nvGCB5)VTKj313J4v0MZ4yoP)~0i2mga`N_Xs}ub%!i
zw{klOGR{W19H8(|z@9p)p61e?^c^=RTf$difrh_SrX%&TZ<hFVD9ZvpRf(H!WnR-A
zNZhn1)Z@#rsY<srr(U`Pb$dd(werVK8tQpTDBUsF>*^qyK+-Mkp}Xd*or#0;@mLBb
z5g=Sy_JErSFQw_q`IK1Ll!Lt7zlpB7xJ`3Uw=?sYSuA8u0NukO2<*ii;PVUaq(f;S
zEwB$gy@%}m_uuyG%$X0l$?(zr2evdjuX%OHA9-UVE=+cJm*NnS?cEo)pm}#;alvng
z<m*3kE-y#-XkPTY`+aUYY|nO|+1k>IJ$&%cZe0J?Zr;2lf#Qihc``2seO`EC^c1GA
zw^ss5pABk0-P~B0|6_k23qF8@(i2mCcKMAfHY0(8g+Ct9<C2w6Kl#i)`ps|bA;O7x
zGKvpCrw61^k#_XN!Om5&hgEjNV_m+RphJRLXD<)hjoQfgl=N{M9Xo0(%Nus*=3W0n
z!P{L?&f+P!byF1i^Rg8^>MQD!_80bqtUg@4Z757G3i?+_LD^1%ato;FSC$A59k=a=
zXg76NL6xoy1750xR^}2fRV&CoQ9_BE<Np2xV~KK$LQ?&JqqfZ#ZFZ{gMZ&+xNtjSm
z6R0K;4)7Olhxn(S4*82$V@nfXBD%!5_VWH)Z<q64hgeuwC-d$9M4||N%#?fFe8y)x
z1Uh`hxBcUvzVWk28VS<G=vT0Di&H$j!a*A9OvzB9x~xwPd1P$NCDNhc5q~&^3gffh
zlP9w-Ot4STlkbUu3FpdF2@^_iFkx&_X%4zxST8DR)6-091XV+GqYQ3#okCHX^%E<h
zsKnDxWx?G^xxeDefe<98i<DLctSZRUqW?%sHOM$k4yn69ZtarP+mt7<z&5C?9v$hi
z(?>_`$XKri6S2$^pq^@QxI4RKPuHK@+FqM&9qO=WszBJ@JyinCI~^-j<_V^xNTbrh
zp&0~#=uFf?;KD#Nm?F5m``(-O$AA1scIEPAUsbvJ?YDO8=1mPwe1=`NjkPtu26smt
zoU;hS!^3_X197j)j{{|SJ|wTNLy>LharG_dCU~K|Q;m&xB(8sZ-4+(syhCf8@p54s
z7-^-kQjXr1B<YfN%7pZfPY&7Hv&Ze?<%@RV;zgUBnwDVNW1E}18W?$-WYymQ;ksUK
zr$Dmy!`sX8DmTDqU;j+4((tWt^b?hvlJ0fnD@B?ZV$UtPpv)K+BTs<4svMhNq{8j@
z0qQ6Y%f%kwY=f@?Z;M1yL7nL+{eValZK6pxaSn<r>7`1jn@hVx;Z~-?aix>iXnzG>
z%e3uhHWEC1m0zH~%?Zf-0=Olem*Z&aKuM#7-v?(O2wf@kKg?l@KkA$OH_Cv&G#3g$
z&%BmysY<`**ws31O~3s(t?<*0d}p|HEAbC1*9?b$sw~5#g{r$Hl;#p}FY}x#;k)b!
z8E@(U<}y7Am|MkJs)V$^3Ln43E%kcb2jQ2HI0<P-Wg6;0JSASLvMjKddP3qRv<ib;
z#*vW544;0!-K2uLUr8Hi6;}dYxMkd>!kw~csGJi8_$N4j!Cd8IeYkrp&~Te+ZaM5<
zCU|Fp`Ns~$0b-^DnD7_X4I(}!F+~Vvm$#mF;ufc#o(|1ZXY7OTeIRCi(&tNdU_5#F
z*p{9y*xE*%ZRqT76GPu+hdFS~8Hmnyzet6f44Hoy7v^kBb1-8FgkS8%wsDo7-;Trl
z%v((lA3U}B`AyBq;=OiF%phO!hpfMs+mt$eg5#D%78CsuzJV^AIMQ#IuAH<FK77}2
zM;sa%w*LM;<@d17&CT0yKmMJ4^4aIMy1woCrlhK@zN9BmepEG4isiHE8Syx0PuelQ
zdQ48($i$?LPE6Ubgo)lh#n(G5+U@c1@4ZL<RfTO^eoLsB19%JPc@zorbpmDV^$rhI
zxltzGu4KmF1rMc=@e+Or2p)x{&t^Cv+mM43NP~VWnhq6?+i|B}Dm#i;8n^{GjRL=f
zbT8Xz3F&^2f1R$=N^|3<2~AuoRc%RD7FFWVb-RC^$oIy;q5boiF+u}hA^Z_LrKL@N
z+{(lO>`u0@!B|T9`*ZA|ab7hqB*IPojUx-)?RHr8+a{r<U8?Oz@16M>uQKm!)6mMy
zDq|{AKRmm`mX}@QC2H<Y=3ZU|4jwKUmUPNhp0l&;6I}MM5_^h}L5Iwd2=?N04O<%-
zQ%2*<M5Xv+vQ}njW1a*rBo$=U=$d8UOkX2>rP4igeuGAs5R!r+cPIHfE>;iQth9tI
z1I8Hgix9-YsCHFM+aU=|FLte4GX0PSk)vb%cH+pe_4V*YM~v_Kn$0e*+Qa!RTh^ek
zws*+3{H(+w+f@d001Q1CU;i>ejMczNx+oJm51lBOVLD3S(jmdUbm^r1*`NIrd;9IT
ze6@qaSl?d%)&+@8G2dGo8(u%WjlpUL%s+L<tw}u+C>U7%V06C3O-VG$1z{r2bRZ!5
zxru=R;{m|y5?G!-UG%!9fl{_4vE9!yM0v~6)9W2m<k<2wJw0h>&YrSMuU)e97cXca
z8nC$q30)js)8IV&WX|T*mNry|9GGJz2jDCY10(xdNlUa7$|e|k7bX-FX^8S<f~gu%
zRJi_Q;9+QTUH14g;Yj;H6f4uqKz&edO`*71Uc{S#8*!vx>OiFV;5dL*;eb}*l3uBP
zZ%F(q%y-f8a;y}knR+WAe6yTNTKnk-jR0_g^egewT&g-W(-zPQXCDX^oD>#-q@C$5
z&HC{bOroHT<+xem(GJr7vZ75i@qseEF9Yo;Os(=&(kR_hr9J)A4ZtmFreEf>Y|9BH
zZm9zqzH|fX^;yOVU_K~p2_+s7eSCk|^sBq2dkF{Ov~th%Voqq~R)(oVE4PI7tA{P)
z$Z||bKj0wRucDXGiVj+aZMyAGL)|~2v_m`Xp{EMJY-9B_q(YF$IkCh6m<O59m`fex
zt(N6RVqgm5y#_80;p~KWT4$={%fv2qDDTb34jpHr#Ix`M_TCH4PhIx<8?X7Th&{c%
zwy?P9Te_F$7bQ&0+s3AZ4$<XK|HB<(oTUzF9`-FZQcOskAK2X7@q@|hD@&S}*EJ_^
z`MXVg#qaIu^)K>Eng^Gb7JYuDSbBR8+4*y4?85mAHZ?gVMvpJ}yT&80b8}B^Y<$$-
z``}&spZ)0{*&qL-KeYG0_pS{~kYIs~T_r38uZypnou9YG)n$8eSp7x33wOR!He%nz
z>>skeeuNTs9`xDx)UX{pF=;1GYMwqfW0$X-vbWy3WN*EFRde=5n?5#UL!%?sKQwFu
z;{Eyu$3$n)db$V27j@d~qo;QN&OP5&&11nZw(xy00nD`s0lwG1s!R07f}Xns1g|Dl
z0Zo89%(q5pH-1@+proarhQ0B%QlQ%KddH>d8sP1r(o2<Ybx1c)bH-JKOSgld#QiRn
zEKEtU;qQL|#7VjM;xocTJ6p_T=ZvzD_(H@vR&F*!m|)TO5W=+hx5Em{<8wU{xY^>=
z!}cgCgz#SV9iEQhfob&KzCP<6?6<zrLF*kIvYs*NW9UQH_Pw`GNtod3&Tapyu*Y-q
z97IwkI5Xj`Kpl;haJIf6Sj4J0lPIgioqo5lUkkRlq_FZ=XJk7Jos3tVcb{`^V$UnL
z`S4e>l038gQ3>(itPsVGIg~p@fCdmC{2Dm&&H~4b@WNJvKnF)@2dyIGVFbxcWZGaF
zS%GG+rLv>^szMM<UdT<tTetEz(sS63jdk0UR96QFsyA(JWzFu*t=Z$%=eE4lW*d9$
zw)cYT;n@2K9acpUBs6f57Sg0Yk6UrNS$dfcBy(^H$N2*~)d_X_fAXh)Y=7{RAN!ex
zM-Lv_wXeUiJGbumwjf^f^ubp3wY$T+58E2lLU8Qu>#^~PF~3Tdl^C`=?MaV2o;5fs
z!_<?%{UHVO+}E#0q{al@8{gjWm6Y8r-ri7Is-7ree@IUGQE0jG<wxbztA^R%YlGv1
zHgoEjUAS_|E=ZUd9Ub#__vpbRTT*&gG*GW;Fx}c%6$!<=yRAX4&p%P~Quyrb6VDgI
z9bcIcB>eQZ;^z8VcUPHFo+(Ldy7ICl9i+?W6B>ogDp2KDI_B^eenONz_AmutrZ%V=
zXi`RAKUG~3cbb7rPXhK5H~q>mK;oee!U0M=prn~lx7VSh0hN8`LIvPg>M~V!lxiJ|
zA`N?Nc~d>fOfoVl&`&7ymnxxdPL*Q_s#W>mmj3Bp+SAQJM!B@aOS^tTr>^BjjKtx>
zeP&9KVOl|u!9QuIYNj=z%u^Wmx1$o$z4T9dNfW=i8+y8@opJ?mPgO#OFI62%w^!k`
zatAWb^xvO`Dx91meeR-aQ9pF1GyJ{JS8b%shx=7@0o>D`dL1`$(mx^X2e|`vT6J8Y
z^h^BI4}wfn9ZENF5--g_9S5k>s+<MQwE3Jw`v4c^_6Wy}>VFF7xcUQLjw@Bnf$5K*
zF9M0diSXIKoBcfTQo&H}6DI1QAgtc=@-++=>2o@s1oSGEOvqBqR{#rEv2|9$L~pmf
z{q_|*f8mViJ+r4`;yKT-xw>I{n_}!&H$^6eiFVgt`2KWw2im0w!az$nVe8Mf1c^N{
z+03;(e7m=|Z9`(txN!&b-p-cNvgg}PPMkPufAS~)$bR(WpV)~roY!c#CvyunD<S5|
z<5|&fx1&do+PSl5ZRXff7o1qk<t8cSdcI)nYOe3;6n`>4E_yS*karq&{*>n7^LGAn
zQ(b=ZlD+-@oA&(=-?tz9_yd3N_Wkd_V{gCrrd_&vS#$SE7a-cZd&J+dJAhjk#TQCY
z+LhW7Z!s%D>-Ozi_Tb@t@rzqt9-MD%)GuWr<@1D!+xZ{nVa#6o6;$=@6(m3Ej{LI4
z9^Qv~+!G^QwugjlFRrlTB?I;8#GvN*sesfY%7i*?z}rv_GQJE<diigpqeU3#rgP%N
z@C#2!X%H_sp*I#sNsko#s&9RVhH%IqX3jdawXp@BuM~1nyM>o}0Y4-j|2dz$q;T8b
zTvuH@lK|0Sg9E+38>Fwl&kxskiAU^K+w2}v`xKwpA-x;HVPeDvrYCG@X3F}HOxVEW
znDtJMSljpCJpD5|kKd=Ph+sO==}c6z>Rl>M&f|{aJp@*}X|UUBM4X%85i}T54(;wq
zfbfI7l7RVR>kLdOgFDPcK3LbNP~|t6YA-79+wP7>lw0LTfx^I1Y+hY~mRe=rHvswA
zCv()2P>4r}Ug$`;(2RiP$|OK~*=MP7NC71Ut1u3|kYxN!?lSkbn@tXO+VRP58y%Fi
z*S2Tt>c~$P*CkAB+1$F>ngod*g;CW)(K`|Ypx^T=hSxind_6ZrEfvfOX_oSI=tus1
zUZNe4&WVX3`_T{IvH$GP{;3_An$qC%*uMJmD>3=sO0awCUnLlrsYB`vrkyVgYinD!
ztokBv<6|Q>H8tUP5VNmyduvnm9@2j1o%B&Y=yYPFQ<8d%eVq><+UCZN!g%sjxA7S=
z2T+oYggYoHc64&oW=_u7*-Pi`q6CQ})6*_^KDz(V=AJyUm4yY{SXs86Efw1Hc<W?q
zdqaXmmj<u_2@4&zEdM)q?rN~)Ob$|bJEf$!Iq)rE0%k&G6DuH0GBMbYm#BRp>Y9e+
zgP7;ft*#m%I%y?4*wdUo-kDuE4nN{5+fUj{H-PlumkR&V?e~SGo%BnZ{}3qYl=S>p
zq!{?~uN1gZzH(UpdDslEKg!<oMw4i!E#oNd=?0{^bOXT04VvkfG;m8O@zPvZKxrp$
z38lM!n&$AVbT6KUAF9kxLb_GjKEHs<qk_Z-Ud9dox_hfIt-|8gOway!@hii;Djjv)
zgmf$QR(Rh9Wte@mvTg07ne8F^zvES7XC2(H3QPV{0l3%gWt%D?>7+`xgwhPW3a|77
zGX4@;xs_1&FY{5tLGhICK=|Pn%Ka0^&6O@tOx}W%ny}^7?4JpdCiJE~`!QyiLj(gC
z3@i79#L4*x<@2cclCVcQ5-{P`*WV*yqN`fa$cRPPjr?f-g<pZ`kX1D<20RX6GY9cy
z;_Y{@*qL)DjkmH8CKl$GoPq9X@Aj+H!Q+M_lrN&{kM#R|C^IX-2ovPyx%jM3F?le0
zoV4JTX@8kp^JFewc4<C_v11$XpZ(c?Zg0K+flVJj<>xK$-g{`b?%uOG2^H(>yyvj%
z+lm>x*kQr}YaXHX-6WdxdEIzqXvj{TIN?Ial`B{6t+(H@x88fp-hB5R>F?O<@4R7G
z-@Iy9-?(D0zj0OSioJH_vYkD5){dVzW|NZ>zAK@<t6QZdCjJ?`=nKu&d%pE}ZF$YN
zC36cE2gq;TykYn5-*rB*Q#=DaA9dm9Q&fK_+KAVi<^<+zY9i_|!1bc};0JHf<>z&~
zyzQ}#$`^(S4}ySplBzkL<F{W0ex+n4J*}Xm4J7S4yzCx%0_ylpdyeDi%wOS_f&_k;
z13jqGrW}-}ctzbUZjUc@GK=r-Xy@HA+8LK7?Wk@xZD&JmVq;5f7eRK%`S)kfwj@CG
z+LVNX$*D;j9Fefl-)mj;o52C=k-b~0Q|5LF7Cn587$31gaMdpQ#zw7gWY~I!2d(V~
z@0|Uar$q_z1i`SXBT-Rw6qry1gl$RENWE@$cQxGN3NydbOpG*d6L3fgVS)jjeoH11
z7W}+~7}Qn+bQD#kM^XYrpkARs+WD=D*DZ}fC~Ckxv4;vLC$4`QP&CG%5f!^1+)W4^
z$Z8Bidpi;hcKu<xnejde6FfKHE}UIkT3NTp3kVc@wk*DQQ-a5{Lml?~5a%TnR7@_n
zT=|L$aX0>gG>MG1B<YgDg#!r=FEl{VU^GDXb=rk<GxksZ^dH+BS6{dJxjFmdv(N3D
zufO&KSG*xXJ^ElPK_cFh;F?#qs_og*>XzzHjCyaU@39=48266P{nPA;i?XF5#tHys
zN}ZEHIy3cp|K2^nQDSY4ZEeDqAGoZ>FmS`nQ_X#nbjK#gZRQxCOkcAruU)q3sUwn-
zpWE!i$9Ct&En8k(wB-f9NUSLi?&gLma52=ndL%KcA+v9ka~lsHJXGAX-bNXCeUk8M
zY5sY8b3u(dqtivbMmqvmWh^^w(ghFPJkJ4cs=@%6Xe()UUH+^JP(O*9G+MPI;8p(V
z7HNDLuS{RMqo+!HLMy-0KcOC1D;(-SRnke7&?=l9r7E)^^Om@+!jyDd(iD_9K$!<r
z-CmkYH=yBG*#WncU+~n{v1k4hGF-h}5>m)2A@jfA9Cq|Fd;)HP*YwM{N|p3Vw-UF6
zX539E!<G1@{UE(g?;yNa0XV7Z_R@{wD%}zqb-?WE{uMsv3I}`Q)xpEo+|yp_uTlwh
zd%{6@b=<l=%eKT%^(ttEQ`$@a(!Jg`T7uA9)jO@;J^epuHc@D&l~LgMRiAPu(<i>{
zQ~T3ij|+SR2^K-iFM}lzysxj@&obl}bN~9lKS4ZLSe!QhN`mju5M6i}wg9ok=k%GA
z&I9pPU}0g|&qDaC(0qxg7EL3&N<YF7`4>GD2cda~OH-rL(r%+;yv;DExsg|-JN?TE
zmz8j_2wRYuWBYph?Zk<b_TdkIWMfm));}`h2c;i8dTjR}JrqOT>DyR%`FeG2#pgN>
zXCuS~Kej8OC>DodI(xc%ZD?S~#>dC)*vyQb6z%cx2|J>>n6DONqCYe+;J5o|Mw8I9
zE9P_4w#3iDJ8+51$_kI%ZiqH#FEu+$5y!TK3U;(0EX+Q6>I;n=Cg)2G>7*W23|_yS
zBVnr(<wTL^BC={gX3B3r6iq<o6FV~e@m&S-od*OK)Y}JfmhGp65-%WebF2+GD11WV
z*71^7LgJz0*HpA~_o^VmWqA65FL>f5-ZO7hay}Fc;jtLXP7&VnfJH^nBRE0?hw?d$
zy(fNkTkTI(Do_vQ?5=1%S4|I(^xNrkXYAaC^EPwhxP*xj>*ni(gvSAyQGJ6FDEbDZ
z`mMKrP>Nf<Bvi=WDFL&yyT?`A58pZWGfx;lFzzyu0VEt&&-04cZ6sX@7=GO=oftQz
zZi8(z(y^~^ZrbwNvNOqKg7Xm+gx?gTiYN0vP^eXSc?r;cWw)|N!SNf7&C?{kst>0T
zs}muSR!DA+3aTo@6#*hlC@LbH0raE`5F{ZRfCJ2_nBPO_%I6_Tb3N^QHZ|O3Cyxx;
z^mwoJbV_1ZSuL(?xG*ud_RN;We<MunNswS5*j2iBR1TCMoywa@rW25vfe4_SV3s+Y
zMWWf}^o(m2^gle*XRlwqV1M}2pBkL-r=NUkUwrYEza_yoCK?QQY)jj$9;e%sP{?61
z^?!yve?x-9)kDJrcI?<O<0VV>YqHOg!%`s;BT#gDS!}Ft_-ai@wcl=jZr*RTSX)ti
zDi@8Z{sjjC$oHHot<-=2P_IoMnY43f&)B=~ylbaVp7guC@87y(-+pt=79P*q`pSv~
ziFF^CS*79Aa!+4}^$+%12X`fFGMJDQ%p-ijwQFA}{#75mI(YkCnI+wRuuf&ET53#A
zu?i9(E5#B@d2$@wWG<`Qp|BHPt*0=1Re8ImjzEq!8G*e032+0l{Y5;@c3MKZCk>z#
zUfSz=3ACLuY~5UkgYaDOhg;e)S5T%k&FH0jLb_-8R(h$*Jd~=$E#V+HAABowsRJ1%
z6bwLRuHuFxK!-+rfvVhJ3LC1Cch7!~TUhp&2e|v8iIX%MwA{_LMz}Igpo!OPTlKW2
zj(hr*_R<X~<4?ErM}eP^ctH9mq&dTs{%Ll5&HX@7RHc8$owQSTDCtyY{4#IgmMZHw
zRpKX<`2<QoAni#%%|QCK3R8CjO80ua37N)L&?<bZcoOP<Wt^p-yI1`Jb1DG0#Ke^0
zN*th-J;zGMa29B&i`;7Kf;sVNTz8UE&X{UG@|TudfTO8+*v{Ln`2$)$0f<V(ohfw}
zpA?;OlM$YWIhq-IV4%z1en-N@=@WjI;jx(I`GqBCs5{#`#Ps+07k?Jx6x{ReH#KPv
z=DT?ZTW{sT<tKcRIDUM>re`KKCyv|1(Mg+{Iby?OqyEMc+jKb-!382cJ-rfE2JM9A
zy#C=4JEVDcVQI-`7Z)U4OxnNrFaD+d;0HhOJl?*2)3;uoJbv64>Uf3v;k}1(NPK?D
z)|c1ZZBuhNFJ(V{!g~<6?as~HcK<F9@7=Mxnvc=%OFcl{xhER;?9u)EHY)*!1xwB?
zu1cU_TRQ^A=DK*lt=RV7+0G3>9k#r<EI#Cs&HDlDU4L{BVb34p^Z6DPTdAYoqs;y5
z1l1x?ZHJ+Pw4uk_XDU;wQ#65WKj<iLhm~ptrGF@QZNy)P#7n>&Z79qo;GW}1LWVC@
zwC`Am%rp~*c8WRa!(R{scDe{U?KAkZ$ZzoK#6x8ISJcQ`e3|fDow|AFq|3&}`s~Q`
zh|L_IRJ)qCv*(Z7<u|X`<u}y6FPyW<nQ0pwAGN`;G3y(a03qSB3xNUE+b6+SI=8t>
zfH>?sK{$sb{)d}2rP@Ax_x#UD023)48O7HFGD^iUvGOhgsbtm1w<v)loy?Od8V9dj
zudT1D!-*;E;%xyw@I@o>L)dCqWYRxB%aDal!6o=1C<XGH6}&KMa!r5iKn^@W`bH&H
z6N#Hcf#7BgqUn|7v^`Ydgt37e08WMPso>m*D)HI2wIABG!QR6*HP&NCCkJe7Pz9@U
z;LVN4^L&}uvW3mVaf{P)NrCJy^!$pR{<bvm1YM7?NJE%aL_VO>kTCldU5K3((|-Ku
zxV`<>>vmQH&BJ^5{Q(=U<@JNsYG|y4K%UhVgbM~iKNrImDMbU*ul6T$&(+z@Mn^|%
zM2g4Z{H6%ylN6Dbfxcdsei^s{_AMeIuW3->kk$I?y5CIU>6Ab1h%YF-DY2tS*n4;B
z(mDI-AN<75ojz^b8qjWkd&6#B`_>j_x#40>>E2M$Ja-{t5Mg3~17W=qL}<7jeo%a9
zbj+`zo1UEX?R0!jU0s#n%D|*J`D#Jj4&;qNO9CJTL0Lq}M|*52m1mmu@BXO(FM}{j
zPgBz$VS>KW^l3s4Za~&=+7n89-47sKs)Ou68MgFqMH5Jzx=OzU%-&D&Czl0YtF)u1
z0vb$tMY4<|Rod%N$1MkEpdL2yN;_fF4wU8+GMtZ1RXpS&>o94Rsx>IQXEx%Aa%h%Q
z#$To*%c>bB+)Cw(AW$yCB{ahaWLnWvrF-fbCf!;=(n!F3kV^lA(hQWijj~G|pA`1v
z1N-vjWs-W+y_^T2>yxwU&{wsKGHAvXX8gdftDMt3jH8E6oT*BCE5N<%`z7E8&Gakv
zR0)}m(!HLZR#3Wwk0L!8HuX9jM7s><ia*NN%}u~gIL|xPBG1Gq%OmNfy^M=I!*isf
zK73oW+yQuv&|#orp_7{s%XB0V4NZ$P6I_o;oAB@o8*zG0q-Q#+xk2=NrI)!we%zMy
z#v8BM$<xO*zpwe>>xKD6XR05IY32*X((;Ng)<Iq^NkWA9Ceb@AK4wpI{od}j(s0;%
z#n6542Orq$@4RWpPMx%2F>XU*+-6Rkuq#*JvY~-tKZLy?!2p3|pnpIN{EA)w_NM*r
z)6eXY82T{@3!~%Xc64SY4n)h&7NgrYZ(4tEubn=5$`3$){^_Uo#b=+nfUvW*X}ol-
z2~o74+5-s?cW>XaTQ_dV|CZEkdnCc(@q_#J^ua@W`dGrlBMAhvvl0qsC8*5WQ-$Nu
zH$n+-SaE<H;mi3%F_M%eLLYnqw|er<2HT9iAXJY0!Eg92OsYHTKt`vb|FB0@c&(I>
zc669+wq;YF+5hYPE!!ASy1xnnE!uDyz7=@euT(;5rc4qVxNu&?QK3ctBi-;<B%r~+
z2oW=asGs#Earhq&cW<L&2VIE33|*liK(GUBX!x+r92@qnc^596wF~FZ+Qmy3{kG57
zUVBaE3wGh+1?iXUjQppL9<jcm0r6@Q2-Qy7#B)7UJjm((CJ2dET0(H3FIL;uD1PC&
zzv|xmX=GS~7g9701Et372r#uB>K>hyQblqg`4BfUDXf}?q<FTvh(YF-0Qqz39>OHU
z3&wGYka%3HKp1LZNSFvDk!%{JaQIVjM}^EmP!I&o8>W6@s|yxomb&t7Kfz!|C3=ub
zCwwTZ;Z)J2FcSwS<(XH!*?z-kN}iMN?Cw|LciHOpAzR*fZp&LQY+VfXw)zrZ0-{nB
zzw!z5%b>*|7@zo35l_S(^+5?Ju1Z%hPuRrRkX^le!LGh`NrTmMb@r$J;LURvA~>|n
zYR#VSi9>+%L-V+Ki;D(Fy;1Map6}TF!kk^de$5`-zweEjPohzF6dGyat%pwQ>F&`0
z$AK&8Zu=RAKK9m0@<sNO+~SI_1>9IgzB@F49yu~%mzCC&Cyr~NT(EDx{K~F<`L(Zl
zEc@FLyVfmQy}dker?#Uo>~kF+8M1-F0UMtf_ebn_tAYt;^4Kx^$shcIz5d2!8yOj-
zWW4;^y__^@%8VOQdoYJe90N~aR=Q)j9&HrJ@^|A)(8LP>)uR&l_Pt8w#0OA^B{b?)
zD&uO_dzkTq1}ejs?ioJ)4uYhaP`4-4X&r<IfK|qy;p)1B`@f8zIK<N`uF^lDRrtn0
zT;ZiVP~rh)dDi_7f}~w4?-PP14JjJG6q5zNWgM;;Fu;>)Me|jVG{U^UOzPpmtGAof
zOQ?rQm4LZy%cV-ZR&kfAgoYhmN-bH%RgJUyaax$AQn2@78@x(A2j7+4rK104UsiZ%
ztB<=sZX)0lw1D`_^k;fpXvAEq1E36-`A&N)*cVR`j`mZw+jOhj6CYKFq>V~?bsD9r
zhj|s0c&+do^l0xg(zf~I9R#^hfzBxUT<&SYOn#dAa2i;v%5<xKh%4J&tj;<%KP@5B
za<D(iX+l*V-mqv<kyj!~9Y<jV9{z<QPdJjt+&RFu-v<w8?e?9!wy>~hdlD)RzYwE)
zsKd+V#h%K7IoZpDR>~Ety>|5YjF`N$cINzfJ9X~7ojh|+g339~lQT9nG@^O6TWwYR
z)1h|F>sx-I$}fNZxAyI|Z>+1M!!BMpZ<oXv^6mr+dm|&Geh3@g&l9wF_=vN%ye#41
zi3EpP>Gy1RYu)<0+ihxm$fhSpY<#%i26{TJyIpB!g<t%eCU)@=JN9gQ%XYUnZFh6s
z1tEk2b}_K^$>)C#k4wP^Fn{wx_Tj_KxzEHKvYUbJ+40r^v#hFAX;uBw<tPaX0~J4B
z#wCPnv;0$0<zAi=_Gn-AHj25`ICYS!L%OxXNpr?oSLsLm*&p{oB~*Fl7@wRMedf;j
z<&AJG2Jf+GE#?Bg+28399$(-%r?M-Z=18xCw-;+BKJ<`HO?BC+v!iz9^%M5)dspne
z_usa6B(z?=dc_bVXftQep0!h_PN=<1+t}EM_4W0{mB~sJ#}rm1B=RoGocJ=%r?Pux
zgX0SvYv6c;1moZ(X%CSu^5A)qz3oTupZ^(U;RPav<~1kq;-CR&p!Q0FIV=1D7<3hl
zGs%4VV|6>2Wo25*CkNrN7oUJJuxs#fLO>3l_y%#zQJJB~7!dZr^`b0j><7>qmG71e
z&tQ#1<f9or-6~bgGbAbF#qO^`N|+cD!#h1OW=D@qN$T&njjd<)WPaTq%&*zwrEQy&
zP_edm*tS)GNO|5+R0cE_7bGGJL0=kiWSj^PC{|_YgwNE8+2X_(g-aLD*dP4(2Wljj
zZDVcK&m1f*F8PW8Um*AzfxtvP@Ku5b?08`t`edDsjgZ;f+b*VElBhboUwNy6s;i^R
z8wuxRNGF4u-^L?F!`bCiuId?SnFC&Y4O!Npwdn_WNOvd%2?g~l=H7qr4SVyA*L{Gy
z_SM&R`{s=hfYeYqe8%;!J$+rirKx{-zy^i~{pz`~2@ZV?+t~P|1dDP1#CEt%1EmH_
zUTIwL8!evr`x6u(hp}PmDL~58+Z~-2T@I1|a6=FGx{A8^eL&gtI^u;@Uqw@yBxKtI
zvQ4GAgv4vKy8^GY1DMN!@*pVv%XHKs>8Gkg-Ocgn|5PyM$~^)9jL&~TfJTW};j~Iu
z#@8}lDe?}~!<YE=ut2&comA<TknUN=bvr;GmTUA^`mYS>=Wa?@ED$J-?kd=qS8y|K
z*BTIXh|_KSxLY&+W`K+ny~IhCekEj?B~BTpgmlkxM%8itA>d&>e5x#~&=VT*VXjn!
zt3bwvo0Fm*f<j|{YQ%+O<OSxG{)xHFb1KJ2MG%KF?<mJBptPqMyo99PNK*wRP8|;N
z%lKQRpY{lpaim{q{w@VT8%TTVbs#+`pGWmC3Y7<oQ!q%GCU65EEC-AT5fzYrlnZ`x
zMu?!kB5vY|@aT?Pt=>fZ%%_BhXblt`5-7e*uy3ES+8NL)JrM^;(1X5cK2u!MgQ3%e
zCg!29zti8D;LHSDY3Jt``~ddS!m`ahnX~nEG2hIUqQ?q7U+tN%xv6Q3Gp0Mdu_aze
z{=+<Md-Aw{i5M6j7Uy%=c9e&n?tU8|ov@|(WxM_DP2Yw@>fje(=;js{{OyP<SFYG=
zmoE8r>4#O$D+}}X=-wS$o}aVPp#gjEy|?Xzgq78m6}xl$mfgSiNb~=$@Z0SCg|l|`
z^~-klwF?qxj@iWcm@m?eOO1_<*{Fn(;h`ZvpU5?!Y~SK62dZ04XP@HimmgQdbK6gE
zKi7rE?T)CS!6Cn#1s;K215c!$i|OQDt2i`_ph=y<gvJ=@bFRvp8u2>NPgN`X6p(tH
zBW-?#1hO5KkZt7m1>$eUaxl<tQndun43xkLilO*xg8MlgY?J}*JleB=aX=WAz9W7P
zGx>pULs)}v8yg$6Bh%w{{KS-<IeW}5UOHp1aU<8|i!xuZQ>RZ!z&K(9gM+?huBWF@
zyd>A^O2AVa^|!bb7pY>$4Ll#`Vc5>T#nA;8qs8O#pdx|-TiGzfUwZ#Qu#|vFTSo`C
z?MLq+Ohh)bay-|m05?||_DokLhLRr@OQ~>fVoixbg2~mVXfWmsEy%=`1EUurNDlJz
z_!%T44r%V6Dj@u-Bs)ZU1Rf_IVQ;t#F#F1c`=`R)H3w+&DT#Do_C0Ct6JbNhh=)H0
zxLS2|)W(KKH6*m#{KA^upIx?x3+pz!wrvYrV(4ESwmr3g7gU}IQ7H_7q|qr?$-obV
zmCtt+gEOiOuF_czkPM>^-fnNdb=AK2!Mlok#O~d>V~_7Y^iQrR8boy1y?BgH0tH_u
zGA?KMRaM<RZDQ29Gq_6wn$jtOLl~mp=ldQJDEOk`4NU_Ab;cmUY7F~By<GLrf$cS3
z&)(x5iZ<((VAAQT!;a33+naA)w%1?3>Yrb4-@0YjuYF_lb5GS!H@zIYy7{`(qk*;G
zM#e{MbaLE3;qoZr<dJC?A_j-Xtg}ahXB#&-XaLb<)YaAFFM2<E@W8K?rQst?L_6e*
zSe9uuSjmjOPiVj`?UWD>|3L6<PO2l$k>q6T?My2BR6?VV74EHYTHyhuA5ixPO1JOI
zPvW+s1C;&^S~54b61IR=`F1GWl2#qBjIZ=dd)+;uj$4N^Y@Hda`=h`~hNi>^NV_M$
zMbY9%JEA=`^60+-Nhjzw&FQya=a(@OoMw5~b<$T)Bar4yN9sU{m*8pFAYI~Nu7_)-
z66$UZv+F)^y8G_~C7n8?U+IszEDQA9ch7;%JhYiOsS-+Bp5rRtnfvrF@e*+Ru5tph
zJky@g%B^k(NN>_Z5o04(++5RdfrI%U1mI`hTj(g-2@6YUPw@7`7|T5A@e_tNg)ix&
zoSJkpjIVAAk1)w30)dlphR$F5p$H4Z^%7KBN5ui`YMYa%vALqHmg49{g?q*WbL9_n
z>5oV2&YU}8XV0GYGoxI8&jD<<BdxK$X;WBAgTD*rVPi4QFRww6(0sou_qNW%Ha69-
zxo*fGedA`Njm>RaU1eLbcnbZ^K6+x`Ub|)w@88#)w(V)-U^Zt?xqyU&)01Q4{w~0V
z1c_VMzP9_fZd!*TXL0fByKmZngbrHAf&`GKb91))Y{#ZG7r+0#_w0@L-?d}s&)Mja
zDb2SNHUYiqqju!zj2)5L)zN9qy;C+ZH6_0zO7ny->@hEoN&p!d8<k));tPF)5+D#R
z*flUdHleZ}@UtOo|9brJv9Dsgpry7;x%wGg%2s7WarxZqld;ScALa^Fb^9Rg)TuAD
zs-5^5Nzq7EAH%Sh;C-S7``upluR}R*lyNy7vLZLAI3>I*m6V&&MOfMxi_4fG-Nr5+
z_?&9o=j#*aeny4|ZG3!K0^q2fJUL?*FP*d3u3V6SdC4wcjjc<k&YYHTangl|kzwBd
z=u;VWh_vV#&l1oFLOtiPM%ur0zsO)a>Wag?v}fwVbvIL%D1@A9QJjA3{SPw-{au)-
z4qz+cbwvrJ9t9T#;3b2Sojv26kzl|kGH!U{$}$p45E2pl<f3Am133hOo=4$@S9EUu
z(AdCp%^?h4VIF^&BW(cbmgd%Q!|IJHQG8`2Xl7$`Qu;}n4WUYboOTs9%+DoE9JXE!
z2K}lm4j!#(K)Cntsoj6FXtT?kwzT!kR#Y#~HG!#M6jJF@P*z2}lH|fr6>&ycXBIHi
zSzHkgx96M-FbOXG_PQF+nbRkoxxabiTU(l6tR%{~t%*vGFA`fW)Oj4rPcY7EQtk0b
z&A?#4pEKYwL0+e34`gq5pAQVYTutL}iBrM|4F<u4Yv<5uOkJIw?$2rspE-N`IyJcT
z`jvK5Q=@)p?TQ9CUgCW6=#jt5oq_}c8~aB`Cq^`2joZYL2^&8$>4&UYDH)%bvhk^-
zO6RZ!;Xb9mQ>CCuhPqQe+B&+dP36U)b?5e74ajTC7tBHkwG^e7v3_!baI+NUoOb=Y
z+1+IJ-!Ob9c)jR{JDqb|UrvtD%mEbTeaP`cFCptc%?APWTKOmax?j4r3R8y?w^f)%
zKf&Gq^;6{ym42ALZ&%@yj-Rut`~h6+akR3hD)XK+fOJc|I;2~MFI7TmM{kWA{^exn
z;C^zKR{D4tIn?Q(f_|i_4)|B5%8b4rFL9DyGfWvipa2j-@4hUn1l*`k^t#IMX--92
zQ<e5K`@~+j`I&@DPZ}uwgf3{f73O_33RObg4^>Zt@2jqa2i#ND!zHAd(#p18ug}D(
z>j32ilz0hgua_}UhC2w7W-1_QxjDm1&osN9?gGax>~(;y(=6IWLL;8Sy)I0onX*ek
znN*~|w_$Y>{GEC96h@q0-qL+hhV9hMnGT7EcZFZk%y7PC`4z&1s^8Dyab|)Q?ft?;
zj#HWUOiQHGg&euDFm~$nw4FPD#^=4o#T9>ij)Tps5-hg22p97VLI{g+TRYE8@_{{9
z6m1=ctf#+SOn9FiI{{NZ=GUyk%<)B^$L!X*wB`OIyMO1dpJRA%|DN$`c5iQwoj7q^
zOnR@<_*{&-=4s9K%d@j~SM%w^JGXs%(uw28?9%J6*}&A89q!?<ZNKMnXKzdM=L<V^
z?u>o!!|&Vi*DhN3;GiF#?da>X9ti^j!y`5<0fn;|2o_T_(>8S!eMUmaF`GVi+}%ed
zPz;TX+MtA$p<(GmBQ`KFWPQB@O6P!>*KxlwYGZBF?%#Xhw>fd1L<!eeaM*c31c@DA
zP;Is!UwrX;&3X=?YP@<Xz!<Fncla3fakUqq(H<(5bTCI-+z+K&n$cgSN*o~5pLQVK
z^y4~h!rO(X3tZ9_D!xKUxsD&NwM%fQ&W(<aBA^Z15%EUwdlxUBxAPaz+37Q9?9ywm
z*_GE{_j3>D&Yg1sVw6ME66U(tj;{JW+{V{9wlKA+Z6io%T$R9Tu0pur05yUIyI$Do
z)_yp4r*Odzi^_ayW4oR2B!U}NJ!ZdQCq|sHmLLA=Uu{2p@9fWFl6+MGRyqTPXaQMp
zFmYs{o2#Kjm=40KcnA<xK~laEOms3)RVQ|jFh`srgro{T6y``rG!j`HC(_XjuQcN2
z_(24}NI>{wXS)uAgHJL(075(Hr#a~on2^3dOqI{JAMQ`?P+@cB+uZD;-MRP3<~1m6
z?7px~VQ*7VU0rZe(Rolxw>KhaLAQzzsG@16K^7nQgx~3hd|%ka*ofaMF*G<}D;mfi
z-hW^Va|`}ef-O#d+m;5i4H$mDP!L}6d*T(ot4rzbJnX0YQG6Nblp;=UYUor08X6q&
z^AVKI`g(lXSW|uS>Ly#7R#*8FvF8IKb&sIIR{<W)8yR7H(wLpUaNhUlb$50-!~gB|
z>l%a?jQyA0if?Ru#Eu-Dl$x?*r;hnu)~t4L{)Dd)<C90MdtgNAQp3+113VKHj@pMB
zRHxjyPL><2*ltBSW1AC=T{yy~lSW;mC}V^wFXoCK?$ky=qYgrIsAAm*(ZT|Yvg7;+
zjEfHr6+P@gv>#zpryy65_PTrGmjJC+cqQ#uL56LGSNdf<>7I~zN*p))ZxiY?N_)eN
zKLsrD64IP`qF(|Yu2OX<{pvWa{7RKjx)FD2P6f8Qq0AM2<_87^zZ#v)n={lc6){&~
zpw53f3d*>F3|B%cw}cD_lxCn#JL@MI?o_^YuTq%ZF89RGFzKEOcP}3l<Ggek_N6z?
zrAm51yP{M1C6sw7zuIK@R6gOUR<b`5F3o8#%OL$zm!aWD?3os~6JJ$lB_w_w$~X?f
zNyxka;HEv4J3HV<Iuln8ucTf2HNq5tOEAw(Mhv}96FX&*ekH9;Pt-xP{{U{6?h9$|
z9(Zx}IBkTl67dy@_Ow5VSiHnNj#Q+VIgYXYB@i++P`)p`@nu4Duy~U2Q@l>KN@pW&
zGGi5OzB0hzoj5t=2U&@e_p~?^%_83X?7Vn+!gvs$Tf~$3JE~mE+0%R2MkYG!<moYc
z?aEm@b@r4UKY7wu^*OLTuX$@;^Y`P2vv%{F>-OcRpWB_Aw|pD$`uds;4iDIe5&}+~
zI^o-SKmX(t`}m`e?D|(<+uWl^HYB0pqJ*3aZ(O#~Q^$S1jg+);6O;sn3C+bvBpe)5
z_{r%hrDfbYI=hv35T#x)Uuu5U{4J@+^+TMw5Z}@+8VDR+{R0v@c<p*bLIhtUhOJle
z_OivPf6(p2L!;sugeyjLb$QM1-MQ~qxbuaEodlE*Jka(oPY%R_C*|yd1Z7v@gkTZi
zV`pXF2SQNOiT1*Nf8jw3L9X70fwGNaPJ3yV&D&!+o|Gyf?b%K|o@P`LH*Iq-+o*Uj
zZ%5#gu4dY3xA4adReX(LyL7LF#<9^6{{nU4{5iXN<%+%W`kVIlTkqJ_*Wa)+XU^K>
z<djWLP5BlzwwZNNSKg08FnB@ZlN9@Wk1r0fT?#?s#WU<`pR%i%JZ}8BTG`X5VUG&~
z5T;cc5+*vu<917Ntup<mTAWe)VA~+710`Xbj|qH1JM@40um7p?ADN88stQzp+4-?Y
zVMjrN?}w}uH@s?BPq(jVlacG!zO|dT?z&K+aul8<AJOsK5fq+^3!wzF+(pL^qEvoP
zRGFb+(EwR_P;@>BxK5fx(@{x_N&#Hb1U&{LuP`a{0MN*QoUCX7gkgY*0ge1MRiuSV
z&-{Z+9DDwS0b!!6Lkzi?Wu7f3{V?%dpX;B26vml(Don|uQ86(2Cq6#?kya(c<B^KV
zPmHc`aN>&8cH-2~)P^p{?yf^>T(8+X@4Rgj<D>o=^TFNw_UPUd+gRK1EoOXLU6W9;
zp-!wC^+v-07xhF#Kq`F^2gjv2@O-F^JU;WQ-e#tb*{Ks}{r+k2cO;y!jfsOf2rH|8
zYsIEF;-k~kepTzRgn;fI36i2sJueHFI(oLfXR8ZK_T<q+dpi4A^mqWM$0nvUz#Tne
zQ^!~dnX)0}g@LI<0!&v|ul4q5pz0Zrpw=mxol2huISqckZ*PlhXrI{@f|JxYUwvu+
z<G=s+_MiU4fAE_&xUx)@?1SP9h2eVIU6uXT_O6!$bweF^`=yOg<}&Y<^&&UwJlcM=
zf`fp%V#SMt=(Hmj6lv!QMV&bvsq9~6A3~Lo{^?%#Pl&b|{z*65@4>Wz#7mWKK*n48
zc^OH0e-V^4{lJ*=-3svYRf{S;fcH_Ai%%9PxkX&=9c~4EAp3E;CoS~Sy_J9Jsp?R=
zW%z`~7$L~=1w4$5(O2Cl6?p+NO|f8r9&G@~`08o#_Cmv~gV2<6$M<L}oQw_KY4H!q
zif;PHgb?8Y%ncxI9!{=RJ!M@ruw-u5pA=!j6fsnGm5@7vR>q+sh%zWtLSZbRA|jo5
zc@u!RbHN7&1Tz4w%vS<+;^EY0oEeoH<E_U{N`>$F^KhhxQ2fxAs<sJzm2-vRejdhg
zBGFDQyuwzw10I)D##8zM5qGmq)%{CI8bHv*&8_>wMZ3U5@m1+@{|xKD`ymBffM3!B
zFh@pQQ3ojOnchqj_A~>fIUw^EX^XJYXCe;MT-NOiPe;Y*LXU+k9!+HnEng^d4#qFj
z1;a^Oh=n1|s~WHUauJQ`#GZNelvK&-6oWQ8Hlq2Y+v$4WWKv8JUWSVcT7nmOrg?8$
z{0rwDc<}D~-@ju2i~sVUi_dAdd-reKmtTEhUwrzt-T7u-^P=*lBJ@&(7wSD^BjX)*
z?AU}IJ2`D9PaU%(N2Y9`Z%DjCpLKT(_*v)2kLK*I=5MZ0<q@%^$BVvgcX4USZ+q<Q
zYO{0aChY(4zx&_Yn{U2tvrp&j=Rg0Yee}`qd^^g!Z@g)L@`rzD=g*z9fr$xgA035p
zASbT;yu#kL=I~A1T3b?|r|h0dDA0V{+g}|J2e(brr5HiYwJLZ8rSnRuyrOW>sXti`
z(H5$Hh`p+3wHRrtuZ=DH_Umi*mw)kBO6xCdc6L_t^174ZjgAe?-y7l|(5Xv=8k%rC
zHGm-Cu^b3gEBq$g;d;8e{9O&&2JHrwd>#;OCfcpr!_m#zW>JnOH~5|OO9&ySnWDtc
zHVIYwAVWu>=|}$*BUEs6X8~@$u8Ut4g`fZEh^^s#mE+6_H*0dBo-b09lSh29`h|o8
z(n7qH8~r1AInk$0Lkqq~v3fd`CXbs&DgWS)Y1^EgrM@B@sbjo^9%g?zTV}>XUe2b^
zK(^|>N_UO}%8L|5=6zL}BYqbqh`=Lp#85utkRVbeC<mB^5^>%n;E>Mn$fyW**w<fO
zv+LJy+A1$FlU^4LI8PA*gEJJKeaQ3YNbcvZ#OsAZ;nBf`L*dZS81LOoyuv1lqUZa!
z6j>AsG^#XLlMH_0-(+wA3No07$aCs}5;Ou7`5_t*iJ&t0%%e`|^rB2iE8Zk4(ns2*
zFiZEhC72kcX!FsiV&K+~)1-ol8exK0FcBuWU4d*(j1Sr$|IttF{SV%=9u1DSZrv0!
z{f*teaog{9rZcg34S|M-={7ZBiJa#}@{Zp?MgDN}a|%-3JsrwF`HgoP#zrS>W_rfP
z#>W*$k2jLVrA2$pRdI71cwW??zhOf|-G0#L?AbGR<?2;Cf8m1H|B3{X2M-_E?VGpl
z;qALNKl@aJ@`gIfVe8hQJUTYyf&@2OOdg%G{t;f)M7oX7a@}1$*3;Ez?T5Qm+z3!T
zDw9LLZH}wvW|a=EspE4m0u|e=KKb}#2^9a%?%chr!M=x#`kmYe6E4`P{FxNUcMMuG
zdmD_prd{}?<lDJ~BzN4ZCe+Xm0?O3)<th)<op)M=$DI8Hd*CLM1pVl5C6spjTS3xH
zNIG>|K;ordw*RD4hD|eyG@`tH3KAbkIvK76+_K-fzrtjGQQm(myd0!T+CWJ=-P4X<
z;+O6Tt;`4cm+=CmJM-KU6qX>KhRP{vWV-0bw6TnZ_^MQ>Y{;8igcIe0dkJpSPlU^`
z6<|q+N_W97iB`af0z!jE%^!SslWLV;4_AaCnN4ff0qt7K$%JP$L{|RIatZ$?z5Nh=
z9;fo+S7J+N@}yt#fGT6BV}g@$c^qzZH^y=w&qb@z=CYQGa6qI<x{uFH*l+|5PuDUo
ztU)VuM?4@a;r}vZ+@uRQ2oGq+8)aAp6@)z@R@qf|iJSgafL1Z!lk|n`ZW#}dezb`k
z1Mo|C>}gI&bLh}&h6!Q0nIGtKHg9xnR7~~}$K{Ln@-lBX#5sWsPrYP(B+nNYNt3FI
z{rHUAEZRk^0iKd7X?%%?Ib$xPiF;na*L>&rRk|INPTzST;+#?B%mVb;B6aok)Anco
z?4Q`=)Tlk4y>DNA{e{oFcdk9K&CQ6rM~v<G<hUI_Ib$c!%-HnNF&m#4@v{y+n#Z2+
znr&@7x5fEYdwB1O=F+GB#h*EQXLHL|=9cZjgM0R5e%aQyo{6XGkPvay{`}wkxxMqw
z2Wr#pcKh~yyMFD4ZLF@_sUt`2@`Z~wBgS)JV#?ab$0FYyEWXLWAxrTOFJ5TAd$z6Y
zg)refk9e0CyOafG3jd^fQYM@!lyb&eBpAoqC1kJ~?FHz(ngaQ7QIt45o2<`PH|)#L
zzOcXgi@&sA|J|=804a|yOlS^PnQ}7_f&^Qz>Cb8nN>kLU*R5zI)zs_2P{%DtJ>8u)
z(BJ27j1llZ*;WAB52}Q;gOhj(<k|VmhI1Ia`BsoL8?@C<;ZJ?&DDX(5h9Q;RVgGu<
zmhkSLF5|tEWZokUgo8(9Y-6h*J4d>lzhkG5zwyOa2Jt|#;1P<j-~!&J6}K;<RJ^Fq
zKjek#s?zG|m0P2HL=QK<@MJpF%e?+Wh^UmGV`4n0%1yf0fy|x{h~X~P_M`XX%tQ&q
z8-_|RfIy82S7tIp0kC)Oa9g`Aa8L5WvMQL4Px?tB0lfuQ5E)|7r%;$cn0(z$bZOKi
z)05grOGQU+DjM&Ejw`xRpbS*pIuw&m#^D7ZmEeUcDD%<8DRG-&_#B~XXey<JdLnO;
zM=Di~N;>3S5_C6JDrWrD;<)jO`0*o}M!c>^I;bF|!*FIB8DCG@Y+_>6e)QuX*}FgZ
zo}D;z+P6CS=Or4cm|3pvMPNX{V`W3KoPQOGbcBkPdgX-<pw8^8@lqkgu(HP|E1s+G
ztQzMbzf+p~rTG#vuYquBiL)3yUiVDtQb*zCPSrb~G#BQDBcbE*{Rj5o-aT7eSrVN?
z#tjiuM<(s~i4*?X;xS3xY->8KIOr%H?cGvc*4y20JsrLFV&{deudLhJ!m8c9dDlMs
z-6!^u2B+Wt<|F&~x4-o`h6V>E)C~C}iVq$=l6n|I1ZAf>r2t*v^!!t6Reg9~i@Gn|
zW8#ZCOrRcK4Yc9-WKT!*`mgGl^2_!HcwN_^bZg~bhr}t{ZJmZI&zZZpiC>s}dAiLL
zs{eFPeA<DVl8*a@PrCVvd-|7uNe6$?_pB35MCCwD5Uv4@@@cx2{=aYe)%8|T_d5u;
zq=m}1B1<z&1IK?WbJH)wx?Adypx#_TS4wS%#~smQ0EZ#2n=4%dzaK14g?qzH8e=@g
z3?7H~V9fJiYtTxwl+uBotJS1F)0^OS*If%dH>W${Vq6Tjpp_8#QHHb+bdS^R;g|lo
zy@@p922lU0fO^<w%c=mp%E`m|-+nqM`g;RMX8$TFv#Sc4_)R=4m}T&PR|Vu)5;zeq
z^)k-i2lnG(&h%8Y+*@UUTlnJO5cYs9%Y9HwL$f^hL#E*%|2i~rqE69qi~5bPM~s!s
zG0@(ExfDkZGl)IHWEt-VXOQHKBL6OI$XuqXVV1#Huq8;a4V5&&kU_WER!hJ=;*9xP
zxYc~nE5?3uYQi5>^)rTIikW+P|LB;6hqD(>+iO=Z+Usv#wkvPEW*07;^VQ{{!C~uY
z>$1(Y9h-Z!U{4<}*u(pe?B4CW5-R2u$F2m8cI)nB{$Rna!@37Vcc{xorv~iQx#M=`
z;zhs7aC2J>x0u%c-XR+q9I@xyd%ne)!`uEEv6PqqWvflRiFhJDlgR92zRU;(%&9w@
zV!C&gQ|Dh92f8#CseIs~)uep;i+BRgWOC+#SFU-t$`7S$u8l8#*;hnE^EZF+XS*9)
zc2{}(_WHN>@WCVB&EWi)yjiGY(RWXDLP(5ypgziaJqQ}LC4+`JAK$1~fV#@Q1yI+}
z&9;c1s)W+s^pE^tM0tBjC0}$ho;zX~%YCfvY;(c0N4(hJP`{0g4%*oGs07L}-)=VU
zw`YyZJnlO`#>d9|>)gP=pm7@(H!g9p3}ZIiq}VC4C_Z{+nJwn16<2ICM+ik<UJf^2
zCJ$BF<;RU)JW#m4j#~Ex*$scdn0LoEU15fA<BTnSsn*wdGj2oi@VdH`ghVb1SXq;j
z8MU;$ZVQWRHov$k-f-3CrP_Y<!TJ5dL|LISXMsl{HVfU|<%Bc)@F)oC6Jdfqg|Tle
zJ0=wyQd6_Cq<li9vD4U|$+enL@RM6qA_lj3=XSdQNj8|qhyx%F=(r>4bGL{q(+^-L
z-sZ&QuI~2|$eEBtLE}e&z!9^1dSFz=NR~=JvHBQ|J(Op~gQbK@_)Mo$Qnvh)lO)18
z-_YR<_)q@e$9DOxH|@x=8DDYXCCp<dPuR)RC*(G5W0PanrvdMf7<fvdql>+$5ade<
zWvsYeSI1^$M}x%<r}%fkR@o}@`0U2^EzSWP@`Kq+iwnMTw4p|}%Roe3*ro&tR^tzQ
z<9_`3iQT(<UqZxvd;IXBJ$ZCr4Ru}Zf*YIqY<lLXojrHfU-aZ!JZ`w+%K=}r*l*b_
zA)@0@m+fvnvxO&fcIU<&`}&J-?AO2emHpyx{?<PF^>6I!FTb`2_wV}Wz>}v=N<cZ{
zX?`eS;@exdC^RovgcNUYDs0L+>poVV!c3Wi?scZ}z|SkK!u@^0>q>-Nm{5BnE(TEI
zrfrq&4)C&T5l7ZP(8|9A@Sq)SH^Km=Ie_+5wW}%&b{cfVQMq}zP>Jh!xK#E8_emUg
z-yb&2O(2$Leh9rEUdpNvpi`DXGp;7o-O^rnPbl5%<r~(R7&2tyd-;^*T)Ed(_T{V>
zgag9i?q6yuMjkhGR@eM}k`yFdAgCCBCMUOt8C?8)mB{^)M(8n)MH*7E`q&UyNMmIN
z(1WoLx2jYs)hwgTTS93LT!6Z&j0I#^;z>QrCjCkXH%v(^`iV1j9znQgSPzeTM!qi*
zxYM1hu}UiGG?fNIFw?IwE<_x-myqEqoBzuptOb;KrGAhCJTC>*?Nwwig;gmC7k~Zm
zUxjO=3i@(uxO<qFq4ZDn4}lEdOgHU{x@Ofg%+0yBnI4#Ogb8$*OrS}l2_&aX6K3Cz
z#JEisEB6Gq>xVwXDlaR%kZralSwM0`peR=n<S*Pp$mHFGiK%fv57@7{tgpY{hDXOV
z=N`99JOp<2vR$}zQS<OgXCM)jVDdNDG<VJ|+mm~<cJG$v#=8$~_R*{u#Z}F(hiz!6
z*G7l?ZFEeGCx>lkj@a~xDLZy%+D@K7;SbwQ&76<`bJ&(72yAZbsxP;@KydAguk7=W
zKl7KbVF=kSJ|ceLu;wAY$}_*J(<>XA!^I$8|N6Rp^~vXU>-tT5NID+R8duT6II^W{
zm#_CKPq7_d4hRo2*q+>Xw_@AZ<|dEdiU;6}kP93VFl2rvot4YA<yE_R{e}w@Tn&qm
zm|KSt!hD-|Rd4WVRE76ZuYxS~){s@D9C=+b?y)rp$htcSvdv}Ns_O}kBR%*}necZ_
zHXk$9ma8$B1)#33ZsR?y5E>*XNZ_2Fp7QgcY~f=U#^mIrKL|W9#Fe_7GvVr8@jKlJ
zxm`{p6>p4mNa$lxm&4O>H9uy?a{5jvfb4oX(q7>&y)Ve@V_7lQId<TaHx$q)FY;LN
zv~Es5R@KIKeS}LEpm!uxa!u{FctzCa)}9nw>z}#W6dze^a}phXoN#~_qEsf#3)2UQ
z$dF5v(!0BR#o$S5SB7rizGrvu-cv*3j%7YSsR1cv3@qN6p;=98Fxrv3_@s$DgANtx
zEBb1nK41tf&Q5^mLH*{Upi|+>kM!;OBPybxgn7|mmde7z>=SdPGB1F<_JP~=1O2^B
zYmo6q-JFagKA1@I=;?JmIt@C)4`$9$;IF8w2t-M`rx}}mcw2&NR(m-hEr}QPy${~A
zKmT9<SN6tx?@C}8pg;my|2#6Yw6tKnyKw*BJ-hMkw|491O@FO&adFX>BsDHCN#a~Z
zc!<aB7`!PnlF`+1SPXx=^-H4WHl?YlDI1m4jO5Sf-g}~XPj!k!%2i{XY)9&Cvq=qJ
zaV1>8*ALg_EiZ*yT$mH(=fWAcbLURm<*QfidmsM5rlzNTfLYg|#J$ITlFYk0x_t#?
zP15e{!zXs@+gtYSH_F3}TguA~KXb7zNt!K714`%jKK#)B_#gj?U4H$F=xo?u{MBFD
z|Led1@0~6yUv$hibwod0q;yg4v?H!3_U&dW|Ee{3yWo>K17i$u!f*J8UNxPTK=XmQ
zV}S5UL^OENhYn5q@-|;1fLjS(_mu+b_(1vr<)obS%P?+sYm;W;1NdjS?9Um#j#rvf
z&-m+5$4w~ht>~rMz4TKjonh*6*6ALEm*w3Ui1h~s1GIh{bi6+&ol5t-I$i*{b^2vG
z9ONA_2;2`_m9HRO&<J{=DyItEO*#w?dWJ!7#E6^zi6$B35+Z~^{y`1M_{clq)E)>6
zQI7+fftzub0DhV0(p;tyI%teDO}BugQKyxVxc)`g`)~A3ph2ZZL$&4MGAs{;x*vZY
zbbmaKD?R%{&KWO5vyG&_j~0QdI6b}w&VFd+SEe`NWjntSHq#2I=T^8<==yh@Dju*(
z`i;0NJzN9F@hg>K>fx~GSW&vwaT5<c@xmTa;~qkAvtQw#%9(hd_fau_63+Pz8TMdo
z#9R10rt!TMFg0@SfJHaPX5x-HPqgIk=~eWR7r7Ci+L#c$T-fg9Tb&fg@Gvh|zh^)A
z;RkkBa~WG!IG^}TvnO*O3vl#EzV>gfuiG;3J}s}=>f(y6Ev>qFWpP>a;<9h$8W`-=
zJUHR!2RUTTR;a$-VLOB%pt-~}_ha7BT-`7C&GlVdU)i$f+e)YAsmHhO*w6pNf3(MU
z@7mBnza5v*apmo|>|g$G{@kX{oU_iJe&tP!w(_>Uv2J&7+^~Q9-~B)Rwd<9IMKb5F
zYmbbM`Zet6+%m;{J~1&Nfk}Dd3zC>x-W6cpr0lkz?T8QBQ-0h0izf>fER?j1+3oJ=
z^7;Q)zxbv7-QWG4-MV$luZ~x#czJJZZ;Ih$SHq5%9Xj*1^KL?*etoW%r5Y|;0(GM%
zVqKl>;z5|_<EAL;F2^vhQ;#*;A7Q+{s|FZ;`@z4?<6nVez}tyFJ1>0hV17V|QQ0uJ
z)2{f^Fg!f0wDkIRrp|5&t_qEy0}gTVCIfsI3+xH;b_2|;u*tLMA=<l_f$a1Ff7d}}
z4{w>{AUj;*`r;T9{f`24YUEjE?&XMkm{C4n=#MhM51le`A%(xnOx@xhbfY{eTgo}&
z3-C}?k>Pge|MFk|XYwadC?E?JduawD!f+$AR8<!em82{_Ro=b#&{wF}B@M&y@6tg%
zfgO>#f}UqofAQcLi6^lXI0d2t_re9205DDwrx!AD;~{#E>xHkvtun_jMFUAGEz`|l
z#VRr?1~|e;-a-#Ey8Dq9w^5)?9GDIlB&0i34WN_?Gfk1^YzVk3ZRFR(MtD!3!Z5z_
z$#bZ`PmPt8f6C{D{iA>MC-!gt{9oIJE0=vA-V0X0+YtpwkHR_;{)t=CqB=6$jvhUF
z;LqoCcQt1YIR7yB^r;VctQ_!plTU2i$}oB4m|yqCLnpo6-M$jBy3FCY6<b<a@{Toq
z^r&5U?UE0okDtt{j(42K%F>E`^EK~C%-fSEk8M?gi`u(QOb*!x-+#?syL`p|@Q?r4
zZ<%0lLh9zl&#sOh|D^Z${$soT&2{_g^Dph%*Vpa7A5L5FfrlHdDE~c`(ePlqz5D$S
z><2&k(B63GE$dL{{mn<evH#nD```MfJK9p8^229czvsFtdp|Fsj){~_pQd8~)DfLE
z_TwrRIr{{qvZSiLw)r^#s6zT)tTuEzGQtF%o%RL&Y$Mq&QYCoZSL#)e{v~cnC(Y3g
z1CIf%P8U3IfW#@&RMJOxKP(ka_M32eISi2QKxr=foTsz!E5kE(yI#6gaU4Xugbarc
z(;<WYtGFw7@|^%SL2xIHSkjE{;i|YW*CBBdavVxv%VI*rRh6qMD#Mjz8*#bfuL<rW
z2gardTgDyXB0LxYA&@ll<5Pij(CYPo$#JASd__BP(6jvj;hf<~3qdlTfHG`d0njUH
z22?niKB<>+nlwFbk3Z^Iu5uS$_Y1r_fa|IB0Q!+C)9tj1U`-lT+A4dj91?F6xs0#G
zMX$$E+RJc2`0o$zG`x{X5!SV6Q^7=3Gz!}8mhJ)%UHK(A-6~A!m+thLY%kdB{iAfp
z9QBop7EPMyFswi<$~C`S5RNgKaXZ^Z(2nFggUdD?7J!(PTYyN4={+oAf*XhiJr1_%
zFvi0e`9frsJ&%J$CMGPId3X+{4q<}xI^ErEzJ=*e|LH%n?|uJ$Kh(XnvTBPfoCn|m
z*KONFX)fGY)x0}5XDjndwmhf#OLH5m_r&e*F?AgBTMTDTOxwhfNoV4xXO7$O(6qRp
zb{8fNiBX2lTwYnTt<60NA?voW!i&^f5>A%v+Gn5IU;cmpKlW^MUGrzV=uO+j*RR^?
ztC#JM{^_6DkrQXFt9!t9w{|5++_eAU-~L<s_y5EH+vaDVh~b}C`Ry4mXtzlyU>5?6
z@QCs>Gjr6x_#ZnqV@D>Zd^-+r6>uY@I4fVMWm_?i+^RNId}<Gx$GHWPt3=s$#SVf`
zKKaDA9dp<k=3Av}>*6JNNqI*!5F`+)TnJ!XVz5`eV~#~vs5gNolhsU>QJ3&YT-%F4
zn(Yo%0&R&ll#p!?NOOkaI#h3aQqg~6uHwcbE*L`hvkX1lTpAA{@}*$}LBh}E&~9@q
zCr?f`(xoooZ46$-d5z#p7?a7UCxIVy0l1{wci*UwlO7kU^YFXd72X3xAWt*lP*vaY
z^JX%;(5*7}Dai52AEAje`egP!`XRSNQPQ3lDWQ7Zcpj7)FEhvWs2$JM|M~y;XYz{#
zp*sOV>VQXF;iNluB4ASL?C4abw)<-H{reB?=FMBSD2aS)GbAroxI4SLU9ccygk>uV
z?t}xJx%PoSCJUk?@ljrpAqEaA0VdHWD+<7K6gNHKepn1Pbi!2?p*SO7QCJLyGz1g{
zNTFxm(9;Yw$^c!e=^y39fE0~LWubIsTup%j;U|O$?<7=)SBmmq6DEK|^5=X6TbqUl
z`>n6HQw@V#3U=(D{?Gr^{`}AXT*Aa_F6*x?El6P7aG{1g`5_+~xdsQeN4-$qz<b7X
z_&b~a^<qcD1AFaO{L8_j1d0{g-r9D?yStm$EQd6JO!_&B>({RN`2=Xu3Ky1_tgE-j
z-g@V4`{O_T$JReSZu1gM_+0w%;UoL%%dh=h!-{C$ym?pkuYp5j^YBQIz4`WW4Z=UL
z4?p~&3lQDi9J21PH4Vo1@7%XX_v5{YTN?QI0>X89>uXB0%5-3$R|8h;t>w(e@L0FK
z_5M5d)_d>Rn@Y>b_^5sU<(Kx~{9peY7n0O~RBrvM->$06SsCE*Aij`MPt=Lrd=*3;
zRB`xTPU=tobTkg7Q7uu^v<t7E1A!*UUEuvv3~t_u)p@=1(S`-ImuCA3D9zrUygjsr
z@A6BPaj@z^n~zC6dVTt(0ti$3r#)%OpfafN`Ip{Uy{|zlHz4Cer>!RKR3*GbqhiAA
z?nwiv(@1a}>E6c~4P?)}mE{t&gGSvy%|Pq)hT<p8k<P%X+%Q${B_yr=e#BAuNedcP
znHByT7TtY+JqHfZOlyP*T7-yvxsN=2VuD_U?{1aZ@ub2UbW(59-)|1mLBEWrm8v|$
zuD<8y%1u9>$Mg^9{pS686{~)T7uSYgVNSpfe%3Ez2SmJ#s&ogGW+2i~s%H5|Tm->A
zA<a1+BxF3E-Wom%cjlN(&ORKNF0AtOGz$;Nx=YOcdGVw*bPregqi39Tw~Uwmn|Ay%
z-<Zqs4k&5GMIS*k!pB!+bZF+5C{l$z=gm|>SN>`cEDD7%5yC;UM4A94k+G4{AzwLW
zUS(W{&NDG(MCI{Gku?~w=b9&&yOkTa^NKZRGfD69lhgJu|JA>+Km4Pg*!aY#Ev>HE
z-2Af7FS`gG+)}i*YD;tTHuvbUtu8LxHV^oUiGH@brMw=p!I6HOJ~r*IVDpy4I9Gg5
z&)Cspr!=PasUdbMdxvdP^EGoI3v;(_+!2$$<nt6~DR;Lv?CFC$_Qh|1Z4YkWP#f6w
z8x>DqIB(zo!#}k5fAkYOdG3k?gkhV1GH<{Bo4>XH?*IP(u)qJsuWWr~*;n%Y#vbK~
zFDjJ9w#r2%B7xwr4fJ;V*PP==XYBZ~<91|vTEf$W4Gs@k|L~w+1xii$m?{Hxg)qU+
zg2&4DC!c(3*RNl<`9%p*60jth*`DIymdULh@gOQ!Zg%2}kGB_<y|<xi^&dAL^9$bc
zWuVIZ79<H>9h?)DIoo3i*r8jt$83M-slY>LYY{?I(2KqUZ^@!_yLhWO>%%h$-F}5D
z+mzt9*y`<lf_5cE474-a5bn^8uM~W}%C;}1{5A3@t&vXorC{J+$`VC|O-m=OKF)|3
z_N3?aL0G~S3U?PKBown{-T80)pr`n0=Y&&fI8DI&A~~R5R3~zfYvn0$e9*lNqy8|r
z{pkJXnF$<HPje;~sJkVeSAhN^(R9KcA*q5?h&;T)%0O>lJm<{-u(^qp9#S0FgZY5L
zaxt`&c#lQl`BRQOV={Q0@!4A)OO=a6=m&sVp$}9~3s(zz0~I|!#pRa>8bnlZ=t{Ea
zk#f&e<ST6Z0e4jZx+{UFh&aJ_=LnzvK)PkcCVem%Gzjkr$6oodmy(+bx;b3*g0kFH
zeh%6D@4aVlfAF3&t;_QZ>bMW>>EkE1^mNXiK6+wvPoCO>%!^OwrO*4dtXnIqw!5)u
z&)E{Dd~%g-pBVX}fnf<Nqc%1&Zc`I8cH-C>4IU@$=*$TT6GwdI=Yc<<bJIITugaP7
z;)@7}&f+QkK7Z_HU}#YFv1iLGD_*yg6O%SJuED$eh4l}%*~F0nKZJDd{CVLHtAiZU
z0J~#P9?jZsfAc&0<=_0mKK<>dcI~Tg{A#>4rFoAAF3yKT<y}48XG}e+y|urvo&gPZ
zqBp2Xf<3Ig{e2Ru*6iA~ul*|ugD#&T{ZpNnX>?S!JkVXe?A5lsuG}nD)eWAmNky&f
zH<uu1hx9{;h|izZ09VvG;PqRXv)|OAm0LpHFKM(=w0)w=PL8U_jY>#7a7zEgaox?L
zQTU~QLK$D_SGv{xfOM;eM@2zbI(bmtRgMm=@>Qo3_81fB`>9HQ22koS%(2J2RNy6G
zL<vn}j%Y|x=Y$Jja3<BwK^I#IS=NNfbR^!sKRJ1}BaoY!gRsX6Q{|7l>^auBxe}1m
zE!%n8ecY|a-uQwi_ey6RcR^cyk;MhqC7`;1E_3XCZmO4PN_m({Z>1Wzl{qF@(ooT)
z&m!>4EwbP2L#=e5a4S2o&x|VUK+<)<Pw^)Ny^L>v-IWk^w!aTW7}@}T0ZmhcY1Vg>
zMhPWOsY<^vD_+H&?Y6|lPTNgCx2L(lq3wpb?87i)=){q6dX%2OG^ZX=mRW>>9%mPz
z6%b_u;7<De0IQp${0NhDLQz+wO=-}dw@<ahWX9t6Wx$<9L-+*7W!i@tS*4=?k$z8E
z)f5o8gPE%SUWD)h4p+Q#$G>mu6+XK!HrF?8-52PVH7_sO%KV~lZQ2z>?2Ch%w|aWo
z{anP<(J?#DyGo}|*|F1S?C8n!Ha2z2I=e^haC@&J?h>QDXLGajcJtaz`{Xwt+b{m=
z=XUe!>$a_Vd{hkm@#A9dM+TK{zI;EoV<(Q<6<$ETa>>{+z)RJ9O@N7ARhq7U@r8Z-
z_kZua0bdH37uj_H6Wrd_WzUsin0@*n{2N#LuSihh%GyVF@A;PvwnxElEHBKvz_cU5
zfR*m&!WRea<I1K4r~CIG*!_F={i<KiaHxiSd-$&S18#FdXu-_b5miU2s*F|N@N+Bx
zb}C)*^`oj)P#9;aCn*+TIj@jB9Cek8vJIVZ1av$7gE5@lFU-xVTOZRayDRL~EBVkZ
zb@_{cr^gqwUBwmiT$8`Hw(bkG=o~oTR6T91Z`&5<OgNLO`b1H#8=DfQ<i^4!=^`zj
z9{QXVX5!>4NOw=S1j4?!U`J&_efi_i2sWL8cS!RU`BV*%VbM@o$_;)PUX2BO_+!4N
zu=|7?$Ku>Zzdz*wAIG*VZld&ycSPUmLMKSO^0Jp(IUdlvygy;`KKjDie*FH0pJk^9
zFqe8FmXIAhJ0su&v)sDWAduR4;AC`k)TSmUUAkpo9uJ)n&R=Jiz`^9fU?NK1Frsll
zFcf~iFV9t^&(oqv$v7`PvTu%z=BG=g%m+*q>0{rY3mxpEhrVb=-f|^H^s4+g{o=O)
zcuGN?pg9op2WkvbQBI`-B92sYcUKQ4P<RLm*Zt&(AjBQYo$4^$3;gWauG`;y^G&<-
z+C_ExU3+lvuHCtF$L`*`WA|>}mb#^mf8Srxe82!8`<+`-((m26ExcQf`%oSI&aJ!B
zAKQ}$PsPyBYv5Y4#rZW27EgTuc_@aSaJ*r$CINwgm;8}_4pD7xZ`j(#x*BGW1cfOt
zho?`U`nd{T;pP+R`3q<5<Y@^hs8h#n=J;_57?XY|jTINpLLfl=_BVfTH@^AS9zW1P
zwYV5UXjF~3bF&T!5<GRS?yI_$pw+3e=~u@<m>KN1iK$5&9vSwp7T3SIW=ji;#y#TH
zIkzH26=?F2fpa@pT~%crb(CeFZiK0zG#5}hlT?7_>20`br_OLmC1jh3J`?J@pv0-;
zx3ce3{8i>Ulx~=lfk=C*Y{Lw6a`1e+Ut>_I=-?(*(g7T|g#xlXfYSXSNZf;f@EO1B
zm7-#}+7P!KV^~-~FT<rD=2WE{c>B|f0rKI8N@%9F+0IHuUkY9J92Y9Z*`qXvVhl+}
zTgv-lOlZ(aApqvv7jwMb1=HU&haeElJ8{4~$X&V%2kzcfLAP4za>kAD!SKg8CjUxr
zrS{=fG*gv$>E95%QC0c4UFImmC{OfM38i17W4f`xE9$O<bfe}ll{owBJDA)3(Co_q
z?EnBT)1Lm#ma)IR)D?hU$H5Nym)(K^aUCWi+iLn1Z8va>wu)P24&~7|;N|=Vkhdr;
zC*csdspQ_sG4TfN9Ag2}f#9C;X1Wt6%RAaZxTQNZv%TX7$k+8&zy%;IXd<lm^}n!Q
zVxD0R5tuv=3Z8KQq+cm_cj98l1L1hrfHMVSW8==OZ*W5nOeD<Y%CfC5uh^E{p0Ri*
z0liCc4h{B702vVvG-1;-ll~g@$+PEda)ygYrmdrEM3lP}X@~ft9e+9c>(9Tkk3agY
z{rYcyVb{L++7}9W^mcM`+%8=@ule+d(zhq3u-gy%Uc7Y4UVrN?J8|iv4Gs*6f$US6
zcG&LLwmp6L$Ugt%Gn;+<#QQPZoB9UCNGdNK;w#88igNIFp)!5Wmmjuli8)>u53s(x
zY-_8_%I~W21_x!?FP>^hbhuTqSBxs>CGOw9@8>qzF`-`Vf4;2=L1J&mg#^EvoqF{+
zRnNq$x@X~;+mK+QX$$OH!Hq39tn6clL3Sot40W5_;`L?k7*avUUt@;HQ@ita<;*pL
z0rWKPyMQ2f<-y&FLw?RISM<~>ycIw<XfhUf%L$=EZ7&5c%4kbM1>3elagDXP(6GIw
z{HiQ@3&nL7XuQqR_M<Ily9*T+L7GFb34X#CxvD;bpPQwFOBs_cia-hVw7P|~yItWu
zT+C(U*@YF}_2T+jnc<BocLWc1xa4A+$4CFI(m;4`Bc5(){HOvc6Y|yelkZ*nSwsv9
zhyi6L6M|5H<I@=A9UWeFMM_2U;v}Df1_lS5*+ZxplFltWR00*iCq$@qi0Om*BNHSx
z(h>#XQX2(=@<`){0wc2n1N~kIOt8?UU>(EL8F&<)H$)nY+{uq;9wyn-UFA=Dm@2#H
z&m(L=kd5>?Lgf~XFK9syO1NS6RTQ5{i8H8s<IKQPm?=ZXX%1WWN*d*bZJ2n3hi*xY
zufKl9&YwH0d~VwVby@};ZgN->bG|&k=m)&wRn9q^efZcO-GAr;0qX9Zd;a#rjc;$;
zx7TmkP1kSRt#9wz?OXTl_KiFC&6nT!Eev0M_Jw`(#aDj4D|<lORra|GxxpEV^)-9Z
zcF4|MykL_@rpbhV-n_xqw#6k4v=cTnbIeYiKCOZ7m><v{86FdpZ?>THAxM1w@#pr%
zr(f7ZrH#ROPaO|B3=|9?&$%O8akQx{*eA#x*{s6QuDBI}HpaOQ?kT5jv8Cwxwd<Y^
z(&-&RZOO}3ztne}u~GS>L_W-vdDNSCa_R2wsR_LmM9nw#1lmOk!=!_9iw1HKr3jO%
z4hOlVdn^A0eDHP@*Nb@K-#y%_ep^CG*Hd0uGZzj20?)y%YR?VlMu;TgVH<r;khID;
zOGx(&<97ZFblhU#iviCEPPa1vvj;Q>#>&4khYEZ7RU+x3do+sM+lksxK$U)JnHBU<
z{HWcjztiDe2vE3SjHKKg^xUk8rr|H$nWRbu!`~DH!Yx*1-EE(HF#Js+qA65>lkG9%
zNELB6?S8lq;f1kBX>uIKO64oYOU$82vkL%HbgHIbf$KW?4S^`?5ml9yn?vzOd8f^P
z9j6I(djjs!B$^GQ(MAX&GuBcmP(u%M)gQvuX*yje<Q|2puyDp40X?dQKW&Kdz~``V
zqiu$JO;z~y_UGjA!>!WwqbT@VThgj<GJKi|pEQCPzYPdCH>T!!ZUS^uk2#I`xas~f
z5Zdu%3LMHY+66SCud!VUKj;OaEbm}YXgfJjhhIf1+Gy2hqMf`9nKT$*=1(4%1y?OG
z`dUU#KF}Dnh=(>R6%c|8U)<Ri3m-8)K5EmNcUYk03QoV}aCOagHa6|q=9aw>L#Y|e
z+B9$V_IBFXqyz~u|3^;D*!0O0cIx~^n>c#PdIu)NEcPqHPNnCt?eew$-UIvk<In8l
zU;f%Y`PD~u?bFX~Wq!eSBouUN44ax7w+k0f+nLkHY;R{><;a%aUYj{~%+6oFVnb6$
zt+PvlL`RQOu2g8O;7k9F>(}kcqetrNdp>t|^D4N~&m7BkBk$uXlfx<#mRJr`?gIGL
zZZ2o(>yXgUZ6l+-c4TT)@t?Gl%<qbKVrt5UG{>t(#b-#Ecr@!*j#Bp;JychFwj~~B
zOXCse7Aa4!XOBT;uDbX3s(ul?W4G!oZZz^82I7No_pcUG+{}d=i-l@L(e4ra^y_?L
zHRi`WAYsAHAxwl&!Q4=}W0oKIo@1rMn?)+;JEMJgdihb?@bQxIRqX=4lP|Zfc5sux
zpmL`Sw$&b7fMCp3-u?0$+LFTNVq;?gNOh2WR*u{5C*CG+!SeR5G<fnV9q{r~o)`zn
zLjw{)&@h6ld{N%H0Kxp2LK?zFLuiw5=QmBMPB_W{UqzWj9~G@iQ}(=G+kX7Li$9A_
zRzxUd?`$XmiOmiV1fw5=7X6NGBP7!8(9%G_V8ACSwyLnVjjcQCrINlk{9rGI$)Oj%
zWOS;cBcV>FK`RF>(%~slJnpUr1Z{2|Aj1?qUpr{5G$w`}&oJOLgD@^J5Yfn=DG%V$
zj4E7!aE>3=44dH_=A<t~aIm5%w~Qak%|=2atni5!BWA~q@X(>0_}YSS1uM*<Ee4V|
z-+bLJUOF%A=Qcn8)D{+b%3pr!cytT~KyD52#-~QRrNM1W4VpvA+%deovTk#JZST6x
zOP`-x67#?82dCMaIX64&Z$lt-+`4(k9zA?wTkBia=by+pV02i51p6MhZCXOYAN;{j
zZD3?fow`Hmk&tlfu04IaAUqmfmkSWxUA?}Sk~dk1=icpm5@2pf-Hz23l}Wog0kjz)
z_>_0JgO061O@ag^$gAV5D6vnH$Km|NZ8dQ2@Sd8SveDrY-<I|5x8K^$+qb;KvD!d8
zptxy=lzU7*QAa4szR@}L<0}86asp@vQl%4sv+I>1!^D6FZ3Gx)9FMBaWS9hRTa})w
zgtTW{EAdj5&}@@o=8s6hixt9pn80zTW>AO3bDI03ao9%(qm+}VoS_?i0bEp!7r_2B
z?6(si_S6$RKb2cU@mHB?L#3eLV^lRxk@p58RzM08;SM!^D^SAeQ?wf<ijv`XW<&qp
zp1{%M;bV&0Gj_+V`n&76SIWbuxuTa_c~CjAJ3S^5^=V(Z7hVVuq(wND3T8gyb-WB)
znJKfR9sba7D#ky?Li$O@-3%8vo>>9Kgu8xR`PeDl%Pid#Hp&ox!iDm53O~!(+oN<K
z-CTJ+;o5|(V}P^)WnD%*p=Wqk;Y{pJ+UQcW2O#OVu7BEt8*nCk6(;+S`{V8Sa?AFE
z!Yt(i1Ilp&$0zR!*Y!d<PNndRG(;RtRpD2R6IrJHGS<`q{03y%RA%%nF9&G5_A;G7
zbKffM5qG93$q+R21~lhX>LcRIIKe6N2)*?8n3xM`gY*-^RB^b0e{kt@!twT6(FMp!
z<R;PtP&}+`qkLSJo6iXqZSY<2@`NgG`ljgP6GV8>6^4%OS1)XMXh8GMguijLvaxPU
zi;K3szG*MSBpz1ZJp63ey4zXoeBoy##-_(?>iAKcK6Tn=&R(!_2@+l1BT87eXtS01
zu!N9pdvf=_efgW;*(X2$JNx|aKC+vie`U+_F~>8G4G#3#ks}j!^5juFclLys)n(1!
zbGEg+V?(23Hgoc{wRK52X%pQ;@LJ-h)O9&0IrsF5J(+zhp-4<VJeilpVgFLVwX(s`
z#w)jN>NLz<<b1HN-HuESNpLtKA?BD}xpKnZeCv|E`QCeW`r>8f<(T!4j$7X-w?IwV
ztsA%P@q@><xwfG>UrhG%9mQw1p>YdgqH!)Uir$si0OjvYEetJRX^1__KLi6789Rj&
zjDLf1Caj0gb{TCna6&cQQWb8Zu5j@sf?*~2u!tLf5}0EviDs*~8jdS}^sQ1xn_*E%
z`9zRF$lz;5Uc{0QA@U__lP?((GB`I9UozPG#CZm`MTu{dU-RvUP3b7%z}F$L(C!s~
zjPI17(n?-ce&j_7S00s5#v6nMnLYo79%jl{{|XX9MYuDE`qKasCitS~+oJmV!cFyq
zKV|M=#H&#!{xyiOKYaiE&$2Un=T!tUDPdvH0RlZ#B&2DMiu4eK2ZD?erv^e~!r4=y
zGBHzN{y>c?ps{)avMMnPDj)_KCx#6ipB)rV<;H<3zI-4|a614u4sEIuNEh)EfJQ(#
z&w+yaB+CFqBPjCb;Pf-LfPMZIWZp}_&fpZ2uxRMVQ(C!&qC7zcpY|9rIXw+>OjfL(
z@Ku6a12`waqc$(ZO!stm_?L-`7cXd#I&4eJOA-#2{T7Kn2_9S{%Qbjx17bfNw;b?z
z3R{s--D(u6q5&wL-hMyJ!v0SrPpWZwWzhwVrL|34-Q4kp-{03^XU?7Rvk_d+%5}dd
zPo1&%-~WMi^$sdK?Y8iA$?o5IZ1Yc+?b){KT>}~Q!uu0EsZZT3O5nJ8<EA}+^vJ&?
zAV6|!2JM4eGXU<<r%|f3ys>WYa)=v-=a35;T%*V5+5Uk6>+b3G+a~<Wga$(njB_0>
zb<0;MRvdhC@OA}bue$KML7<Gf@Oo<KsRE}CEu0#V%>g~yN6_@nUJ=VS6aFu^A)u98
z39bAy-9Skjw>p0Mr5VVutup91uA1Hrd_mGnm45qhP|zrJ{8QB-X|-Y+0r&SYQ||gt
zI*HeyUG+H^Xrz+4-mhXj@GG5R)EOI`VXu^r6;j|<y1VV8QTa&)^Utc&{xGRBY(uY@
z_4LoQr#)4cK>DZI_3GFEWJ#P7uc1V%a@_AOMLY>vmbl>uw5o?Xl=KqP-im$#_A*@h
zmv|)&+)I_<c8#OXbmPu{ftM;L)SOgbvR4x;bTEe=lXHyKftO+HG<?FX{7XDkLb}(%
zaVz|KShtfu>42xFGMCUObM(TV<}!@abd%g&Z$cd}>7;vlqnEe`fyY%Tpph=Au*VPf
zvb{9%0*8FLkni(&g(S$Fl%V4B?K8nZ%i@AW)vl!o5ouE0c>I(rj2Wol5f6-U+|q|Y
zF@!*IWXj8bQvwBtuhr>AE9iB0wp*{5f41RHOdqkSnPYb1%vn2f?6l^)ZqZcxfZ3NY
zvAgrk<{mz^U;WK5?N@*M3;W$~KC&Cvui4`Ag6)}@eC9%>Z%}a_JvL>hP9N8Ne9AU9
z*X-Vddz!BoG}m_9<k92SGdQY5EA7btVk~KgZOS9__@jppZA*HzKrS@dl-jDsEiwHX
zhuRK3Q@X?q@2bp1W@dWCu3o-iuU$H)Isd4gxp>N6f9Gu*JAcJGhNhI!L9=!aj!P&K
zUom^{j$g{cmSyJtP0jOc5Zh85vAP}$AyKX>ckfG5q!T6n$9YQqwHkYHOZV`qDojwe
ztx{#%0ZP5;F1?y3D$o6XDi3t`YQ!Dwvze9%A9GE!{Q%K^c<4~PjM@)E8#jBwv+Rh!
zVV6#vyI89pnTxX*TM{CuBW@ET96Yo0(8THKiZn2n#5@&k4}OZbD*USsZ&ZjE_AfK8
z!pv76zBIrm`FUdarD|-V{C$C6{{BU%`U(Ud3*0^3J<fx%%ZS}a@M7?h2-j^t`XJ6s
z03;Cbif|J^r7&m_S?GaZF%M$aDlBgpB-jT{;V_w#(cP->$QToC6jU_izz@lW$(062
z!=vz{FsT4O#i=mGsqXB`ZBIgjm~A8{G70Uqb-8cv`6@k`M+%KpM!vlA8GNMV5UTP6
zvJ8N91H3FMKio@mK(o=hz#tdQrE;dGiWYE9WRezx1uOJj%1f^V2|mkmcH+4Pv7K!V
z04mPcE?=|@7tX68w0XI3kZNRP$Q2LjOwUYfusJF<V>2f;;G92e7q7&FI+tI6-LAg*
zroH{{yY_<*e`w$P!4F)$_rd#i<?2<N)WEp3qQOh$`|O3JV3p_APTclz{=!N7@lSqY
zQ_}$RKn%Y}{YHlQr4^O?piNI7vw^-52@W0h^zowIxOUg>-+p9WhsEso^lM;H{qvc#
z>PU3NySMN7;c(7Yz_cMuAXHK~)FXnDO3K60+-Nsl5~QYQChXLi6Mil2@W`ll7H(3Q
zn$kc$F{v`zwJ$#V!e8H{e*IIZ+JFk!cz9(`WsfjH`Ku~~6Lk?P>Im?<s@&_)!l?n&
zPtpRY)?ge&8+$5wq^iR~Zgux`E8R;wdc@s?^v}Lo;sSx+q+9x@87Rvtq3-WCbfIBp
zyofDF&3FMXYdS>*^*EhYWlqJ{$dZ|=gQrg<qrU}xAK+yU#yuv0RE$kNK>L6kLO>4A
zD91yu%s!B(a=Mj2epQ$zy&9cYD5rIhQrgKa=6$$wldbXRfF?h5G-{w;R86{3_Gt&Q
zT+v&(*CG7@%$dHlm;TWBE(Op|QsI_*s?xp0OSf#}b^p}Mu$WunAEdwsQm?zY(18-o
z67WkJlu0XE{5lPAE8X2IbK1*tCGEJUJ^j)SVDhk4_@o1*ncup5X$MN0rGKK;!=(Kn
zH`I5v{Sr!hentCzgbB_?FlWY%I+PdVO3;Ij<M4-|Asuo_8z3z8r5QwS2o$@*dgf;Z
zMuvyfkJP_4PjOa{&@6~?>*3JYsEtle*!0PhcKqyl8`pe!=rHdPsqT$O^tx<&W6$n>
zd)NNr-~CVa`R_iquRi(A?%cX*OUsLj0pWx`9{FX#aO%joojrfX=fXYlLpN{TwufTg
zpNR(>^{)fdYLsH^#Y^}Es&H&wg0GmHed;?9d}Um`7y91zmUW8Hcp+Mco{M*4MDYpG
zY+|h4E?qih7ccVA?2H|ky7<Or8$Nr%+W4|^s8=cLqJ-?3_^zk7uG@{TzOcvl?%B!`
z=O4vOC<o7#jy1)z!Od728@c}8D*%wj{E+S4^IMfW<zKcB%xN#(Ayu}i>{}@Ofl{y2
zNS@7Ocf3#t0rl{8DC2D?xzm(VL5sP^7p$bm*N=#w`G&<O-%)~6J4%HR!NNPdZScbJ
zC4={{_(H+;`n1LPVv%2u+4{&9F23Nfh)>&vuZnGADOeypz#q{c*b&45Y(F<qotaQJ
zoWG={CXlnpBjXM?Y4H^)zEoiMhg8|-1QI|dqugXyC97&EvC7^oEFf@WqH{O>xRlAD
zDLf`C3eYE4Dd>`D>_}m3y`dou4-VPb=!m2nR-OAp>XF35)*4pmITwLc-apu9J$)SR
z=8#UD$zYok(kz9JG)G71=nx~$)*lMPlV4RH1CKWx(j0loJS0S6MGn)PcA&Hq%EMOn
z==A&PCH-V>0MHvM|77OsJRV}{W}lh{z78?t+~!0_?^2~*z4Dq}xqQjLL=Zc-I!#PY
zN)R|A0pg?`J8{yEN=+X>A^n6MKYhlIojGSSXU^L6>C--_9XWp7rcNBU$)nRYHFMM^
zj~$c!m>rQ2vM0f1eREw5{EqUX!And$D<PL(ziRKi|A7sUP1w`LMZ0_dq4IgyuDt%H
zO&mR;GCeF|Vor?wJ$rEPv69wiN2iZiZ$EX#TM`3)!_!HniRa@HBzfb4uM^xV;dgl}
zPWD!EFis`q3Olb#e(=5b?GOLxr}pl9@7ZysA7Nr?dD#Vuqg<;eA&1KN>WeS^CJE*T
z{MaK&$Bips5hz4ec=*$i36u2$gc;}1>(Iif0n}d^20Ox_x0hPTvOOh~?W%-yOaIcI
zda25Rw3Yv>Ko|$(;9M(|gJ|LcEqS_5ntT%1*pT>H?x>_w((*ddq~Dlis$r<a0UXZ<
z=PG&nDQSnlbS4f}zPnZ7P_%<mCFAcxLZz2@$$$en*|_d_!YSN>(W&CB6nWxL7AVLb
zv|2$Uoc<cCUbVk{Q29fbUy6zmJc_I+V@GL*P4_I@Qcp8bngP;Q$0;HGUai2(c_a7I
zH`{QkY(p`o2TrOw-a+o@o-UbXag}ZfX)D8~0=y0jJ?S<}Qvfb?QIs)I(kabAD|ZiF
z&?xC)MuD5|{H9943=aKBhH$JP;{h<ay+8@Nt;3XI%NR3FrAkP9SvT~l5@@-ENdG!S
zR_Tw7=~cg#TQda_CR3gi_=3P<70eH0QN$H_j5NEW%#kL=Pe2h*FAqZ>7yknE_xEaE
zo6tC~Ic6uev#=+h+XV+lhWzo@iK8=`k59^c)H=Eb6pF3OJZjcun`^sv``T^$)nEV8
zzWC&G`{wg6{el$EIIxHqxd^@jRuRT@Y;?#jTsm*3PM`E02)FOuvHK4miLrd9d2!69
zj~utQ&Ys8*%&QnuWyrtGaIPZ+iaA?eSadtvHzlKp&5(ekyus+JcUX7lVLNeb)GnM8
zPjPhIrl!ZlN1U>$3zw|DcT^<VN~L%ZTof1EdcOVq3;X7?&+W;>2Y$O1Ozw`x(`_-(
zOB?HM_AylPP?3yh03{s)1^qJSX2xnC^Bh0$&{ORP1#%kbq(1DxzAZ}*^q2H+&KuD5
z;DVsPKb}U2M5+LLseHcDAL(|t#F0)4V*CO(;_-RIYgiyR1YHt@JMD?{vRm7m)pn;1
z=b8C(v4&b(^IN3+iw0kTHnx1{h~tVE4xQJ^`5FXWcJp*pf<t_H@Oe2>$<*XHl)I-_
zWtIza;Q089Fe2q&J;Zmr;uwSIB8as8`2CANgBY1}DXr?A9tpXwAPX$P3qvX<d^w~-
zlkW7GT)jbrJCVcqudb~4>NE)(9vYT}KjOjxw>@#|)7aRUI_kJO$CTd=bo}^?9TW3A
zee{T5&o(|jW@F>y)nhd6F14+%ZF(FX?X2n}osn^>FbX7v3G!9u1His-!a<ipPjlrF
zSuE_XDXhogd2KetG<&&LZb6QQM|@QymlsK8^#oypYhgA1sb+ak<ApQ!7tWuvt5+}k
zwjRD-@L6eiblir=kJ`Y<l=TgbS?_@S22_@)!BOiO7`Cq70jYlL>R~$*SAj{<n7Vtc
zvque!ZBl(bzUsTQykgrLU^*qpaZ}T&GiUtsEVobe4-eax{N^MGaK|=ZDo&g{W!vn@
z)F73wZ?k=gTPE7uD02<68qoUG(5_y&Ed1j(C;^arxfhof{PW)W`i^bxJhvV89IA}e
zMEovm2J6cbpnm+5AKKe*zhyHsGb#rjxLdQ?C$rwTr;kj#Qb7CsvoHL;3+kB9c(e;@
zn|+#0JO~s15rjICnKIA%$+|%kYCt$Opk{4AqbXF1YNJoY^~BMx60&_|d_bd()gayL
zDxsBMs`M`>?u0VG32xVa9oL~|FL4^W|Dx;9>0o1kmfZ(Wl%qua`}#xpHvxC~yWRwF
zGVQ47YZbl14wQHe6&{rebEKni_jb_?6G3oGz3G<yz24`6%uktb7bGgY6go2R8MfhH
z(M-40Q?-I{-|rW3SCgE)u=<4Mdc<8<rC&nXj?$g@FlYNN{c%gR-#yKVkM6<U9<-YH
zr8|l~pKhfpp|oSJr$Ikq4g|+@sLHh-rVd&D>Aw%BGK0~op6dAf!V28KGJANbR&+}6
zxKQBY08Ob3<#Av?AU&>2W!(^G6T%-z+?VNu{kzf$G~%oD1W$Vv9v}Qsr8~fQj?Ed(
zYUP?Po8WXS(oQFgKDTkJ6N-)i+CJI@b4T&uqM7FGrwF-j_Z8-69=@li(|Hu;VgwA*
z!r|3k_LvV0*x2Nh9pM6!siW4@Kddl{`?;96XKgk=yJ*+G_{Ki|&F}2fzyHL({^BdU
zd+VmHaJBjKUGJ|#_GPtPbJk!bey88goIPcy&z{!&-e-3;-`;=lPz>j;4GoV<063-j
zoJZzJukuIvuo})BD25+Ff`@D8xh7T&CpcYgWbB!>t6W&WM39)8=&|!>kK36u$8F}A
zc(9YlZ2I&$>mEO$!t7UJag8lPk@}<NOBMvb{7Cuu>`Pk^FTj<ln>%7|cQ*a-^@<o<
zrd7(4L*R|F*S`xA^lxXhm2Ty|@@mvy2mlfOU`XtM#PzxiQ>B`2B|TbJ2s+i+n@UjP
zVNO-r-CX@79W2peiPJSa`ElL<uruE4FXxjO|D*209)7gFJ+-yi`h;SuRG9OO1$=cw
z(2!t?ut8lTg!q?@-Ri(Nv}9%%3WvgRk1e{ftyZI;x=N&zQ7QRp$}R+n=HYk*3BP{%
zM=nfM`Ku=REF3BqbF)I46+tAFP()lo!vGCXqL2VOG=)OO$Rr0#C?LOiM|3D8upv+9
zWv>?Z>+*Dc+_TH6a_sRrhP!iN8YvVnMDHAAH1>b78o{C6s1zw*DO5s4zn-7EY5->3
zDr@9KP`QUO^ene5yHZDycL(SxNKRN#<&VLfJacthw+1~Gmv?d+C95B2PM@|n-+I$;
zPw4LJlag@IJ7}G~qt@0jV29d!>`+^m3QxMs*2V)k44xfIKa;Q&r=ehX5n`%z6F-9w
zt2W$++}_#k??+4>o$;>`2ob#+9NW1{tOsGF+n>ef(0h*>(xw{j(&B>OYQZ6+si_GG
zh<&!XzHATg-?pcZA6c*J?aen{xACLXzK^tjpif!bwIvA@s|vrd&Fh*7o7~CWCZV#=
zj!cb90J-FYC--Y}>la@&u6=Xeg;74$9X)bXc)W|T=a1z)dGf^D1E2CZ7(}<&bz$O#
zZ+B`6%P6#{n^58G!bVjxO~XE81H1&I3LP-28ok|7zq!&1efMva`O8rHVQz(6$H_KS
zniJAqx@S6o64&#ua=|TC17vo8GFiAg9(q`-GD!NR9_A=x;)%hrOiPp@F~oSmUp;-P
z?^o{Z{ToQUI;4As3rKU*uNh+<ns^Z|2m4S>nu03bOp=}ltQFyXcr5_`@4`vgS4TDQ
zx@n}NN<eC^n)O_^p(cJ4r}QrsfM4AohCkh#e&9B7(yg=usV9`~^|%fK{0=I^2EEFh
zx&wBpSH+QT`}D*s@LS^50Jp5~GQH`}I3z=()g_IC^7apfGHn>We%z&m@C%$~nUv;8
zPjl=5Qcw6UzswUp*r@}co>zEEoEWR(LX)bTDo`vz#Lc1bB}x|RSfP)_4KRW}D$>U|
zUD5ZSDxkoH*s9NG7)5Ra{e7C7hJA4kK?0rYHo4D#T*AbJ1c`ycQRSyg4BQLb-q^L}
z`DMFx?WTSD+fSrEwrgL1ZTIfn(VVs@l6!upv8_XC5p&Ojaz0&=2pQO=apKeoJ9*}`
zKO*<w(IXcoSdkv+AFyLHC*(fjSBgeCsC+a}?`-Y(mg1#_1^@cNt%v^LtuVwT_*STP
zSyVhWG}vjUPfht1v?nA)oH%pBj-Eba14mA3f*w;t=~o$YNSZQH-HB&ddAeX9{q?W>
z8rzkXB{8Vmwz<1yYg_BKt30f2ZAU&74-B$GVIhHDAB{5;QnA45H5ou11ZtE!=E@#8
z^pU8OQsLf+$F1QOdehEdxT|ifu{`xaOL_rmPq!TZn{xr;A9%8R{Z;t-bK$`4La4Ik
zVQX_q(gqyW+h<cbjtVsCl3w2~g`k0ev5jz1eW8#LvA)iEj1^mvT9rbuU{Ti3XCP##
zeh5pNDO1&m^H=^gkFyoh{SB?gmkAEUvwf<G<9v5Rn5clnrk)8;JJFyj6hY}PRd!VU
zQ8+Mnn2D6byX!piqD=BNfm;&T|4$}iMmbD0H$N|Z-WC=X-TYKi^z758HY<tusU++9
z`2|1F%cmZ$Dn%m8{bWQ=W=SjK5lLwtq>PG+<VD`xB)wGnad(Hx9HGdoYYoV}17$-f
z-7$ur+lblYiPZ;O0-C~v1|u=w{z@hL=OlH~(9fMcYj3~zo{b)vvO}F+5*Q>TAUqsq
z?;@X1+ts%@)WwQCyC2cRjxa$3@zK}?3v}%Imtt=nw`g=CNQ{j7K-fPpEJj?ODo<b4
zSNU<1(#ca|?#D;f_|#z6R+TsQ{Ozc59k#O<&e{2k=WIm6z`dK-?epLMR)g0aU%feh
z_Kfw84B3Ff9^ng<gbw!E?LF7Pv&G)OJ++_b*4x`|qazHAN9^Q@8Q=1y$o+NF)#X+D
z?2}JamoKcZyVs8KLg$eqUiY7T`iVV${MgUiaFYtc1Z9gbK?jcg?s9?eL`SRY3_!!|
zo!wmo4Z}WT1EN*blzzM|ss8vvhMwT<OVErnCbEPyd-xWr^edU3(l7P8e?3lALbH7~
z>6hRubd?#OeK>ejff+oZp9;FEuzNqO{30N<_-%$s<*qHvr5jMU*VB$YamYtRE4%(0
zSs=W-heJ|Gx(zQW%ucJyljF$UX{B;-%rI7v<%Y`o0UGfXdio`PniJB!^iP`gy2m>e
ze!zbJ2w&={N?Mpx0qN&oS=G)PdKrdz-LImV^o7G%?b{o0(;uT_LWWOOx+T=jC0>>#
z&<YQim(nCvro9#HbFa)9Q~G82RCP!<(wGXsJ!z$z>n&&`9)3<QbwP$LA^p(n=EheE
zU|(3=!mpJ|AbeSFzprWqrC(V#04A(-V+0A`>I6=B3kWA$jWl25SWWe-FPTpS(9K1S
zh%?h4y9@y5CG_VDEwYnG+6-GzHpIuUb6{e8!bUV_va5ne#yAY!C&qqkJmyfoNNjEF
z*vjIX&Cf2_?Qiec^{>9MZ@&E6u6^~D-M)20jNOV7saacdXnU9Xg!E^J#2~U&shK$!
z7>*wogMaFz4M~`IIzML*9zPQA#J1lKJ34j5#;1?CK;mnL<VWSf+|7;w=1*>H<UNSx
zl~v~iI;2>*?CCrtVWmU!@{nD;c-}5(9_OX*8O_^6$4=UzzDebmuM#~fGZ9x2v2wk&
zx@LF3{KkIymp`{Vw{B}L-;e;YCt+g8);Vx2`tW&i28F}q(g5_TytD1#MvAGzMz*P6
zlnNqwxU_=>1FQ;@5x>wI3V$D`GhCtL?h%D*2ol&UNPFpCcSAMj0@@?-DoN6tTk$G9
z^oZ3}wq5)Z(j4?c0hCX3LB^dNyu8r8?2538m#yM{{z5z?w??gtU*nKBUoB`ed<o;r
z1vf~sIIMs)Ll4V>tRKXdgh`)aoTlxEo@OE?1+)jX-yeT)$%TnXY*cJkZZjiQWmgFg
zy~>p1C=tbg0?$fVm<KXLfpP%E8GlJk{z*k0l20srQCO4Y#GTD?Z}YMQi3JG{bN1-b
ztUY|l_L;|a@BTyS4<v~{u)B99T|aqZ%S+2@sFHfs$S90hfhWCEN|p~eipQ^4^oE3A
z<{<$4KDRVW&-{}=|80~<6W{Y&<uw}>5L&EC?_Y^!@Tzc}Oqltaf-uq7+vChP9o(Nl
zS0f<p7cXA4w?Fv6M#jb+pUU%_8(1ZXAwhJRm}ycFriwK%>CY>h$e!V@G(DH@{f&FH
zU)ZjgdG=5q77km-2KxJ*7Sb)Ws4MCWcixY<^444S*6UX_kd8|5>6Ck~-}v<2k3O^$
zr%&19!w2^1Z+>mJuYF@H5)2tEM<qy%4zk^<%esg9ZTk3(p949r0n2Pdb-yl!g#6t4
zdk^`SiHWfh>t@AA4ZHJjhrdU0Uv<ww!lQ_%G^n!eZF76mKK=YNwUPV&Kn@3uSe>C{
z{cdYIw8D5jc)96c^-)(@Pf;U+hGCzv0a40LD9wH-NrQT}4^#=+?&@tR?RCErTH&<9
zO}C6c(@+oZ1D+4|F9RVcvqpIRd-HWAx2nxM4ZftgbK08&bZI6mniEn|(!Jqd!K>Un
z4;7sfC-YQ7!ynxf-h>Pnv?_BI-fii20D#&&7@k2hzmC#K4|<M~=n176DDhJzly0S8
znxU0(lqzx3FYT=$aayG-6~Lszil@8rSLx2A=by}Ocej9?ko>c`><ym^ugpUUPU97B
zWw^3ElzKuLF3qUSb3J^k^kmCGPrtgJ;JBjslDpg*?M{GyBksyhIJcK!WyW5rgjVi#
zw{$08u^kvtJVjiQ=l$)eUJl{jET6g`?yUen@>$|1lx}6&$itK8j3wiXEKz>)(qH!N
zl4Kf52Md~<!;EY;$0UF_edU#}5cK7WabsNE-r90|XSF&@{F@RGR#%qAfDgJbF;)o`
zocSCa9FcG_Ec#{(b4&K<!DD;;Kzz*I2X^n)UHe8t#Fw9cE<xg^KX%6=3<AU<l$iP#
zZO@${--1u7hGp`>BfY#Saq8@88<`lh)r~cKGW*097Z%x{W#fwfgoF(a$C3xnhw{Vq
z@m&%oc&~zOPCPKnL$g4qvfr&Zhle^f=Z@IvQz!hgmlG#X*vzR@HZ*h6+6N~zp$&+h
zgb8>F78wOEw)bq|{v-SN7k_78e*Bs5P}teqw(V!Twkf`EbN89;sH}En7l-8a;$`Z*
zm}toCyrpP)J;^EN=w{t|Ju7ef&Y)4SFzwKAbLHk?<wo-G!<%sKh9#6==_>r76L=`J
ze+Wc+M8j7oD<b^Ly`1k-MJ_$3o=;zBclRc|)E{d0D8r_iI7tKNlOrwkDR#Q3rd-hn
zrTbPr;#9j*J8-_c_&O0^GGeQgFZ8R;`&O#0xUQC$-uZ&zf(f@vNkH_&;8MOlN@Xfx
zBEC#`XR1Cm#Q+c~4nd-6PsmKWgd%*b6ij3FQGVE|KbSd~eiBYX^CdtAfH+}}5THh;
zKkjp8F`f?%LK8jjCt@;X-!%y!UM5v##WpGKAr59<jH8%GZdM=x#7|r>!<d^D6$1qd
zchsvPajBdr4@i|X%7H2Lz)$zoBV2@agK#JZ_SnbOTivcW5XHUFyHW^5CpAdC`{r9V
zIyPp0ut4SJw>J1!3F=1x{VJ`R0`w>P-qz*DY68O6rW%77_pOyRTX;Ncx2|2c&ovOR
zs_^I`2f&uqXg0k({W?9-LMLsU@JHauH1rG(*oc_=F@LybT)2uub@17*|IV&``l&6=
z@s>nY_QGG=r17qpKR!1$Y@HH9W@e`C?RVa`H{Q5t{e9+lVn28wfpY$d8u4m%WnFw(
z;p+tDh-A&#k>ecdmcG2aV!!>}?=(o>_YS~kQ$D-#ygou^eiHXVB&x{utP{*pM^csD
z8*l+NtAk3{6@Z>Nlo{9jQf@)N*~ZZEuPY#Rpmfi$b-a=;koI&#ueZN+2cScHrmnr;
z$UPPK(WmH@6;1vN|DeghU-doVJ6;8)xrA_+QZ|xax@W$Ve#FsCSK<cj?|b#~X$7L-
zdd8FB_0E{oB5cx*aj^;5eLScb4`6>9!@X2>sQV{Q>OjWZr2BIE(|(Yu<CM^fZsKHq
zN*o7STj3^zQ`plz^)eg^cfM43Sypa!z>RqV+>&16W?hstN;(Ib6H5G~pHQa@)Nx9v
z(}#J=ape$`y1c<CX|#;9qSFw@N>$4zDgkq*7l3wY2Qoen3*$*$6>iNuH`{y}zJ$cD
z`#bHzt)55h`*28t`m!4<T$P%aofnvcTo|Wa73B(@3aN;)5ezr3Y-OtuG<>|0avDTT
zwP;TD_pTJm$3*ttI{g>EMGD4@+uPX6LXmB63rA+QH;qk9_!eY;Sz2=>XZ3ibb#8Xv
zR+d(6V{OA9mi_H-e`~+^+rPDMu74$FZrQgOIrAq$g140pcWRzH{K6N-Hn*TpY6G~A
zlS@Y~U%g_};s>^Nw<HwIiiw|B@?Y4<;Gj)SO<K3+<u;|kb5G!1jpbtc*ecAs6SEH=
z*yH>6oHyzh+Q@L19X~c<XU?4TFB0d^pS8)UaT_^)(mF?`%sK`|q(`bt%oydQl2-XY
zxqaV0`kP<aumAR!Hm`YmYkR}fw6!M&ROPfSUM>V8F?5GxdlU5q<4L(wXU^<aZ4AGL
zqWr0wO1VHFXtWE;Mg1yN)KSxKzd3bO(sP>FD|g3flynm^T&iaHXa^pu3ZFo_qv1u1
z#G7*;+0GNc(e4E$4R6a*$%|vo_)%#uA$VxU3c}+T3Ol9ii`8n|zI{o093zO^n-Gvy
z5Y8vF%L6{0t>g%)bDXnzGHbJHXUj`V{v~5cI_GPbg#+)kva;&uU)z54{<)tuBpOeN
z5?uXJ5>zyKCV`>T0MHc#9CgJ^d{(ZXS&(92;CfFa8un^&RsxBJj!Gl)jw?ETz)6sg
zmLOSo^>8x+w>EKmlQJ%QPhW@M64XD~qYf-PZzpi@L!HDWP%miWRaz?L4~fXnlUt=f
z((5U9yM7vaLf~ZsKsEDEJ-Rclve3Awh(I7O!U0bv?>zF)w0XH;RuwY<ILID{#KfSB
z(e7l0LgmUW5xk<vgU};GqjvW6IWg|jV#YNf@|7VUr=htp>Uf7!SeXxL05NsuLlPcd
zNO;(Mx?qoP+_dYTeP*Bj<|F&{-~6rp{D1lj`{iH#wSD^0-)lh70QdDbwy?BlU430P
zab(OtxqklDXLjqxb-Q`}nqB+sbNlAAPi^kOJv(>onEW-swukb^0vVt;mhE@H`^awG
zzG3Uzo7OkfZ%2<F(SW4^PI+2iUb6MYMSGz(!NKgVp<(MD7_`ZwGj{S!TzfY((rfFR
z%WBB;cIWORyL<1pJ$`)4o-WPWrUt6{<;4pBtlGk`Ei5kCM<4yx?%uuYWqp`|RLnfB
zdRJx3*N%8ljJlJ;!vUr>>L&&`+AyvGlqPocSAqg04ox5gY}$7)KREy;<N&e1js5<R
zw4hz;xMzImbyd2jnZAes=ASv0i>mp$AK>Y0v{S{K;bn#f?c2i%5BJPV>bNDX$hR9C
zI#3UjfW0iMbZg*=Eaoz;t;)8k!p~R689=LS$;azVo-t8MB_jv8-Q6?&6(01YV~>1>
zIpW!$j)W|aGR+C48;ZQVIzP>@CA7jr<>b_g=0V{SO0$EkxRGMzrgAI8g>sx~A}tkA
zmP18B(DX~FhXJsc`KY_mHq%_k14{Y{iBtD0!(mQ(`f=UAu;5;qv+eNf3OxQq*V`Vs
z)qnbVmjkCVI|vuMGYQi7sfsemgsy{YVb0_eKA}Nh=}QPIj8}5=K!FqC0<WauiGE2n
zfpRPBCsn$q9VqeYKzq{2(!e1lZgd*VpTV%jVn?I(y%KWF=MXkbIk!NvNC(sE<F)>r
zPY|A`2;QM$dI7c_lT=Te5)k8`;+LNNol8G@d-^qp57?IG-Hmm2Vz3Il>4%@c{Ni)_
z^b;1n?)YvD=rFGMwj`!;h9vcs=ex?^j(8_lP6gzjH1%p;o<2Tduf6e_&73-Eo8qNr
zHLtTJoQG<M`Xy{=t{oT~wRSOy9uofGkoY6zhlA2wt-m25W?}A$tt#BaSg)NpF=ZEd
zS^NA6J9g@*9Xoy8&o_3CAF)F{oRb((uDew>(I(itG=Fc_zWMk|`~2fi?Y`#Sg~d4u
zKwFYqc3o}n#kmR0ns%jB-pZTLv&uJh=HqGA9<z<)*8=BBeLgPigpwY7rZ)&$iPyoi
z4DT&RxK&-{A8z5l{|_uiF)w5JW6p!ep{n@39pKMzggX#uD1o>i3c3+z6(5L>%TH59
z^z!wE;0jZ2(Ikl=>`@lz9?Et6;xEGDSk6n$F;TvdQaIW(JBjjL3%*EjgA{@S5AOL9
z1%DG`-FFc2-r3U1a%`CrfA8;1c+^fflqbm}Pj@g#Pt&b(sS+8sWKc*bgCHHAtxfFH
zVoTDxq;sTU8Wc46^b$0qpjb_AmmuMfWppDsaZ)@EVDl<;Z$A%aNO<5*TZQF42F@7x
zDs<KPMXoUjI3w?FQP_3Jf(-^D3pdLk^Nzb)8}6R(3ZR|J5gjG)aPxFj<>K_bV%0w@
z+z5w2g)kAfIdPVNFB5!WVX&id^R?pS$+-S?V0hR&S{#B>V_~}!%;3(Rz1ZHhXKWdI
z^3<N*y>IufUAM2*;XnEHukE)|zx&N^>`Mt0-(I_BcW&RZC+xXgTJX=MPo6&UhvKG=
zPWlQ(Z*P}<{pD9S_vDE^ynV;+-@0v2?%lJ6$B%7ib0c0b7vugyac|4Mv$<}I^RssQ
z*0<i!Sj}N8()%C0Z*RW&hNqwMTom)q?N7_gt9~AX{hS=~9vL6A>0?Lj^x0E(>f~|j
z>+kcg8FTY<_W0?XJ(*v!wVibv938V`C(qcCqu>lkQeL*7|NIyB_{n46&DGY~?q5Ra
zb6gDzqwKE`%hT(GGFPEz{rJR09ToKoINI+4RHajlD)k>GpA?dbV}F}!niER1w~1E(
z{&k#mEB(;3{iRC(j1RY*lyNH40WjG)R5U$<4U}-Bzn2Ocr2^2?FYR^zx;ltnLsbb$
zd*UAyCZV*)iWhd*nto06fC}HC^pks(X<A>-=Rwejs|BQAD^>cJfW6Fn38h=M0ieWd
z6*kT3UaAu6ZfP#{P;PU#m(n0T%dvzE3uGHkJ9^Rq8nh}X?RC8LE8XkxDtew+sT3a2
z8+7H4D`t0hj0z4&$37gwRl*<UghtRRZcPpi?{ETPfB7Vo=6#?r9;|{^cu6;)Ko*cj
zPp1I9R0(NMx^*1f%Q!tI_tFmodSwTPc<8d+I+HId|K1={%(L(f(1`^kR?Mq7)V7GP
zk(SB~t~0&vp7F&>Zj7tJXK?ngySvwhhK3b}_n@|XC6}{&Uw!$def0OgvAcH>B$)RQ
zM3}}!M>7>?{ro)w>A8J`^k`;xb`W4bK6X5=Yo9rB+_uE<&1qg-oSXL>68n^|QQo(h
znzXK-9;KJiNGa`3%is6fShw|6F@g(owzfQP6XQd6?!qa%c<G!3i5Z2Ru&Lw6t!L_}
z9qJoVfpJbkf&^QskqRW=Y%HwWz3aE^E3ULvz8*ffA7>?=i4jGRaA9IcimwvNr$SSI
zE<~hKs>(Vq=O%9#t_LIU3XmQ(fIZxpwSl17baxu)`a{@A|G)|N#DQj{L~gVJO!yH7
zg<ECcM+>(K_L=pU_)WlEDuME!CE#J2#o^E(;fnZ(yGjcS;tfxXUdoXd#sqf^Mf#(y
zN7^ZN6z2TG!a^{3E+l6jD}NL`Li7B^wm%f@FLrN9z;J<ME#8?RULXgA05a-Hb6Z(h
z*$A>NguSu|V4_@@PzKnW#itOqruZWq>MRTbK`&O2BH<C%w>`DTD}BC_CdC&FZW}_#
zz)ZoDHa^jGcXj(`Md+}1%R4s(C+3k(j&k}zz5%4UAxsp~TGKC-uSiNy_f`<;e>o0M
zOO-!|N^{y_bSLb*=<Rm^D^DHDGyaFwDB4LA;iZlone^*bhld7jZ+p|XEA6ao+UDY_
ztv+3_#fOjW@vS>{`^&HGtB*giFMj(w`{Xwt*>6Ak$Ugb_cP>D1ee1&Fg6d*ZgI5e-
z42qlEk~}paa0u+!@oBqs@vPre@$FY%+ZwN8Kbf-?u9TA$8aFsC`r&W{0n)~mZyeNl
zwzsW1<HNvVf7Ivvg$wq<hacLROP6fw#3}0=9G1-U!e-|d?Ed|Swji3j8gwb69tj`A
zij#LRLU1~6qvIpip>lCKo-a_UkJINa*!0W^8=si6&aNJN`gGoYAz@-cX=2dljR{qp
zI)x@3wFM@gXa`EF_eIqSWly?QKQ2h%SwPMDO2}q`p7jb3&X<V<uk%X13hL>o<JbKG
zaBxqRkoI~zEYnhk$+lisK$?AGE&Q<4Pg9kCB~B}MR08G_4|8cxob+=y{@pFo&bY$@
zMKX6@8HK7`Wl$LXDd$y1SLH&bp;zHk2kLHFW(lQ#9k&&fcnRrNs)XOiy<T1=q<dN4
zC1jXag-z88+^<=}4*SEDss#K4w-wagX~#giXI{!Y)${kN`~qo4ug8a-@c&Sn2`|z7
z|FQR<&zdG_dLNeMj;*RwzN&rqc(-nLu$bL}H{1o5AVd5>Fcgf?H;MqLPm%yopZHHm
zAw@_*iK4&=LX1#=A_2k)*aa55v$NxSO!rJro0(}_UEUm1Rl0uHmG|@H`_^0CJ-Y)(
z;65kc%)Ik<d1pRp9y{=@qOa28V{QUlHWWLY$|cibwCN$ej{x&@zL9=Rm)t@81oLr<
zo#Ckke$A^%kMfFU7C7q|as2eS`i}mY+x-nIuOle?fB>=nx!sAeiVXT3$T3HB+fjtg
zo=-fknc9D(Q(tbyzSnp)D~m1Y`Now$rmK0J1s8MeD=)v?zV+>IwHIFaPP=vcrWU|`
z!K%Pa<Tnlyn2Wt=%i;|vuM)U2!K+$7{j;BGPd@#$=k0gie!E>(;P#~r&3mWLoNnhH
zyWlrU<vS3$<iQn}J?@9Ub+g@3xT_7#?VDHI*|R(C!ykRFJ@wpE{#y6xi<jD|OHa1L
zJ3J@BEn)rRbpEoo=8~KH?ah~e)Lwl4+wD8gKi^(?>7{mA8<T?veG{Rur**}y!ant3
z-r>)FR?V4!JK(Rey#a7FSB)Vs#J;;EdD?^YyuMDtytHS}>xWi_xZ}DQ0oVndJr9g{
z{TOHFN5DAORy_89WN4onOGZ3kJh=V~t##qIY#8P%8nCMjzslF)>Op7+rQ?ILDWQB;
zYXjFyxM~x$lct;N5J*g!I+D0d4&A~2JzqiNx#S~%tZydF*aboOFfPc#5o}bj`R%mC
z9DsBr9_(P(!fu8#4jK-a(!oLFci$?e%+;ttNdJ&I&qPmNh+gCk2O9|nGrj`ImykIS
z`Kod@67wRe?1#DR_gjqmU^owoRX0cz7cA$2l3Q{3Gti9sCs0}xqEWffl%SXU*y<iO
z7NK5@8lT&T*a7mM=_ek$*mkxzy{Lba*S!7Uwf5?FzSq954a9f8{*CrEZ5+P%&wi_Y
z^~+!O6^IwV``z~HE3bIaxqkhc2ERSEjW%MNT>V+Ip4^{#r)hX@Cr|M%>u1_0KK7Az
zc4w!3`|IEIW`d2&fd-@dbQ&4Yt!NAuCr+OdPhKB;`W)Z6;2RYe+Ky~JbLLcg^4VwF
zv!D26J9^@j#-^Qi^7I95C{Al!I^5pThT^qXUTbf^`BuAi<*Ki$@LUG3q21!vrzd#V
z^+(zhPd(W#J#nc${?y~`_^C7P(v#0<lX1}>>S1j9!An1AU(v>iw-dNM<x66@-Wv(`
zE2YVgeWBXNXG~n^=J|=Cx&?7h`oQyB5PF{ih!^_8d%y>!TZZ(%SziI^vSqXh2rhif
z_EubF<A#BJ;qvTKHcum8(GW)b`-P%gj+cWSPBh_6+Hs3vgvzVsl_9?<x55b3CWT+l
zk37q`<+A1Ua;UO`<uYKztqLP7=SO%HZOM650Pw+tb7KUmoOpp|!gm_7#-(iR_C;HO
zobq`=c!D3xigy*y*cV4wg)3Mkr}87;Xb<TMz$jzF_cFlcBsxBP@kw}Ro$m*uUSOPu
zM!aE3W|fb0quxrUyy}>{3_=%N{G3oYAavNHF}K%GSjR~xPMGV>e92-oXGybA1(s^_
zXn<cl$BywZHqQg5A4r47|5`7Q3#1Mj+dQsOp0l**<JRh<hmUF@bx&&>?zZ4cLujy<
zZzW;t^=nt#n{U3+-qC`ajRbQp&kRyd;B7faWD9p&SO*)WRBHnCQM$EttUdO`+4kwr
ze7ZgP3?GcW?+2-wKksU8-q0A#_qR@+KjYWOA6w%Cx18+Ln5_-P{+(TIB5rA&bX)V{
zUfbWj(@yLhZ_j+_X>BAfwlj}E-p*)lT|ad}iw~ZY&_?0`53C<;+AFksclO)OH{NY8
ze)Bu++h6;7`@xGZY7Up~y}RB-+_|@x4aI|lwx_z&KI7Gh_)qOlom46Jhibg;^w2G>
znFJ+c0+I@_t7rARe%>QMn`LC-gb+9{unuB}UML7Y{5ITC@6Y)k@bs}+-YS^zuqZgI
zA4^cWqk9m-WsE0W1!eq-?ylZIU$7rQSRQp%{}%1S$u$=uGBL==;#gpAS5IT76XNL5
zPfpfl?2LXDN`_j?F>A`{3)pr{hZ5igyz3&SuDoawT0q&r?mZNXIq>!7Cg^=T5=L57
z7$~bJ$N3&Wcd4Ks<Rm+N9q<@Wd-+~Z0S7YAW$+OXu1X*$n~CtxjuVI{gIsMSaw&OS
zB&~u~_>xOJ>2RZj)&mUo3XJP?(gG%YO3^sH`G<#M_}aB!nJ7Gqb{dJxP<QU!(jat9
z3+PjRBgM^|H`<Fjn7{hfueUG#-tV=qe*LTM`EPx*efP!hwjaFmvhSn3AsW63vG21u
z`+k)!PvUd!l<#*inDIVq-V6QUu-Zx^#-%4Nw$K0aFSTF%?9cmi``^@{ar256>GCUY
zR`T1XPHeRck6&yb{=~=IXMW-5+p{0}NIRuLOvX3WxE<Z%6|*}U*v@LRak?Elu7Qbx
zjk6?2H`?~;bKVSW2zTrHjrP{-Z?xBc_`~+&*IsScwW-)%-)JY!o|n($ldbJ`@yRFJ
zM?U`X_S}a*tU>R|wxvzU&MC=0vEzq?zVyW}weQIny!wm-hx=i(nb6Kz{?Apa9OTs}
z0Q{c`qz`%>|N8}+xOxe8<Vso=*ubOuRza07$E$#{M}_5b`gZm4^!-^bX>f&~`mvzs
zmTAWP7~G#bf6*<6_bV$H?EseL7NEECrGKO?yG9^C$|#wITP|A_1EiNjL8aeI*DCx7
zk18h~>r`bU4du)A2CDqMV3j<e<c+vWJ618dRXih5ULoV9Mmw|2`c}yp^)J&=Hu9{3
zq8a&t_u~VW>+wN&3LeEv%(Ha>vjkLQ8L-4^{2V$`0NLoD5x_rWLyv)F`hwzF#XsVR
zYaC_H14cUwJb!DE=`q;+U{$Wg3&|<3Kza@=u=|O~18xc83_fEv^qD_Rm+V5_MlfDK
zXyfEZPDE;a79+NoO^G=JoH3KY#)m8U*se;MkC&R{P73Cl`|Xa_AiVaTH4k@JF!oSi
zy7%$p+wJVxGwu9E&3$K2x6|iNX%0Nq&Yf3$X4|i2T|1`!DZLUUlSo_DRG*ElqwVo0
zcG@RC^GU@|n1<V&_-HI2k6n{3ytaMw_*UCIaa?o!ws_f0$N+78?(gol+qbTZ_Lkb>
zz-tROH%~tORD1lHXWIE^Kh%z2dP;LNS0pw?vo0IhNE~VhH}=}i*Dtr1zVkx+=9j<H
zzW1H)w71`SLmP=Z?cSk#?Y7nsceRnYt2M-)HWPebV!!(o{}QvAsN0EMFE2{dkE@R5
z`}cyPAyoewaU&$J0GS**(f7Z(Y$*^;>Cjbw2Jg6vI}Lc$v6d)3#WTXPjRA07X6J#9
zzFqQ#bf5Ct8l=Zfg6KUDGS_uU<he?3`p6&+@cd{!L+Wz%a{@$sB7nPrQ0dJv$64P*
z=LCRar88nFR{qG52U_dsSYb0UR&K6%D#YIiRngFOJ}CignDAV#n`A1_0mo%OLMkSl
z`SU<d-tfdLWa)4jGzeY<%3cnx>5V*+%Yenz2Y`5gSG!@!;F1fK=gNmSCUOkNKB1g%
zCwDu6u|CV;g8=qa5R>bdMva{eG7BKQD0<re27H$fgz)+~Iud-nQ!33B9<By(8_?PF
z=i2u6c6(EU$Qzg6Z9jVZ&Gz=SciZ*bH~rA_Eo}mP@1pNv<e)zA)rZ@6Z~ILV+b6c$
zaSb*+__e!#r`;6aVWm%BINN^q7k;+=;m`k}_UyAy`9mk){noebK;AjLvvabYyL7&t
zRoFUpqJ8cQztTSPb3flsUc96VNY;%T?WOO2$5$(kX+v>D8;_%Wk(0M=@i6!i4NhEo
zRlBUMZML)Gd;I(bzmoR$&71A5H{NWo|LDi<nl=<$8t8XUp7a}$<Qxyqo7>0RrAyDW
zr=EJYJ^AFbZAY7mAOG;j?Kl6?Z?u<Q{GRXUV<QnB^)30DOUN`0&5#KrZ4OLrq}c7(
zVMzW&K-Vsye{%+kNvrf#zk09mUUbW4b-XR(300rULrWN8qyZ{VJmN~m2$dh{#xRzX
zuQae+K86vN^ToGJzf1>|ZwRFys6_)X@-Nd>8Bnsye?`9>3XoG_xvXHsEvJ`xR@t>m
ze&LF~VxZDv{eh9c(uAt>qu`Nc#&>(9Yn8qcvWZ`9&sdL2ud?e=^#LB021Y($IS-7u
zRplQP;H|j;DN{gZw`=Za+4KXH2RHX;An|-0L09%w3>1DDr=JPOe28t;RyA*t4qmj&
zAtJKas9-$hD(mUY;VFwus}zCsx}U^;^6*sbg-qA&v-NBsdLTc^5y(4Y8xTv-HB8#*
zmz?BbGr{fC_%?RW_La}M0>NDaKo(QIJ(%N7cewFAAzHTHgu_R|I6=ok@YJcDcImN8
z-b9=}bE=&@eX^ZmGjZm$;xoQ2>9q2XU%JpPYW_XBv!(fqTefe@p4NCayX-%CYOQ_f
zqvzWvKKU_UN#Nx+moLB5u4=yIWi32o$w|p$o4lmvWZT-_Q9Ze`zzHp_g+zZ#b2{@n
z&t`0GZhD>c%(Kt6CqMD2cJln=T1;|l)TTNk8xhs}-l4X8{Z4!P)gQMPpZ`w#+V6e2
zefRn2wf1<u-QK;e^~GMhr8UHz2YYSTw>oiqQ?5#|ap0LxB^>8iw<#6rZZ`k`y{Sn=
zK~#(FCx22h{f&h~d%PdG4U5+s2NH@=$Uk~Unn%EU=_>w`G07^N0t<3)1fv-f0$u+q
zfUf9SJCqGRHYI>V!&K^TJm@bf%e6S^Bd(L^&@qOv_TZ{o>MgYTlke2<F~mU{-_Y;r
z4}2xvFD1~Lkt<hxSlAnf-tS|(f+~0ap31exWFx_chL0cLoUcp(b49B%RlNY^;EUc+
z3V^<vs(}bP&a%5w3ha2iNpL62h5`1(+J)($#Ce6g&qG#({(4siId3F5a-|B}ki{m#
z4w56?cddiUXR<1LmI2%t3M#!^UZJ4!$T!$G^CaMURv92xI$30hw{Y?sn~3?jketoN
zn{+Jb8K^iz!e)YZKtKHqU!>F^CR*O<yr&)FJqai9Kv&+NAn}+ZoHZ)#%d2a&1(OpU
z@SZmr_vF;`m(I0MefpE_Q$P3Vws~T!J^%HuwjcfIhra#lnP;ADKlAfH-#+;>pKd!B
z&b19~CO-Yy&-%5n+%xz6@4ndn!Qc6N?JK|i#rC!q_&di>wv#7Mscx!^4Dkkw1Hcc3
z5pxA$T?5|ncKXz5fA9yl+_`<n4{yKy)|>9kYe)Gyw<cHB{jfGAhqW2unTVrG-`~IA
ze)RHd?aN>M-S+AaUUv86wlFpnc`fXHZzgDh`?4W@LP`|I_E~5n0e<=Yz{E$JD#N|G
z^m}xr%>#7|Er%*ASXH*1FS*M)mSLnDVHJPLFBo->FycnJg(n{6j!-fyEW^C8%JBp&
z(?81QqN(yx|43h<fU;G1V8o3yg<FOZzf3=t6|BOIu!?uA*D4r!if>hV^bz^;{Rkrs
zK*}_XcZysx+M{$GTcd1%yvJpYISG>O5^RYWaf!Ewp`3#}<&oobFP~@}k^|5?hOsWo
zX<)Q-EH_ly5N2B*4%k*Ncp#nmt(S|hAUdmUX8l0fO(<T#kwVWpPI>YL_%P7*5{7nx
zf%Mp3p^3cc=r{-RWlN{gm>%01=T)YFUr^0?KkA-z=RvLh+|E+uD-Vu25y;wrjYIdF
z%D56S@)jhGqz?!(5j#BRGLLAknEXoJRY=~qi0IE73Lg*Bf^tMqE*y(G?Bo!rTR&Kc
z)vGmqg+lWzZ;#qJy`xRTnYN)c4@cKmvybxiV9jmoytek_@pk_Fd2K!}wu_g}whI@I
zw^OH&YN2+}c1}HLPds(9ee9DTX`lF+Pifx0&~Dwm+unZXvfrk|B9Mw<+tH0RZ6<d7
z(Y%dKt}@85L-OOnt~L_aY#S$ESto9+ZE0@*P&<A8iFRz`get<VQTYfi8;Kk5Tx+kr
z@PqdJSH9l9_LVQU?|<*b_SQRZ_ypwLgFEf^!Ci+v$>&)Ku1KUErH{2*_p6Ul7C*8N
zFXBp}drZ4l2o9h(`3Li-SbpwPf~bY?Z_@o2fF9bU69c6OSjN8>RQX7^3I`yzmkbsG
zbK3Qo0vO6tPBt>DRzPS|3Y2UGAE+EBS@<6jiMpMRMD^!;S~s&X&>HdH{+`m>Akb&q
zJ<-6+iFx{hXtW_bbm%}~U;X)>)-Q*&_B*Q00xv^4=C_Tqrr|2Yi4&Xc+_{~0@#2~G
z)Kfg;`b0bO$38!QwC*8CIUD1b9aaFIlAP@)u_%W{Bnv?n!Z`zCYL`LBL2U`*TRX-9
zwjtWw>cfD<t}FD&g(<<2evp^THaIi+bWUh`te;i(7k&)mU|NG~;VK?^!A)Qvq;4Z$
z1%O{CsTyg;3`{KWc>tSRov8a<Yt>tF7??n8ZS(rK$J;Z{J=ad`?6f1|VNl^!bKEY5
z9-mE-9DJGw-t(=8OlDlIk<IFC3}{^Dq+DCE_8Dyy{>Y#FW9{Rg`gpsm!QyLQ{;~$2
zoBp1|r?ipy@TWf2&Oh};+c|f-ojiB8J^svd8VuLl_h0ya`v-sf@3nvUxBqT?^*b-L
z?|ki>ZST&mHa=&3I}^7z9okoIxzy_R$4_bBqiejaYg;nUx9yWBZP%?kH`~?gS5&8~
zsy83bd*G`Le4(2GTtm4Ifct)U`-SJf-JbvI*W2sb^c=&FT-jlhrMm8^?t9v3YEF_~
z`ML)h!2alM<bh6T{EN?j66nossI;3`8D2w3J4~Bk>3J_$F0Z~c;wxQc0JsY81<Sm|
zqnygW7j8KXly36NI!1WE@+t$C7qp}+zs$Q_HkOx6<c+YL2a0#(e-sqWX!8i8tO~0D
z{zsLKe5>?~xK%Pn{t*gKK*ktGsQic@aWh&vl&-u=oOjxEa~&(6#bqBKIXD6AfrilO
z2n9)X8p=9uOwSOUQ(Z>Ukhgy6T__`so!u`pjY&i^+Go_1p78!l+CvRN1f{3?9kPJ5
z!^3^Ocv8OTJA@iXvVzN@n&lRh?%<GJe3cN6dh(}yp$$dXflb5+W0`fAtq%NUE`)9x
zZ+ku)1s0xt!{u?3bj|1RB-pMl_oo3|Wx&+IO}$UjG8A-OihV*)N!u_P08X@;PuVaK
zBMKk0sN3gGfIE51&E4I*elCJ<P^@okXx`n>+@M%#ZzwncCT5>_T|4GCHy%H+-A<oA
z-cFv}XeW2pwZS;uo_h9)_Vjblv=4vmW9`)0b6U{fZI>@!(dOWm7K&O#s%?&GbAnv1
z;ve5WAz53Rb~%A3pL46T)(L#1%x{b0_44<RwkIBcrkyx*u^nF979D53MZfo;y~@|R
zzwzz%Eo~&e@wKnD7hd>oyL{#Cc7yLl?cHj(_HVa4Y$Uj?Nh=86+(f7ipidIh*7{S{
zFt#7Qoc%ecCmaqX>an4xmwA_Afk#Y>dCFHDSAy}lSp`*Y{w_DE87GPU=f`%Ep)AC7
zD5rCG&Hz;@h54A7ZL?VLtR>`UkDK&kmHXY5+RU+mAbm(}a$lPWKJ?2bf=xwE%CipP
z(+u~unHKF4Zd1Caet2J-h6io?_|bOy%$7E>o7&i(RKGjZ9((+p)|}_tV~;5&Jbs})
z^~6PQBA!-w^2x`AzmNy8%R#eKV?0i%=MsR?;UkapauA-Xe3ijLKKCB-kS2@5HOGaA
zj<XkBnK*icvSX5$D-;ZLhqdq|&`_~~!7mQa{qXY{498t>uZAIffwHfl%Ble?U+MP?
z1(vH&c-z;*2-E{wbBe{!Oz42;$^?GtXlmK{xkN}E?}_f17U*ZykWaM>moB!g<Hz&H
ztgBajM&yB3&b(oX+mbZIsUxKmF;F9eD-y_KFva%_W-Q=2d-5m$)Sqa-^vj?3TbN#d
z_2u@JFMUZJZmoUtlb>$Sed3et)TK*pO`Dl54RqVu$ZQ|qX>Y#qcKgkL^qcMP{`Y^U
zz4-NSwJUGE(|-KgkJ_7WzAc9yZs*TlXs6jwP#X;{j4E8M@D&RtF~wYI*ib!=@0@5S
zPHwkj8%NshySzT{jy3~#z1e{l-)wAZa8!Kv=I!=fZ6tp0x4-BI!ERo=uEDrJSf@I2
znRRz>UqN+`T8=v!^`Z93CHe03pAg2bG^Tx<Q{PM+u%S~J`^qX9%a{8oP<AiNS;o0<
zDZfevP<Y~^eJ?2bM=ih$D!pt=!I%b?^FZ-ZwyGRhmbr><nP%DcvAld$0ovvAF|IQ5
z?mtU-U^%@?-Uwqku*@?;;g{vCk}(4Gg&SpzbpR+MUtu|4f$_HJ`4h4Sr3ckTKc*`W
zIPG}`JrSYb9wTTD;4;BhS;YyxS{_?PCAuDhcI@L`IL{LZqJAlMy7I+meJ%j`KnA}A
zN+;!A4?m+~V9fS1=CnM_>zvSw!L|r02C8fUlx1EnKa?FV=y7=BuhFOyHA$pMMDj+m
z&O7<AUUSX3vUAi`J|)=49%C!!?twn$4rODy^7@~HwfuCTYB1?!P87oTrW@bu!`wus
zAWTk{O)Tc<w$Q=ix5-IPg*`3eILXJyz5EiIy*;mEm~%Oi$A)5kebaM<zpl+1$@5`u
z!P?r=reJfuZEtUAVRu&Z{e^bs{JHknlTZ02Cbw_zwX0Wev|Bgsw1Yh@5cRjdCYc+X
zzMFuJ#OBs<Z7$ZaRyID!>tXL`PUVAp`&xG#T5HF*b~LB+*2Hbi(GS}G%{%P}FMPj!
z?Mq*2-~5U;5#Rbw`_XH!waZuE@vTj_wTZaBr_eVNoQmAn`hsUOb4!~-#rR78QH#fY
z^3y}zY5-`)@?~249yu#unXl+*&t*A_xY!olB4OG#IAUM1f{wq6dzoVh(JMYj<OMEQ
zptDsvQUmU5ahCfPfz6mVWa^`QDnJhPMuInI60@M;qrSnX?2yuji;lZ>_SG*$$MZOR
zdVsL0_0{(AV{K<=qn$dn>CMCW^T*r8OFQj}C(g9zo_nl4@#Hywf8?pBE(&*{UAlC(
zUA%CnUAS<%$LHGlb7ykvp4Nj~>ZeEk_~$?2&4e2qBdC-eq^b&E&P$oA&5wapCl}2y
zS=&qcQ5o_e@j=4kRR;;dw>=5R;@G6<H@(^18GY$az~t^t4R$i%Kr4H^kuV=2j<=oW
z%I|pA2OU#G_Ar8RJ&hqqu>F#qIKhzW2||Y`%d-IN3t#!P_@MEQRG~z9GogWqGwPI^
z)^gCFK^}w$W7TnUbF)2m>2crdc;e(v+r6tn>-vp$`^HUwyav`>;m9R*Zzwd7;b1>c
zAcDR9eGhiB(2L#=efZh-tAGC2+SAWH+iq*{`|dZs*}nD7Z?(rRKHfg_v%k>JTzae>
z(E!CJfZL6ZDScQ25ns>#hkyI;x3B%qm)fl>*R`Qx@&BN`@%Fpz9c?VQj_}O0&$i>I
zPWeqxZc8?jl0)0j=Bl{8xTX%t{iJ8k?8xqec62Sz-*BsvjtkN2Akg5*)sVMef3tl_
z8;LLd&KKL8ufHL`u-lew4aD5y1kbMOy|2yHeHJ$k_%m%ma36A*+h%h7?+3(M+_W*9
z!`e*va=38OQMv+d?;(B^l-<B6YZ*oz%c1az33c3-?vh!$2=Ic7j7Qn9oUZ(GT>KS3
z45CvJr87wq%W`YsUpS!hRlbZHgUiszzw*oSMq3L9jOAlJinnkx%PeOgH$XxFZSpQ-
ziFGB6y59@RPSg8;PN)B?e*9rDXw!z5e8u7EvRptIY&9Uqk)hZ}K-&zoQ5YtjmNAdr
zRc8I<x{nH<e8~xfl%JvvE_GX!Q=xDLMZZi_z_6lUjb|!TuzKi*ew0uP&Du;r22Dcr
zASqt8pf1Zn$XckAhsTTWFzLV3QP(ggz~?E1{0QVbe+O%JHe~?6;G;W;r2D+apKu9v
zMS&Avib-oO$ef*Vw)z_Mq%A`M8^Qg4V7X1+3yKm|Wh?Bo3vHH@r>TG0W*!k?pD^Z)
zuki&k7~5+zL3?rsa^_jael`?uz4dl`^Nn|WyE1yXg0P{D1kXMm-#+d+oo{AsD6rY!
zc3s~l%ZphwKOaANqV1eM)ppLFZRa)juWxL(+jqEa>2|xroU6?Q4}`<Nd3-w?mCY?J
z8jt&mh$`m($y#9l&Yso```%C-YDcuW*mK<u_72+Bw=cKXzV}l5jo<i<_Po{r-+$r7
z_Qo5p`?jV%-qxhR)d;?Fz26_S=C&rjG2v%bS;W%zgsG3DZS6OK{f+w+i1H(VMxnd*
z1Hs5&e9QdgeGEFyv6J*<5mC`V3^>A=frp^(!Q5U!|B>sE3UD`<c^!kdFvsv5@riK(
z#{x9H^ZQDJgN_IFP0}1cY$QC3(brhdsI0C^n2&y}1rnbU;2E3swZq;pY;CP+p}*!Y
zoS!|n>CM8$ODEffiznKp$4<2;9zWNfeCk4bn(*|c_RKSnwP(26@XX^{_nmJSw1GXV
zwb;%Hu5xW^!?2<C7n@t9x3<*R*4@vkH^+c)bD9rmYKDU{z;R%4aHOmj#ndV^9A*qG
zDzQ&;*yW%D<T{X0J9eZ(4lo8I4k+GCgi+ix2;8T0a_`;2uIRY6Cr`9<ARxj6nPT5R
zsknj%hQbl|P@SOiBP?_hLg6TcwgTB+g-$f_kOs#95k>&Y_^bxZg78op8nQfa%|l$%
zHzlBR2%fKk({J$So!fWSaUQgDXD_smfAUjp?f8jyUwph8?#8ui?Z%C3e#62Mbtqos
zplzw|liR<$D<??^dO1U(LFG2znz*;C&bHnD<e&H>?GOLyKd#;Cwg!{$wXgo}m)q6L
z*V@NF_R03~U-)d>K6P4rTt(7=q>g)(hpF!$Zr}RiSKHTq`*+*xFTd9Ic#vN8Wedm^
z)OX)~*Kd5-+Bn)i`tgspW5>7hb!v{tzJOp5=QgCeTERwx`%2H8Io<YT_aQbON7vfX
zb<VP=uByZK+D3cnyDzqHeD!PXrSHF#Z(gV~AKO@K_a&PL!1g56Z<FwYBk7ZF2Zfpl
z8Thy_mu^qHxSuI@Ux}yQ7kY1*03NrgCKlR``*Lf5<yflYsrnB(&DU}71yw%tX2G%y
z;DM1I81;f8kHx)MP_-UI@y{|!4krgnHc)!Yt|}`Y1NZ@FI%DPl&g4TdRrl4XGqhc{
zpvo+_2jC$Z4(LdYy|y{!kAh{{V?F0Q6x3_AUne`P9}8xAkht8z#f}-*u+zk#lqZb3
zmtiFynF`4A)i?zfeZ<`ohm2WD3KF4zrT{#V_H+rZ1vP3VA$1W|D5cX;_gVLy<#>V3
zT%zBpOsD~>&f{?eZQ(0lG4cS?)lZAQ@Jxi1g(o3=p`GK<S12Cy)gNVj{2-JagvhZJ
zkK^Ud(mCoK1D4_w&#}}~d3Y=KSxd_kb=R{T%m+~isy{8_r@zRctkTx0luy#h=c5f?
zKu_frz=2szquXhCVHOZqdKn&p{;7ey?%x9Jz{mKLxrep|T&G@dY@%(`7O@e3&wWYl
zM?RV9gQ<Q~m&$nF<L1pf?Z>ab)2?29*YhNT_~7fQ6DQlbGv|CoVQc$@7HnIZFIluH
zuZ0^gNjb)cd$lRpId`sIdg3W90H3j~H?G}iH?M1vf0q-5jJsMuYGZO@=TuwY+HRZM
zTkVA8v!TEaUcPfg^WlLu5Ag2t%*ESpX`a69*Vlgk#TVOGe(x*opZ?ZAYv2CPciYRa
zzS7>l@>aXPd%fK{xYh37yX%Lnc|P+%ZR!WDRX6Se;0;^!lRS8>`L4IWy%jfkk?kS8
z(7jvYrpz1xXzaQn6qvS%k*73c$OOH-P~1L7(t)xnpk2jZW&GGIw`oF}6$JyIBJ4XX
z!7Sd@E2e-zZaJbS9PXWkJmUOib;XPr-b$$KzShCMvLGMU6^4mlB*&yz%Ls1})TL4v
z>*vJ)+K{Rb_>LBCq1!y#c1|5@XV0&<i;r!$C!XAC&wluf)?F7Jo_X%P!iDzKGw0h=
zPoMWyho_&q)Si6uQhS_dbk3h|XHTDPCy(#6t@Ulm=FS$a$+eJEbIYznzT;(s+tQC|
z6T$e_>tmZ0rnfoO!R`T&$~q)p2Thy|q>Rp@>8Phrs9$90FFX91pAKf(1R26W&hD9=
zxp(JU6rWVbKgsPg%CpGY-M!oH__eY7-u<E@I@$4O7d+V}xitv}g)YwCWFV^vfHXJ&
zT&1xKhz?{RE3%7c#1(zPRJW4TX+0oKye_gG1<+#$AF7)sDd8OC120NB+phYv(c)9*
zyv2b_q36^A9)Ieoc0#_}ldpJPEFTBFeEG6<9#Q>|Y6HNob@vW_+?CI?iMVs14M#5T
z?wnZHZuH6aNB`&-+Ot~lA5!^?-};u%N-$vk+|U0)d+b9WZfjcL)9m;5d2`lX->SyZ
z_PiGK-~7r~+Z#W6T^pTUb$Zo9^;$p5V!HWjpD(}ia@$dxKK0y(+6G_dL^1uN2S;|B
z9He^Up3SYzc80e?aS!O)rZ3lSZEm*<eB%GX;r1e%iLZU5z5LS4?bgkk>PzbU@)Zk~
zc~Ham)pkC9-450xNZZWq0)A>IXoGwRu(iS}nku7@bK4c|hOW7f&<8>j(7`yRABNbx
z2qkN|3|v8_$9PpaFqS<`Q+7X0Q}Hq^<H@h~m1OF#s^PrCc#IS<K~;H$4+5i)iie26
zaW?nAz$)1@-)h}sc_2DNn+2mMrooqhdYfY$FxvT0n-A(%^#^bd+Yq44L4AGFt9*`+
z1YqPl@ChIEbX}CWTw@YrRCQpZQy-6CQym${dcBzub<8vF`$Y;~#2*ztR}qn8dQK>q
zr;9c|0euzMI7vRH&E<u|Zj#{ND~vp;?+jJ8Oh1-YIr_ZtoC2ucuX|ctla7)z)6Zdo
zkUUgg<CZd!y+lK~Qc5%KRhY{I!2yf-<i{A8agLlHgLMx1&^Vgne-%J2)>8p&%N)V$
z>oecdG;S>=s7p=;&2==0pca_m(k8PHfa){!SLX1;hadQRQSZF-PJ8|J*F8`3mPY#h
zF)h+KX~-Gh6DLn<VR*WoK6Og#q0{Z`xeM*mV^6ndpZ$o!$J*)B7bR~~^Zl;F?p@x%
zb6E4E=E$uR;ycq$oreFkU!T6d&WGnT8SfsnJ2!5&YgewdEAPD9e*ELt+bb`>(q8_-
z%k8C?zTaMW;l=jd7hh<v{qVK+?zO9a*5lv-H!<DIt6sG>VAGJ8D-qm6+&2+ieOEu5
zHxV=7j&hqLxsABe_98z$bhjGZ7E56S4HuJq4he%|$h^$xVLVX?xUC5_-oVp?4P!{`
z8vw2(s3jv)!F@=0-&#b!;0_t~0@)AvsuD7>M<fmhrsM7zt>L&j#N8tDrC)KZGE~B&
zVNq)at)Diu2GUaL<j!HOi;t)sc|f0Ala94>=eD#NINLt@(Z|~-KlxPqgu=%@uJF;P
z+lN2=bo<bUo^H=Qr|`_v?U|>aYEL}=M7wn9Qaf|zY&)%uCSm8~j@OM`5%XOv>Z8;w
zLF=e)Z~C?D!yg#gX>wK1R=bbepI`XgN2bk04RD$C#M>PYRN9VLATM#o6-4wRlx%`0
z?v5wnoN$z(o6CC~OvknG++5!j?uc)7;tGXdF(wU6WagYri!HRGb<jCx`;8x<Yyqmi
z%V3%$9P`UWQ2<}@mMmbDUFi(u$(K5Kiw+!hV&RzqM-9{EQ&rOJ0(mDiIQ+l>bo=%l
zZNQGVr=EJcojH3}e%o%lTIgSX_p&z>w{PCoZugKD_3Le2J4t-vZ+S2P9N?GicrRTz
z)&B4=|5E$2zxHR_>5CWIRSj^wL;8)^-)t8zUTmNJ+#hN?+O!-|+i+=(&-Qcg=LWBQ
z)j{;NFMg?gLz{`W-gw(@b=uhE^|pEDg+&B5vnk{aO`LJz+Y{R-j#Dw5f(M+okL-`|
z96H1$UIte-nP(I<u%0~Kc1~Ytum0eb_NCwXQv25P-}2>S4oo&3_K$?*fqV^m9IRY<
z;9#(>xqX0X@K<_3Pp|^@fk0+>pz<YZuDJEd++VAd)YNb8LlrOEQ1}s+c?(8-m5u3w
z#eDoR(ie<r>>^L!Ed1z4aI5gZqv(|A`5dbS;LuZi#iM_6`70b4Y#aUzizW;*QofMq
ze<%Z%^COg9#1%%UEb%Po(eet+W|B@%437i&MHuBxwlLHQFAuuhamgVCj<oX~z^SkL
zSS2v%m`4A~d-2Z|!`}AUe36EBPJ1J;fbV5YQt-fQIoU`oZU|CdeL<=GA(vYMIvo?1
z$H$(pdWSb<1!)H1p@j!ImS6Rs18tQDi|4?f=}a3)TPGhLQf135XDV~N*heZ&JT5+$
z^QPgC0gXKW9>+xEI5Hj`H12uaS=2EHG{lD%-eIH}fo#HJ*^GZU{Hq3O)ww+j%HG_r
zBntOu+7*1Y3zRc2W%b(4f{9mwm}8*)MSo}E1`Tr)bNJP(SKHffzvE5C?(UxNI^Y=Q
zS%;mITJY_h(OPIn^PU!gn)l9}eN3B-3pysZ)M9J0>u9@sn_HN9E7P{-*OQ(ncf43U
z<&$rm+PlYx^={s3@4WF=d-dhlyn*<EHW1(Ww&wXa-tgO?xE1UAjqB~o^{eg1tsA~=
ziS-?ui9Myc#i=$B`EC^pVa47|D6N2>+~?Fki2?lP;9P*TyLC_d(5p1~9u}Aea9NLW
ze+3O!EG6b~LP*-ex7gF3w4W5~H|s~4$l5^r$mz<#LqT{SgIYXuvRpHh75f<Mn+O%9
zK0RpJp}@(S<N@sCj(#=@TB>-npk;2`)>7!ysY7bpHEj|$+l33;?XgQIwE;NO9)Ikt
z!kPBuQ|H^$&s@?5;qms|hn~=8;R$Uj9~b_jHVJ3^MkOASK7Vefojs$C=t*v!+wxsM
ztc&mgwle;5D?Q^cA6z6a%N%k{6D!~o*IQ*v<06oH(y!_7Gw#T*XfxrQAO}Ow7`i7q
zj0a8)QW%<pr+CHF<9R?sPR=SqG|I_=+G+HUES{s-khi)0i32hV!@SnYjV~S~c=C`e
zcVcyH6nnyOMJ?V3G5U}X9pH|p5W8k5y22BWwvrx0;ee43sC*W(!Hs-HOCB4ipvEpd
zAqR)-)se#B1s~5y+)>!vT5p%M0bqfC^3+aS-(0t^IP-Ah+I8=o^Wi%tWS*0_+xED%
zX;1!=E4H`R+9yB$T>Eo>;n&+IKlfQ30sHNR=bvvcXpmv!@f5Fv{p>HcwG$^jA@dyv
z9^k@{Tq(G#LFg-A{N48a*S^u-ediq+9e>wz3ta9!s-uTXg1p&+Hae-@@Tt>$SVu?F
z<d=S+JCLW7(>~l=y6O8vH#Uwd-fD-m5&Gh9{Z9Lm^uF@)tL^qJZJPAw_LNR;e`Ck*
zw>T-rAHCTPXsEWCLN5@z10Zyn_nuJnK&6Qp(CQLo^vp(#{h)w6bfA~Zm(%X|g3yO{
zIgDi^EaoZisR6%0qqP6(7?*BJmvsZ+R>4GTyG6^t`8+iDz+Z6sb(@tmd^zeOP&n!{
z4zRG^rh*B-1jSqO%qG2zohb^=u|?>o==4LZQ7g2CFB2$FK|(JF&=F#C((OM;x_|j2
z#?QpYL0RP{4qg8PEK)irp5*{t_qPy{fnTM^$E?SQPOG@yz*l2Z*;Bm5ub*C)#roXF
zDI=c8z{t;hk+R7Zy&k~SHv03}XI+pl(o_Iel>vpf9^rEJpma!or>l5^K^_d0Lt+`O
z&loyguY(sKowmw}U8XqvjKP`rDpbDcykJ=>i|k%Dl#609?ffk7QIO)QYsu@`4wYXe
zY8CE-0B!p(8qnBV$iH(w^{6%pB2mG(p~gq>acrcl_fv=Tare6(*LYpL05NZ9<iTt<
z74KfT(%yROEk9@Y?mL&;)pxJ8Ygezgt5<G_=B_p$`)z-hlVpcAk8-zxWZpm0Zr|K(
zcNBbIu@-;4GM(FBNbmBtr+2RSTTHLK^m6<5^WSdY<XOmXeXG6r!i&BlaplU@c9**y
zv|+fjch~bYF`I}3?u5`Lg3Uy(La@2GuQ@gw2<BElC(-+v{}j*1*brQ6yTl7`V8jzU
z%@3|44aAKLp`)F~wH}5D#_Lytk&P@LFU%!%+$ds`0Lbi#u9npWk<~+|sQ@3{#4$KQ
z*}#T|CESKK3nx$VD*v^1`pkMee|}R7{uAxVCr`B}pFZ85e)^0y2<O@}PoHbgJ$tD=
z`^@9hGYd~WrHul&A3c7tT{zEX;k-5t=i5bv3m49|b9|%Z)QNWD_=Yx&oPa;3jR_C)
z^Nc1ZrMbPDv0LM|jKWTTr;Yl%0`Y^_(%ukp#YBByYb)q7k)`j}v(Cu!2_pvL1%QrN
z$HZU#OMlMpfRSLdG2QWYXF-+AfIhgPsQAJ&8cdi6w3=*c;siTi8kUS+3BSEbbX4}L
z7OZc)`Br;V;ffBfySrR|7mX{J`|oiCK}~Aes6#J0eCvx42M1gV6b(4r&<6#@6$)Zc
zY&O`exqPGrmgy-+7r~dZ6ymQ=;F<@pg+*r`xXLC9&oC(YGMvJ?Itb+i2A~_)Z)mZ<
zuTJ=6`!j#$&$NH%*M7Bq^i!W~N5ucyYp=E6{>^{d{@HK-X1n|bZ&bSHTcWtt=H~V5
z?ed#%wyW>(mCk$Wv=`d1{raD3|Es_JpR|k5J=5NN>HF;;|NVc^zW=?K+L=>l+s}RW
zbL|iRsXyJ0Y;u{jwVOA%NArdbluf_N_08|U)c)rG^KZ2O@o)Wq+siNgz>5S9F5b>W
z%$ZF(_c6W|p_un)fBy5IZ-4Yp|C#pbU;LcEa_z&I4`8UX_I<zojf?$u0@eSJY^NUk
z8uZ?fp8xmX{BPUu{MJ8fKYHzlzUl$c&K%JUZo6u0uAcB}SguY?SC!~fN-_xM^@AA@
z-F}p#lR-im=m-Tvasg~-FRV22QEi}j69*p#;u1RTR{bQCPWoj9fc|bHmY~Xk>Z^rY
z?u!M><*Vtzt)`n||Mug9<JS3<lLr4NUhrdDKct)wHngcXz_A~B9512PpV~F}P{C8d
zqlgMPUaSgy$SlV;RKCZ_=NI`F+=WklWeKbcS-resz;#;!+b?_qm`;Oz@ao@upq(bD
z56@y3fFcu_HT;Iok4$WjpEAC6azn`E*XrOLSJ3dM7&_!KZUQw%mOPiD!{Zo|D32(l
zM;}_il~y3fBz!>ne#Up8<i&vQn+_$@a+%TnA338fK<Ne|>R~y+h+l?f9>)scD*zRb
zI%oO44^6a=ATS@Ewk%i1zbRIz^rN8k0!|M3Roo-}`$1^OIcgUFC^0{Q_=i9(5+5wV
zA3)>>g1N;4X}4Lc4hzp&%va3I(}r3BnEO8qch)|U!T1Bjk0np>pm6N3#H>lOF5yi(
z8=m`*Ypy+c;&?lC@|40(ZvWmn>5tHH>k}Wk+dgrkuM-aW7T;Uf^Wv3zdk0p>ty%=;
zHs17Dw?Fg7M*`oO;PtS))rsSsm$Gm??unj_1h?U^6r&%rfynczIo=pQvfk=fCKRe(
zb4;L)@dUBKz8dlZ(r$tR%AiZmG3DL86Bysa0+ejz5-GMwWlB%+lpQ>aa7=+s0Q<0N
z%c=*{6SS&w>}9E;?vN`DJky#Ntz;pu_1v1~{w>~^#GMuEZJlRcHnQN~+TK)K^EzOz
z5QtWSIC;)nGmr4v`#LGZpBjM^u#Ca1;W#Oqdc^Rotur1%@0z(NY51|1deDw8l<oA_
zY3wph=zMZQsKdO<zt16=2j;Q9+Bt1KKfeYX`fGpr&%-|nhnN6>>q!+X($s(%WwUYS
z?}o3EMXn5_&d8}s*m*P_ttWaMNXNZ#<9d7R?RT`H(5_t@3f@G(Vo<A9HAr?$UWg(D
zZz;B3$uGMObkNqqkqwP)C?|9pplmL?`K8>8)ea+lLGe^sIXs&Qbk74x(U8ZEI{4{~
z+~P$k{W%l)0}pJ+R9<Jfd1iry_2&9|`_v~t(SH3e{Dt=UKl+9C_|s1}UeWf0?|r}h
zz5nrVw{L&_>+R~h?@AG`lvSNoo2!@Q#~;7VQlS0Z&wr-<FaF|x(0=tl{0~HPzkT_)
zeyjb)Km11;pl-KMf9hx2=l<|7x5q#EDb-y^jo`}V%kBEL>+S5>b8UO)bbI0VzS91?
zzwzI+zyCk|-S*=j{z!E=qQONSOm=d_vEk!>yv>cZwxI)VQwP=Kk3H6Y;TM0g{qi6G
z6YZ1QtZbY(C1+9Lek6-8Zo-${=N!d4sBUT?{Oy1AkJ|t6Km7OYh3|Z)U3vGiFTMI6
zZuJ!|wNiiH;<PJ#zOm8U+`eQn)lB1$AStGe0Chc7TK6=FxTJ&1W7qow&C4c`fn;k-
zn+f#!s#^L0ZB8Iv0dT)}>j%buLa6)*g&WI8`cXzfa77bb!AJ+#A`{VjmgEr7k*v!_
z;B*1>(IYz5IkFTKr)-fW<z0_MaPwv&@@g<sPI&o~!mtWb?+S=<Q3Nc2QJoGdk4Fk2
zE&A9Aj=1u`BHawwR{ALq9e8XH^D=MU!YT1q)K#EOz8e=TkyHGsW8}^{kw43>G;|XP
zm$A6oqI_yA++OG$dHo?O>#5^(2Ixvw0V&2lw-@yubk2GuuM<}uD4J2GEJ$p+g(GwU
zpe`N&8HU0QG9LxzfUk60XVN9FaC4rYiAA9D`ZccSW&(rj*wZjqdYKMbz$rb|Z3#xY
zRdPm1HL<HBI*=c6rD~P9_bU6K@CRW1Kv14iuc8&Y&HSY9q&UocB2RM)C-8V4&u>8#
zFlPXqID>)Xqvr6r9eOR4t#QC}qvj){mB-f{69k@Tbo+GuS~hB<a<pw|fyt}Td4PJG
zH$EwBXg=YjA<ryAyLN1?9n<>fzUDk$1Dacza{`TpK6y^K@eQiny}>+l&$o|bKQ?iq
z5sSI%K)^4&=gC(cM4P@7n~}W``Z&ydnmQ>^l?v9_K?dU@OX~&HVFdD4*V6^GN#%*B
zeCg1S^PPyux`fA`0+w+!o!gi-m)1oaYe(7^i~UU=VCOl<+yTP+VSkS`Me4P`k1Xyu
z;I;3EwP84{w%PLAvix?V&Gl?3)(-nB0Vfl=yN0)M@g1>zlsH!o=zp=9e$&tRKi~v9
z8wly^VR9mookoA6-zx38Vq4k*1Z(I3@~uT8(srUW+Wi!L2F6!(hi|@hDRIW?xo_9J
zdqBUD-mG1-89nsZ|H`jp;RNY80R+HYPy%h(65SY&tp{D<sjzQ*>V_IWRRU}(<=roO
zRZasSjmvMB-@W25K60PowQIZW&aP-wNaG*TfoEB+cpQa^Whb%qOD+c*2RFCikT3g+
zHV#uAgLe5i0gf>8Rsd%`3X4}J=JL=m7|k-M$Elqib#&<y3M`$*m#l<XmoL0%k_^sv
z?`eQoSI7L7FMOf>`oH%V+UI`xm)oiH7yR(?*S`40_8b4df7rhL4Q(j2fjA;Lo9fgD
z8d$Dex!lg3J=4DMNB>Cs)Bo<DYtQ}6&$RouZnnSoxBpK2*4Mw;wzp2S&wcI>x6l64
z=UY3zqa#n-aLIi8t+%zYyV;(4;;D8_gWOmC$v<g-^S}OoNzXrSZ@>9wHeC2g_PW=|
z)(s8BXHK4=W&Ke(-Ue{~!UaFr{ilER*V>1E?iboI4U+PR=%q&;lZ_#R9alJv=fKi{
z`_{`Zx4-qb{y*)X{^mb!ufP7IcJtPCwJCJmhH<3b)zLm(nb6`rgJC*9w%F(LUEVyA
zft2>33?MxXhUr=kK*b*f5LZ_wwl(<E?`kuV^0^;XpRgUOtD_Dh&7)wHRZ#S+@G}mY
z(z_a<rPK>Y#rS5KzT{1shVr5v?eQDC(&;DCI$o_$$dPa2t=KkG))%PCc$31PjPL>&
zJ=PUHozXv)MTY4+tvaspC9@!9v1PI^2;-}K1aJ(rwGa<Y<S+9(uK-PrG33kFW))O<
zLGe^5IRN>J9jLt1#wyUqMT?9<XYp6sysm?>!DCnN$I~R$`6hi~FjRRq6iN$6soQ1*
z3I<sttiqYWlj~|Iy#=F90Qs@4m&-?>2(nf3z}0a*mXA<)V)%$hT5zNOvCQox7-bZU
z7O%2@q#e_%0sfJ`Cj{@y+M!JNp!BQY55W3?02Afyxg7?LwqM!uWEPw(9C<jJyCA4r
z+K>l#E1eBRj?aXCEF(YyLvvV*`rWCoSccXh#^2r5G1WIIlxac9F~*BiScG^?W5K4t
znq_@+!zb$4*mDT*A!Fu9-_tL=$7{)F9mP0Ae6W|-vQrP%63E9s=0I->M8lSXHO~X#
zxf;Q<5gfO1kt-kCj;Ahl3~^gAWJ*pk;Ivp}PW;;O0PX>A==h0dpaWNAm9G3Uz&4fZ
z$7wkge={d<XcNGfo!8cLvX5sAc6PSfxicr*32h8`4wVf7n<zG6v>86C;|HIycwSrA
zMsI`ffNc7U=3IT?>I56mLx)*Ys(z-`Vw=7o1e*cc*?quu6p=#4QQAaV#T0upp*XLR
z<qzf5gP2~Mb&c0HqqAh^G*4RxNON~u0rn_>SI=~P^87>~ZQ3(}+Dy0(2(%71nYnW$
z^^=@G@=G6E_-Ne-g$KrTLE)wtJ*o)#F#yQ6kt)X+s^NpA2aLf9*vfrco`~fz&ZzSs
z$+jAuv-ZT1NvG!QbM1<0a5f?1SzxO+Ot$pcIu*t`SK4jtRG{!<T`B-2+nr<p%Bj=T
zxB}=Zoj5YNS54;lwwPQ}Rhi0B$>K;-OiIJC%Vl><$Gm*`T|Yzx-Ng$R{dH`<ihc3%
z$JE)^yx_j_?%VD5jjQeMtsAmsukD;U(LVVzpKPD~{O8(7KJhW_q>pOD@pgOu>)-He
z<}O`)tUdGGhuWDZp3b#Ab!hIZyLIcf7WwymWkOc9AAI{e?JK|g<@U-CUT$}^*!P(n
z9b`IcwV~h$rrr#A_r2+%f84&w+a#{arYr6GwQKDRue{Ud;wXIfy6UX9_n=6B$EW-E
z+qJhYw{Ly*>+LJQ_oepQtFN}(w{N=sd1(Dm+uhsEZB4S1z+gysVlvi{C#lGmpM{*>
zpkV*T{|3@Bav4Q`ct0>}56NfH;X$Vi>gcC11l4Cse{9JCUO?H{SBkb^<OhltsPrfw
zSftOu0hBC4j>kDBh+@JO9<W`V-#WUU0&wVXP&)kT-<Fjti@wK`fw)hurE}Ef!A~(E
z`QY;4FnvYA{KV>Oq!}bBlV3TLNLl2S+&Q-W-QKE!YlzMvt}^1Xld_3tz(sBk))Sk3
zvY_%kkgp)-1a~}SS=<7QJe5zZ{DVPn4~(hA@dvcM|C8!%R`D_ztw79tWFHyH7@M{o
zue?_IV(8;>580FqL50Yvcnl+MrbEV30NLv`24ETYlZHv>O8g>jIh6h}9--O}c&NO~
zFy65v2m7)^(Tt&>lElkl6%>BC&;8E<#wy86JvD=dl?A{TUO2h0kG}24d57{@Eg%`w
z>V+ymoye@?mT^xh&#SyZgjc}k1gFMAE%<q*xW_|eNFEE(eRcAE*>YI4taBJMvE@MX
z9`g%tib97hH9eKlA#a_=iGNxnDR81KCwE=DK35WQmBM-KR>&ea8wqqwR~w;04tt^Q
z1AB#^K9u1mhmP*1O`-7c8z(;EqXeZnN9jPB0x`jI#8yG5DTbGeNn4w(ojB3jxwCxf
z`?%kN^z5@2+lQWeqJ8wE&$N$y>_hD%A9=1l^UM?eDB#ZNlWl8z!&e!&>cB_YE?hX@
z9#dJKS>V})3y#m9Kd<t$(;G<HSkNZ4b-n3HZnT}m2ikD$@3OIzzYbI;pB-pz!%JfL
zP6%HmKiJorRvSN`fR?ZEJL_BmCy<GAYZ;4q98SCTHgJf}YKsEKetT!S3Zgknv?<e%
zlC-yN@A9%~jIPMWp6KDpbc=zOKBVBy1Oj6ig6D~@pwi1Y(-PBsVn@a(2Uy0Dc9|?O
zXU7r1_<0Ha8K-XT1J4H(2V0&mU`NPK(3gv<r6ke|Y{mO(th;x)<i`dh&f$Rb0p0cE
z5I)>i2Ri_rA*Qc{(O!b<G64CJX7Qk#{ZM>FDl%{u<ODlo)dQgZuAhJgmDI@(@W=P9
zUB7<2-Mn@~Yxu+J=#RCtmoB+pJEwUqoH{oTR$af=_IL>T{(gJnsVCZJe*S0LCw}&4
z+PNnm_v>C?df|ok{TILAHr6(DpnbSK`{9qa^^>R7|F|mgpzUdgjz5pCX*Yl3q;Y)M
z?rXpI<@V~UuW2#AuQplpC0N>k!Hj$Sc$YR`_vGF{&W><3g^%OV9z4*!b(}YMt@&U)
zrcIC?LZv3h9BMbOTx&0U<N5Ya|M73M?|l3DcKzxV`IWXyU&XgPMCMnnsg3P(;b6z*
zl#}6)Hef@P^0Xi7__1pJctVg|WP4M~F$kRri!R~4fMWPFh^HL^lC{W#iad5D|1cCU
zHWb93hYcv&s$0RRzaY3Q@S#n8=U|-mD6hU%i}6)-Xv%$=0rJQvc<||Y^LRnqCr?kh
z*-Vry=*k{wYLnq|$*ILS6Eat?a!foNN`?{+-jE<c|6Hz?mgE(Zj;-+vXmG2+buRgW
z;wk)?w%t8F2K|^$iHv)^VMDR;gFNgapkq-F{(v@p2$<uC+dm<TiCWxLAV1=k^A%<}
z(G}Vp&*Rz3;$N2GW5IFc1ov=yITS48#yo$*_@MVh35`OA(2lqMlyfH6We%Y&yrGv(
zat*|bb^^=3E*NP>{77HvRdkE|j@(lT1KCQT{QYPE)IW-DF!Nt~u-`_%>n|?@;P_v3
zfN2%L#y3k2QkHSZNfV?T6!Lt`E7bTv->xyqARf;}Iu^n6x@4{<Op=u5M2R;f+C2E=
z2)8OLOvkVKGHV!4o?^ep0_k&Fb1&0LzT2g>gNlqtJ+iCbwnN7zb?9YUG38ca&G|g*
zSBuH&iyT8bP<*Vg8ParWu<Sm>Rvdi_agURZds*dO-h)Xyz4JpD%TZD4*j%h@*|e?A
z!<o~E+oi|0{H=u#fB0hi=*OOFAOHBX3Lo;TQ=fe5v3Bv|*>?6E4>EHL86OHfCVNzO
z)svHrfN#;(27%Yv^XgvSjFeZ^^UMSD6}Q)mM-E{9bN9|}yM6nv^Y^q#-~=&O8tV3?
zgMIuZ!OmxMLEe6}f8%3-*9iM|ro%9ZnAb|tGG5PWU27aZbgR!m-huX2UgGAqGl65?
z)D(ZhkByy<0NV5kfyBYZKj5M(;i0!VEeolbvSqLk#n6<&1*Kq2kFvZVa5>EZPQj&M
zvg={dOhL3z5W6_|X#DIR`HEydIe+pbA2wrg!6rlp>0u6F=+h9{O(`vBV6Sb^UlwtS
zn-1Ekd-5xLZF8J9E()-NQ27FADxhPYOjCEnLG74?StE!0>9{S3dP}-!s?i4MOxT>k
zN!;p@*RS33*Qa@o;&E*#`1sA5HeI~4dRq(jL&wya<%?s-*V|*yKG8n?v!7|te&j>m
znC;)V(O&)jOJ1a(KX+dEkG6A9Jk<`bZ>U1KGQmQfy0GZ;w<45&{wrT?U;pw~+mC+u
zV-IM&3YEc*`tawo$CxYo_IK}SGjpdM?CvSca}E!5pmMPC;MtllC1VS(2kxVwo9y1r
z+wIkFf46<<pZt^d<uCq@biL}GIkGuj!5Nf0`+IFyI`*`4_8AxH&Tcu*c3WYFV)s*Q
zQ{MfLe{UGnX7not3)$f_NBGxm&O$HcEUJCwOmQZQ$Z%K&1_R-%uT0>iaP*7bKTOm6
zNQ{zAT`;Va=PVNaFq4u84D_2md{cjPpC;~fssr%@r{SY+EL@ZKLYBbUd<IHQZZ7Zj
zw+~c*UV4CU#<Tom2tKka_JUvXh>!B1S(h=sC5W0=JC)rI9*p@<Oc{Ta#U^w{M#Zi_
zG#txH$M@ucDpM#r*$pgj_EV?C$hJHI6U1D1WB|~Gmo(**FNDjbZ2AHGp|5oE(yt1-
zon?pd#!~Zcr$IVNcbg$f{Yv+okG}Nz8i%qF$+0}+`a++$`Xd2;a1~D>askpaZZ3N`
zFqR>cTK3p!CP&5AGvmiJmE>=V3pdb#gCadiS?Mf~8DJA`jzb3|)iq#i#g@%+hmDmF
zE+KsWO=gOL#k7ANCqzbY%b{pOS1=_1l?O~M{}aDo_=#a+T9z~i9fkg#QyepM&S@e;
zoa=1ImAi?PPObb0qx!n9cq+|bfU@YxttvfC3rqc5AnEYchNg3aamvTIgL*YO7*m*&
zCaNA&2A#_9YtG>|C5{k(_krUPX~<VFUN~+6_K7#q9LONgjr<YoSUlk5fP#-L0a7^D
z*jV_rr_{|n0j%%UHc&^BpfG@#pe0DvF6H{s5YEjEP8VN=5y(Tgt|f_<e%?5oKBbMr
zrQ_|1C-Ro0C;gVB^X)?)e!}4+A9<!d_4MO@Bhs0(r`mC?g;-~-X}#fFzL<yPH*Q-)
zxsNv~!d<zVE#ETMOSejFAO3TjD#H~7HeC5`M9$ntj>ZXmPul*H??juE6_TcaEJFH7
zYM41CfxJ&LIh{OziHCkCA#Cd4@tT%?f($58Fmw2ibnH;D3`8KK#^Ts-<7{I$RR8ve
z?UPJQG9X9o_=P|8(djm)ih&x$Dh8J6Y)CI2=`xW<lW96VG=vd1L#I*qgeFGg@X)(b
zsw5A^&|sX+%<WH`enOu3)Txv0)XoWSDo$vVapHtxh2zJ$w{5+xZ}OqMJYd7we-_?+
zD24@;8xI{)g}NzpM+F{Tmr7UNNpn!vguMXWRhBi##xCxO1ntXM;v?OECO#q-kzUx+
zfwV2(-Q923uix~;>YJ+51#Jp=<HHtT!90DkoseJFw0J*pdZ#_7O~i*j{GoRC%&B%?
z2g232-)?Wc@rDIodi=3=>FH<NHg8%w!oohA2?icjG7o$y+`qfqzW%#kZr}dax7+2n
z-|>NR6q^)MZ#pMe0aOnzm1)6$SRHgtb==xM)=r+<Zl})d_!cW}jjGKAuYWyqR3pV*
z)#vin_Ua2Swy*s5Z@1t3o!@D%y!=wTb?3IqkF*D>J8#2d1L05mvq(}<@9{ydfBVVp
zgzLx~uY~gUs9oCgYBS%DDm&IUHv7UigAI5G^%(-2xC+(&K=oO&6?0Hy(=@Q4U*N?P
z1B^n-VuSgR!|`Jo$hN@b>0juDb}6LqVN><N<h=>%c!T6rzJlp`K$d<?>Irb5F@4u3
zYY>)zOqJmi=sZaG4TFR8Ih$j7-9CW9G-tJE{~(hj`NF3JU5V`{r<JROfV7TAQqIF$
z<qQ7oG~mcGfF_-gGVDwYVADtwx{MJ?1AQFmyyR_vRZT!TxTVnl3}PiU%Hgl#u}=|z
zN^=Y@r->6XCRLy>I&}IE5=LDF<cu(ud4h6)ujl(;Y6#9;Lw-5!RL2DtXgVJ66nz2t
zNybn{x-7x^Dlt8{&X@jIWpkJzbR~N(pYxA`nRbTCzZc)X$}pNudo0IeC>YZp2quPy
zrSO|jN6QG6t2ov!*xZZ2P5jEZJ$j?C5aamkBdUtiUnr-4b3Aa;%X5;_)3N5`0J8MF
z9M*mOK+_QOW0fR<UY5U*IADYX_?~>Ig>=fcO18$y0*-S(ONec@MRPTw=1<Q_N^6~!
z+iW7oLFtrpOi+i&Ea~M7OyD_=$~!Hh!rA6A=!Y12tS8wpaElV>6*v5B10QW$Z>LXh
zwF~D@w8tJh<?#69r`uCcp7pmEKKzj<+jH7HJo60SU%1pB<F%?6&bQNNc_Y@g=XK^L
ziJ$92oAfz2`Y`}oGk#+iIOv$49N>Xhc-D4oCcMUzPiofTqzRR=ydKy13d}7S2o^tN
zPON!0<g>?Q)kKq!c?AKau{rZBb8zSD`a_Gn$P$)}>k8P$P8VG20$uuS%8{Yk5BLhA
zUw)etKtaV7%GhPFLKNf&K{2M}!LL95>tI_3>rjllh#`p#7pGwrXw*F|N@+mf7NJE7
z74ZYLYETxrEL=$6)nb<q!|;kw<a6oD_pFJQ*P`z2(gAn?>5#wKaQn`#c{tGM46cvQ
zano_IN!Hm1>giB;kQn{K1}F>6EU0<<pEoJwLvfLp3aJ|I%<2zz6Gg~b<E;YynF;Vd
z0QKWVD%>IjPdU44ZXsINVw?>D&n(=$u7$R2__?3`nf4d|qrccb|0|zw=bpUOj%pKe
z_wLR1=IgJttJf~K6KuRrp3tUat?kN3umAYR?e!o1SO@XZ_V|;}w#T0NXxl#Xm=vz5
zR@`!RMDlNI!?5pXA6R4`T)Ec%oB#5^YJczV{7>!0?|fIfugOjg0;&U-NEr=|$UZhm
z>&Lil=ziPMMsAb0CTOF_y@=<|Uuat=PD<x9?dfMf(w=<s!|m*aC)&}C6K(g(&Gx+)
zUTR<Yy)U<~f91>V2QPhJ8>MTKsg0jx?rRfKU*){Z>+ED(J#bxT4)8dBXQ4ozGWrYP
zKIe9|^S$8CW@1Azo2VRop+i#AwK>ehfF1^!^uLh~Nc*L4fO8w_$H$uT#zAUaHZ#N_
zGD+tR3M9M<i+Sgkcnp*yOfcIGIylHkXCuS_Wi%2*M@J<$>#0ysewgdGM5o~LuD9By
z$|(cP4}FI!t1tp%1AqQE%VN0;6qd0gP=;79)cf(}4g5nou@gOAX74{Ee#Xrjp{@ZP
zQONW7*P-N(an*nH10pR)<+DvQA3%DPvkE3W3l683>$wQ!|M!EEQDMP$L1dJiSzho3
z@Zisxx8yCuGEG6@X-TJ+ct_ew`&Yz^l!1_7gR>98QZ*-7MgJ%nBP`4NAUyx7!f5)k
zn;zk|iDy-{SMe;<P5q)fu6Rb?sYJBk-Nyz=z2^uXmQhrsIg{=&NEzO;=!b55o}W6Y
z<Xe4p>`r~tG!?q3ij`YQ=X3~qy8psg%pE$pfW+Ky$sr4Svl*zxZ`L;WfN?K#qnHq{
zh>b<=+t;|*Wr4TP$QLfr0+cBOndd@mG%wI&a13upj%xkGT0!d?zMHV7H3JVyZ)|YV
znA@PY+R2m06;3$4sd?JBPs{(jF{xg;N?T-%qdu$!Qcr*roCjQS@dhGw_u@cM4@qNC
zVj-!k2;v*4QdR*!1|aTy=<>YTD>!*0Ksz(%7!RrFc)l|*_XC34rf+n}dmfme>X<P%
zYsSTX#N14MX-Dj#{-W?2v7q!|0c_@1W92H;yw0(?wpV%mrH{`t9N0{h@>NiaBxpuM
zfx=gM8Ej<7FY6e~D(!`otoA}1datyoV0f{G%>fnk?HH!PuH1Xg?vpgt<+htWKE@(?
z7HMnheB8z%)^=ALiCeeWJa9JucDr`%rWV=P+w~haz4*r#JF{Qj#-`ZDt}zW-jxN7I
zM_fK4&uiJZ1UC8*S*o#k$y-#9xiZ1)Vb>%>gA@yK9T@hAesBeTA>oIxwW#J*vD;g!
zmuPr^=gzI$?WXELr}?En^o#93{*V8c?U%mr%kAuA=i8BWQE`p&!G62(F5jQfK_%OG
z=QJP9xpL)7J9GN1cJ$A*Q|F&(M>bDM>M{9}&BRe{xNf(@^2dhk!ijg@dZ+!@|JVO}
z`;C9_58Dr3e4*XEriK2IL+$w1db=$jZk;^Y&RjU>s}F3F?%liF*0rfQCOg<nT)22a
z8;EDy*>e}#`HPRYot^V-dt;~F-#gUae&b5};VZAVuYdKM?Q38AT6^t>ue96R4DpcK
z=GIobd$21VciRECHL0C=J*=-xFi7-vd3ZC?HyNTAQW*Ts?PLOe)ChTKmxHqJ?7<De
zq{x#j;)2q-j8iX=4F6Ann1#H5ke@#6#_mwMfa+Ie2S9n@fdv}}TfmcLAh2KNB18IZ
z$<47?8~;%s=?Vp7ok>@DWTq~bXCD!FdWjIF%iL1>ALWj)tZ$*tFgh#)C7~%BD;}Y2
z1#+2JF$3mAOMSZUEfajo%BtX!S}p@z$3ZuCCEIa|=F*2j+Eoq_a)OJzuGvM&J6lA8
zITgSzwSs?_VPI2zdqNQW!;+TkwG49^@uO)$+rLYFe>Ht-@G!uy2w6dda*#o?du#&d
z9;)jitTdy;=qKR4X#Y)t(FIHHd0$xNzMoVEP&^fna!J!}V_T8eNMQQ(nHoK2An%|b
zda9;bpF#KAB$+$Bz)~SZDq|f&+KWfWDi=-0+7QKr^nEbSK}Yak@<28)6P3#>!9pSn
zIW^XT$P$WinRJdRPVBHIsc#>J9&1HM#gM_tS2hw%mjt01HD>KIkd7@0fb|Ul+gTrQ
zHw85J?kg|cnqQ-ONE?Ks?bNBwcJ}O<cKY;5;qn?)ULwO=mv}28uUYr2R@c_FiCb4#
zYkOL2P!B4K9%9DtdX9^$L#~%AMqQ{s7C;Ol<JfX%z|Zgj_F0d?GhdKKK5{bGWiihj
zYcr^K*HJz$Uawm%8Xe3xlCNM6l~r)LY>m9a1#FRdVpElkH0bn((2b}y;(=^XeaHFJ
zb02B_yML*TVhii-JZw*&_1z$rkv@|O#ILEpc=>4Eql^I%R*o3y0Meznfc$cRE(+72
zC9}$co1Fqa5M%uefp9cl9tg^V*@4+EQ5#Rij~s<32h8T?R@+$Ll5?6CrTbcN-_Zu*
zR=atND;M{)iMXZB1YerG<!1xrZ|f^NVpDV>AN=SSaFr$i<gu4P-MT14Coxo-#<1<G
zD>%QmSeqkG4OsTnc%oC71}`1rr+0RIyA$uL<_p!l5#a9a+wJFm_GjCF^cVlb_9uVs
z&$e?<JuW2@afsdY{kD5U8;Ez_YVTfoS3cPF=Iopn_GivL);5oy*2W{B>_50qyB(EW
zZ6*%gZ@gZXTb8c8_FDVD{U85N?RS3TpR_lB_``NjGPvUL^pls`+UACDL)tukLYs<h
zE%f)b_`jv}mgJvlPdvo~QJ-iZ|HP--ab7*UenRz9UH2Zew_g5H`|aQOXYDJ$`_=a1
zcfZ%(y?UkH*}v27@eGVMCA@Z)&;H-qz0(Kl{yyv;(l*uiR8#%<-cj(zDafCIpNSB^
z?`Ic}FZXH?)&J~YwMp6~wB}PxDt(QsW06}>d1)kPT1+i4zmU)*O+XK446tXGSLvB9
z0q}xfQ2Bz9e}-Hwqc1H3^&kMYT7F<q1dv_&EAK!$H8Vg{vXKQ%#Z`uWAskBHXy+)C
zwk#Q=oRJ<7R=>-29pyq1USK&c9f0p?6~4k`QMb2j2oP7g%d_b#@Izcs`N)K0lpUC8
zRGRd|^)DFlk4ghl*I~6B0G`FVu7cPNJc_RH?=2^QzGayC9ep5}nm|wzi=Fr<5cO{q
zyq8}dZkhLnQPxk&`_B{TVuTt37rI<G8dQxzI@)?Zup9%LA~b#^O&_kYD)R+63S5}6
z%=5c6bRN{Y@2iuu@ON3^3>kNb0geeMvav8qka2-AmScrmUA#!{^Pmzc)7ZxNYnrUR
z!c%jmXHy4&jRf;x<wuyU36LIn@d=x(jV(>>Yf5Xwz=n$xfm>VKO0!Ym4uxE8V3WWV
z32s~8**R&Pzl$X~Y(Nrf-IbGunt9vN!^ivvr2T`OVD`MFIhS<q24OB&CFCeyfz}`H
zu*FLi%LYR+1if-;CqmjRb6D1S&@o3roH?r22M>I;#JtuQI%4CXu@SvoDccCA@rjGp
zbEfEOjshq*jkwW>O{_DE-(WqxA_DqLuEr>%z}f-%er`l-D$*)-T+bnLRR>~DTyvrx
zU_FiKx`og5uD|NM>I>KfR)FEE5G65{U)+#IL)lvuA*ON*6%&+Fpc)05(z%=l$Y)VU
zN1}sd!8k7#OGkISoCPf(3E?dW-la;<+WMBSLcI0XyX~#F-)+~gYtyhtr#Pm~!~J&W
z?yfc&cZJhNMIw3FNy58)WGmZiUp7`9$P-4J%eL5Bb+t$bM07ERO4o)08Qku~gUJ*X
zy_;JCNOPb;=eJvlj_)^MJRh{<b|kJitRFqrKKF~CZU6qi|L?Uw`K$j<JN5Wu>bZMj
zmYjQRIBvGvH?R9vA>PuoxxLdiwokPq$N1uPo|m}qhtgZy)n<VAJggs6C)>Nz-u%H2
z+JE_f{=c*@{?>1`8<#J)?Txi|{_N@Y)DzlFY#j4uLi-TMM>jUwiBqTA#mArYD`%g4
z=DBv}{3Y4FEmJh=9Nm<LqwVIKm)m#0{FU|(|K2}nzx%~6wX0XIG(PEn|A;n{hvi3Y
zX84B2Ep3>v8N={82hwui@a92jh8MSmg8mKkUiuSljSLp@#wl=Ma|I(lPuqG!0m95s
zrtiD0dV3KHs{PS!G>S`irv*RR<9s$O%9*$DK;?<Sp||p5Y<P&L0{+M`w?HqbK1ZMg
z%lQo~(iWZlF%*^j5lSwQa@jxnO#CInRQz-VZD~EDZQb_Segu%Nd|)mI;0>8W$1c)D
zHD}|1(h=KjnPog0@YpCX>5{VwMtqWs639%LPWvcX1<U2D=;wO8zpVGx7ePhxu<oAt
z31QjKv;4&G2ajr|-={hM#=t87yl0pH=K!Uz;?bt}Yri_i-wTlI{?+M70N6k$zgt)F
zDh$Pw9gMcjw-i(vFw!9lIyMO!ag?vIFGsP*FwKRG@xG;|7OBJ{tC1IuNGKRnb9$PO
z02$9B+iMM_IRhPWqEQ}7PM~q+)$^jt#?i*ddQSt;L&IoTcw+Fem9bqjwgOWYfjsNY
zjZJP-;$=A7?eysrZRgYp--6_;34ZxTo}U9)q}J62o?&H^!4(KuXnSk@<P%%6(Q`A4
zSmCoVk|o$~!;r|BRqHJ*CQljj2<sSm8NLb3WLos)gdS}|OuO_sM-x?zZ_f7sz=IsO
zuk~ons*qcmWTNE89-9%F<Yfmqq|`~lhh=9o5xF%FA~Q6ksS}$h?t)3(Xz!xW+Btlw
zJEBVmeQ~-qOyFl|@PkbE9X|GnHXNeaRJ)L7t%cXD$##~nWlKFXu?!f5u_o*kWdwj=
z3>`NURW_zafCs4ZGPa;nuC!t|j4BLmU2;)IJX+AIL-}4daj|Qq@*H5;<V!O8<JD$2
zZr+xo@3mVu`B>YHcJ2D@cKbF@-{S;nl*7k28(;-(l1`91q+>}n+LTzfpRimPplA|@
zNB@Q;(jiZ0<ZIX2y|Y>HR~1!94ip-j280~DH~_gl272z_M8UoNJ#7+hX~S@@o!CC!
zKK1cWv@iV17uqL(_Gj9qi|5<p&pq4Lk8eqe>a1P;gZsQ4NrT)mbr`}SZ6sLaAL5|r
zprEtf*5Yr^XB_tT@3xno|7QCe|BwHwef3Mf=U2|2)n<WrF+YCkydU)TD_>6>_a@@(
z#Y^q%g-dPg#Hn_C=bXw;$!={}?yD2uQ=omWTyHOZ`#bHo|H*H*fBH{<v%UJt%f5{%
zFN{`uYeV6$TJLivLIVne26a^(>EN8L$UxHl?zCdHnFmnEs*2hi-}x*C2=ygj4O0DS
z1lq^_tOK%uYWGSL-2a6mu7lPT!iBO9_!)|K1S|?Z1CUF5o;F5S`p&Gs;t_9vPtnXX
z^A_0xy<kCJIGr!LUMwg(JI^4dXiHZ?$p%1&AAe)#TsA_<7<mmA+hr568ySrdjJiiC
ze3e=U$7ILYmacb_P=lMG_!n&wZpu!Q^IQ59jl9Lbs*f$hEMs-~2&-t8akK5<-&-cU
zA3*IyGtzz#zDL1wlLLPcdH)uEda$BtJ`k*GzE!dcH}=v3(jQd+qx3Dy2dcjjRX?lp
zl0&{ifz$foR;Qj!n20n#^wcCrV_w!Mi~~K-+@mx&C?|GOoe{e-3v@o?`dB4&7KxhQ
zrnwKK{{xj^&YO@!xnN8KH9GMv;A7e(@chE|mKOWlTUz{Yv~_I`*0mVOOJg{pdCV`m
zVI#rjfZLNcmFC4atWR=ffcY$&3GAd!F2e`ZN=A_PCW3jLHJO4J-zuv*Gk14{Sc02D
zc{@k}kt##B$KGuB7TO3pXneIp1wJe9HVIEoI<k70b(9^M^CJ`R+Vco{1f9HS39%t#
zzt@GVPbdrNK%0El7m_Zf8oN`TwPvk*NwWr`3_y!FM14JB`(Q^FIqWch-ik`vYb5(Q
zn>ch+Z_2p!o=-7=Ll2O;XWd&H+4tB?2!@iXXr&Q`^Q+V0lCgXl9)%-dlw<xnj*!bO
zanyKw6T$oraSS**Vk-Ec-MW2S8;9$@3UT|^uJXJCcwZX~?y=YIoy~+gmz~%{tsRh?
z#cA1$4wa)#@_{(McM$8G7Y9WHR9%X`L_#vBskvl=(7)>~92*HX6kK^=vtS-I91RXL
ze15GN8x93_nQT1n^LC@VdkFNaShqK}+p|wT(?0s)kF=+sd8&Qv)1Pi1{=~=HxyLWH
zW4vRH*DPRx4s^c?mW{*%7T4Lawg+0Q@840(?NASV+tRnc{N?r=|L7mL*MIaQt@V$#
z)7mtg+BvCsrybWO;?$Yb?c|v=?ZoLb?d0jx+CZF)Y~Buq{QEjk4&=96`|Zt_Uu&;^
z@B8iB-~2}V%2&VKzWv?j+wHrzxLDDStn<M2VSlja4&Rd4+xJGow<9s&9Mlzw_}?zn
z4;xJm27kOxPL(UBGU8+u%ZGW}k@{3W_{O$`emSNcs@*}4eR6DjU@6=AdS(C)e3})Y
zID<j@HT}|dGilHVq3}P$aj#F8S^7<%gk#f4P>MT^dFMV|b)+O=hN4-H9|FlpKVIYy
zU~l|3@>7q>+HSF?Jn4#;HJZd5iW%lQEW$FLd>xakz<3u$E572t>~>x-S0i~brmSa^
z?0M1Fh7tLo9;F{WRW{O$>B*Ky$XG=yY>x|M#}W$C;Sg8}(7-qHKEjNj7$#E%KP{R6
z7XDx`G|5B9;iKS#G+oJBj#t4?q8Hj_{oo!cqciyMBnG9lmK2(MIg8F*%sk9@26!vo
zVSnR+yBc_L3omq0R*T?SM~?KsXbb(@ecVA~yh0xqCa{cRTCND717jW^_;GTAiFLrU
zUtyXea5aItG57gorGn?dzCdM7K^-@@k818_qp;r2oI26YXkO<QY_23=H=2;gE!k`|
z{Qa+fV$Q_^&&8~#G=HhK?I_9x8Go@+b(zNETA(KHn0c?4q0jWLVa@q-CE1{%xNx3Z
z%!fUR4JtNGj|naoKXhd~1hT-X!cDT(XqoqmhxuAWUZ2f$6;Gi2SVu~0e6Ewu0I~}Y
zfDcdB!K`;VnOrB_6H3-19{>jXN-i;N4$QVh2Gm|>iVpd;wq~v6(r#DPgYd}B1W@yI
z8B;M3!?YMr)<!s9L2Dy>9wRD%t4hGmQiD~W>Gj8Anne|taQJ8%i<GKV?H-Xz1JJPC
z*Tu%;#?4#p^5x6z-FL708HwFpz7DBHD2vzr0TDmhu1?Q@KxeNS<hR@08>C~5L#K0;
z_xxBFpm2oQ*^v&2)QuV$^j~OmmK$39ifW;K(E2te1=rpD@)?^7(V~M_!{RGG8pnp>
zFbzljq;p^O<}r!QW1H>Ni5)NcAA9Vv_KTnU#rBzB`1y8Gn~Uw8<8AHuwhYne;!#Z<
z%$o_NF;JU_2lscC<`${@zV+z)-+I2i^Uk}X)qt~cTw%){6aB}JpJ*peoNVjceBXjM
zLiNfJIL@@0)9#?{-ni3VdH%cYTVMM|`@xGZwpU*ML3`~-ueNutyse5;dDU(0h_7aH
zRbqEv{ebq;z{DnkwyuMZe0JjUq4Kn8zT#OMid2PxDH#Qz`<_8$6(GA{OuIV0f54wJ
z?I{@Be-XkDeo>a{sgT#pS;WL&GN2z=1eJlfk|aX6Kq?4t^hAGTswcVJpuGPSceM*P
zj9`2cQ@$%Pf7NeHqkjrv6G3<@3qa2?u*e&L7Tjdhw3%3>TZU0q$rwXmPEV)~%X()X
z%6g*s9fU8e@)J44CBwfnjPeUg&#3Q%!Xi)P)<-Z4AStp@^`q1+(~P{2;`z7muOBS-
z!uQWF!%xx`O4f1=kY0rcmh&S}RwoAm>>*3=R4iZ}<td}nu?FUWVCG!Ur|U<wa6a0O
zv!LGM_4C_GZ?`Q4Ug~oD&YgDU>b3Un<#*eacW<>D*Z10;+xvbQ3K!L_RH5|s=%IMD
z6<kjmoGwBE#~I55-%7)rqdcQ0)0c0bQXRZ}PyuTvaLmiZJcqzbUXC8sdO>q5w<WRo
z_uG-QLEv^4Ue)dw)^J;rV#>MFz>8i^oH)@=p3quBn<>^FyseVJtKVy3%NrqC6R-|t
zJYez3+cUie@(J3kkpRZ0dIlmVNWt~>-~_`M?f)uco@f0-0|s1YJ49o(^02Tb$fk!7
zoaJ<xwNVLM_{J;V#)Vbs;4GuZ<HFp!#fOx#w}8CYsg#?paG8O@Sy#salT4SvKwlEQ
z%E9xzCVpqJ?_#S&N92pqq3TB;iJs^t!Nw7Y9kGq)XUkroY%mBPJqq|ve0jD)B;#fR
zCMzNBz+%{J7EtoYSOHLq^bc?GFQ*G^Y&l7IIuZ`Tx-5_tsFeQb&>S$lVToIoki$pK
zY^VNkKsEyGp7Py^8#iyZAN}a{_WB!d`Ie^LJvu!LHFo0Mx3?jB&MIgjs+2o9ODfsG
z+ndUP$i#}oKz^&<Rlll_<=B?FjzFC?FCV~7isDXgT0TxvGKv?V{>L;3a6q$JVpoqv
zdBt1qU*oETs-(RD1H_uYp8=huwg-H!{jTb&c=OnXAJjgkvI`e3v`>HLGwm0C@#p>B
ziHjfms22ZQ8c6Utn~DqyGF*X?kq)!3fZrap>zCirpn9?$Ti?<EwkCXSL<Agt_EUW4
z1-!BzfH;eD`%c@tdAq&!>W|vr`fvZc_WG-@w%a#vv@2IGw=37Kwgc^c4sWQQ8qD{!
zk-4i41ntE`KV15xjdDdIe#JyK77QAsvw`TF2+=w%OI#I9=|TUZ4bEV*IoOa>$Uxyn
zd}<i_%2RJ|EYyY_-OCBFVU$mrVo~>eD3@W`o(W9fnM4vB!Gnt=b58nbBk5PgeFecd
zN222xs5;VzQ#Xf@fF7bMFO2vi(vfYU?-vr9afTOk6D$WXt^Tvw4kmg?7jLg4Fo`Qf
z#WNs#*?-BWf0R64*f5s?)rQf(D55fnCmmfbk-esb(GY>__L8TMRNAq52U!Cw=N|==
zAG_Y96H`?BfmP)Ng&SckCk@R=SGa#uq5Si&vJZwPTnS_OD%wZ!FXNY8@P00VzhsOt
zAtRQ8emG>ta(=-3Sj)c`P5kg?nw&g=&Ob8bkmiHnHnbF2-%wgIwLJ6U`oxK~wzG4*
z9Y3+91vKBOI?+yO^RThOt4FzoCoyO8d2a8jHWTl>bGcnsxN`YQyY}w&cJ&HxRm|(>
z?;S8#XRO5s9G8_IAtMiFVyB?!v_=3$o%oIuuTR}rKkQ8oi+ARJ7L&YihD{K!SoH(X
z3M~4)8NhC>3rKT&5*sB>Kr;SpXw7u|_;KG8&)UjYCKPfuTAKz=zH%>@Y_T04vNWG+
zzU9q6tV1v|C-Q2{#=mS};Lk}i|4}FAaPiirf`w>31kPP4UJHu0+(0^GSv&^bMM`jI
z$>c9VClRi963T>=HDnYd2Y)&ZeYZ)-kV|HY1hWkj4*5>=BR<V-@UULPUjX$oAG8?M
zg)k4AUg^xMTAT5w`Nv{KXZ_tK_&f)io(&*Zieh=@XjQ~y!oZ6Th)H!kZBX_luLALT
zeZ;&hhTJA4f0f>D+q@2@XtoLYBvDeE3DZjUp}+c<e(eFwC^Q3%?!X95FslX5i(-JX
z+4(b+@}hMth0vpCu3Vg#1viINDvFWpq&Zl;7*yx;<rU1+pU=>%oO|b4TpBAoIKXde
zGx7cJztUcP^|I1@0UMob;-^uLdDlhV`7Qy2&3#TG97?^)!C6yaqv+Y{%TX#-=;hRB
z1&n25{k)hKw=ff~XaJXj>%IhE8xT;mH5)4~CFTp+eK)BClfd&7bToi778{~H!2qGr
z@|fz&hvRs=)83t3r|<9cb?~F@^v>z_<P%S|XP<qxedgzWu6^QXKGPn1;)%9>`b;~#
zxfOe87i7~BV`=c@yZuIEB)+1MI-5RZzO=x%J+9nnKm7j7?UnESpuO_a%k9;de$amS
z+N<s6t()!Y&8zL!-COP6;rnfEd#$Z)t+l)N_S#L^efKWk7U>58n@{<(HeT)@#4^Gg
zEa}c*9sRf!8hq*4SmBhQjIau-Y6V3%(nB-SRXX_j*r^zBf9OJrhSJ39t4lmZJMv6W
zJ|1WtR5{-^)p-DDI<*7oN%kZ%Ws{Q(>Vq_`bU|-Bmu$1;!r6B84|+*YdXQbb{vBj=
ze8IAgF<z#9KN#s}IXS+R>iI>xKL{+_I0MIcQjd}e@5rCvtnd|whJFsrG7|$u3yf)C
zq#NN;_@5LqIsp|Ucox27&O<Pk7uQe8wD;EyfjYjt9)vkYu1c%i^TT5449k4Ti%fJ(
zZPrQpA^W~C@-72qgx)O-u8*L2Ohep%y~uHDhDsB|HkKoL%mdIc!jEQEIWhRLtkR>5
zN^3}Yh<+KMDY=@ez1Y|KUW?{TNdPCpxdq9uXVpTQMKUMVSUYTNuD4_Bxyr!2&%NKg
z^y84`KIpiWhu4o%?&lX2ID5ajx#<_6B<7Z$tV1|~<++@HHK^bGNE%$9kGPBC=Jgxe
zTyVAGO1q`G4IFb6>lLo7^D6TL7SUq%_@IC4gB_f0#+JRk1Fc`~Hcfw;@0)Bn<PW#)
z?3`++PoHjQ&z|!g0Gry}tQ}kT)dWryVheSqUOr<jx#%EGeOO0$+!4-;RBs%z9)s?{
zn}q(5bN^$(503gxW0ulc*D&tSi~5Rz8rJ~u-f$?7>^jL>W7{%JdPHv;#p%l(D4%rB
z^vn9mmmQTCmD5>E<k+q{m|lP8?{VSHiWNzy{j&C)%M+04xsV`0VzMQp>?s^QJ@%q;
z!kH;Pr%%v^6=1d>X_sCP5&I7zk&(>=!C-k^j{-b%d0!t1U-S#S9pbN|P1_aSq5u3p
z<H`gK^Q2`k7NVY|mfZ;j2%!2iCQ!y!8c<VB=}CDP9(w8af`tX67Q1pLn+D1zXNkSE
zv6-+lY0xrK4!?2ZhBgygsJ?nd=^TiD;HPg64js~FLZlqb?7;6o(15^s0rc1o1-k)@
z=i$tHW#8zds?P``ZmJ&#-3t6FfXe0ipB6m&=b&dG$a4|7PmTe|uRv2>;PLYl3=j$o
zu3H*tIJ<%`_jk1*XHb(b?`dJrYWmobe58#l9_P+pXdn6LN7{#;`%ruOnP=O@#~yFT
zPn_rjF%Pse7;;P4Jq=cC?d*l~?Z~!9c1Cmi2%CwIwENO?2%BU#uc^Ip<yw19n+F0L
zh__#Vvt7G##e?mYEAO^Ddw1LJfi^ma#IvS@Wkc<ycH{N0x9{Bc>)&V-I&e08>D1&M
z3n8rggVXp8Un(8iwCmUmV|p2=#&X#zS~n9_9$?I;d`3h@YyzqX8lOS)-;5jKp?<~x
zB8j2;3t{9LiDx`G9hXxP$8&*@J#$C{+PDf@whh4knk<mNSnuG3O}+<Uv~lEF<{#rf
zC`3+dB3_1NHLKJvyWqV{`=I<qS7js3C}T`7%Nb#n-hWvznz?LpnfcQ;w|w){@WaDx
zq6C5;A$z$_<|Dwef@cydsI*x$$eT?(jBTr6<b}d~Q{XRWK|=?{!*Z4}qfDUWmHdLj
z0fl$I7zSW2<_$nzJBdis*u<+TOLH721bLk(5B1iW_hWuukPnpcKrgr3@UhlBe9DG^
zx$dYI+Q)sO?P&3?iDzBH4l_seWBq~VbU(M(4Ul~#X`6ggij$8l$hkW}8w<^C`M{Uf
zMh97!L>3mOaz{0XVk2kqnP(~F`A$BV%ZFsS4ajeaLrL#EjJJrGHWLEg!k8-wd=E>T
ziQV0H?Z$O&CU#xF?d`nvhz$Vi8OBFnnU`qnWDaL8PkC;8u3KQH6S(M>zFbYf|MG=F
zcBo#=rCw7pcZ1dsbE&ZAqaQ_1){3<j!WIHFRi8kw=mg4invs9xDOo_tE4nQBBhvj^
zwgBKpz5?>4r(}hW6K$&7gpa%lx;&MTa7oE06v#4)7jQ5~PjXHh3$HQ!$dB@-uL^XT
z`p1rx&bZ9G*x`byErHt5RofXjacfA7%!1r85&MARDf{D#fk80Y*P+Un>3~E3hyTs5
z3qN-Rf(7_5u!^w&hf}(&utE?<J|t(+Q0e8BrRuz#Jx9BDV3ZVZi6bpGu1cuHGOR+5
zXEX8gE3dUzU%ArWy+UW<U{k}f`_@3hhT?%b69+Pla=`Nd-d;p^Y{aJS0cEY91X<Fi
zNxFdaRF{rU{U}rD9k=kyF!4LgI5)Wd>#PX_3!4hRy-8`)u%S@Q!IpzvvJ@Cxk83lr
zsX>T=&R2N2JxH4g4++x6B}mRXtZ6_xee!fWbLw<Ed-`lUbNWm>ar}e^uw2r`a&9x?
z3*=JNKL3SZX&?XWXWQ1voo=&ys)OUYHVAM2@bz~0`i-`C>sGs^_}$Czv>(6rTDx-j
za=Ui*TDy6ZZ+hHNTU~E=cG*bm`D@n9G(020MuOX$Zu4bp4a^MW_`nB^eG<PYO8pt|
zy03hdLTUdI7Y!nuR<Lgq3dSZOy<Gk%82Nx@IU~H6ygn@A+N2)ATX^>^PlStv)HM|x
z@*U<B_{1JqlT18{s{$QYB?on@`fsjEguTj%i;tL;Lr>2N3W{$|2TDfL#02}G>jK8Q
z0%N|CGx;Lz2uorf$`_QZN0k?>Dj&=?yq}JO_tW(+3YN@zU-+Q(Kh5futV#D^1b)Sr
z<<F@t69tqh7RLj>VJugIbn!FR0F|Hkg(zGB=&7BD!s78fG%%K~RkB8a#&iq6;i$6W
zUygyX-1PeKhTv#o@+@X+t<2h#TXH!w&O(*9{N$`RSJrv%U_)~eC&QT6AE-PvV|-x&
z?1!AS0pNrga}Eo67S4Q}j72*OW*&g%gc*i0U&GIw$DGJ~$khe{i)k;GCEN3b8v4H0
zIXo-rxr(`7lY+1cb*7xR*r@HiNLQr<^I}Z_JPw(hk~iJ(%2VzD@Ip^G_bbms&}Uxv
z8D%yca-J81I7Prn7;!92S)(wfDJNT)Yk7MiP;c%5nCJX0DQ!6Rd7Zs%%R)bXwH-b9
zSq24csCwZy)(W+mh$T}m!*~Vd<U2MNK4B~Hcng*ObKInE=ppZoh<%hAd2FtLUEo}j
zIsoGsZTcam47&0mdWgd_=OLf*fhtS8Mh3;9Pa6Rf5A`kR^3oQIa*ID1><SKu4roQA
zKXW<H+6+Dr0eEKnQ_wx4Kpw%+rHA-WMUjg14%Qjn$@dcDf}a9N%bd6u^cPIBI#j;s
zmgx)nW+Du9N+Y|1Vu>s|?5wyHsF@rCqmH=&Ej$Nh=K)HGQk{>6fhN0Sbm45uM;K*{
zUpK7VYVO>*-QIZf?e^-c@3c4Hyxs2HMZXH%F?4`wP~ws&%Qs)$@EHjRs0qzE%co(=
z?x~Hs{?U#~yFMh9pX(Q}j6S$5LTJF7-oYs2@FP_N?K3eNR2U$<p;*fUPc?9%JM2n`
z*Kc^@YK8;1Y%x;G4h7DpFev-pM{O>6jqI^Q8h{lTu=!j)4^=ZrW-yfw1)e|PO2wx>
z^Xc|!Z6==m@JHI|bLZM2`St3}TkXvszusPd?MLnA<ty#Zwd?KXb%mSP+m);Dwq3qj
zeX!s54t6yVY7?=q!SMdUw3*=h9uH(9t$LS%^d8^+P+h1RZj&7~Sn(|g|8*ZdyqSpq
z^xtVBBY2B?2=HA2G#>;;o<9hb7<xSLesbNvI`KoaWxL{4)hNFNW@e+7WX=3Zp=;{b
zy^tKJl9_z^|E#OG$&`98=<+*E>be4=NmmsvLD2$3awV)1G4ec$f3_=e!6@&&bVK)k
zwiWy&wErSth0FfA13G^a7ZfdXiE~RvkN#X}vmO^qg;g+>>4(_7`WZwUmh&T&Y}?QY
z3rBV;uec}rl=p!EmSvd?AEXJK#NbgmZxj@mdRc_7X#vk7ow)}}O+Ui%aB?30^a9xP
zDlg(V>Urkans_wlaO;hyaLpOMQtnrUZYa&TAe=TBd8m~IFb@T5{_&>Zn8&uvy|iA=
z#vhe|<Z5Bh8TKrYHJ^zmR}oSl#swDeb=-Jv5&l5*2+w``HRor|!823d3_yu)<`M<u
zdo6$f0ud<1-XyT;QAJs=IJkk*>oN}FA2t)+XQFXH5TFF#pg?ctaY^olwI0e9`K%GJ
z2OM)Lw_e#7Drcj_$wF+fshys71UM-N4*Rgzb}7xefP7uWFu0x%`t4Za!ye;iKsN1W
zo0RuUQD_q;+g@JZ)^M77LMa=tL;nJPsth0Zbr2<;4hTQ^lmWKi{gQO(3yQvExE)mt
z&oZEw-vaX@x68oBW!VLYBTI3u@3@6rSf>e4m!FDvD(iV-uXPH*Q`C=-6j=$0DMkhm
zIgv$dG=11Cis|ESm*^bzjsd9l1IkXI%IPDOt}>wC=2T3?iwdd?5JosBh3q`NvIQd#
zM!6AtJdjNW5h6cMr$IQ-tD{8qY!9@=bZ}ppa2|o_*auo1U%!4+8;Z;A_1CXx5xb`i
z2KOmqk0kkQhdQ_7`nG@#Q$6@ABV9LaCnnfB@Qw_yCLfxB>{Pn!1tP;KGf>BX>eEw{
zn_TsFJ^AbUL9>YfNK;RrabQE?0Xt^~ELWlE;ea1{==|`N24w~gKYOA2qr<i$oX#ho
z(xhxQ){m{X!}kySSqm<^dXpzw{KwlLc)sFepZsKd^66*W(Y1~C&bwFJ8*jbk+nR1^
zBXQ^Y4Q(Lqs1s|$u&0fLH-Gp2YFS?qk|NT3!uvZPYWD-#OPXg|IJmGZb)#z4rxdH6
z)Xf_Q;qtI``dMuz>|f)RFB!rT3$*~fsV0DCgyr&rWx8d6@BL^1#8&_wrEdX516>8c
zbyVmr3MP4H`AHslFT1LU*t{i9FeN*Yi$#R53ub-jnen6C4+2%Mf~pJgd&wB<Hl|m>
zvJT+=c!1*jAhiEF!wUCI@Gr9J)$Z>Hv%NVAYBYFX5OH`j=D<n5MqWncRp5T`J}Oof
zgNwdErO~%aq|5cgdQU^T^VJ8(K1|w+9Hkli{pxVWeU00e*_}w47rqKA(guO&1$ZkW
zXY~*JV`C>zp7g`H+-B>At{|I?oTTy<h<-I7XWd!wvtTC8D=T^P9xun>$^~+ncQN6f
z`KYBvXvdz{v~gi>W6bbZcKaOf^@9E$@GvjqIYDVZ+^jKw?dZDX%0?}$@A>(<@Yo4}
z>&aCV)gArtOK&za*h9=2duIc}K?fZ2YC46`nqD%Av5WD~{#2c7GXeOw=RMC-58|b?
z2{G4cYy<Qw`S`u&Jpd9jq}brOk-CcxTwc&J`_4F(!Q(p^B*b;IfXwN<(sf#wQG{F)
zgHvqZ5CisyvNiuYFFZ7CqG(V0hffxF8r#yr`X-%SG2)30lqa4c>EaEsxLuzSFD&*r
zxMaf5puEb8C$|HsF*-!%EcYQWs_WMOkt6-~F?8xr($Y74P(qx|ASXu2F4(Eqc~C}2
zFVLwbd*(X9lXe4APNK#!K#3GPGJ0n2QRk-(`SURX0^UU$>O(p<sE$e>`s;rsU%(D@
zVdj097;p?n$HNkpggCAE=rA~~cFx_hNp^)9R0(%7XmAb(fte=S=*(uKIxawG_M48>
zUOX6e<Hnu#?z=bI+q_}n>h*U0=0UrC`(E4Mhh7$o-W{DY@4eob!UN2apB=UjQt9Yc
z_!k)Mwham#?1dfeo$6pb2VL@jt?A{!+$lH%gD<k!=U@jIoGIm?_dzQAbLJ=mGj@cY
z5tB0&2a4}|;O2@0_mVPb`2JGKL^D@72=b_O=6=c}hmN)lzF4XFfKS%n-R;j^@ansL
zzXjs#`SWdOhi^!nklYOo{tx_i3*NhZ=hn@3=k{%%J>kJP->b*5Br^~0?fYs3SB3l>
zMsLeJ=*8I$`VF}F1^3Vf4x+1euY9F*)k5V;hcm5Fn+c~ClkVBy7nH1@CZxs7;zxiA
ztGw#Km`8W10N&kZg^iR0qy3}{zYL|r<(-BBzx0f9z!iQOO2?zhe^Rg<#kPUi*hj-p
zL+2{M4q#dSPph6kO?Z^8?**M@WuHV$iTX*ZI>}lV`cv}PX!9z&0psP6aon9jV^xnC
zMfe?MjxpSkrGI5egIgta#1(%fCK-y36pp-pn2y4$MWHAdOIhr%uQR6hGs}zteH=iD
zuQ2d{F}DtB+~>n#TieHd%c>WVsyDREA8Xti1g#V?2XIU5ny)Mz^TWry>4yiIIRRDv
z!Y|D8b(K7e)dxP|w7aiG5-09h$Y6}{HAZJah24?DX2Ih=YXs&UZ1Q||O!EYGGM4u_
z5EhmYG66q$LF`?DbbH)~9UhUf!4SDO6T&ecpbsI;IXryrx|k{RWdOQ4E-XO%LJwaB
z<a@4BsPzi6W_dFdPcpsj6cUSG3cMlHoJ?NSqF24TEm0podhz1g=8yCwTTpL01PYH$
zZU;|*6zE@RZ`eeSY}=(jw-I&FKQ_&F&Q~ueM>aC6eZv#J0`sVkut}tdcFsb2xOJ)r
z068Tqf9@=Xs!Kuf6x7CW1ZYZT3>$Uv<D|>0UR-5UV$x^d2y9aiz$|wWvQ`sKmFvG`
z6Tp?<EQ>y<ziR8?f)C6(iof)((mw+>%r<kZ7Juolwi}`3SHL#^V-uh#>wIVvpgHu{
z{xWZKiZW9vl<-nK<_o}0&KQ6^j;rzv6x#83xj7(;4!z}MWYVGMj!7sT-dIR}Ri04J
z=9heR^w2$MyIS;em4dHeU)6@<>a`on-)T2*A2h9Z#g84_>lYaenCpJKL%(&1U|kGC
z(rf?3_AwTbqdGu<Dw}K-fU3c~pk&T882|&tn~ok-(Z$aM?DZKU)xS(K&mbG$mVS7B
ze`)s@XPLIP`6Bjx-?Pf4ZO+8(@9z8KcZcq4qTxX^1tcEq@7Y%j=*N$r)a0|~cW(1t
z58f$_K6^{;d&Fm>)QL5K`}qg`vBB8g)8@n*iGEpDJT=Jp>=QEF2Er@gVro9NKW$U_
zq=T<ppcdO8{K-#EwgBK)LHV@G-;Zt?9z{Pwu|7h?1TNQ4$wcQEz{@2M=~U4p?Rzgk
z&N5%=8Fh^Mmvt9Z`5zQ)PnT8w0x*6B{4{i~g3|ZTOa8wsC^tUBod&90)h34az5MU;
zWLL@^UHE?G?**l&!g~SbA4G1+S!E;Z4%UM1i&=9#5{jP~9`kD5L0Utk#$%<0({Qdu
ze!uNUL6RNF^Sy(eos)jx_~gzBm&20<smPFfw4Z1~kbuSI`ue6e3Y(rM{6MqjInpew
zF^F4;c&@J6aMk<;6n+Gczp6Yh1IZY!wC0;>p*>byKgb{)9B4D4a`WIkiLUvK7-P_u
zGuqP8=LO9?{`1>v^v8VRv%LK;g0gsGeie(sat`%}tZf^!ROTrlcSOJ{DI_|${(Ww7
zx^Nw;K2=W9^2TB=3!m+z2pZc;J4%L8!OaV{_y`|IrhH=wsT=d2PuO|`5ndMmrJsO5
ziG8+Na?)P+{IEE|8xU@T625E%iXR*@=rdkxR2vD3o;)XLy%-mO_qxOO#0G$N_dKui
z(KaAFy>_Z+m0>RdFeP>=D&0WINK|+-!2{3`%%?RDKpNSli<qL=l2CkUlkmY~IZMF0
z%6N*127aJ;mhsbp(s2{L1wGs@1>j3o<pZjZ{;1FB<Fb1UBaGz*@QyNw38RcdfBmo2
zW+ELlVTO@-glr0`0*gEalw%mBEIi4Norq8dBIXb>MQ;$IG`d`k=!IoJsDSXQ^5{_j
zVo&6Q;g%T=GQL!K^Y)!~{l?99<tm$s+gi->Dz|&>jy4y&d!p9=+A$vHT9=-5-Y(TI
zhDC+j8(XnysCU(M8Nd?&s4*`#1}G}qKL`{jGUO8bfWgtb|0DAOPGwx?ya=UZ)~zzd
z3^=x1lfc$C4_Y7b6$d|b&3E)w9R~0t{S{+41D#yv<_%JHnVT-lV02sly?yI$e8wb2
zo9WMOt}Ji3$fei}Vhk$fi}GKE8K*Yy0lDQ*mnrxe35DP~PY5G}A52d+dHwY4B8>Q+
zh;;%IRX)-U(5o^-Ct5{2!XwyL!bs~wb_K)_JtHU{>?<2r>gdS#v&UUwSx3ozAKXt9
z5L@A229zKE=LXgO6@O4Ddd3KYd_%Hav`XGU`kr(j)XnfNhY{XOPxV*x>DcnnS;MCy
zh9yl=&OoE}pT@t$Yy=Wm3};~oKXN=itB{xzFS#O~+dpbymWPcsl8TS#3QnBZ)&^kB
z>AD>VrCdeuTYtEctZ}sT9o4+W2^$aEfGdIFnKylAS;k;TEePRd>|`vaUcPOpkEf_3
z_I#@lKq$YNcZ-I==$txowL5aA#k&6RUmYxr`2=VIu+XJ6zXiSmp^!ze%3Lgd)r-Gw
z&3qgx;kiP27PZ8gU;F$g|8llH^M&dOu2%+JKrtvmj{+|4cIpSe*b2$C4E;v}b2AIL
zUO&loIXZiO3c(D023d4E)kF5VEex?lEb#++?HhDZhEDi%J6he6R=QKi<OzVVlCSnv
zKJA#c_N`KGQ|c%?_BCI#@e{844}F$99Oy%}2C1tT_`TL4sbjYu7buD@`bX?2m~G@I
zHjlPXJj#3hV(fqyK$=i|MU~ii{grC+@FJtYeEMIsF5n`|?R8Ta^8jnE$Tc>7S7~se
zjO<EVr~psZ)fI2Pz|;59Gi)X-r}Gq48Bpoc={8Y6S@!!XIl$-(>mKMWYpP%A3C*GZ
z{6G71-b{#8h%k^YC8My#^kOGpMPEQ!4)_{mQVxC!<Z&@Hl5IJih7QbuLnr6$29#B&
zWp!Up5~CX=(9>#^oT1?A!UMknp*9q|yKE%(+O=!96z;UW-2=&Gpg(3kERg*G7<wef
z_UH$jfQo^!&bECi-xK0VvSNxH_?uY_3wnk^@|1ukUAyIKX#6!cx@r0Av^j&%PCRnC
zO0dPGrD}57n=2D*YfNI?`)AHcZUzq|W<zpB0bujR8z#6V%7Zz2MB{gB@9kMuUX_<G
zSmP50G5g9A`0Z`k+usAC(0&VtWa>ZkowwSZ5Ip5YW1*296DM3Zc>$`dV4_WAo=N6X
zyh>gHG@Wb#MwVs%lKWl&kLMi0Dto5GvqR}{+|!dD#T~U0ARA=^lTL+^xAH{?K)Xy=
zuq<yCtdiyQQv5+-lvnUx`X2=!#P;`t-;az(@%%yT8C^1_9|c9T%r^#LEC<M!d1KjV
z@-olwqdJy>KUjxr_}92K2j-1T)}%8gGOl?M;B?Lu`)x!l^to(c=c=67i}I&~5BfZd
z(BDMlMLw_TWT7s4Ua7265FLJ>zXn(~2lFi>Tm^@gw+wPHXO3hJV!_77gEtuY%C7V?
zX4lz&P5|9I*tgvEjWu6g#$Ji_tt$r_ck2V6Sxnm_3RER^_WYotjG^f_=tsW4N0b5A
zJT$_zpzm@VyO=_G2jwo3t(^q+j<R06vFBNElQ*8m(7zYLeLP2=7mesZg>d#$&m)|+
zpdD<hQP|JS6Mm<98G2BdTblt@8auF?ob;9*{B-i3^>5Mw^l)O#8>R>|uXR90-7KR5
zX#I$|c;k00<u+vc1SDw}M5a7-i45%2ys8C0uK=wLV(`_c;N^rJyx9bThabHJgDZmM
z=yY0{^s&@S8mx;~v+r?xv*zyTOr78%?ahPp0>xP;WV0MPC{=`i#E&A+{nTx=$r{T}
zWfjV&%l3QiAsyt)PXqv3>Rj!*3@SBW$p!4^LB5C(v_O-#4&MlKdHC(`jt`DPli2Q(
zG4c5?DF8G9IlMRk0sb5Kf)36S6ixEFuA)!-r_Szg3bUNzDOodLrOP)!we46AK;tVD
zGQ))(-<3Sv0V|zsCl-`oV5aXCu`J2y2Xr+!5Wes_<z(q7Jvdb}fC_Ph&vL7vcEU7p
z?Wh59O))C{aT*-J&R;Q|m@jJI_QLkS4`<)Fe#@JSYu9f2z0P}k-2cht+Bjc|+F>2+
z=u>SX-)g$;MBP>=%J3B1UIM(?dFE;j0BIK0Gn+vKq3^=;?|eeh_dH2*N^@{?=F9hB
z>R`-(6FKQvT_(R0vVXAW2d*F7KhTCEXV17g!!s5+_;V1zn~j7yG2C`UUQh&|*$|4&
zTU|z_4Lu<F3`f_kz{F=AhIwG1{mO3|+bm1}4&F!*Bq!}3I!jQRG7DdV6{rao%DQ{H
zNJCZIqC38$CKzQ*Wzf&bl~C?mhEYGz%`z~5z(;<ho01%6Du>5`9|84PkUORafV_%a
z=br9j48xJ^-%*q?ldiF(gUUu(hE-%sM37AO^s)uE6*5LJOXnL!FZiUxDpB@%FDTo{
zFQk*4U_nohfsv=+QM8q&d^`k4yni}k<Q-v@Kf*Hp{pcQ54vb~*M@L;)@RPw03<b+P
z_^JB4hPBWr%~-@@BjY<~OIet+pw7vZe4B?wK7m^_SZpyqGM2H)5KfDKZvt41Ge@X^
zP?IKOX>J?hHKV-!j*})V=rx9GkUM-RR~1;y`)ohFd^`;sWn1*Huqt_(3mKz4Zd+e&
z*I{a`@lV{6$2gpqW<bm0t2PSxSQX<cb<A^fY$j-VZXflPUp5??8$CASFhvZ#HCB$}
zwn_R90R6d-K5W<$8h(^@8hm6hD~{k--3vd@N&+Rl7MMCO?I-70pzxF_Xe75D{-(_a
z`s@=(yHwi|BfvmU-g5!8lRW774=qsZ3FKt48a+^0ulrBzobJQNj-(8x#ZTU_F<()?
zzJNuTAZrKuKy*-l!1cq<Y3{0z>p~8V>=vu#RvpDp|3i_;CtT4JS%>;7JoT1J*9SgR
zp*soaE!pY#3V_G+r@#weu1u)U`<3#2jfIeD)7R_Hi(AM~ysAuXh}l$3dhmtH&^GJj
zC%Q?G^#;;U7NOQ7)~QfeQ-G?E<(C-rSEAy=i7xHY15D;88hqD<jkZ-C0Q1TQj&q>&
zo4Oxc6Sd_lC~1xICvkK}cgWI*qcVEO^6+LX2~<qV!sfI9Ij*bH>EnbPHx#A>w$P_7
z$~O5Nf2b(O=G-o};jjLsUsaq%rkw#&MC6Id7gV}X7-J&@=$^~dsFeF`kN&D7j`9Mj
z5d5~kAH0ajfd)P_@RFy3-b@IOt{l+0-?8{Hi;%&Z35NqsgM^6a@Jt5q@vt`c`Lns;
zxr%q*;rkTYRNUen&va4_Oxc6b2VBCH1E)Mz1*l`{Rvm&+HVX%h>~Q<(KsUl=j-~*(
z9vv)0F=fW<$No@0$4-cgmn<0S%q?4<Wcnlt6LH$M_%bjMbJmM84+5Gzxt+<hWk4A0
z*wiaFRey2k6iy`=V7Z(ft1}(6rw{%vi(e+j4CwTwK0v8FzDu6l(Rj;K)z$51S=erc
z(W!C|Zty!7dfu)@sYMW1I>l3Ll+tlcCU;~3RE+?xAapa7d<0SEeBlI@ED=lW0U^Pg
zb>N+7k=x_aVY*IGe6(5Bg(3nyz&69vji@}(qY`Cc44mvxdLmZ=Lrj1^nv%i*Q@ZmR
zYI5!{%d147>>i<t7bFyskjD4IqH+BOaMp#@JAg&IM@YRoC`m4Q!XHr!R6Q-63Mn6L
zaC`Oi)E3E3GG=m)3(ED^=_sQgnx-=reaTsdRq{vL!UKh$r->Ox{)Yghg6pK9!(YdB
z@@{gaYd#<b&&<Gg3D&jH&v@P!`zp`E40;ygibW?9Zwl&5$MEE#RV}91HdMBucwPCN
zIb#vSMuEjPS6nj=cAI1rX&ww+-&pr8n_Pu}&x=WnH!#K=cv0l99`Ek?i~PLbpEik|
zyxbxm6k|MN5zp9%eI84B4h{LTt4^pe_tZESo%8(5m?sP68*Gv-IFz!OV+b8{YU)uw
zg_jpx)SQ^{l=&v}h{rT@1ti9@{9Q^Tj5dw{xTm|26yq{XE;4wNfN}vbMRJwv<b28L
zbcMrjRR)ZUlLGXKNdK8bz^JMdjQpn4oIo5s__P;|23K8FUUQFrCRx>g;Br*@`T=cO
z{Cj>%d%-!+gIVO(cd!aaT}o6z+DCn*+kw7(_bcVDk8O=1;0ehGhNSD0=R3Oq!88sL
zideDJL876L_2N{Z%$o`H%kA`Y@Ub^Er0Hi4Y;KjWP%@^?ge?-T`vM#%Jb6t&dSXcQ
z&pM1zh+mLb_E%gs5Cg#uVE--h29sYQ6WixiqPU}r2blOfUBQ&M&AHt#fLDq1Ym&KL
z0pxMq!V~**5JFyRp`kwj(UU%DTlBwZN$^w1jGqDln9{|%sp!Dv_?o4G2N=gxpuBKu
zx8kRqTc<)t5Y7MgKlwA3Km$i;b$$qgdlW#2;E^WjD36>BD0!<;m5)G$sI+(#!3PPQ
zI2)BRHoPv=4~;P6(wVAW;K;M!1d)wI?k&Y3IP!o&p94T7oOR-1E(ZFmS8uep-@e?g
zT)n0Z2CqZqUCR$Nf!uGqdpu(yeK1K6`dFx1kL|EU{HV-@!O>5Y#bl`$e|QT_6E@RP
zG-XZK2ic))_D_zNMf&K{Jf;bWs}udK*wh}-$?xP9qd)DKPJ+L3KV&*fhQSU}=Cg9v
zIAJGndyweula0enM}BHYp{o;TU{2b$sf;|gn(b+`r~QQS6!@$Nn?czEeaXikPU)v6
zIP$I+`UsR4sF3_5K|gTV1{z9W`#7E`rrb0=L!^@448a3ksN1EN74#xQ7ZHHICkw`6
z0KV?l1)U}DQ2-6KnQR9-6A?J$`4Iv;3Xi^iK5D^UHO&N*ew8oySb4Jyv7r-`4*O_g
z3aASjN}j_K4ZLMjmGyCV5?AHDC#0Z9=2s}7N}(HJ&iCGBfG!iP`iD3tc~Xe!ybZ4M
z1@8MjZFFTij3~<|r~xdWzApTv2O@#nHOYti&oa=K?vp)VfGv~%$P@I#AwZb&`pKfT
zV#@rZ@l|ccqMt>+&-imPLv);p7Xj-D7WMuP0b>%h5}w8WT3ge?xt{rAyy7HCT`}N`
z$8|yk85ze|h;W9P@eCs==QY{6`dshZ$4;(5c;i8yG1;DC5yA(-a;#@;L0ra20ys`i
z@bNHTxs`L6Gw$t+UmeXG36HA^N~4!e#vU)1(1M<YUFy+qLsI>A_x3!l5rFtIy2_VM
zQ7`G3$4A*Y(RO?37ap7_PdXcjepZ^c7LLUw<m6G?o7!>RMnT7*rK%8A0sY3dg<_&9
z@L~|#Y?l!F;nv`)7m#t9BI;BEs;tt%kF-W9U-c^pC;c-zTwhhrton&f>KDqpOlat!
zkMux|7eMJ<-b~PrBrHcEZKpEYM}IXB!(&=mM;maq+;q}uJN;LBY>{K~qzWv-zGQp}
zj=0jHfM(<hfIGOK?{A^geh^gM3T&%_t$;@Vq#Zh~c~b?qEilR<U30d6I+i}hyqyg*
zX=3vBn`B}aKK6#92lDp2{$@$Bt#sN<{fwb>EZYE>SL=jI69BItCYj(UOZ`d){^PHB
zDhA4)ikG2$KI){Sg}mgDndKuZ`UpmKATuN8GN}8n|0Ao$hADW4h%K3~Sij9znZWC(
z*M*SwtK$lwtmIMVb~HfKWeB{PfEeb22qGUPC1@F-!>AY%rIm)>i+CSgT9BwSQ8i$u
z4_!H+DnkrSZ6;C{TRN+7v%U;6X@tW0^`fdYyx8CcvN|n!mml!)w0+@)bKT`=*>V4V
z9wz3BMV^`9RkQo;#?4!P!1&g!I}<)~xU#`z<vsbLPfoJK?i1|y4T$Ct(ZG|-`luv8
z3xKDMhzqI=AYJs)9eSys%9yY@i&s7`VXCI+Vi$CrEvtjrgM`S^!LT2B{eV{>hDKB!
zXhk~atfPFMtYXni8?A_FL;0P9*d4O-7_Yz5Uu{|a5t=)9@3p(~Au3DlDwz8C3P74A
zB0~w}V4K8Ge8wqXSE5km08fuGi*D(*5)N1Q@has#Mh*ceSq0=rTqj$ADkFXr{YX=E
zK+%qT70N}V-wR3>Fs4f`V43<Uxhof{JzmC-U^z)nY3R#VVB{}aplAzL=~?D2ey0av
zqx}PVqN(;QAUn!h1>ma;`ceMKvx)`)Kb_|qh?d)tm;j!beps?fo`!~ny!&xa7q~se
z%viFu#kU8PKE_6XSBxGLj)#J|1(GWYEC@28cs%i<dA+S~Y=|#c2{=)6Obcnt)*oXE
zx8u|%f&~QQ3m??tc`NSGX50twul$RjTQAYU*kJp*BNB6RfpMGi)QOWF+`^g#Lm&6_
zV_wx5Io|9k@3fd}Z1Q+yyz(5_im97#`NU4)7@|~@jE!t06aaWCpf>{!>BOYC5<)!I
zfsl<3VZINXm~oMG(bZhZ=91gEXhF(7eoh;S&=(+we8%1cLzf3|LJp3Zjq!T6&t=9{
zeMX2J>R^5iiY%(*pDbXO`GO-A`30UE6aeF)>%OoL2RX?0?Ys)LSoNk`B(Yd9(AT2g
zl0sx49~)vzLTD<TwhXS;Jh?4w_Q|~Ipbqo82%X58eL?*?kL3%G-EN!I9op(!@X%&m
zoI%peTVP#%A?mZqic4>mmp#CQue8HV2cBaFn{qo}cR+_xH^8x4ZAV*FNIg<-o}Jv?
z-F4b+g+2CRAa*e3totfqZ7`}H6$(g}9G7*x<rYw;zmh?juiOkYKJJEN*nZ$o>5;B<
z7p}?+ZyQK=+=yS^Ohm@waZ39+GOF~JoLOG2omSaV0a%r$J@Y)&+|P{fx&g`xKLWV&
zci3n7z@h)$|N75bN;wm#0SiXr3I$aOa2!|#RjI02&R!8>AYeISw}@{4R3Qe&DnJg$
zM7O1&bOR)lN1rj`p+o23LnCP0!kH;|Oa@gtd@eUrZx1q*iN=?*_wV@(QZ^$@e(Koj
z+#C$`>QmI_R;j+@79da9K9Hv4?%v(gCgg6rdF!@sr9$tT50qIaof%r6RYJDp`CeDm
zp@8}!SEQ~J`X6ZH!@a`Xx68ybZ4jz0?Vr|e-niA?e*4{a{l;#)dz%6_SDgLD&d4Ul
zhVDlDOd^HPATsndmGvtY-5B%qge-KKZ*S^;0bmC|jUCv=pV|Wd#h3Dd2{}6iedzFs
z=XHfaWXc8Ls3)|llJQjr#8*8gKa?5Kqn@0g@{kj-iK{)@e_B5BX!5@1ux@--wi9q)
z0>178t7K(z82YMY=onwLTR09ccPjz1*L=n|{g}6G+_|H^EE?BeAd>WZ$-$Q$y#VSW
zTJ&+%s?(Kk9u>MBY_i-wD_}5trOpwceH6_o=TY!}@|VkzS%Dw>SVXQHxDNpEnC%ff
zQ#s)gLMeYgD4LNLam&2q=@0Zr+Ozr@IP%=0$QXlRIhHg2ZEUP-%wddROjJKQdc+I%
z+>X=V&US>c|6beN(#C*Wba=zv#?fphw%5Ix*f5@PfWDD2k#TZueN!PXZ(v-dzwi<U
zPQb9hhZfthkFg5dJbrqSUJDV?h_B`a(jFsZ6TG7@?Cbj+Q^N8gn@+(oP7yNaNjfnL
z=UQ;r7z9ukaEx6ndU#pKec7_EgJol5Gj)>SjQ_DhM|Aw?u`oD*Yz&1~di^ge`vApX
z{?k6Rhvz!Qp71%~27ty4l7lY@jQODFF%+KY0%DGT%uicne!-?*hE7Nh$+u+n6At96
ze&l;7ocSi`%7hCYN#T^JxnqR*u4e{0vwX%v7A%%%g=6~w_NVkB-<wnQqw-1BEg_1`
z1~cM}5g=Dl^kO$-bEVV9tP_YkNv8$GD?NS?I?@A9(x>3V=B~f<j{HDn-R|m7{(`mI
zka|E{3u$Bmi0oA|Ec@3wgra7a-Q`n<$rsRc#l|}&eD9Y)>fW#JK+vo5C>a2oI?IE%
zw*|p%q%wfxn&*x14fFj-R{@JS!8@&w#P2nl_02wsp7~`w^nq+<bNrMoWvi`SvblhA
z_(za6v8T%kN5uiUdC4k0_zXVMBVHxWM_%#G_5qIi7)4or;Z?yv2k&UxqhO?W-2GCO
zpd8;3d%Wy)<cl{j;{l6sz_Q*4W%r@~{eSwaLeog>I$)xLb0;mqv!ECh#T;~0%0+^Q
zD+gP`%nu*-^^^x<%Eg^OX!P$6ZHgI6e#tX{Bg+Ba2LKP6eZYq{i#h5M8B9z*aFwRh
z`eSaINU1yDr{I<+=qRfXuJ8~r?u)Hb*AEEFxVM#~U7B47z<NZ3EG8!&{ATlkuN}DD
zFZroUI!O44V=H*%n5O4LE4;{9*mE|^uZHD-Vx=oQS5vsZmxGskeEHtSyO*!Et5<G%
zLzGEHb<#prNy$hcE!W^mtHOui0Aft=xMP2^aE_he0e`H>e6hQ;PI0ePhX@=O)TTq;
zKvw#(37abq$w)^N<N%)o<=9Ew*_=if-P~7+GS-D&#N1}V1NR-Pat0gDl+qU%xCrS7
z2lAyK{t`2J+Or;duRg~I0yzuojU&peUwubBv~}7tx6`>g7%08zI|tg3rZ4j5EEabx
z2w9Br@xU9`Zux2%4_&6wh~!JjJ&+DQBuKpo6w3rH5|6YKC)V4^liT?yRX>xA{TvhY
z*SyV)kaoe2yXsHiZ8P$ygXKemt?J)iH1eDai%e)-m1zUF+<ptt@xYLL2&xQzyGqcj
z{=Q&x{kWeP69M=z2hRZiej-L1(&#pBnGvM?Y&1U=1n_YD5XyGrL{HbCKdKA4&-rXM
z{ZVaAzp4J@{wN~)0t@`L_1qG)!L2kB;4ckxMqgtAV*<gGvh-x^>65Z(>Ff0QET)8G
zJQ59kfj-M^I><}EM|W)WGYFEyVm%N0s<9=51vi@&D#dZ3E=&zoea3}4&NXHVo0BMV
z4V}atgZ?xh5AZ=3K2EN%2<ZbBC6iy7B?yeAa-?+UOk4V97eX#c4UqOdKZREs-{1?b
zqTjiD*KeQW7~qXfY$mumkrOr8)%6Q%v17B4Ryq}|X8>jgXh$GlVFG;6lg7GCK}h7=
z=7~Ze-{GIC%QAq&EZm&&p%+MsPf{I<PiphvI;gA`8`f<I-YL@*%sOd;WM|s394Ow3
z-G0_1oa;eCze0h9F6E)G7@R|4{1Z8(7i@%I|06ER$f&w{T*C$x;wuJ=tR2cXVI0S@
z0{6Q@qOJ!agG%Nh&{64r&zPV2AKD!M6Az-h90t@;R;3vmQvLK1rv@XTE3V8x-7W@)
zE&~ksY$jA++2t#8op!Vd9D%W_$^p{&5P;TcY}Q}oBahPXlpN@ogUb5kGbSL%^DASQ
z12&*LS1~=#iDt4xQKc(J2I;aNAPuj5Ab>Kq<RppH*wbYI@T93{xfR@h2YlAI92B9i
zP>|yf+Kw1?xq$zSYgMO38O1o`V~fd^kFR~pSMN7t=$!hUd8%yI5%qv&Nsje)dGH_l
ztA8n9z@~%L0R)sVLUo=p6gnDrH9U+Ls}DvXhB!Haw3sG806L-zC@i8EE7xr#)60B`
zPw5;ep)qa$DHT}|by-a0s^j54X``Em=qUCe;`$@DI$?Z~1D{D$6Rc)f(;!!!(3=`4
z`^f46ydRL>KuE52x}!xlfb`heeL{9Zii!gp9X`MkD1$x=j&xeeeK|{uJ?g_E@YYQh
zp0~X)Jn$PDz_a*cqK<y(DdQ}H58}R91bU=IcHX?#9Cfy@`-9uV;+IaWLegZxBu$9j
z{^$||gKVdqw{rO{B$9z#{FFD9=wOdy%r7Unocj9glw@$Wlmj_+;)+ow0q(m-28#v2
z{u3Po&#^TIr(7*SENxwjFD3?b<h75yS%m>7eIgHvD%0x%zc99iHjhihuNZk*5N2|K
z+I8qO`t_f6c0B$iy{~p)aerTn8j5@m_Obp&TKVX<+VhSA3pxlg(B?JDY$iC{g&gY4
z8DiC53(zClL~OO4QzzS*)2ICbEH)L`iU0R@^HIUtRL~Z?!m+?&!{B}=ncxZJ(E&Z1
zSFY^s?<vs7dU;J;O45DY-)Jk2mB`8Fp0x4Miw=6B!iT5yLm)$Je5HHrGDQVG1gngE
zB7>UJfp&(UcsZmoyG&>#Pj1#CGvi_)u_uPZYWE|wCw+wlJI8q@L9R-&xumc9il>j+
zEM!@OBUdQ(XZ1tMwKzYbMQqO2Ne0gvK&$?jX9|!*UJHJ29?+M??UBRU2(YPOfsZlv
z^}cM$d+j-y0Uh;r8<me$7x!W8R<K@A+_V#m`FoOk*nuZ^S$wmZu>a$Q$dtoNW{pig
zIZ=%!I`;`RFvoTl0zyNh1Zl?-mhmEe1Bo6BJfT5jZsQzNM4P)9fdR`vsPJhv^w8Ik
z8NcT0G~;4UC^3ejr)~|c1s>ntDf_AqkRD@qEW>Pr6@%`)Bub|3Hfo^nknx~+0@&VF
z<80?C^_+zo^9-`gA2~T5L5h}UG<`w={O$(>auipP@}g0Fy6?v|Ldh6y@?6$w+;Z|k
zLHQ=IC?}$@KfyUgDT^HJHeUUOGSezwy#7rz@TRVeos5H)9frxL)XTEeWw6;@2M8Hf
zxjc5np78X3t31K7k>fyDnfp`m0~N0V^aJos^yqZIiyY%zgkD914sQ1$Zzx*@g-T9&
zwNc#~0}cMIwkjR9nWz9%3{Tb<;+2m<i^}#;R>4>mdMr}f>oDm(V0=+bnRRPCB4j>i
z(~}n+_<`<jha}qml2;Ht1xbT)Tj94}4qwIw?3N5y1*^j~$q}NbiWO$_2U6$ij{x?}
z<pS|_cM1kP5M@qS=QQS(B{P)G0JPo=vStLY3gA!srQ1?AR16P+{y+EKQ6F&vfGo-r
z>;JF)<v;Ht7&D>7m>iH1G^WcDp$l#X4CtB!#Y0q$oG=>)%yhj1f?3A^az^nLtt@jM
z80A-9l*#mNfOvxG^oL#Px^!+1l1#996=?+F*w`<7gvvrklc+*A4dr~O>3A&8g|{v`
zJ&RB^pgRJxqythh5L-+4TksT>yxISlQ*GplV<EsFpV%+r@!!CUq70Nh$e(^&l@3-f
z`eYw53qls2K0}~<3@qsFf^pbq*}{pA3Ha__Ub!0*SHz^HA3wVeq6ftTncQQ|l_BUX
zV>gRde9EGS%|I6NshitDGDtIsapsVTep7O=1^s?~qi9%I`5tcW&lU}aFgRg@`B_j&
z2KHeQIB#r(pj}O)L5f8@cxoV&ZOfI0UQc}A{|w@<#j*V*2wD8qSrQCv1Z0Q)00siv
zq1cOUY?eIhK?Z9a{75_Tbh_Ahuytcybya<>r_0XRLHq1!(YNP3_2Eo5@4ffpc9W|e
z$1I1onfn9_Na~W!2{w3;AJPE7q-&#=Hdg&*sOj*h%H2=V0j(Cd(`3QfZPz7wnY?)J
z>0qDsWFw?Bc*}*CpfT$n{*=frQ$-m#6&G;UwfcJ|wX%bWC<}P{5sr~G`^Q%Xv{?26
z(~V^=>q)=t8)1%RmDj?XKH&wwx|&}->cu=~@zsZf<d&J^+sE7X)|NM@^eyC}Cu*cX
zv2o&KGjZrh&h{faH1sFgK5XXkH@;(0VoN#x^v8z6<Buxk{-XA0GeO-vHi)`DjuoG#
zj<~72DD^`LeZiA@-iyzhm3(#(ei2S>ZwmmVXI}9HW*LOJ)cVwDs)>pj7sFG!X{Vad
z3A7g!w26-cXu3k5EmLe&3f2D&u!krywg}vx`WZ;t5Whq&e!!32mk8|OjiJYa_`Se=
zs)N(0=z6LQ9bjxl=S3g=04Tj9&`v;=jpY_^T-uLpHbv#1ktepG9M#4Pm!^R}6eT^l
z_ybs$lk%cTxqfhvNormy{({Q$oQ%bFD0v3sv5!RbQdaX>;Yx?=E;}gC<`;ib|M-P5
zg|xJaPLZH=S6s(3<4m<{Z0a_fuk?yue;Ly#H?h*OhqT87+D>I&59m+Qx}1(9nA`$V
zAV?mFd}7KcwA1PT6j>)&;j)<kN*}g>tNsw5*l+ep_-Hr&ED1J6h{eWnGm)_>@(Qw<
z=ymtM{@f<m!@(FG9Oo_*omJ9T=|6OFA%c%`L}6#<i0Bt8a-b2^F;p7~@U~Ng_Wc0G
zQ}H5TVCk-M*{P&`)-eUpO;TX3#1vCU(y23btUB@IKc($2`Xs(bh|7RQTVgW;psRYN
ztRTKu5IW;yu}2raq>w*oS)VNS^91m*r`kDU?FWJ}BTrA)W&+21@d#HZEGQ<R$`j9u
z&|sQc6n6zlXHkgZRH!O20`jJVujfZPOOj~t$g3E9ep8m<tYF4FCHm~vjP`*<ne)i9
z15?N(<1!U`;fP&b_H@t)`1R&NIE$vwxC%wvk&<H;H>g075Ad$f^qt2tCOjR3gDw2g
z4VH<O1xFUX9XSJ{i9|yv-^OQ!n-@dT&w!AD7vD!Nav}_>`i0sHM1?Fk?|CD@?Nxat
zUOIGi*dbjHA#F>b+*b~?u%d`XC7TJJErE}P7=D_49$uv2KKlUwDE9r{igU0<7x<h}
zjQ`yKe70W2u0^i{OCVSTUKn-z5<){;dXd*BE8Erh3T@Y!k;k+uupbY!z{B<#)EC9)
z$=-fJDf-wbP{&Ryd)+qn-@MWHLRteW8zw#`#M{J<YQe~6X;T}C&CM-uCb(sfzTx}W
zwNYc?OD*^JwNcyWEoGeT;eoijlF8ZT^hfu#{>K}=IIDg1m^K<(1oPf|7N=Y}sxP8W
zplaYZ*UkMkeU|Yc?S_i<+4KhG?I!)<U&$i&<fa(EF@d0mennl4yU&Ir8x54=cS}c>
z1bRVheY4)k%0?v{AZ>D3Kq?K;&-}4Q)iImgEVk*BEcoj-NARwrbXLDk9cV7DB=8&o
z{(>*x9Dt5^@~IZ0Y$$lh5MRVj)kT815{6IH2P9RFO8aYPq8M_@@7>=|o55~=`HMDD
zpP26HM+6_Xl8{`f$Ry4!czz8fvS}W4()YzcIYO&W$f%7Q_^!^tc%IFS@Gjs7Ng^OA
zB*$EhTY%#dE?}xi-Ssb99}Let0F2t<<p=9zzlzewB0qJNjqD;Lszc}tpj`&{7yGl1
zSJb61!pDGIY|1zgS#CR}tB;vZ;A7M_i3t?6l1s*-Bm-L8-2-{XlVyBCjoU&-b}zFH
z1Q|k~MmISYRvi22g*JU^mS@A1bzAj(`l3Ln&?;{~_l`o*pc{K~94J5|pu|_|0gikA
zL_gNcMds6@#V0|c%k>+8zF~<rHGDwyLR<1uR(;g=DA=Fs_wE~R6V;aEUdIXSr7w1o
z3NLtXFog?+^rd>{1EzZDSgW~Vwkv%Ain66}z@!(u-1j4E1ZYY}>NonX^bqHG@%X9o
zvZJ8V#AOFivWf?5XSvZx<(*f?{u|bYqpV4X0(O*~l1-RsH1^HBtoLLWn+e6H(SME|
z1+tdITCBGrV>2dLo^Y&Fx?RWxn-h49?X1T#_h+t7Ta;};*#&qUQm6C0B|Jqg@5c}v
z#aA$K3q7UXaYF~WB!TQfzY983FLYNQD_$UW%=D3#HlIVF=vWUpsW6||2bZ-S)ua%;
z<*U+D!F1SOf&{@;yQC~JrJ-4Qn^S~DKtU-2ie}Nk4!r`97Co|A=wY}!#R@&Bs>Ysi
ziw7X0!P7aAVJ7fXg;QSbVZ7y(z5Sq4EUX6&aMq>28mM8f&TQR6&`34lk3g1#(U<>~
zVlv9&QVk}aEX?UNl3NFb76a_=-N_XrWU@oSa`lHLl?SFW(NYf<c1*xQ7aesCP0{(F
zr5c^T&Xg8_uEn40Eu8uJQyY^GHb9<w_>88Q-JbfJ{HIX*E0|9Q0%^+QhdH>erUH^0
z<v_s(gRj0adHDfO{3p>Y3h7$_K4k$+oO>NozdCR@%MRZCK#+IB_XCQRig6l@XfLH{
zH~8wzNEU|-Q22{_9BA=~9laf}SrK(DUPAd^WL}(dJTT!yOgW3gdN34P`ldIuB5=R!
zg9x&*%?oNRj%B<z6MPqh+gvEa{+uD_9r_z>YkRA0Z*R7ZjbkEb(YdQl8@Dm-w!3%l
zbf3m|)Pq5qD?4c8S&HrL6YcnklkNEN6SkY@f9~A5qlLZfb*-p&+J^-(X}hBT(fnyg
z>Y>bi`HooN`gGqYVLpJm&>1Tq{<N9<3}XZLc^^*S0YXEppb8OV1;IX6-LrV5&(X$d
z7v!!T<qFNVzk**IS>8g#^CY~Xh_<3VfQ);GywTys3O8@wa+-d(xrv{7{sgtsZS~4f
zH#V{Ef3X)ov%zt{)|gOsGtlvujZxYln>zZn1t*9uw*;mh>eX(4_j~J<|7w_2+M83v
ziw{>G=5{Q*lrvGtBkdpaK<_aNn}z3$ynU=3{y}5t!y~TPkMGGUrXJ-_Zvgni7YfuL
z9A)6qN#LsREEB2V?UC4S9>h~W;cX+A>SkLk-*SXO?+~D&JbtgmZ?{9Sb@Aba4!r0f
zS9d6j{GyXNK5nef$1C>c7gf<W17m`HClc!+A@H$hd~nmasC4<KBq7uK2l?J0bKEQK
z+d|b}J)T=$`a2MMH+srCjz48FJF)VUKa7rD&J(*lHWyuvOOCVdyWUnn)d#3Cl)k?V
zvA4??Z%(>IXF<`d!U5>SrgZM&rCj|-F(y?#C`x(kG@c}7jv^DNx)Af{#Vp5L;d;5-
zZ3(v^2VQ6f-hyaNyOVnq<XEO(EcTg4LD^DmRXV-NLRZg&w^n6Mr$5W?V@yG%OP&zc
zZw_SiJdnQ4F%sWcu2{SV>Q@45%=4TlLi1Y&x`W54oLou&%~kmt!?A&rWqhrB&(9Nr
zXDs$DP^=de;4S&UWWPm54!CK9Q^ydcov(sQ4X(3daIOdaOuXi&L>-DE^GIH8z5)Ef
zu^H0XR1p3y*AQ8m2P3nf>JiBlU~^)`pfj)1hqJHWdYhK5WpCOlb*wfl8+>)#x=MCz
z1P=Z6zw+y%3u7?=BT>OBoO%2|^8zIc9x7S`?kYXx=`>W5N*kxLnGT+wRJs}?=D@{Y
zWWY07DIQN?-3DR;hdPdQ#}#q7XWOjGv?;ToUgj)#%AKG{x*0&LGjv`G>nuOxMtDR|
z=b=+67`oU3*TFKn4v%iZV*=DwK6XqC>5WWKu@8pnfH9sklggVInUQN^g32<%ht3C;
z_2vKpyfC9qHXKReMfxO%WdAWm$61%zWVg~B_@?K83xVl?B>733Vna4<>H}=rWFNSF
z$zXx&#e)JI$Ul1In9nvdh=Al_>0C7#{npzXu#M?Q_>2<j=|Nvb?wf4>ppssd3Xe2v
z>>*%N-0t{}V66R6$Ac#XxK)@5m=IboB*l+tc=9&`IPTlVChO>ZAMmlHa@(#v3yaN-
zb!{p(wYlKx4Pse*@bRs?qRYkvyDU$G_wF5SI5bdiu5WrXvAuoVS3TGW`I5K%Xm=@~
z1%9qdp;wmp_9k!C>?7&3PJv{zPLmg0ANyK5IeVVV=kYsrseAv6myiuue1-sjH=x(M
z;z01UQRJGf2l^vEPCs`0YJ;O5W4nOR(Ek-Yb|IgFlmYYs`jhKPTM2D?1=_C$eHI;@
zDf6NnoBEhA*=xD6*Y?%8po}-2YG)S=kU<?AXT0rI-iv<y#pdX!p()2P{PaVV1s@O*
zb~wqe3h7r)#UFFsyFlklNbld%ghO8t7T*wep2XNW)6MpVKCf@=3mC{dZ$Zfbstsow
ztGpn7ok1DZGmDkE9c%|S7q5AgxBc{w1tA9L3yP13G`SQbH?f3;$6@9HN<V;IW!s1c
zD$f`Knl|M)Qp$Lb-D(5+A^df$l24zU<r^pzkgYK*I?{hfko0b|YbTo8P^A5Wr|z;{
zf1+?drG0z_$aa=3@%32G5y*leM09vpfn_Pi{>eU>kA4dfkj5f>L(XHk!>pUM`C+5}
ziX)G5>f7m55090KOP%q`q%A`0USOF-j72$#$2>;8%MZjOz?XLA7?8A17uvD|<NTi(
z-S9BS$p&c9H|v~FxDl$J=I;Ox_zKYz*#-D<0aS+lZbwTCZH+mB@&~-VT=|ScsXIC{
z@45CSn`|kajv|9P+AhKX@OzD-fNTe~mB+h@H|;*PpUOF5lV>aAA77;$`mX&Nqas6p
z_`+*Bi`I|F@%X`}2_1kSan@5xI*u&BNn|1>KgM}X)1U2S{-jJ`|G6ywqn=sY)tHz%
zB9n-D6#dZ#Ea`VleI!qK%7U{g29vvCV<$SX!=Y|-GGh;b4MpgC0vg1_AQUZDuq+4T
znB;olR8HB56J_WXZ0<<c^COs+n2d^w-Wv%2bsB}S(hQb#4mJ}sJ2;CFz7~H)3lK%N
zDryvEkq!8u?`36o%2?3RVSA$rXSvon^VdO_+{|B{Xs%l)09&%U3<F7LZH3CnLsR%b
zNYXi-hX7p;r0Jv@h<uPJk8XJ?Mbs&iJlvKgTy<31+S;i%2bD4u!3W1Q06=S5nb5i^
z0^xhv1o&293HJ3oIU*|$OYUZoOPjJ-6P`sCx8`_5(a$V;pplG{AHQPctZ$x76MT@?
z0HeJ74rhe%gZz>~Ds@0s@!1xKsd5I&9D5Kbz4V{-6`*7w4;+1@jz`<8z}Y>^N_GLr
z(@rIi3Dj@=;kBL{Yi)fUUXBB<N}xwNutg2YEk(C(-fnmA?l!){i~p=!LeYmWxlIRo
zY}7p9s;m~(Y$UuG<mjS$h;_RuooZSS^6YEnkx!_4dn18PY=fK%K}a4}e4?~@dbw#H
z0`&tHRwByc8(r{WGv>%f*~AqC2I@NdZl!p{HuXL&hVmz1GV7&o;FWfXV(Y?Z(jy*k
zDC7ViPf%2!P*nAB8+1Dj%J};Ru3^UstP?twkD>Y}SE`D>{L)>QG`#LgDVqQmvkf@e
z8IoBKV7ZQ+o8(aLHlghd!t-zXF3TdKtYq?MUx|P|!NgrYjmF4^-?EV7pYYh9BD5si
zP_$_;p^OiN7H-;ev^}!teCdS2{dNh8W|TuzX%)J>pqdsuHUP`|fReYWtkU?IT7)jg
zL+~jhHqVj_=#gI`IGP&#NZaRw$<T-@n-WE4$pviI;H>oRzQF7CDxMKWIVDTE$^OEP
zP%`NIK;-wfkPV{_r&Xt%)gSfDF)-<YwbPBhC_YKDzaTFyT)9el<!70U$K?ysK+P*@
z%c2**(~K+ONB437N<Ki+c?IGupcBY5Pa3a$`D-$~L-dBfXwgL)ux{ecSa>*dyr#VY
z${6<mugQ#`6;keM3^rI^*H?YTIEat|L(xN8Nz(+1hyLdJ!#64s6ito+(qL!7=%JtF
zcFt;hO7IDGkmCnN451`8Ft!@obp%AKzES>jnRJg3+VCJ>zoGY>q`s~BDf0#r@*8>a
ztNT$HE!%ykWRVOTOg#Y}d4nAh+VO$dkhI&xwHxN&UU=xQ{iR<wmn%02K&XiZlojcx
z%k3~9T+VgD3$g?;=;5-$8NfqFG8og2Gz+{0l^HkSRXyYL*Kt(ZBf04Hhoj-cH6`D;
zY6L;yt-}_D2XvJImI6;l6+eh{RD!eNCFykRbfNyIg@*ExPp`zL&Tr^-8v-zC%yBG;
zou$)C4N6$wEHeY87zcR>BpGp{)n83gE;Jw`X)kVi#{xsW=enmpl?ETCl0!0S?5IJ|
z^_$A1UpefcL(`ctF!K7Db@7$_x&#d?gF?Lx2Z1%{0j1YKg$FyWUje@H5-Iw9rh*CK
zP;Pw!YVeDna^|NdP&>u)&>6kN#si*wC3{#D5P;&Z{*k&Bj`j%c!hx+2ThTN3Pab4v
zL&3KKjw#LSJw0Hek4aVy$rT(n61??`2Z-UVTbsC&#4{><lYj@ec~--(RpyF;z_;Dl
zFZyTD#b*q(y}rWO?(}K-*gXt{*?T>xPx_-d93c=ge3UwYD4$Gpq?D*ILD(b4@`DCY
zT?>BoJmpO3U^%^gt-IP1K$hiLLh9K6%uajg&vos5K6BLD%ns?Y(zmjis3Hn|yBAW#
zlVguE+0hwO&m1$Pp+QFs2yK}I%=giPreEO4fc@18%~!EP^6|$E94poD38_aW-|!Y6
zP%;5po@~V7C6K}qNtR{GZvZs(iINk2DsypgC7akbC<KQMF6z1fA!c!gqM2k1kU@PX
z{vKC(`64t`*ICvCvtqSVPYk+K4}tsTEUV-Jqu)tm8xWeAuEuDJO{XyOmR#zYx`vKp
zs>Vym=~L;aOF$cS8PKA;Zv6tlM>pv%3xqP!L4V8GS`dA6J~AswI(@XOTYzHAZD^2p
z+aN>;vO{9xs@E!?03{=tQC<8<Fg<;fe}|eYsAuwd;5*#zi~3_Oh~D^tblrjjKwtiG
ze)u(rwQMNkN(Z<kK_&v`KAfw^rMK`v#rB%$N+yx}68y$Mi{{V+MK8-8Bdaz?)EAxL
zdH!zIW}^5;@SGuB`X_P9!B2BThZ_GaYY2t2{nA;u5oUUh`?OEZ*Wjxz4z^#qeSFRS
zvwD0&ux-#tX8BTbnbX~cmrSU(2!wu0_f&MGU)e6n2~V$+jy?a0#x|+MIu%?8C=1V@
z5|LRxsQ_K@`mrwKtjj0^JXxbupQpeUu$lOGEeWCsUxX1iGtWwZq9J}19vI6CMw!cW
zl?RsPns3M#EpDqaekU$)3N`MT8b0MMn=<8bA{`1w%jq500~+-D3wh+FtmD-fQ@)!{
zJlYqzGrvWcXgo0Z?P)rNJGZl<@13{{$|M^dM)f=b5oa4gb+q|VC;He_cJu=sI+v&S
z<4@9vVUgg43Ta{u1n#toA;`g=fq+1N%)x}7=*mK&>jnjM`tnH)M$wUSz_tr$ALLjn
zyi!C5_d~55^wuM~1HNR=>q>KZ(hG0-hctR(LpQ?q!zsMf*ykxRtq;`V5kx;}o~4PM
zHG$A3Om3uWaI*F2=rZhs@&lz3p;~mBUN417cQ<Al^cbX-uFXXJ6~6&^D&7iULtWAF
zGaGBX?l!L$&G!T}sj=bmS#936AN#nn!);`D@7`^<@7!*;Z{KeF+E`#4Z;#5u^L(ke
zU!mw;a~rZr!FKv&)kCq5G46Gp?C<^_>WuA3sy@Z7M#wOY{!HNtPL$#6ywC=(x{IeL
z1F)K?xG4`o=XNSNWXeW#>yLcsJC%zNNF3W-M1x@qWgw`R{%Ub&bqe^v<ie$&Q2*Xu
z3iuJ>X<N=vVHbS`c@)Vv-G>0>*-QlR7tqZaAO^Txc+1uQM6bNl3dFH35F2_B31!d_
zD03gQEbv8d(#T6>L~#57&?OUj3?SjJzEgTQ4uFU*Tf-gtuCvnv-3$XbyKO>7##qnV
zpv7VVQI|D=vSou|NS9uqQkXU33+yA<(5LcMGyrt|8|qhph<TxtJhY{~Xd}PpL5VOM
z^pu;c>fUjQ!AqV|gA8(td$PyL#E$TVhMt>#Vp*meWz6}M)4xJrP&$FaSN{Mg_p#XN
z7J2!d`x7C!6|@e)lN#lF>?U{t0bh<oN%UOgF<JpVIfnBr9(e#6v=!qnIP!JO8zwUz
z)}Ofi=_2E*zOhU6`m1`9&%7BqrIU2z#U}leQYc-H3`JJ}eZ{igu@~Nq{gRJ9PCgaB
z{9}+T@SKnf|7=U7mOTmBQ{@$aiX$sF0N_cx{=Iw>{18M&Zb@gpo@JFh<N(Meoe%!X
zj()tc24K!-jpXO8`h;C=k+w~2e3uKv)*5#f@dc2_rf!S+orgyd52`PNpLLm7b;I|F
z%33#aOLpw&zNbE*^#i_aEPBeoc{9=R-F1po|1W&*qyI&+4?_uDEjR%&js^(csU9P%
zCnhKvJ=<w#-T-=MT{DoMC+1POaMSM$t^+ZD(1&P&m*kYE<EzoC6B+L!BuB41Q@1Pj
z<Y4hYTr>*a84`EeTJ)9u`4hhKbczCWDp;27N%^cJWKNhjv~!#=0kX`8j?$xY^)&1!
z%(#*feJ+;`HDFkV{9@0Fyw19oHZNfO=i;P7n!17;Yy@ZdUB^^rdSsIdxS1R*#UVb;
z1K8Z2bMy#rAL{k#2b_LP>j1yv)UR}fCiM*taQ{)L!N4N&Gtii|=lcO!^hib?KDXW(
zu$waI8I(wSK=7IJLw@@YlTK*F;2%qj%pOWdd`w!gex^3-dXRnKXB_r?^&x{4lOXc3
zuZKh0AhbL$z_SY6?&R0U_I^`u3F3CaJ%5yuH$$;O=xu@Z_TypSdP7M4yRD8>o$_yd
zmihtax7`ZRV~f(qj?l(D#GVZh?OuV)!q73;+WlQ$>@K~ktKT4m?u7JxaIvfC9jsf{
z+rI=8AWixdds4|6tU5l*=ivN?9R=dFRR^VAH^tPw^pQ?qt9*1uZR&t8;hT5~U3T%?
zpS`SbOf<$BsFUQ#e#`G7s!uN1DkvOii<OCv^)t&1Oj+U{j59-`$sXDqd;8&1S`+iD
z^3fKu#5TkU2luauZrP?eZF@x>%q+Lakwm)yPy&5@Fv&ERlGZAn3BA9XqJwEVvtZdz
z%kZ!tALb>lcCR)fs&*^z!nU80p8JCe=ufs%fq(IYZ;pY`m;O19e+cA53yi$P*pvLC
z4CpAceA$&-M{_3Ii(am7qL=b;_!pS%c^Foe#ijto^D#hK%5%&oV3+4v82p3GIoM|X
z<rW@rk>Pos<A{w+7l)jZ2h?pd#`Q4)h!g>G5aB@ERyzE^obFs?m8`l`zk<|ll5PN(
zc6nHC`h2%f;}*PLm<kXxhxnoTlrWJyj3xBFvITGQH$ms0s?hB~uVp(|HpQ-@DX20a
zHUtW{P(J)R5*m9lZZkio&Riv8uFVyrzQ*ZL{nO7*^t#apz&p6?z?;VOVd^Nx>I_P|
z%H52z!%5QsvAIIki>StJ*9kc;DLx?de$X1*L_lfDh$}{VaI-FoOcTNdm}t7u$;&qu
zt+f7*{K_w8GcgkfNLPm|ST3Jq=!g^207Y9kDoVwIBOl%haT2KD12f&6o*YVE_;MC7
z4fx&F&_Fb`70}bjpwV+csIi2zDD%*H(}Ab693|2nFt&^Vm>N!@7NcIIfL0mBBYo_R
zZ-jLk=wbx$Lth7Yw~4><b6Ape2x8J>p0uTp?V)o^y23~w`DMVYGr<cc-g~bYIn;Nu
z-TB_>>p(ITT|xnR$XhS@{x^Y0$1~PGczba1ARj;{i`lvjs3r_zs^ohQ`=IVR3PyiI
z>mYo^lua~a;LmAj(E~<w{aA=C=~wnQF?b<HImmX2(hR;e>7;FVGYeM_c$*KmS;>CO
zV?oWrlmRbqt{QDLFLD*59~5}gnTB%NqQ3(TfY5MOhG!*so6!`M=d100Xx(yBM}R99
z{wna=n)^5W(AG05K1)NN6CdxB-{0N$2NUn!x$CPE$g$0?B_oLbTqkZ<;306%exxq(
zB_Gq%=0~BfYEX_pctcm+PpYhe_HfXK0srR@9SOdr3rBz_?=*b$m!W^rc0E0Y0=H*R
zTLbd-mv)#Ltf}i5`*E(Cq%P#`<4%V}+T6#Q%hR9HNgnW|Cf`^Fyv7FbjrArwKy*~S
zQCFNu28`op3DVbNhkagpJ85j2W1x5f5T*|quR1A_wA!3>Ez0@gyW7gd#VK7*=MQ@G
zyFUnQi&D-wEB^(zT%X>w)?E;(6r_-{Vk_>FGvx_gz5^IUy1jaE-vl}9dI(fhGD91f
zWDEf)eFY?k>_DIXT(Dp(fj{|KKa@`Rg3<>}X%`oMrYFuEk&wl7%ELR$3Xkklzt0N_
zG#B^SS1?R38q<w3W&jsCB`+`}J07|mOB_d1*~{i*A0X{<R(M(FztQFy@(NiIhR2Ih
zrEz1gv(Hxh>a66823h{Pyv9BEJ%`A#9;FSINn({=isw3)ONsFf0sYJmMOU=I@>~{o
zlIEC944z|<LNAiBubV<#dXi7<anF83SH>#lvjWrUKjkxD{3{T8wkgMQC<~Tyofd$;
z<Q0A_2NrPphnip<?L%Jmbg@GE?9At90~Is=aDoUw*4%xdIg~3f+!}oQc5Ve`F5LG=
zTzj9EKTvEp5Hi#mXQ^E;I*JCcFDSrM9anOQ&5x};J%NR#9#fn5I<4~84AD2z6of9c
zBa3O_`>X$Ho|#w_I4fa-^Db);Jkl~?kT2y$S2W4Sz`6~DBJ!ggpekKZ@`|>~uxHF$
zXQx}p^Ec2ffE}CHJIWgW@qsY&Lf-kp*#|avh}rfG9wmFej0i+$=omS``#yX1B6q^h
zR@{IO;AIeRu*9S)q@K~ISTk%_*>&d<AIP#7`{?+9r7-{ugFs}>vi0M$C#i$-U3NMs
z^q?6CGiaDj5IKS17`zGd-~vE5wWad;4Nvh!6mf3@K?b}W$b};c&n%04bi_}wRevHN
zKkJ$0R32Nw<N%Z$r9*=p=qkjoK=t3~%qEq5$pk7t;%pOt^snBIqA<-oDW|^o_7D8p
zR-h(5F3naQB}Q@gl>#!R<H~)t%jZY!R&4@(7UbaR7e}WFX?>H2t@FA`0uv;Qe#D{6
zXB7|lzQ}>E8c=txRPexG<Yye=mNWVRn~jq@C)<hR+pZfAv-=~1Ja5tc#(=FFV6Oeb
zhNiwqOrNc@M~7sKYGHq=vAS)G8?1fF1u9?lpZbGKk&!;?cCh~+?q@(`%mE{b2;QvG
zR{GD`q3A$&=|`6RjV{GiM%|SXP0r5Q4_bWCVewn>l^o)tBX3&U-Gi-N?B`VuACR~o
zm#J!pl)HAlKb5_JBgQ8Ch`1j!Dkz#|7;ONmP6%^a^8+-DE#osvz$mj)KJzKO7wRA-
z7JED{m#Pl4ZfLum&Mgw66F^>V0AO_;XA@#i*Np`D!aMtyF#5zMnq<~#IfAlpwlisQ
zPyEY39|vZh*b2<@6tJ+f2hC4>Uf@2vqOULOCXMW6JB!BUy$=^j$sqW6=>F%5ogCQ9
z65}rdCdO0}(N{X50}6-Vt1NNZl71b!v=85GVGiU?bpT^%U9m;1?a~;67;X!dWul6G
z1(AoI3M0<c{&gJ~qxDCfvepQ#i%p-mAdURdr(gYe$cA{S&ImV+`+fol$I!S9!fJnZ
zC|&~T;)7$>4=S?YRm$a$1Qjt)=D1tz`;|{xwN*QiFBnvMUIxkz^L1He?-+_^QQllu
z`j-Jb0jRzsEA7isPh{IY1|%5=oigEKeN2DqSJWsEd6z?+zGqvyIq=43Dq|x`=*LVX
zpgTT`Eji&AyCFhQ?4+$>v3xj{E}vL;;U~EX&`<S8Bz*eAt~wq^2wh~&JTuLaFMRG}
z|3wjv0U#cU$xluZP-bT+EW@&#z@n@fKj%kW!7BV{2T*oQIH@8<e!}a&x4{f<KnBv3
z`)80Fi>2Rx0c1tqOb2E8YO<koj8p|hGf^t<g@{A!4NhtO*~f$NrOWYf3etQ|P}<W1
zIywWW7i)ZZRDBZG@$wq7;3uwN-?)r}FoQxmb@)Y~(!Sw<PQS+Yz%YcaEF5OM%cQ$>
zQcq+J=9p}T(Z5S=Eg+piB6>L9T}PEw@WAAh@?91H?-)w9H;RgZ@&^#Q*)PNa;^0O+
z^4xv`c*hu!G{@Pb<+hk{12O5g+{Ck;p)qcfq$~~~?>fY{BP2^bI^MG7BlVG7q1f*a
zUElT9gnIZIsb1)5@M9C<56baeO5X@f9+ub)l3Fk*%|-w@H?QBcEbL<<&j5-Yu_@_Z
z4{V}M>2K<@?zh_T(75&qe<;%r^ycsR^aJ7y@M;wc3=ab(eI*km^u-_lb{n*r8o@d#
z?*NZ=lq~nZO3O?iC-7K>e*C{62R&|cZO)LGV_)qx9UCmT&=i>bgRv{1<XT55?2sZV
znLdi09uI8uY*S*_Zpv3#>2=<?;2#Ae-=fW=JC3qROcLqavBCPmcXqkf$8z)ov)rfv
zzbXKUOE0XIobW;}oY-->Z0y<chtH_XZ?<Eyg><0`*X8P1EIE!x9fA=zml4xHQ#Si0
zFs7lKZ0owGa%D+ZzU-TD6$GVYRah<??Wx!>HK=Uqangm)?TWJ;#=9E#5^^kHS7a~9
z7m$a(%4XX-LD!YClq1&y1vCBP`0Tcygv@%0pz{?69j|=U(kI($p3uK*m_f+6(oLSR
zsEbP3f^6cNJ7yn&cVKg%Kd)Da37xD<NX!rTUd(aA_+-Hyy2=TVZ@Wda99+&~)q=5a
zwlVQkG(q8A=6vd~K-6Ku7NwC}Z2@!<g2<>)P&NXq%7Lm2<J_39(2+K#cva%qMhQ&o
z1mjPYg^e72h!bpl+|}O%(j1w-$>u2=0k6L_hY@TGyyjoDH@T`yt##4{<xBodwVdwu
zSK1pqk2^+BP;>>@;TWTqXdewBEFXF^;gp~hRAD)LHmnZ5S=+ei0gAWiM%sdsGeXH&
zML$Bx9CZSPoA86}z$h0JJJ!E6CQh{Fovic}*PKfIMqonm<R*xogpL|y39tVJA33G^
zD_wrc;M3{(70w{Gk~zw7K_@X6IQ|vWk!sNZkNMmo8I<PsjGThdK^lDO!o10S^_6E}
z^S2JV-KJqt;$t>+Adv;$*e;Jy_T;Qs(VI$<Y$6Z756=97ioW@*Y}rZ-0e>F!;DW`n
zJ19-gvWU%-xYDUh^aT?rImDHQuJW)GDj%9b_j^H~eCytN0NFJFMwU}5=t;wD#|WcO
zi&rU=lTAopqqG+FmhCSm>rXQ3tQ{4zPZXeMf;u?hOW4{-@bNEgCSn1<otLqJ0afWN
z;8TjOJZ~X7gc(YHa#@Hg5cl8?2q<zt(;xiEL08Fyz-K_=*90D|P-8<v{2F=gZC#H2
zsQ(P&*cSuKzqZ+_;5hb6X6#R2_bnTi1s@!d?a3Y8QxR>@+gy2k<BhfH(Uq$b0qmKN
zwMRkO(K9^)rx@G1&GxVSL+o<s{{Lm2NumQG2t>2U{m=Y)H7N1Zp+PJLjrNdDjP@Cu
zF!7=6#>jX*wnej!nKD~5$896kX@gbIaajp~jtYjVl!tco=2rIOi|)DEHsuel683!}
zwYf{uJec*}-~QMLNq*`|9eS~#JpnBZ<Ba=wYCRE*{cvJ|2>#lTGRaqR3|$*~c*Q0h
zOgu0$iyBM(*tX6*u=doRxPFk&-d%N~muq_&(pZvEHtEcf4?fHpl34H9cdSZcyf)-p
zQ)59}cfXlnT_@7#a11~Aoqo-+5(g_)?CaWBe<BG)BReloU|NgW`whu{a6gXC4YWDH
znQfzgicu(Tz4a_sK(HLdNSwxlSeuaJ-Epcs*=b#T>ue3xAzw4f>_yj0OqEh#=p8el
zRoahE_pZC0I{yIwQ~i+T5;alBTN`xzy5HxsuP-KJw!d=7C+#_O57NHzy@(f1{~+b-
zwf%7@YdVfpox{qFeOUWl-&e4)T}lDSHD*)iI*x-m<E8%LJ*GnmRZM+KZ`<{LK7X}A
Vc(ZIalym?9002ovPDHLkV1i;hhx7md


From 65d694b257c6e54b031aaff8c1ab2b3918e2d50e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 28 Sep 2020 16:12:33 +0300
Subject: [PATCH 002/173] started to write the nGraph class for
 NonMaxSuppression-5.

---
 .../include/ngraph/op/non_max_suppression.hpp | 122 ++++++++++++++++++
 1 file changed, 122 insertions(+)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 3bc8c152f7fef9..33daceded11c40 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -235,6 +235,128 @@ namespace ngraph
                     clone_with_new_inputs(const OutputVector& new_args) const override;
             };
         } // namespace v4
+
+        namespace v5
+        {
+            /// \brief NonMaxSuppression operation
+            ///
+            class NGRAPH_API NonMaxSuppression : public Op
+            {
+            public:
+                enum class BoxEncodingType
+                {
+                    CORNER,
+                    CENTER
+                };
+
+                static constexpr NodeTypeInfo type_info{"NonMaxSuppression", 5};
+                const NodeTypeInfo& get_type_info() const override { return type_info; }
+                NonMaxSuppression() = default;
+
+                /// \brief Constructs a NonMaxSuppression operation with default values in the last.
+                ///        4 inputs.
+                ///
+                /// \param boxes Node producing the box coordinates
+                /// \param scores Node producing the box scores
+                /// \param box_encoding Specifies the format of boxes data encoding
+                /// \param sort_result_descending Specifies whether it is necessary to sort selected
+                /// boxes across batches
+                /// \param output_type Specifies the output tensor type
+                NonMaxSuppression(const Output<Node>& boxes,
+                                  const Output<Node>& scores,
+                                  const BoxEncodingType box_encoding = BoxEncodingType::CORNER,
+                                  const bool sort_result_descending = true,
+                                  const ngraph::element::Type& output_type = ngraph::element::i64);
+
+                /// \brief Constructs a NonMaxSuppression operation with default values in the last.
+                ///        3 inputs.
+                ///
+                /// \param boxes Node producing the box coordinates
+                /// \param scores Node producing the box scores
+                /// \param max_output_boxes_per_class Node producing maximum number of boxes to be
+                /// selected per class
+                /// \param box_encoding Specifies the format of boxes data encoding
+                /// \param sort_result_descending Specifies whether it is necessary to sort selected
+                /// boxes across batches
+                /// \param output_type Specifies the output tensor type
+                NonMaxSuppression(const Output<Node>& boxes,
+                                  const Output<Node>& scores,
+                                  const Output<Node>& max_output_boxes_per_class,
+                                  const BoxEncodingType box_encoding = BoxEncodingType::CORNER,
+                                  const bool sort_result_descending = true,
+                                  const ngraph::element::Type& output_type = ngraph::element::i64);
+
+                /// \brief Constructs a NonMaxSuppression operation with default values in the last.
+                ///        2 inputs.
+                ///
+                /// \param boxes Node producing the box coordinates
+                /// \param scores Node producing the box scores
+                /// \param max_output_boxes_per_class Node producing maximum number of boxes to be
+                /// selected per class
+                /// \param iou_threshold Node producing intersection over union threshold
+                /// \param box_encoding Specifies the format of boxes data encoding
+                /// \param sort_result_descending Specifies whether it is necessary to sort selected
+                /// boxes across batches
+                /// \param output_type Specifies the output tensor type
+                NonMaxSuppression(const Output<Node>& boxes,
+                                  const Output<Node>& scores,
+                                  const Output<Node>& max_output_boxes_per_class,
+                                  const Output<Node>& iou_threshold,
+                                  const BoxEncodingType box_encoding = BoxEncodingType::CORNER,
+                                  const bool sort_result_descending = true,
+                                  const ngraph::element::Type& output_type = ngraph::element::i64);
+
+                /// \brief Constructs a NonMaxSuppression operation with default value in the last.
+                ///        input.
+                ///
+                /// \param boxes Node producing the box coordinates
+                /// \param scores Node producing the box scores
+                /// \param max_output_boxes_per_class Node producing maximum number of boxes to be
+                /// selected per class
+                /// \param iou_threshold Node producing intersection over union threshold
+                /// \param score_threshold Node producing minimum score threshold
+                /// \param box_encoding Specifies the format of boxes data encoding
+                /// \param sort_result_descending Specifies whether it is necessary to sort selected
+                /// boxes across batches
+                /// \param output_type Specifies the output tensor type
+                NonMaxSuppression(const Output<Node>& boxes,
+                                  const Output<Node>& scores,
+                                  const Output<Node>& max_output_boxes_per_class,
+                                  const Output<Node>& iou_threshold,
+                                  const Output<Node>& score_threshold,
+                                  const BoxEncodingType box_encoding = BoxEncodingType::CORNER,
+                                  const bool sort_result_descending = true,
+                                  const ngraph::element::Type& output_type = ngraph::element::i64);
+
+                /// \brief Constructs a NonMaxSuppression operation.
+                ///
+                /// \param boxes Node producing the box coordinates
+                /// \param scores Node producing the box scores
+                /// \param max_output_boxes_per_class Node producing maximum number of boxes to be
+                /// selected per class
+                /// \param iou_threshold Node producing intersection over union threshold
+                /// \param score_threshold Node producing minimum score threshold
+                /// \param soft_nms_sigma Node specifying the sigma parameter for Soft-NMS
+                /// \param box_encoding Specifies the format of boxes data encoding
+                /// \param sort_result_descending Specifies whether it is necessary to sort selected
+                /// boxes across batches
+                /// \param output_type Specifies the output tensor type
+                NonMaxSuppression(const Output<Node>& boxes,
+                                  const Output<Node>& scores,
+                                  const Output<Node>& max_output_boxes_per_class,
+                                  const Output<Node>& iou_threshold,
+                                  const Output<Node>& score_threshold,
+                                  const Output<Node>& soft_nms_sigma,
+                                  const BoxEncodingType box_encoding = BoxEncodingType::CORNER,
+                                  const bool sort_result_descending = true,
+                                  const ngraph::element::Type& output_type = ngraph::element::i64);
+
+                void validate_and_infer_types() override;
+
+                std::shared_ptr<Node>
+                    clone_with_new_inputs(const OutputVector& new_args) const override;
+            };
+        } // namespace v5
     }     // namespace op
 
     NGRAPH_API

From 6ceca93c963e8e937ed8d3a5cb0aabb19aabb5bb Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 09:44:49 +0300
Subject: [PATCH 003/173] Added NonMaxSuppression-5 to opset5_tbl.hpp.

---
 .../include/ngraph/op/non_max_suppression.hpp | 19 +++++++++++++++++++
 .../core/include/ngraph/opsets/opset5_tbl.hpp |  1 +
 2 files changed, 20 insertions(+)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 33daceded11c40..ad93e1ab48722e 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -396,4 +396,23 @@ namespace ngraph
             "AttributeAdapter<op::v3::NonMaxSuppression::BoxEncodingType>", 1};
         const DiscreteTypeInfo& get_type_info() const override { return type_info; }
     };
+
+    NGRAPH_API
+    std::ostream& operator<<(std::ostream& s,
+                             const op::v5::NonMaxSuppression::BoxEncodingType& type);
+
+    template <>
+    class NGRAPH_API AttributeAdapter<op::v5::NonMaxSuppression::BoxEncodingType>
+        : public EnumAttributeAdapterBase<op::v5::NonMaxSuppression::BoxEncodingType>
+    {
+    public:
+        AttributeAdapter(op::v5::NonMaxSuppression::BoxEncodingType& value)
+            : EnumAttributeAdapterBase<op::v5::NonMaxSuppression::BoxEncodingType>(value)
+        {
+        }
+
+        static constexpr DiscreteTypeInfo type_info{
+            "AttributeAdapter<op::v5::NonMaxSuppression::BoxEncodingType>", 1};
+        const DiscreteTypeInfo& get_type_info() const override { return type_info; }
+    };
 } // namespace ngraph
diff --git a/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp b/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp
index 145cb866826564..c7c293a00a5b4f 100644
--- a/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp
+++ b/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp
@@ -165,5 +165,6 @@ NGRAPH_OP(Swish, ngraph::op::v4)
 
 // New operations added in opset5
 NGRAPH_OP(LSTMSequence, ngraph::op::v5)
+NGRAPH_OP(NonMaxSuppression, ngraph::op::v5)
 NGRAPH_OP(GRUSequence, ngraph::op::v5)
 NGRAPH_OP(RNNSequence, ngraph::op::v5)
\ No newline at end of file

From c035301eed5c196239e832bf408751745c95ab81 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 10:11:38 +0300
Subject: [PATCH 004/173] Written all constructors of NonMaxSuppression-5.

---
 .../include/ngraph/op/non_max_suppression.hpp |  25 ++++
 ngraph/core/src/op/non_max_suppression.cpp    | 114 ++++++++++++++++++
 2 files changed, 139 insertions(+)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index ad93e1ab48722e..27101831912078 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -355,6 +355,31 @@ namespace ngraph
 
                 std::shared_ptr<Node>
                     clone_with_new_inputs(const OutputVector& new_args) const override;
+
+                BoxEncodingType get_box_encoding() const { return m_box_encoding; }
+                void set_box_encoding(const BoxEncodingType box_encoding)
+                {
+                    m_box_encoding = box_encoding;
+                }
+                bool get_sort_result_descending() const { return m_sort_result_descending; }
+                void set_sort_result_descending(const bool sort_result_descending)
+                {
+                    m_sort_result_descending = sort_result_descending;
+                }
+
+                element::Type get_output_type() const { return m_output_type; }
+                void set_output_type(const element::Type& output_type)
+                {
+                    m_output_type = output_type;
+                }
+                using Node::set_output_type;
+
+            protected:
+                BoxEncodingType m_box_encoding = BoxEncodingType::CORNER;
+                bool m_sort_result_descending = true;
+                ngraph::element::Type m_output_type = ngraph::element::i64;
+                // void validate();
+                // int64_t max_boxes_output_from_input() const;
             };
         } // namespace v5
     }     // namespace op
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index f07a5e39c8103a..d4a832a1ad1e00 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -530,3 +530,117 @@ void op::v4::NonMaxSuppression::validate_and_infer_types()
     }
     set_output_type(0, m_output_type, out_shape);
 }
+
+// ------------------------------ V5 ------------------------------
+
+constexpr NodeTypeInfo op::v5::NonMaxSuppression::type_info;
+
+op::v5::NonMaxSuppression::NonMaxSuppression(
+    const Output<Node>& boxes,
+    const Output<Node>& scores,
+    const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
+    const bool sort_result_descending,
+    const element::Type& output_type)
+    : Op({boxes,
+          scores,
+          op::Constant::create(element::i64, Shape{}, {0}),
+          op::Constant::create(element::f32, Shape{}, {.0f}),
+          op::Constant::create(element::f32, Shape{}, {.0f}),
+          op::Constant::create(element::f32, Shape{}, {.0f})})
+    , m_box_encoding{box_encoding}
+    , m_sort_result_descending{sort_result_descending}
+    , m_output_type{output_type}
+{
+    constructor_validate_and_infer_types();
+}
+
+op::v5::NonMaxSuppression::NonMaxSuppression(
+    const Output<Node>& boxes,
+    const Output<Node>& scores,
+    const Output<Node>& max_output_boxes_per_class,
+    const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
+    const bool sort_result_descending,
+    const element::Type& output_type)
+    : Op({boxes,
+          scores,
+          max_output_boxes_per_class,
+          op::Constant::create(element::f32, Shape{}, {.0f}),
+          op::Constant::create(element::f32, Shape{}, {.0f}),
+          op::Constant::create(element::f32, Shape{}, {.0f})})
+    , m_box_encoding{box_encoding}
+    , m_sort_result_descending{sort_result_descending}
+    , m_output_type{output_type}
+{
+    constructor_validate_and_infer_types();
+}
+
+op::v5::NonMaxSuppression::NonMaxSuppression(
+    const Output<Node>& boxes,
+    const Output<Node>& scores,
+    const Output<Node>& max_output_boxes_per_class,
+    const Output<Node>& iou_threshold,
+    const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
+    const bool sort_result_descending,
+    const element::Type& output_type)
+    : Op({boxes,
+          scores,
+          max_output_boxes_per_class,
+          iou_threshold,
+          op::Constant::create(element::f32, Shape{}, {.0f}),
+          op::Constant::create(element::f32, Shape{}, {.0f})})
+    , m_box_encoding{box_encoding}
+    , m_sort_result_descending{sort_result_descending}
+    , m_output_type{output_type}
+{
+    constructor_validate_and_infer_types();
+}
+
+op::v5::NonMaxSuppression::NonMaxSuppression(
+    const Output<Node>& boxes,
+    const Output<Node>& scores,
+    const Output<Node>& max_output_boxes_per_class,
+    const Output<Node>& iou_threshold,
+    const Output<Node>& score_threshold,
+    const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
+    const bool sort_result_descending,
+    const element::Type& output_type)
+    : Op({boxes,
+          scores,
+          max_output_boxes_per_class,
+          iou_threshold,
+          score_threshold,
+          op::Constant::create(element::f32, Shape{}, {.0f})})
+    , m_box_encoding{box_encoding}
+    , m_sort_result_descending{sort_result_descending}
+    , m_output_type{output_type}
+{
+    constructor_validate_and_infer_types();
+}
+
+op::v5::NonMaxSuppression::NonMaxSuppression(
+    const Output<Node>& boxes,
+    const Output<Node>& scores,
+    const Output<Node>& max_output_boxes_per_class,
+    const Output<Node>& iou_threshold,
+    const Output<Node>& score_threshold,
+    const Output<Node>& soft_nms_sigma,
+    const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
+    const bool sort_result_descending,
+    const element::Type& output_type)
+    : Op({boxes,
+          scores,
+          max_output_boxes_per_class,
+          iou_threshold,
+          score_threshold,
+          soft_nms_sigma})
+    , m_box_encoding{box_encoding}
+    , m_sort_result_descending{sort_result_descending}
+    , m_output_type{output_type}
+{
+    constructor_validate_and_infer_types();
+}
+
+void op::v5::NonMaxSuppression::validate_and_infer_types()
+{
+    set_output_type(0, m_output_type, out_shape);
+}
\ No newline at end of file

From 4ca722106329e66b812de0f3f9be04d32665767c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 10:34:33 +0300
Subject: [PATCH 005/173] Wriiten the method
 op::v5::NonMaxSuppression::clone_with_new_inputs.

---
 ngraph/core/src/op/non_max_suppression.cpp | 37 ++++++++++++++++++++--
 1 file changed, 35 insertions(+), 2 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index d4a832a1ad1e00..180712770a43eb 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -640,7 +640,40 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     constructor_validate_and_infer_types();
 }
 
+
+shared_ptr<Node>
+    op::v5::NonMaxSuppression::clone_with_new_inputs(const OutputVector& new_args) const
+{
+    check_new_args_count(this, new_args);
+    NODE_VALIDATION_CHECK(this,
+                          new_args.size() >= 2 && new_args.size() <= 6,
+                          "Number of inputs must be 2, 3, 4, 5 or 6");
+
+    const auto& arg2 = new_args.size() > 2
+                           ? new_args.at(2)
+                           : ngraph::op::Constant::create(element::i32, Shape{}, {0});
+    const auto& arg3 = new_args.size() > 3
+                           ? new_args.at(3)
+                           : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
+    const auto& arg4 = new_args.size() > 4
+                           ? new_args.at(4)
+                           : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
+    const auto& arg5 = new_args.size() > 5
+                           ? new_args.at(5)
+                           : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
+
+    return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
+                                                       new_args.at(1),
+                                                       arg2,
+                                                       arg3,
+                                                       arg4,
+                                                       arg5,
+                                                       m_box_encoding,
+                                                       m_sort_result_descending,
+                                                       m_output_type);
+}
+
 void op::v5::NonMaxSuppression::validate_and_infer_types()
 {
-    set_output_type(0, m_output_type, out_shape);
-}
\ No newline at end of file
+    // set_output_type(0, m_output_type, out_shape);
+}

From ce514c5e8098c010ffbeb25b30b20ecd0934c3aa Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 10:38:36 +0300
Subject: [PATCH 006/173] Started to write validate_and_infer_type().

---
 ngraph/core/src/op/non_max_suppression.cpp | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 180712770a43eb..463567e6a2342b 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -675,5 +675,12 @@ shared_ptr<Node>
 
 void op::v5::NonMaxSuppression::validate_and_infer_types()
 {
+    const auto boxes_ps = get_input_partial_shape(0);
+    const auto scores_ps = get_input_partial_shape(1);
+
+    // NonMaxSuppression produces triplets
+    // that have the following format: [batch_index, class_index, box_index]
+    PartialShape out_shape = {Dimension::dynamic(), 3};
+
     // set_output_type(0, m_output_type, out_shape);
 }

From 66118f74535cc65e1ee8f4a2d4785ead093baa24 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 11:43:06 +0300
Subject: [PATCH 007/173] Written the method validate_and_infer_type().

---
 .../include/ngraph/op/non_max_suppression.hpp |   5 +-
 ngraph/core/src/op/non_max_suppression.cpp    | 151 +++++++++++++++++-
 2 files changed, 153 insertions(+), 3 deletions(-)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 27101831912078..41fbdb86cc0466 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -351,6 +351,7 @@ namespace ngraph
                                   const bool sort_result_descending = true,
                                   const ngraph::element::Type& output_type = ngraph::element::i64);
 
+                bool visit_attributes(AttributeVisitor& visitor) override;
                 void validate_and_infer_types() override;
 
                 std::shared_ptr<Node>
@@ -378,8 +379,8 @@ namespace ngraph
                 BoxEncodingType m_box_encoding = BoxEncodingType::CORNER;
                 bool m_sort_result_descending = true;
                 ngraph::element::Type m_output_type = ngraph::element::i64;
-                // void validate();
-                // int64_t max_boxes_output_from_input() const;
+                void validate();
+                int64_t max_boxes_output_from_input() const;
             };
         } // namespace v5
     }     // namespace op
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 463567e6a2342b..b440a05f33b05e 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -673,6 +673,113 @@ shared_ptr<Node>
                                                        m_output_type);
 }
 
+void op::v5::NonMaxSuppression::validate()
+{
+    const auto boxes_ps = get_input_partial_shape(0);
+    const auto scores_ps = get_input_partial_shape(1);
+
+    NODE_VALIDATION_CHECK(this,
+                          m_output_type == element::i64 || m_output_type == element::i32,
+                          "Output type must be i32 or i64");
+
+    if (boxes_ps.is_dynamic() || scores_ps.is_dynamic())
+    {
+        return;
+    }
+
+    NODE_VALIDATION_CHECK(this,
+                          boxes_ps.rank().is_static() && boxes_ps.rank().get_length() == 3,
+                          "Expected a 3D tensor for the 'boxes' input. Got: ",
+                          boxes_ps);
+
+    NODE_VALIDATION_CHECK(this,
+                          scores_ps.rank().is_static() && scores_ps.rank().get_length() == 3,
+                          "Expected a 3D tensor for the 'scores' input. Got: ",
+                          scores_ps);
+
+    if (inputs().size() >= 3)
+    {
+        const auto max_boxes_ps = get_input_partial_shape(2);
+        NODE_VALIDATION_CHECK(this,
+                              max_boxes_ps.is_dynamic() || is_scalar(max_boxes_ps.to_shape()),
+                              "Expected a scalar for the 'max_output_boxes_per_class' input. Got: ",
+                              max_boxes_ps);
+    }
+
+    if (inputs().size() >= 4)
+    {
+        const auto iou_threshold_ps = get_input_partial_shape(3);
+        NODE_VALIDATION_CHECK(this,
+                              iou_threshold_ps.is_dynamic() ||
+                                  is_scalar(iou_threshold_ps.to_shape()),
+                              "Expected a scalar for the 'iou_threshold' input. Got: ",
+                              iou_threshold_ps);
+    }
+
+    if (inputs().size() >= 5)
+    {
+        const auto score_threshold_ps = get_input_partial_shape(4);
+        NODE_VALIDATION_CHECK(this,
+                              score_threshold_ps.is_dynamic() ||
+                                  is_scalar(score_threshold_ps.to_shape()),
+                              "Expected a scalar for the 'score_threshold' input. Got: ",
+                              score_threshold_ps);
+    }
+
+    if (inputs().size() >= 6)
+    {
+        const auto soft_nms_sigma = get_input_partial_shape(5);
+        NODE_VALIDATION_CHECK(this,
+                              soft_nms_sigma.is_dynamic() ||
+                                  is_scalar(soft_nms_sigma.to_shape()),
+                              "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
+                              soft_nms_sigma);
+    }
+
+    const auto num_batches_boxes = boxes_ps[0];
+    const auto num_batches_scores = scores_ps[0];
+    NODE_VALIDATION_CHECK(this,
+                          num_batches_boxes.same_scheme(num_batches_scores),
+                          "The first dimension of both 'boxes' and 'scores' must match. Boxes: ",
+                          num_batches_boxes,
+                          "; Scores: ",
+                          num_batches_scores);
+
+    const auto num_boxes_boxes = boxes_ps[1];
+    const auto num_boxes_scores = scores_ps[2];
+    NODE_VALIDATION_CHECK(this,
+                          num_boxes_boxes.same_scheme(num_boxes_scores),
+                          "'boxes' and 'scores' input shapes must match at the second and third "
+                          "dimension respectively. Boxes: ",
+                          num_boxes_boxes,
+                          "; Scores: ",
+                          num_boxes_scores);
+
+    NODE_VALIDATION_CHECK(this,
+                          boxes_ps[2].is_static() && boxes_ps[2].get_length() == 4u,
+                          "The last dimension of the 'boxes' input must be equal to 4. Got:",
+                          boxes_ps[2]);
+}
+
+int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
+{
+    int64_t max_output_boxes{0};
+
+    const auto max_output_boxes_input =
+        as_type_ptr<op::Constant>(input_value(2).get_node_shared_ptr());
+    max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
+
+    return max_output_boxes;
+}
+
+bool ngraph::op::v5::NonMaxSuppression::visit_attributes(AttributeVisitor& visitor)
+{
+    visitor.on_attribute("box_encoding", m_box_encoding);
+    visitor.on_attribute("sort_result_descending", m_sort_result_descending);
+    visitor.on_attribute("output_type", m_output_type);
+    return true;
+}
+
 void op::v5::NonMaxSuppression::validate_and_infer_types()
 {
     const auto boxes_ps = get_input_partial_shape(0);
@@ -682,5 +789,47 @@ void op::v5::NonMaxSuppression::validate_and_infer_types()
     // that have the following format: [batch_index, class_index, box_index]
     PartialShape out_shape = {Dimension::dynamic(), 3};
 
-    // set_output_type(0, m_output_type, out_shape);
+    validate()
+
+    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
+    {
+        const auto num_boxes_boxes = boxes_ps[1];
+        const auto max_output_boxes_per_class_node = input_value(2).get_node_shared_ptr();
+        if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static() &&
+            op::is_constant(max_output_boxes_per_class_node))
+        {
+            const auto num_boxes = num_boxes_boxes.get_length();
+            const auto num_classes = scores_ps[1].get_length();
+            const auto max_output_boxes_per_class = max_boxes_output_from_input();
+
+            out_shape[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
+                           scores_ps[0].get_length();
+        }
+    }
+    set_output_type(0, m_output_type, out_shape);
+    set_output_type(1, element::f32, out_shape);
+    set_output_type(2, element::f32, Shape{});
 }
+
+namespace ngraph
+{
+    template <>
+    EnumNames<op::v5::NonMaxSuppression::BoxEncodingType>&
+        EnumNames<op::v5::NonMaxSuppression::BoxEncodingType>::get()
+    {
+        static auto enum_names = EnumNames<op::v5::NonMaxSuppression::BoxEncodingType>(
+            "op::v5::NonMaxSuppression::BoxEncodingType",
+            {{"corner", op::v5::NonMaxSuppression::BoxEncodingType::CORNER},
+             {"center", op::v5::NonMaxSuppression::BoxEncodingType::CENTER}});
+        return enum_names;
+    }
+
+    constexpr DiscreteTypeInfo
+        AttributeAdapter<op::v5::NonMaxSuppression::BoxEncodingType>::type_info;
+
+    std::ostream& operator<<(std::ostream& s,
+                             const op::v5::NonMaxSuppression::BoxEncodingType& type)
+    {
+        return s << as_string(type);
+    }
+} // namespace ngraph

From 709a12154e325f78c7200e86cd555cf84ff4e26d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 11:58:12 +0300
Subject: [PATCH 008/173] Fixed typo.

---
 ngraph/core/src/op/non_max_suppression.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index b440a05f33b05e..4203730e00d487 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -789,7 +789,7 @@ void op::v5::NonMaxSuppression::validate_and_infer_types()
     // that have the following format: [batch_index, class_index, box_index]
     PartialShape out_shape = {Dimension::dynamic(), 3};
 
-    validate()
+    validate();
 
     if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
     {

From 2983f6b70ac44b10c3ee98b5b91612eed769e29d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 12:10:59 +0300
Subject: [PATCH 009/173] Fixed opset5_tbl.hpp.

---
 ngraph/core/include/ngraph/opsets/opset5_tbl.hpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp b/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp
index c7c293a00a5b4f..398ae8c2b548d0 100644
--- a/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp
+++ b/ngraph/core/include/ngraph/opsets/opset5_tbl.hpp
@@ -157,7 +157,6 @@ NGRAPH_OP(CTCLoss, ngraph::op::v4)
 NGRAPH_OP(HSwish, ngraph::op::v4)
 NGRAPH_OP(Interpolate, ngraph::op::v4)
 NGRAPH_OP(Mish, ngraph::op::v4)
-NGRAPH_OP(NonMaxSuppression, ngraph::op::v4)
 NGRAPH_OP(ReduceL1, ngraph::op::v4)
 NGRAPH_OP(ReduceL2, ngraph::op::v4)
 NGRAPH_OP(SoftPlus, ngraph::op::v4)

From 64ce75aebe68aada164b76322a1b6ba7b14b0932 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 16:18:06 +0300
Subject: [PATCH 010/173] Some code style fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 4203730e00d487..2a9ad1040d0222 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -640,7 +640,6 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     constructor_validate_and_infer_types();
 }
 
-
 shared_ptr<Node>
     op::v5::NonMaxSuppression::clone_with_new_inputs(const OutputVector& new_args) const
 {
@@ -730,8 +729,7 @@ void op::v5::NonMaxSuppression::validate()
     {
         const auto soft_nms_sigma = get_input_partial_shape(5);
         NODE_VALIDATION_CHECK(this,
-                              soft_nms_sigma.is_dynamic() ||
-                                  is_scalar(soft_nms_sigma.to_shape()),
+                              soft_nms_sigma.is_dynamic() || is_scalar(soft_nms_sigma.to_shape()),
                               "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
                               soft_nms_sigma);
     }

From 88081f110f5bf12f705c02e96fe5699d8dc94744 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 16:40:30 +0300
Subject: [PATCH 011/173] Started to write the method
 v5::NonMaxSuppression::evaluate().

---
 .../include/ngraph/op/non_max_suppression.hpp   |  2 ++
 ngraph/core/src/op/non_max_suppression.cpp      | 17 +++++++++++++++++
 2 files changed, 19 insertions(+)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 41fbdb86cc0466..5f080503033bdb 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -356,6 +356,8 @@ namespace ngraph
 
                 std::shared_ptr<Node>
                     clone_with_new_inputs(const OutputVector& new_args) const override;
+                bool evaluate(const HostTensorVector& outputs,
+                              const HostTensorVector& inputs) const override;
 
                 BoxEncodingType get_box_encoding() const { return m_box_encoding; }
                 void set_box_encoding(const BoxEncodingType box_encoding)
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 2a9ad1040d0222..49e15bc52e8900 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -831,3 +831,20 @@ namespace ngraph
         return s << as_string(type);
     }
 } // namespace ngraph
+
+bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
+                                         const HostTensorVector& inputs) const
+{
+    element::Type input_et = get_input_element_type(0);
+
+    switch (input_et)
+    {
+    case element::Type_t::f32:
+        break;
+    case element::Type_t::f16:
+        break;
+    default:;
+    }
+
+    return true;
+}

From 9fcc84254d4810ebd046301bb97d92d5aa96c1ce Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 29 Sep 2020 17:47:27 +0300
Subject: [PATCH 012/173] Started to write nGraph reference implementation of
 NMS-5.

---
 .../include/ngraph/op/non_max_suppression.hpp |  3 ++
 .../runtime/reference/non_max_suppression.hpp | 53 +++++++++++++++++++
 ngraph/core/src/op/non_max_suppression.cpp    | 43 ++++++++++++++-
 3 files changed, 98 insertions(+), 1 deletion(-)
 create mode 100644 ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 5f080503033bdb..5ba5b63b52a8a1 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -383,6 +383,9 @@ namespace ngraph
                 ngraph::element::Type m_output_type = ngraph::element::i64;
                 void validate();
                 int64_t max_boxes_output_from_input() const;
+                float iou_threshold_from_input() const;
+                float score_threshold_from_input() const;
+                float soft_nms_sigma_from_input() const;
             };
         } // namespace v5
     }     // namespace op
diff --git a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
new file mode 100644
index 00000000000000..fcdd34713c5640
--- /dev/null
+++ b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
@@ -0,0 +1,53 @@
+//*****************************************************************************
+// Copyright 2017-2020 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <cassert>
+#include <cmath>
+#include <cstddef>
+#include <functional>
+#include <map>
+#include "ngraph/coordinate_transform.hpp"
+#include "ngraph/shape_util.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace reference
+        {
+            template <typename T>
+            void non_max_suppression(const float* boxes_data,
+                                     const Shape& boxes_data_shape,
+                                     const float* scores_data,
+                                     const Shape& scores_data_shape,
+                                     int64_t max_output_boxes_per_class,
+                                     float iou_threshold,
+                                     float score_threshold,
+                                     float soft_nms_sigma,
+                                     T* selected_indices,
+                                     const Shape& selected_indices_shape,
+                                     float* selected_scores,
+                                     const Shape& selected_scores_shape,
+                                     float* valid_outputs)
+            {
+            }
+        }
+    }
+}
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 49e15bc52e8900..9a45acf4b6bcfd 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -650,7 +650,7 @@ shared_ptr<Node>
 
     const auto& arg2 = new_args.size() > 2
                            ? new_args.at(2)
-                           : ngraph::op::Constant::create(element::i32, Shape{}, {0});
+                           : ngraph::op::Constant::create(element::i64, Shape{}, {0});
     const auto& arg3 = new_args.size() > 3
                            ? new_args.at(3)
                            : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
@@ -770,6 +770,43 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
     return max_output_boxes;
 }
 
+static constexpr size_t iou_threshold_port = 3;
+static constexpr size_t score_threshold_port = 4;
+static constexpr size_t soft_nms_sigma_port = 5;
+
+float op::v5::NonMaxSuppression::iou_threshold_from_input() const
+{
+    float iou_threshold = 0.0f;
+
+    const auto iou_threshold_input =
+        as_type_ptr<op::Constant>(input_value(iou_threshold_port).get_node_shared_ptr());
+    iou_threshold = iou_threshold_input->cast_vector<float>().at(0);
+
+    return iou_threshold;
+}
+
+float op::v5::NonMaxSuppression::score_threshold_from_input() const
+{
+    float score_threshold = 0.0f;
+
+    const auto score_threshold_input =
+        as_type_ptr<op::Constant>(input_value(score_threshold_port).get_node_shared_ptr());
+    score_threshold = score_threshold_input->cast_vector<float>().at(0);
+
+    return score_threshold;
+}
+
+float op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const
+{
+    float soft_nms_sigma = 0.0f;
+
+    const auto soft_nms_sigma_input =
+        as_type_ptr<op::Constant>(input_value(soft_nms_sigma_port).get_node_shared_ptr());
+    soft_nms_sigma = soft_nms_sigma_input->cast_vector<float>().at(0);
+
+    return soft_nms_sigma;
+}
+
 bool ngraph::op::v5::NonMaxSuppression::visit_attributes(AttributeVisitor& visitor)
 {
     visitor.on_attribute("box_encoding", m_box_encoding);
@@ -836,6 +873,10 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                          const HostTensorVector& inputs) const
 {
     element::Type input_et = get_input_element_type(0);
+    int64_t max_output_boxes_per_class = max_boxes_output_from_input();
+    float iou_threshold = iou_threshold_from_input();
+    float score_threshold = score_threshold_from_input();
+    float soft_nms_sigma = soft_nms_sigma_from_input();
 
     switch (input_et)
     {

From e610ce1044338f9f93421f8d9b720dec3114f5eb Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 11:12:04 +0300
Subject: [PATCH 013/173] Some additions.

---
 .../runtime/reference/non_max_suppression.hpp |  8 +--
 ngraph/core/src/op/non_max_suppression.cpp    | 62 +++++++++++++++++--
 2 files changed, 59 insertions(+), 11 deletions(-)

diff --git a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
index fcdd34713c5640..6399b574e57d57 100644
--- a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
+++ b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
@@ -24,6 +24,7 @@
 #include <functional>
 #include <map>
 #include "ngraph/coordinate_transform.hpp"
+#include "ngraph/op/util/op_types.hpp"
 #include "ngraph/shape_util.hpp"
 
 namespace ngraph
@@ -32,7 +33,6 @@ namespace ngraph
     {
         namespace reference
         {
-            template <typename T>
             void non_max_suppression(const float* boxes_data,
                                      const Shape& boxes_data_shape,
                                      const float* scores_data,
@@ -41,13 +41,11 @@ namespace ngraph
                                      float iou_threshold,
                                      float score_threshold,
                                      float soft_nms_sigma,
-                                     T* selected_indices,
+                                     int64_t* selected_indices,
                                      const Shape& selected_indices_shape,
                                      float* selected_scores,
                                      const Shape& selected_scores_shape,
-                                     float* valid_outputs)
-            {
-            }
+                                     float* valid_outputs);
         }
     }
 }
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 9a45acf4b6bcfd..9694aad2c3f417 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -18,6 +18,7 @@
 #include "ngraph/attribute_visitor.hpp"
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/util/op_types.hpp"
+#include "ngraph/type/float16.hpp"
 
 using namespace std;
 using namespace ngraph;
@@ -869,23 +870,72 @@ namespace ngraph
     }
 } // namespace ngraph
 
+static PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
+                                                 int64_t max_output_boxes_per_class)
+{
+    const auto boxes_ps = inputs[0]->get_partial_shape();
+    const auto scores_ps = inputs[1]->get_partial_shape();
+
+    // NonMaxSuppression produces triplets
+    // that have the following format: [batch_index, class_index, box_index]
+    PartialShape result = {Dimension::dynamic(), 3};
+
+    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
+    {
+        const auto num_boxes_boxes = boxes_ps[1];
+        const auto max_output_boxes_per_class_node = input_value(2).get_node_shared_ptr();
+        if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
+        {
+            const auto num_boxes = num_boxes_boxes.get_length();
+            const auto num_classes = scores_ps[1].get_length();
+
+            result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
+                        scores_ps[0].get_length();
+        }
+    }
+
+    return result;
+}
+
+static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes, const Shape& boxes_shape)
+{
+    std::vector<float> result;
+    element::Type boxes_input_et = boxes->get_element_type();
+
+    return result;
+}
+
 bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                          const HostTensorVector& inputs) const
 {
     element::Type input_et = get_input_element_type(0);
+
     int64_t max_output_boxes_per_class = max_boxes_output_from_input();
     float iou_threshold = iou_threshold_from_input();
     float score_threshold = score_threshold_from_input();
     float soft_nms_sigma = soft_nms_sigma_from_input();
 
-    switch (input_et)
+    auto selected_indices_shape = infer_selected_indices_shape(inputs, max_output_boxes_per_class);
+    Shape out_shape = selected_indices_shape.to_shape();
+
+    outputs[0]->set_element_type(m_output_type);
+    outputs[0]->set_shape(out_shape);
+
+    size_t num_of_outputs = outputs.size();
+
+    if (num_of_outputs >= 2)
     {
-    case element::Type_t::f32:
-        break;
-    case element::Type_t::f16:
-        break;
-    default:;
+        outputs[1]->set_element_type(element::f32);
+        outputs[1]->set_shape(out_shape);
     }
 
+    if (num_of_outputs >= 3)
+    {
+        outputs[2]->set_element_type(element::f32);
+        outputs[2]->set_shape(Shape{});
+    }
+
+    Shape boxes_shape = outputs[0]->get_shape();
+
     return true;
 }

From 8461e5724d6973ffc4800282b2aba2607bc0f6b2 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 14:01:14 +0300
Subject: [PATCH 014/173] Written preprocessing of boxes data.

---
 ngraph/core/src/op/non_max_suppression.cpp | 111 +++++++++++++++++++--
 1 file changed, 105 insertions(+), 6 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 9694aad2c3f417..11f4cbf40ee561 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -14,6 +14,7 @@
 // limitations under the License.
 //*****************************************************************************
 
+#include <cstring>
 #include "ngraph/op/non_max_suppression.hpp"
 #include "ngraph/attribute_visitor.hpp"
 #include "ngraph/op/constant.hpp"
@@ -771,6 +772,7 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
     return max_output_boxes;
 }
 
+static constexpr size_t boxes_port = 0;
 static constexpr size_t iou_threshold_port = 3;
 static constexpr size_t score_threshold_port = 4;
 static constexpr size_t soft_nms_sigma_port = 5;
@@ -873,7 +875,7 @@ namespace ngraph
 static PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
                                                  int64_t max_output_boxes_per_class)
 {
-    const auto boxes_ps = inputs[0]->get_partial_shape();
+    const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
     const auto scores_ps = inputs[1]->get_partial_shape();
 
     // NonMaxSuppression produces triplets
@@ -883,7 +885,6 @@ static PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
     if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
     {
         const auto num_boxes_boxes = boxes_ps[1];
-        const auto max_output_boxes_per_class_node = input_value(2).get_node_shared_ptr();
         if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
         {
             const auto num_boxes = num_boxes_boxes.get_length();
@@ -897,19 +898,115 @@ static PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
     return result;
 }
 
-static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes, const Shape& boxes_shape)
+using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
+
+static void normalize_corner(float* boxes, const Shape& boxes_shape)
+{
+    size_t num_batches = boxes_shape[0];
+    size_t num_boxes = boxes_shape[1];
+
+    float* box_ptr = boxes;
+
+    for (size_t batch = 0; batch < num_batches; ++batch)
+    {
+        for (size_t box = 0; box < num_boxes; ++num_boxes)
+        {
+            float y1 = box_ptr[0];
+            float x1 = box_ptr[1];
+            float y2 = box_ptr[2];
+            float x2 = box_ptr[3];
+
+            float ymin = std::min(y1, y2);
+            float ymax = std::max(y1, y2);
+            float xmin = std::min(x1, x2);
+            float xmax = std::max(x1, x2);
+
+            box_ptr[0] = ymin;
+            box_ptr[1] = xmin;
+            box_ptr[2] = ymax;
+            box_ptr[3] = xmax;
+
+            box_ptr += 4;
+        }
+    }
+}
+
+static void normalize_center(float* boxes, const Shape& boxes_shape)
+{
+    size_t num_batches = boxes_shape[0];
+    size_t num_boxes = boxes_shape[1];
+
+    float* box_ptr = boxes;
+
+    for (size_t batch = 0; batch < num_batches; ++batch)
+    {
+        for (size_t box = 0; box < num_boxes; ++num_boxes)
+        {
+            float x_center = box_ptr[0];
+            float y_center = box_ptr[1];
+            float width = box_ptr[2];
+            float height = box_ptr[3];
+
+            float y1 = y_center - height / 2.0;
+            float x1 = x_center - width / 2.0;
+            float y2 = y_center + height / 2.0;
+            float x2 = x_center + width / 2.0;;
+
+            box_ptr[0] = y1;
+            box_ptr[1] = x1;
+            box_ptr[2] = y2;
+            box_ptr[3] = x2;
+
+            box_ptr += 4;
+        }
+    }
+}
+
+static void normalize_box_encoding(float* boxes,
+                                   const Shape& boxes_shape,
+                                   const V5BoxEncoding box_encoding)
+{
+    if (box_encoding == V5BoxEncoding::CORNER)
+    {
+        normalize_corner(boxes, boxes_shape);
+    }
+    else
+    {
+        normalize_center(boxes, boxes_shape);
+    }
+}
+
+static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
+                                             const Shape& boxes_shape,
+                                             const V5BoxEncoding box_encoding)
 {
-    std::vector<float> result;
     element::Type boxes_input_et = boxes->get_element_type();
 
+    size_t boxes_size = shape_size(boxes_shape);
+    std::vector<float> result(boxes_size);
+
+    if (boxes_input_et == element::Type_t::f32)
+    {
+        float* boxes_ptr = boxes->get_data_ptr<float>();
+        memcpy(result.data(), boxes_ptr, boxes_size * sizeof(float));
+    }
+    else
+    {
+        float16* boxes_ptr = boxes->get_data_ptr<float16>();
+        for (size_t i = 0; i < boxes_size; ++i)
+        {
+            result[i] = float(boxes_ptr[i]);
+        }
+    }
+
+    normalize_box_encoding(result.data(), boxes_shape, box_encoding);
+
     return result;
 }
 
 bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                          const HostTensorVector& inputs) const
 {
-    element::Type input_et = get_input_element_type(0);
-
     int64_t max_output_boxes_per_class = max_boxes_output_from_input();
     float iou_threshold = iou_threshold_from_input();
     float score_threshold = score_threshold_from_input();
@@ -937,5 +1034,7 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
 
     Shape boxes_shape = outputs[0]->get_shape();
 
+    auto boxes_data = prepare_boxes_data(inputs[boxes_port], boxes_shape, m_box_encoding);
+
     return true;
 }

From 06b2edda74f910fb260ec15a4cede60be7a7a0a7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 14:47:30 +0300
Subject: [PATCH 015/173] Started to write the function non_max_suppression()
 that calculates NonMaxSuppression-5 operation.

---
 .../runtime/reference/non_max_suppression.hpp |  2 +-
 .../runtime/reference/non_max_suppression.cpp | 37 ++++++++++++++
 ngraph/core/src/op/non_max_suppression.cpp    | 51 ++++++++++++++++++-
 3 files changed, 87 insertions(+), 3 deletions(-)
 create mode 100644 ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp

diff --git a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
index 6399b574e57d57..9f76239f08753c 100644
--- a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
+++ b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
@@ -45,7 +45,7 @@ namespace ngraph
                                      const Shape& selected_indices_shape,
                                      float* selected_scores,
                                      const Shape& selected_scores_shape,
-                                     float* valid_outputs);
+                                     int64_t* valid_outputs);
         }
     }
 }
diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
new file mode 100644
index 00000000000000..249b46825085cd
--- /dev/null
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -0,0 +1,37 @@
+//*****************************************************************************
+// Copyright 2017-2020 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+
+#include "ngraph/op/non_max_suppression.hpp"
+#include "ngraph/runtime/reference/non_max_suppression.hpp"
+
+using namespace ngraph::runtime::reference;
+
+void non_max_suppression(const float* boxes_data,
+                         const Shape& boxes_data_shape,
+                         const float* scores_data,
+                         const Shape& scores_data_shape,
+                         int64_t max_output_boxes_per_class,
+                         float iou_threshold,
+                         float score_threshold,
+                         float soft_nms_sigma,
+                         int64_t* selected_indices,
+                         const Shape& selected_indices_shape,
+                         float* selected_scores,
+                         const Shape& selected_scores_shape,
+                         int64_t* valid_outputs)
+{
+}
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 11f4cbf40ee561..80d1eedb7a679c 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -19,6 +19,7 @@
 #include "ngraph/attribute_visitor.hpp"
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/util/op_types.hpp"
+#include "ngraph/runtime/reference/non_max_suppression.hpp"
 #include "ngraph/type/float16.hpp"
 
 using namespace std;
@@ -773,6 +774,7 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
 }
 
 static constexpr size_t boxes_port = 0;
+static constexpr size_t scores_port = 1;
 static constexpr size_t iou_threshold_port = 3;
 static constexpr size_t score_threshold_port = 4;
 static constexpr size_t soft_nms_sigma_port = 5;
@@ -985,7 +987,7 @@ static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
     size_t boxes_size = shape_size(boxes_shape);
     std::vector<float> result(boxes_size);
 
-    if (boxes_input_et == element::Type_t::f32)
+    if (boxes_input_et == ngraph::element::f32)
     {
         float* boxes_ptr = boxes->get_data_ptr<float>();
         memcpy(result.data(), boxes_ptr, boxes_size * sizeof(float));
@@ -1004,6 +1006,31 @@ static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
     return result;
 }
 
+static std::vector<float> prepare_scores_data(const HostTensorPtr& scores,
+                                              const Shape& scores_shape)
+{
+    element::Type scores_input_et = scores->get_element_type();
+
+    size_t scores_size = shape_size(scores_shape);
+    std::vector<float> result(scores_size);
+
+    if (scores_input_et == ngraph::element::f32)
+    {
+        float* scores_ptr = scores->get_data_ptr<float>();
+        memcpy(result.data(), scores_ptr, scores_size * sizeof(float));
+    }
+    else
+    {
+        float16* scores_ptr = scores->get_data_ptr<float16>();
+        for (size_t i = 0; i < scores_size; ++i)
+        {
+            result[i] = float(scores_ptr[i]);
+        }
+    }
+
+    return result;
+}
+
 bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                          const HostTensorVector& inputs) const
 {
@@ -1032,9 +1059,29 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
         outputs[2]->set_shape(Shape{});
     }
 
-    Shape boxes_shape = outputs[0]->get_shape();
+    Shape boxes_shape = inputs[boxes_port]->get_shape();
+    Shape scores_shape = inputs[scores_port]->get_shape();
 
     auto boxes_data = prepare_boxes_data(inputs[boxes_port], boxes_shape, m_box_encoding);
+    auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
+
+    std::vector<int64_t> selected_indices(out_shape);
+    std::vector<float> selected_scores(out_shape);
+    int64_t valid_outputs = 0;
+
+    runtime::reference::non_max_suppression(boxes_data,
+                                            boxes_shape,
+                                            scores_data,
+                                            scores_shape,
+                                            max_output_boxes_per_class,
+                                            iou_threshold,
+                                            score_threshold,
+                                            soft_nms_sigma,
+                                            selected_indices.data(),
+                                            out_shape,
+                                            selected_scores.data(),
+                                            out_shape,
+                                            &valid_outputs);
 
     return true;
 }

From 06da1dd86bffc73807527416de03fd028f6b239f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 15:14:57 +0300
Subject: [PATCH 016/173] Written postprocessing of the evaluate().

---
 ngraph/core/src/op/non_max_suppression.cpp | 60 +++++++++++++++++++++-
 1 file changed, 58 insertions(+), 2 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 80d1eedb7a679c..6389a3344b4bf8 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -1031,6 +1031,54 @@ static std::vector<float> prepare_scores_data(const HostTensorPtr& scores,
     return result;
 }
 
+static void evaluate_postprocessing(const HostTensorVector& outputs,
+                                    const ngraph::element::Type output_type,
+                                    const std::vector<int64_t>& selected_indices,
+                                    const std::vector<float>& selected_scores,
+                                    int64_t valid_outputs)
+{
+    size_t num_of_outputs = outputs.size();
+
+    if (output_type == ngraph::element::i64)
+    {
+        int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
+        memcpy(indices_ptr, selected_indices.data(), selected_indices.size() * sizeof(int64_t));
+    }
+    else
+    {
+        int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
+        for (int64_t idx : selected_indices)
+        {
+            *indices_ptr = static_cast<int32_t>(idx);
+            indices_ptr++;
+        }
+    }
+
+    if (num_of_outputs < 2)
+    {
+        return;
+    }
+
+    float* scores_ptr = outputs[1]->get_data_ptr<float>();
+    memcpy(scores_ptr, selected_scores.data(), selected_scores.size() * sizeof(float));
+
+    if (num_of_outputs < 3)
+    {
+        return;
+    }
+
+    if (output_type == ngraph::element::i64)
+    {
+        int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
+        *valid_outputs_ptr = valid_outputs;
+    }
+    else
+    {
+        int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
+        *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
+    }
+}
+
 bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                          const HostTensorVector& inputs) const
 {
@@ -1065,8 +1113,10 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
     auto boxes_data = prepare_boxes_data(inputs[boxes_port], boxes_shape, m_box_encoding);
     auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
 
-    std::vector<int64_t> selected_indices(out_shape);
-    std::vector<float> selected_scores(out_shape);
+    size_t out_shape_size = shape_size(out_shape);
+
+    std::vector<int64_t> selected_indices(out_shape_size);
+    std::vector<float> selected_scores(out_shape_size);
     int64_t valid_outputs = 0;
 
     runtime::reference::non_max_suppression(boxes_data,
@@ -1083,5 +1133,11 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                             out_shape,
                                             &valid_outputs);
 
+    evaluate_postprocessing(outputs,
+                            m_output_type,
+                            selected_indices,
+                            selected_scores,
+                            valid_outputs);
+
     return true;
 }

From 0b6ca39a06de3aa97bce4ff37d54be071ff7bda3 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 15:27:50 +0300
Subject: [PATCH 017/173] Small fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 6389a3344b4bf8..b1363822e69cb1 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -1119,9 +1119,9 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
     std::vector<float> selected_scores(out_shape_size);
     int64_t valid_outputs = 0;
 
-    runtime::reference::non_max_suppression(boxes_data,
+    runtime::reference::non_max_suppression(boxes_data.data(),
                                             boxes_shape,
-                                            scores_data,
+                                            scores_data.data(),
                                             scores_shape,
                                             max_output_boxes_per_class,
                                             iou_threshold,

From 78b946d2a1a0cee7640fe91d3f377f406ed1c4e3 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 15:41:19 +0300
Subject: [PATCH 018/173] Small fix.

---
 .../core/reference/src/runtime/reference/non_max_suppression.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 249b46825085cd..54ab43892462ac 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -17,6 +17,7 @@
 
 #include "ngraph/op/non_max_suppression.hpp"
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
+#include "ngraph/shape_util.hpp"
 
 using namespace ngraph::runtime::reference;
 

From 14db60daf34aeca96f7f384893a3e752a6fa6305 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 30 Sep 2020 15:53:53 +0300
Subject: [PATCH 019/173] Added include for ngraph/shape.hpp.

---
 .../reference/src/runtime/reference/non_max_suppression.cpp    | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 54ab43892462ac..a0aa13880d677b 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -17,8 +17,9 @@
 
 #include "ngraph/op/non_max_suppression.hpp"
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
-#include "ngraph/shape_util.hpp"
+#include "ngraph/shape.hpp"
 
+using namespace ngraph;
 using namespace ngraph::runtime::reference;
 
 void non_max_suppression(const float* boxes_data,

From 951815aaf029c7c6e06ceef791f1aa549985b96e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 10:25:07 +0300
Subject: [PATCH 020/173] Written the function intersectionOverUnion.

---
 .../runtime/reference/non_max_suppression.cpp | 67 +++++++++++++++++++
 1 file changed, 67 insertions(+)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index a0aa13880d677b..8a0542b0c8c3ae 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -15,6 +15,10 @@
 //*****************************************************************************
 
 
+#include <algorithm>
+#include <cmath>
+#include <functional>
+#include <vector>
 #include "ngraph/op/non_max_suppression.hpp"
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
 #include "ngraph/shape.hpp"
@@ -22,6 +26,46 @@
 using namespace ngraph;
 using namespace ngraph::runtime::reference;
 
+struct Rectangle
+{
+    float y1;
+    float x1;
+    float y2;
+    float x2;
+};
+
+static float intersectionOverUnion(float* boxesI, float* boxesJ)
+{
+    Rectangle boxI = *(reinterpret_cast<Rectangle*>(boxesI));
+    Rectangle boxJ = *(reinterpret_cast<Rectangle*>(boxesJ));
+
+    float areaI = (boxI.y2 - boxI.y1) * (boxI.x2 - boxI.x1);
+    float areaJ = (boxJ.y2 - boxJ.y1) * (boxJ.x2 - boxJ.x1);
+
+    if (areaI <= 0.0f || areaJ <= 0.0f)
+    {
+        return 0.0f;
+    }
+
+    float intersection_ymin = std::max(boxI.y1, boxJ.y1);
+    float intersection_xmin = std::max(boxI.x1, boxJ.x1);
+    float intersection_ymax = std::min(boxI.y2, boxJ.y2);
+    float intersection_xmax = std::min(boxI.x2, boxJ.x2);
+
+    float intersection_area = std::max(intersection_ymax - intersection_ymin, 0.0f) *
+        std::max(intersection_xmax - intersection_xmin, 0.0f);
+
+    return intersection_area / (areaI + areaJ - intersection_area);
+}
+
+struct FilteredBoxes
+{
+    float score;
+    int batch_index;
+    int class_index;
+    int box_index;
+};
+
 void non_max_suppression(const float* boxes_data,
                          const Shape& boxes_data_shape,
                          const float* scores_data,
@@ -36,4 +80,27 @@ void non_max_suppression(const float* boxes_data,
                          const Shape& selected_scores_shape,
                          int64_t* valid_outputs)
 {
+    float scale = 0.0f;
+    if (soft_nms_sigma > 0.0f) {
+        scale = - 0.5f / soft_nms_sigma;
+    }
+
+    auto func = [iou_threshold, scale](float iou) {
+        const float weight = std::exp(scale * iou * iou)));
+        return iou <= iou_threshold ? weight : 0.0f;
+    };
+
+    // boxes shape: {num_batches, num_boxes, 4}
+    // scores shape: {num_batches, num_classes, num_boxes}
+    size_t num_batches = scores_data_shape[0];
+    size_t num_classes = scores_data_shape[1];
+
+    std::vector<FilteredBoxes> fb;
+
+    for (int batch = 0; batch < num_batches; batch++)
+    {
+        for (int class_idx = 0; class_idx < num_classes; class_idx++)
+        {
+        }
+    }
 }

From 8dd9b5610cf0c6caa15b08fb3a8feee8b637cfb7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 10:44:07 +0300
Subject: [PATCH 021/173] Some additions.

---
 .../runtime/reference/non_max_suppression.cpp | 73 +++++++++++--------
 1 file changed, 41 insertions(+), 32 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 8a0542b0c8c3ae..33f20ab177c60b 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -66,41 +66,50 @@ struct FilteredBoxes
     int box_index;
 };
 
-void non_max_suppression(const float* boxes_data,
-                         const Shape& boxes_data_shape,
-                         const float* scores_data,
-                         const Shape& scores_data_shape,
-                         int64_t max_output_boxes_per_class,
-                         float iou_threshold,
-                         float score_threshold,
-                         float soft_nms_sigma,
-                         int64_t* selected_indices,
-                         const Shape& selected_indices_shape,
-                         float* selected_scores,
-                         const Shape& selected_scores_shape,
-                         int64_t* valid_outputs)
+namespace ngraph
 {
-    float scale = 0.0f;
-    if (soft_nms_sigma > 0.0f) {
-        scale = - 0.5f / soft_nms_sigma;
-    }
-
-    auto func = [iou_threshold, scale](float iou) {
-        const float weight = std::exp(scale * iou * iou)));
-        return iou <= iou_threshold ? weight : 0.0f;
-    };
-
-    // boxes shape: {num_batches, num_boxes, 4}
-    // scores shape: {num_batches, num_classes, num_boxes}
-    size_t num_batches = scores_data_shape[0];
-    size_t num_classes = scores_data_shape[1];
-
-    std::vector<FilteredBoxes> fb;
-
-    for (int batch = 0; batch < num_batches; batch++)
+    namespace runtime
     {
-        for (int class_idx = 0; class_idx < num_classes; class_idx++)
+        namespace reference
         {
+            void non_max_suppression(const float* boxes_data,
+                                     const Shape& boxes_data_shape,
+                                     const float* scores_data,
+                                     const Shape& scores_data_shape,
+                                     int64_t max_output_boxes_per_class,
+                                     float iou_threshold,
+                                     float score_threshold,
+                                     float soft_nms_sigma,
+                                     int64_t* selected_indices,
+                                     const Shape& selected_indices_shape,
+                                     float* selected_scores,
+                                     const Shape& selected_scores_shape,
+                                     int64_t* valid_outputs)
+            {
+                float scale = 0.0f;
+                if (soft_nms_sigma > 0.0f) {
+                    scale = - 0.5f / soft_nms_sigma;
+                }
+
+                auto func = [iou_threshold, scale](float iou) {
+                    const float weight = std::exp(scale * iou * iou)));
+                    return iou <= iou_threshold ? weight : 0.0f;
+                };
+
+                // boxes shape: {num_batches, num_boxes, 4}
+                // scores shape: {num_batches, num_classes, num_boxes}
+                size_t num_batches = scores_data_shape[0];
+                size_t num_classes = scores_data_shape[1];
+
+                std::vector<FilteredBoxes> fb;
+
+                for (int batch = 0; batch < num_batches; batch++)
+                {
+                    for (int class_idx = 0; class_idx < num_classes; class_idx++)
+                    {
+                    }
+                }
+            }
         }
     }
 }

From 28076b47f67341f639f11b33a47dcfb281238316 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 10:53:57 +0300
Subject: [PATCH 022/173] Small fix.

---
 .../reference/src/runtime/reference/non_max_suppression.cpp     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 33f20ab177c60b..50749dbc9b004e 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -92,7 +92,7 @@ namespace ngraph
                 }
 
                 auto func = [iou_threshold, scale](float iou) {
-                    const float weight = std::exp(scale * iou * iou)));
+                    const float weight = std::exp(scale * iou * iou);
                     return iou <= iou_threshold ? weight : 0.0f;
                 };
 

From f13d8b279e87a0d163988c925fe67927df93f135 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 16:23:39 +0300
Subject: [PATCH 023/173] Continued to write the function
 non_max_suppression().

---
 .../runtime/reference/non_max_suppression.cpp | 167 ++++++++++++++++--
 1 file changed, 150 insertions(+), 17 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 50749dbc9b004e..aa982c804986c9 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -16,9 +16,10 @@
 
 
 #include <algorithm>
+#include <array>
 #include <cmath>
-#include <functional>
 #include <vector>
+#include <queue>
 #include "ngraph/op/non_max_suppression.hpp"
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
 #include "ngraph/shape.hpp"
@@ -28,17 +29,24 @@ using namespace ngraph::runtime::reference;
 
 struct Rectangle
 {
-    float y1;
-    float x1;
-    float y2;
-    float x2;
+    Rectangle(float y_left, float x_left, float y_right, float x_right)
+        : y1{y_left}
+        , x1{x_left}
+        , y2{y_right}
+        , x2{x_right}
+        {
+        }
+
+    Rectangle() = default;
+
+    float y1 = 0.0f;
+    float x1 = 0.0f;
+    float y2 = 0.f;
+    float x2 = 0.0f;
 };
 
-static float intersectionOverUnion(float* boxesI, float* boxesJ)
+static float intersectionOverUnion(const Rectangle& boxI, const Rectangle& boxJ)
 {
-    Rectangle boxI = *(reinterpret_cast<Rectangle*>(boxesI));
-    Rectangle boxJ = *(reinterpret_cast<Rectangle*>(boxesJ));
-
     float areaI = (boxI.y2 - boxI.y1) * (boxI.x2 - boxI.x1);
     float areaJ = (boxJ.y2 - boxJ.y1) * (boxJ.x2 - boxJ.x1);
 
@@ -58,12 +66,61 @@ static float intersectionOverUnion(float* boxesI, float* boxesJ)
     return intersection_area / (areaI + areaJ - intersection_area);
 }
 
-struct FilteredBoxes
+struct SelectedIndex
+{
+  SelectedIndex(int64_t batch_idx, int64_t class_idx, int64_t box_idx)
+      : batch_index(batch_idx)
+      , class_index(class_idx)
+      , box_index(box_idx)
+      {
+      }
+
+  SelectedIndex() = default;
+
+  int64_t batch_index = 0;
+  int64_t class_index = 0;
+  int64_t box_index = 0;
+};
+
+struct SelectedScore
+{
+  SelectedScore(float batch_idx, float class_idx, float score)
+      : batch_index{batch_idx}
+      , class_index{class_idx}
+      , box_score{score}
+      {
+      }
+
+  SelectedScore() = default;
+
+  float batch_index = 0.0f;
+  float class_index = 0.0f;
+  float box_score = 0.0f;
+};
+
+struct BoxInfo
 {
-    float score;
-    int batch_index;
-    int class_index;
-    int box_index;
+    BoxInfo(const Rectangle& r, int64_t idx, float sc, int64_t suppress_idx)
+        : box{r}
+        , index{idx}
+        , score{sc}
+        , suppress_begin_index{suppress_idx}
+    {
+      area = (box.y2 - box.y1) * (box.x2 - box.x1);
+    }
+
+    BoxInfo() = default;
+
+    inline bool operator<(const BoxInfo& rhs) const
+    {
+        return score < rhs.score || (score == rhs.score && index > rhs.index);
+    }
+
+    Rectangle box;
+    int64_t index = 0;
+    float score = 0.0f;
+    int64_t suppress_begin_index = 0;
+    float area = 0.0f;
 };
 
 namespace ngraph
@@ -100,13 +157,89 @@ namespace ngraph
                 // scores shape: {num_batches, num_classes, num_boxes}
                 size_t num_batches = scores_data_shape[0];
                 size_t num_classes = scores_data_shape[1];
+                size_t num_boxes = boxes_data_shape[1];
+
+                SelectedIndex* selected_indices_ptr =
+                    reinterpret_cast<SelectedIndex*>(selected_indices);
+                SelectedScore* selected_scores_ptr =
+                    reinterpret_cast<SelectedScore*>(selected_scores);
+
+                size_t boxesStrides = num_boxes * 4;
+                std::array<size_t, 2> scoresStrides = {num_classes * num_boxes, num_boxes};
 
-                std::vector<FilteredBoxes> fb;
+                size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
 
-                for (int batch = 0; batch < num_batches; batch++)
+                for (size_t batch = 0; batch < num_batches; batch++)
                 {
-                    for (int class_idx = 0; class_idx < num_classes; class_idx++)
+                    float* boxesPtr = boxes + batch * boxesStrides;
+                    for (size_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
+                        float* scoresPtr = scores + batch * scoresStrides[0] +
+                                           class_idx * scoresStrides[1];
+                        std::vector<BoxInfo> candidate_boxes;
+                        candidate_boxes.reserve(num_boxes);
+
+                        for (size_t box_idx = 0; box_idx < num_boxes; box_idx++)
+                        {
+                            if (scoresPtr[box_idx] > score_threshold)
+                            {
+                                Rectangle* r = reinterpret_cast<Rectangle*>(boxesPtr) + box_idx;
+                                candidate_boxes.emplace_back(*r, box_idx, scoresPtr[box_idx], 0);
+                            }
+                        }
+
+                        std::priority_queue<BoxInfo> sorted_boxes(std::less<BoxInfo>(),
+                                                                  std::move(candidate_boxes));
+
+                        std::vector<BoxInfo> selected;
+                        // Get the next box with top score, filter by iou_threshold
+
+                        BoxInfo next_candidate;
+                        float original_score;
+
+                        while (!sorted_boxes.empty() && selected.size() < boxes_per_class)
+                        {
+                            next_candidate = sorted_boxes.top();
+                            original_score = next_candidate.score;
+                            sorted_boxes.pop();
+
+                            bool should_hard_suppress = false;
+                            for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
+                                 j >= next_candidate.suppress_begin_index; --j)
+                            {
+                                float iou = intersectionOverUnion(next_candidate.box, selected[j].box);
+                                next_candidate.score *= func(iou);
+
+                                if (iou >= iou_threshold)
+                                {
+                                    should_hard_suppress = true
+                                    break;
+                                }
+
+                                if (next_candidate.score <= score_threshold)
+                                {
+                                    break;
+                                }
+                            }
+
+                            next_candidate.suppress_begin_index = selected.size();
+
+                            if (!should_hard_suppress) {
+                                if (next_candidate.score == original_score) {
+                                    // // Suppression has not occurred, so select next_candidate
+                                    // selected.push_back(next_candidate.box_index);
+                                    // selected_scores.push_back(next_candidate.score);
+                                    // continue;
+                                }
+                                if (next_candidate.score > score_threshold) {
+                                    // Soft suppression has occurred and current score is still greater than
+                                    // score_threshold; add next_candidate back onto priority queue.
+                                    candidate_priority_queue.push(next_candidate);
+                                }
+                            }
+                        }
+
+
                     }
                 }
             }

From 8287813f4f2b5ccfe929652ec8b668c31b5e428d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 16:35:29 +0300
Subject: [PATCH 024/173] Small fixes.

---
 .../runtime/reference/non_max_suppression.cpp   | 17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index aa982c804986c9..17a8a66a216ce4 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -171,10 +171,10 @@ namespace ngraph
 
                 for (size_t batch = 0; batch < num_batches; batch++)
                 {
-                    float* boxesPtr = boxes + batch * boxesStrides;
+                    float* boxesPtr = boxes_data + batch * boxesStrides;
                     for (size_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
-                        float* scoresPtr = scores + batch * scoresStrides[0] +
+                        float* scoresPtr = scores_data + batch * scoresStrides[0] +
                                            class_idx * scoresStrides[1];
                         std::vector<BoxInfo> candidate_boxes;
                         candidate_boxes.reserve(num_boxes);
@@ -212,7 +212,7 @@ namespace ngraph
 
                                 if (iou >= iou_threshold)
                                 {
-                                    should_hard_suppress = true
+                                    should_hard_suppress = true;
                                     break;
                                 }
 
@@ -224,17 +224,20 @@ namespace ngraph
 
                             next_candidate.suppress_begin_index = selected.size();
 
-                            if (!should_hard_suppress) {
-                                if (next_candidate.score == original_score) {
+                            if (!should_hard_suppress)
+                            {
+                                if (next_candidate.score == original_score)
+                                {
                                     // // Suppression has not occurred, so select next_candidate
                                     // selected.push_back(next_candidate.box_index);
                                     // selected_scores.push_back(next_candidate.score);
                                     // continue;
                                 }
-                                if (next_candidate.score > score_threshold) {
+                                if (next_candidate.score > score_threshold)
+                                {
                                     // Soft suppression has occurred and current score is still greater than
                                     // score_threshold; add next_candidate back onto priority queue.
-                                    candidate_priority_queue.push(next_candidate);
+                                    sorted_boxes.push(next_candidate);
                                 }
                             }
                         }

From 8097591307cf0bc646abbbe12bc95c296409ac8c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 16:44:23 +0300
Subject: [PATCH 025/173] Small fixes.

---
 .../src/runtime/reference/non_max_suppression.cpp          | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 17a8a66a216ce4..0197bee22eb152 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -171,11 +171,12 @@ namespace ngraph
 
                 for (size_t batch = 0; batch < num_batches; batch++)
                 {
-                    float* boxesPtr = boxes_data + batch * boxesStrides;
+                     const float* boxesPtr = boxes_data + batch * boxesStrides;
                     for (size_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
-                        float* scoresPtr = scores_data + batch * scoresStrides[0] +
-                                           class_idx * scoresStrides[1];
+                        const float* scoresPtr = scores_data + batch * scoresStrides[0] +
+                                                 class_idx * scoresStrides[1];
+
                         std::vector<BoxInfo> candidate_boxes;
                         candidate_boxes.reserve(num_boxes);
 

From 3a1b8ea3db0b14e35994a3fde0463f29db674f75 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 17:02:46 +0300
Subject: [PATCH 026/173] Some changes.

---
 .../runtime/reference/non_max_suppression.cpp   | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 0197bee22eb152..133d03cbc9c9ac 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -103,10 +103,9 @@ struct BoxInfo
     BoxInfo(const Rectangle& r, int64_t idx, float sc, int64_t suppress_idx)
         : box{r}
         , index{idx}
-        , score{sc}
         , suppress_begin_index{suppress_idx}
+        , score{sc}
     {
-      area = (box.y2 - box.y1) * (box.x2 - box.x1);
     }
 
     BoxInfo() = default;
@@ -118,9 +117,8 @@ struct BoxInfo
 
     Rectangle box;
     int64_t index = 0;
-    float score = 0.0f;
     int64_t suppress_begin_index = 0;
-    float area = 0.0f;
+    float score = 0.0f;
 };
 
 namespace ngraph
@@ -171,7 +169,9 @@ namespace ngraph
 
                 for (size_t batch = 0; batch < num_batches; batch++)
                 {
-                     const float* boxesPtr = boxes_data + batch * boxesStrides;
+                    const float* boxesPtr = boxes_data + batch * boxesStrides;
+                    Rectangle* r = reinterpret_cast<Rectangle*>(const_cast<float*>(boxesPtr));
+
                     for (size_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
                         const float* scoresPtr = scores_data + batch * scoresStrides[0] +
@@ -184,8 +184,7 @@ namespace ngraph
                         {
                             if (scoresPtr[box_idx] > score_threshold)
                             {
-                                Rectangle* r = reinterpret_cast<Rectangle*>(boxesPtr) + box_idx;
-                                candidate_boxes.emplace_back(*r, box_idx, scoresPtr[box_idx], 0);
+                                candidate_boxes.emplace_back(r[box_idx], box_idx, scoresPtr[box_idx], 0);
                             }
                         }
 
@@ -229,8 +228,8 @@ namespace ngraph
                             {
                                 if (next_candidate.score == original_score)
                                 {
-                                    // // Suppression has not occurred, so select next_candidate
-                                    // selected.push_back(next_candidate.box_index);
+                                    // Suppression has not occurred, so select next_candidate
+                                    selected.push_back(next_candidate.box_index);
                                     // selected_scores.push_back(next_candidate.score);
                                     // continue;
                                 }

From 49f775dae415d4f929680a724aca37c3d1a013e4 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 1 Oct 2020 17:41:28 +0300
Subject: [PATCH 027/173] Some additions.

---
 .../runtime/reference/non_max_suppression.hpp |  3 +-
 .../runtime/reference/non_max_suppression.cpp | 28 ++++++++++++++++---
 ngraph/core/src/op/non_max_suppression.cpp    |  3 +-
 3 files changed, 28 insertions(+), 6 deletions(-)

diff --git a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
index 9f76239f08753c..aa5725650f5da2 100644
--- a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
+++ b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
@@ -45,7 +45,8 @@ namespace ngraph
                                      const Shape& selected_indices_shape,
                                      float* selected_scores,
                                      const Shape& selected_scores_shape,
-                                     int64_t* valid_outputs);
+                                     int64_t* valid_outputs,
+                                     const bool sort_result_descending);
         }
     }
 }
diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 133d03cbc9c9ac..d63a82548a655e 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -139,7 +139,8 @@ namespace ngraph
                                      const Shape& selected_indices_shape,
                                      float* selected_scores,
                                      const Shape& selected_scores_shape,
-                                     int64_t* valid_outputs)
+                                     int64_t* valid_outputs,
+                                     const bool sort_result_descending)
             {
                 float scale = 0.0f;
                 if (soft_nms_sigma > 0.0f) {
@@ -167,6 +168,8 @@ namespace ngraph
 
                 size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
 
+                int64_t num_of_valid_boxes = 0;
+
                 for (size_t batch = 0; batch < num_batches; batch++)
                 {
                     const float* boxesPtr = boxes_data + batch * boxesStrides;
@@ -229,9 +232,8 @@ namespace ngraph
                                 if (next_candidate.score == original_score)
                                 {
                                     // Suppression has not occurred, so select next_candidate
-                                    selected.push_back(next_candidate.box_index);
-                                    // selected_scores.push_back(next_candidate.score);
-                                    // continue;
+                                    selected.push_back(next_candidate.index);
+                                    continue;
                                 }
                                 if (next_candidate.score > score_threshold)
                                 {
@@ -242,9 +244,27 @@ namespace ngraph
                             }
                         }
 
+                        for (const auto& box_info : selected)
+                        {
+                            SelectedIndex selected_index{batch, class_idx, box_info.index};
+                            SelectedScore selected_score{batch, class_idx, box_info.score};
+
+                            selected_indices_ptr[num_of_valid_boxes] = selected_index;
+                            selected_scores_ptr[num_of_valid_boxes] = selected_score;
 
+                            num_of_valid_boxes++;
+                        }
                     }
                 }
+
+
+                if (sort_result_descending)
+                {
+                }
+                else
+                {
+                }
+
             }
         }
     }
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index b1363822e69cb1..41814afdf16480 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -1131,7 +1131,8 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                             out_shape,
                                             selected_scores.data(),
                                             out_shape,
-                                            &valid_outputs);
+                                            &valid_outputs,
+                                            m_sort_result_descending);
 
     evaluate_postprocessing(outputs,
                             m_output_type,

From 4ff9f60ecd0c40625d70d6b88036608906fc9b8d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 09:32:27 +0300
Subject: [PATCH 028/173] Some replacements size_t by int64_t.

---
 .../runtime/reference/non_max_suppression.cpp | 22 ++++++++-----------
 1 file changed, 9 insertions(+), 13 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index d63a82548a655e..dfccbbfd29b08f 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -16,7 +16,6 @@
 
 
 #include <algorithm>
-#include <array>
 #include <cmath>
 #include <vector>
 #include <queue>
@@ -154,31 +153,28 @@ namespace ngraph
 
                 // boxes shape: {num_batches, num_boxes, 4}
                 // scores shape: {num_batches, num_classes, num_boxes}
-                size_t num_batches = scores_data_shape[0];
-                size_t num_classes = scores_data_shape[1];
-                size_t num_boxes = boxes_data_shape[1];
+                int64_t num_batches = static_cast<int64_t>(scores_data_shape[0]);
+                int64_t num_classes = static_cast<int64_t>(scores_data_shape[1]);
+                int64_t num_boxes = static_cast<int64_t>(boxes_data_shape[1]);
 
                 SelectedIndex* selected_indices_ptr =
                     reinterpret_cast<SelectedIndex*>(selected_indices);
                 SelectedScore* selected_scores_ptr =
                     reinterpret_cast<SelectedScore*>(selected_scores);
 
-                size_t boxesStrides = num_boxes * 4;
-                std::array<size_t, 2> scoresStrides = {num_classes * num_boxes, num_boxes};
-
                 size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
 
                 int64_t num_of_valid_boxes = 0;
 
-                for (size_t batch = 0; batch < num_batches; batch++)
+                for (int64_t batch = 0; batch < num_batches; batch++)
                 {
-                    const float* boxesPtr = boxes_data + batch * boxesStrides;
+                    const float* boxesPtr = boxes_data + batch * num_boxes * 4;
                     Rectangle* r = reinterpret_cast<Rectangle*>(const_cast<float*>(boxesPtr));
 
-                    for (size_t class_idx = 0; class_idx < num_classes; class_idx++)
+                    for (int64_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
-                        const float* scoresPtr = scores_data + batch * scoresStrides[0] +
-                                                 class_idx * scoresStrides[1];
+                        const float* scoresPtr = scores_data + batch * (num_classes * num_boxes) +
+                                                 class_idx * num_boxes;
 
                         std::vector<BoxInfo> candidate_boxes;
                         candidate_boxes.reserve(num_boxes);
@@ -232,7 +228,7 @@ namespace ngraph
                                 if (next_candidate.score == original_score)
                                 {
                                     // Suppression has not occurred, so select next_candidate
-                                    selected.push_back(next_candidate.index);
+                                    selected.push_back(next_candidate);
                                     continue;
                                 }
                                 if (next_candidate.score > score_threshold)

From a6872da15da56a9857e0e153a014d82f5d9abc5c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 09:48:35 +0300
Subject: [PATCH 029/173] Added casts to float in the construction of
 selected_score variable.

---
 .../reference/src/runtime/reference/non_max_suppression.cpp    | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index dfccbbfd29b08f..9b53edaa21a75a 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -243,7 +243,8 @@ namespace ngraph
                         for (const auto& box_info : selected)
                         {
                             SelectedIndex selected_index{batch, class_idx, box_info.index};
-                            SelectedScore selected_score{batch, class_idx, box_info.score};
+                            SelectedScore selected_score{static_cast<float>(batch),
+                                static_cast<float>(class_idx), box_info.score};
 
                             selected_indices_ptr[num_of_valid_boxes] = selected_index;
                             selected_scores_ptr[num_of_valid_boxes] = selected_score;

From f64fcede93d14591f5b6843c12865b3815030bc7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 10:11:51 +0300
Subject: [PATCH 030/173] Code style fixes.

---
 .../runtime/reference/non_max_suppression.cpp | 80 ++++++++++---------
 ngraph/core/src/op/non_max_suppression.cpp    | 16 ++--
 2 files changed, 47 insertions(+), 49 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 9b53edaa21a75a..7f530d655b59eb 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -15,11 +15,11 @@
 //*****************************************************************************
 
 
+#include "ngraph/op/non_max_suppression.hpp"
 #include <algorithm>
 #include <cmath>
-#include <vector>
 #include <queue>
-#include "ngraph/op/non_max_suppression.hpp"
+#include <vector>
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
 #include "ngraph/shape.hpp"
 
@@ -33,8 +33,8 @@ struct Rectangle
         , x1{x_left}
         , y2{y_right}
         , x2{x_right}
-        {
-        }
+    {
+    }
 
     Rectangle() = default;
 
@@ -60,41 +60,41 @@ static float intersectionOverUnion(const Rectangle& boxI, const Rectangle& boxJ)
     float intersection_xmax = std::min(boxI.x2, boxJ.x2);
 
     float intersection_area = std::max(intersection_ymax - intersection_ymin, 0.0f) *
-        std::max(intersection_xmax - intersection_xmin, 0.0f);
+                              std::max(intersection_xmax - intersection_xmin, 0.0f);
 
     return intersection_area / (areaI + areaJ - intersection_area);
 }
 
 struct SelectedIndex
 {
-  SelectedIndex(int64_t batch_idx, int64_t class_idx, int64_t box_idx)
-      : batch_index(batch_idx)
-      , class_index(class_idx)
-      , box_index(box_idx)
-      {
-      }
-
-  SelectedIndex() = default;
-
-  int64_t batch_index = 0;
-  int64_t class_index = 0;
-  int64_t box_index = 0;
+    SelectedIndex(int64_t batch_idx, int64_t class_idx, int64_t box_idx)
+        : batch_index(batch_idx)
+        , class_index(class_idx)
+        , box_index(box_idx)
+    {
+    }
+
+    SelectedIndex() = default;
+
+    int64_t batch_index = 0;
+    int64_t class_index = 0;
+    int64_t box_index = 0;
 };
 
 struct SelectedScore
 {
-  SelectedScore(float batch_idx, float class_idx, float score)
-      : batch_index{batch_idx}
-      , class_index{class_idx}
-      , box_score{score}
-      {
-      }
-
-  SelectedScore() = default;
-
-  float batch_index = 0.0f;
-  float class_index = 0.0f;
-  float box_score = 0.0f;
+    SelectedScore(float batch_idx, float class_idx, float score)
+        : batch_index{batch_idx}
+        , class_index{class_idx}
+        , box_score{score}
+    {
+    }
+
+    SelectedScore() = default;
+
+    float batch_index = 0.0f;
+    float class_index = 0.0f;
+    float box_score = 0.0f;
 };
 
 struct BoxInfo
@@ -142,7 +142,8 @@ namespace ngraph
                                      const bool sort_result_descending)
             {
                 float scale = 0.0f;
-                if (soft_nms_sigma > 0.0f) {
+                if (soft_nms_sigma > 0.0f)
+                {
                     scale = - 0.5f / soft_nms_sigma;
                 }
 
@@ -173,8 +174,8 @@ namespace ngraph
 
                     for (int64_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
-                        const float* scoresPtr = scores_data + batch * (num_classes * num_boxes) +
-                                                 class_idx * num_boxes;
+                        const float* scoresPtr =
+                            scores_data + batch * (num_classes * num_boxes) + class_idx * num_boxes;
 
                         std::vector<BoxInfo> candidate_boxes;
                         candidate_boxes.reserve(num_boxes);
@@ -183,7 +184,8 @@ namespace ngraph
                         {
                             if (scoresPtr[box_idx] > score_threshold)
                             {
-                                candidate_boxes.emplace_back(r[box_idx], box_idx, scoresPtr[box_idx], 0);
+                                candidate_boxes.emplace_back(
+                                    r[box_idx], box_idx, scoresPtr[box_idx], 0);
                             }
                         }
 
@@ -204,9 +206,11 @@ namespace ngraph
 
                             bool should_hard_suppress = false;
                             for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
-                                 j >= next_candidate.suppress_begin_index; --j)
+                                 j >= next_candidate.suppress_begin_index;
+                                 --j)
                             {
-                                float iou = intersectionOverUnion(next_candidate.box, selected[j].box);
+                                float iou =
+                                    intersectionOverUnion(next_candidate.box, selected[j].box);
                                 next_candidate.score *= func(iou);
 
                                 if (iou >= iou_threshold)
@@ -227,14 +231,11 @@ namespace ngraph
                             {
                                 if (next_candidate.score == original_score)
                                 {
-                                    // Suppression has not occurred, so select next_candidate
                                     selected.push_back(next_candidate);
                                     continue;
                                 }
                                 if (next_candidate.score > score_threshold)
                                 {
-                                    // Soft suppression has occurred and current score is still greater than
-                                    // score_threshold; add next_candidate back onto priority queue.
                                     sorted_boxes.push(next_candidate);
                                 }
                             }
@@ -244,7 +245,8 @@ namespace ngraph
                         {
                             SelectedIndex selected_index{batch, class_idx, box_info.index};
                             SelectedScore selected_score{static_cast<float>(batch),
-                                static_cast<float>(class_idx), box_info.score};
+                                                         static_cast<float>(class_idx),
+                                                         box_info.score};
 
                             selected_indices_ptr[num_of_valid_boxes] = selected_index;
                             selected_scores_ptr[num_of_valid_boxes] = selected_score;
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 41814afdf16480..fcc7b697779848 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -14,8 +14,8 @@
 // limitations under the License.
 //*****************************************************************************
 
-#include <cstring>
 #include "ngraph/op/non_max_suppression.hpp"
+#include <cstring>
 #include "ngraph/attribute_visitor.hpp"
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/util/op_types.hpp"
@@ -952,7 +952,7 @@ static void normalize_center(float* boxes, const Shape& boxes_shape)
             float y1 = y_center - height / 2.0;
             float x1 = x_center - width / 2.0;
             float y2 = y_center + height / 2.0;
-            float x2 = x_center + width / 2.0;;
+            float x2 = x_center + width / 2.0;
 
             box_ptr[0] = y1;
             box_ptr[1] = x1;
@@ -964,9 +964,8 @@ static void normalize_center(float* boxes, const Shape& boxes_shape)
     }
 }
 
-static void normalize_box_encoding(float* boxes,
-                                   const Shape& boxes_shape,
-                                   const V5BoxEncoding box_encoding)
+static void
+    normalize_box_encoding(float* boxes, const Shape& boxes_shape, const V5BoxEncoding box_encoding)
 {
     if (box_encoding == V5BoxEncoding::CORNER)
     {
@@ -1134,11 +1133,8 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                             &valid_outputs,
                                             m_sort_result_descending);
 
-    evaluate_postprocessing(outputs,
-                            m_output_type,
-                            selected_indices,
-                            selected_scores,
-                            valid_outputs);
+    evaluate_postprocessing(
+        outputs, m_output_type, selected_indices, selected_scores, valid_outputs);
 
     return true;
 }

From 2a292533acba776e52e10ca80f61e595eb43ef1c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 11:04:09 +0300
Subject: [PATCH 031/173] Written draft of NMS-5 nGraph reference
 implementation.

---
 .../runtime/reference/non_max_suppression.cpp | 57 ++++++++++++++-----
 1 file changed, 43 insertions(+), 14 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 7f530d655b59eb..e33aacb45180c3 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -99,10 +99,17 @@ struct SelectedScore
 
 struct BoxInfo
 {
-    BoxInfo(const Rectangle& r, int64_t idx, float sc, int64_t suppress_idx)
+    BoxInfo(const Rectangle& r,
+            int64_t idx,
+            float sc,
+            int64_t suppress_idx,
+            int64_t batch_idx,
+            int64_t class_idx)
         : box{r}
         , index{idx}
         , suppress_begin_index{suppress_idx}
+        , batch_index{batch_idx}
+        , class_index{class_idx}
         , score{sc}
     {
     }
@@ -117,6 +124,8 @@ struct BoxInfo
     Rectangle box;
     int64_t index = 0;
     int64_t suppress_begin_index = 0;
+    int64_t batch_index = 0;
+    int64_t class_index = 0;
     float score = 0.0f;
 };
 
@@ -144,7 +153,7 @@ namespace ngraph
                 float scale = 0.0f;
                 if (soft_nms_sigma > 0.0f)
                 {
-                    scale = - 0.5f / soft_nms_sigma;
+                    scale = -0.5f / soft_nms_sigma;
                 }
 
                 auto func = [iou_threshold, scale](float iou) {
@@ -167,6 +176,8 @@ namespace ngraph
 
                 int64_t num_of_valid_boxes = 0;
 
+                std::vector<BoxInfo> filteredBoxes;
+
                 for (int64_t batch = 0; batch < num_batches; batch++)
                 {
                     const float* boxesPtr = boxes_data + batch * num_boxes * 4;
@@ -185,7 +196,7 @@ namespace ngraph
                             if (scoresPtr[box_idx] > score_threshold)
                             {
                                 candidate_boxes.emplace_back(
-                                    r[box_idx], box_idx, scoresPtr[box_idx], 0);
+                                    r[box_idx], box_idx, scoresPtr[box_idx], 0, batch, class_idx);
                             }
                         }
 
@@ -243,27 +254,45 @@ namespace ngraph
 
                         for (const auto& box_info : selected)
                         {
-                            SelectedIndex selected_index{batch, class_idx, box_info.index};
-                            SelectedScore selected_score{static_cast<float>(batch),
-                                                         static_cast<float>(class_idx),
-                                                         box_info.score};
-
-                            selected_indices_ptr[num_of_valid_boxes] = selected_index;
-                            selected_scores_ptr[num_of_valid_boxes] = selected_score;
-
-                            num_of_valid_boxes++;
+                            filteredBoxes.push_back(box_info);
                         }
                     }
                 }
 
-
                 if (sort_result_descending)
                 {
+                    std::sort(filteredBoxes.begin(),
+                              filteredBoxes.end(),
+                              [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
                 }
-                else
+
+                size_t max_num_of_selected_indices = selected_indices_shape[0]
+                size_t output_size = std::min(filteredBoxes.size(), max_num_of_selected_indices);
+
+                *valid_outputs = output_size;
+
+                size_t idx;
+                for (idx = 0; idx < output_size; idx++)
                 {
+                    const auto& box_info = filteredBoxes[idx];
+                    SelectedIndex selected_index{box_info.batch_index,
+                                                 box_info.class_index,
+                                                 box_info.index};
+                    SelectedScore selected_score{static_cast<float>(box_info.batch_index),
+                                                 static_cast<float>(box_info.class_index),
+                                                 box_info.score};
+
+                    selected_indices_ptr[idx] = selected_index;
+                    selected_scores_ptr[idx] = selected_score;
                 }
 
+                SelectedIndex selected_index_filler{0, 0, 0};
+                SelectedScore selected_score_filler{0.0f, 0.0f, 0.0f};
+                for (; idx < max_num_of_selected_indices; idx++)
+                {
+                    selected_indices_ptr[idx] = selected_index_filler;
+                    selected_scores_ptr[idx] = selected_score_filler;
+                }
             }
         }
     }

From 71a08889838841da623523b888044ad318830d2d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 11:11:57 +0300
Subject: [PATCH 032/173] Code style fixes.

---
 .../src/runtime/reference/non_max_suppression.cpp         | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index e33aacb45180c3..2f7e81b4ba543e 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -14,7 +14,6 @@
 // limitations under the License.
 //*****************************************************************************
 
-
 #include "ngraph/op/non_max_suppression.hpp"
 #include <algorithm>
 #include <cmath>
@@ -266,7 +265,7 @@ namespace ngraph
                               [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
                 }
 
-                size_t max_num_of_selected_indices = selected_indices_shape[0]
+                size_t max_num_of_selected_indices = selected_indices_shape[0];
                 size_t output_size = std::min(filteredBoxes.size(), max_num_of_selected_indices);
 
                 *valid_outputs = output_size;
@@ -275,9 +274,8 @@ namespace ngraph
                 for (idx = 0; idx < output_size; idx++)
                 {
                     const auto& box_info = filteredBoxes[idx];
-                    SelectedIndex selected_index{box_info.batch_index,
-                                                 box_info.class_index,
-                                                 box_info.index};
+                    SelectedIndex selected_index{
+                        box_info.batch_index, box_info.class_index, box_info.index};
                     SelectedScore selected_score{static_cast<float>(box_info.batch_index),
                                                  static_cast<float>(box_info.class_index),
                                                  box_info.score};

From 5caeef76cfd8f4998491ba7fcedca2ef0a1bd018 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 13:21:05 +0300
Subject: [PATCH 033/173] Started to write tests for void
 op::v5::NonMaxSuppression::validate_and_infer_types().

---
 ngraph/test/type_prop/non_max_suppression.cpp | 32 +++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 405a3845049094..74f4ba5a448d6d 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -547,3 +547,35 @@ TEST(type_prop, nms_v4_dynamic_boxes_and_scores)
     ASSERT_TRUE(
         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
 }
+
+// ------------------------------ V5 ------------------------------
+
+TEST(type_prop, nms_v5_incorrect_boxes_rank)
+{
+    try
+    {
+        const auto boxes = make_shared<op::Parameter>(element::f32, Shape{1, 2, 3, 4});
+        const auto scores = make_shared<op::Parameter>(element::f32, Shape{1, 2, 3});
+
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(), "Expected a 3D tensor for the 'boxes' input");
+    }
+}
+
+TEST(type_prop, nms_v5_incorrect_scores_rank)
+{
+    try
+    {
+        const auto boxes = make_shared<op::Parameter>(element::f32, Shape{1, 2, 3});
+        const auto scores = make_shared<op::Parameter>(element::f32, Shape{1, 2});
+
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(), "Expected a 3D tensor for the 'scores' input");
+    }
+}

From 01200aaef42e616f11fc9b4f325dce1959d5009f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 16:38:48 +0300
Subject: [PATCH 034/173] Added tests for scalars/nonscalars.

---
 ngraph/test/type_prop/non_max_suppression.cpp | 92 +++++++++++++++++++
 1 file changed, 92 insertions(+)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 74f4ba5a448d6d..01462a0659697a 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -579,3 +579,95 @@ TEST(type_prop, nms_v5_incorrect_scores_rank)
         EXPECT_HAS_SUBSTRING(error.what(), "Expected a 3D tensor for the 'scores' input");
     }
 }
+
+TEST(type_prop, nms_v5_incorrect_scheme_num_batches)
+{
+    try
+    {
+        const auto boxes = make_shared<op::Parameter>(element::f32, Shape{1, 2, 3});
+        const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 2, 3});
+
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "The first dimension of both 'boxes' and 'scores' must match");
+    }
+}
+
+TEST(type_prop, nms_v5_incorrect_scheme_num_boxes)
+{
+    try
+    {
+        const auto boxes = make_shared<op::Parameter>(element::f32, Shape{1, 2, 3});
+        const auto scores = make_shared<op::Parameter>(element::f32, Shape{1, 2, 3});
+
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "'boxes' and 'scores' input shapes must match at the second and third "
+                             "dimension respectively");
+    }
+}
+
+TEST(type_prop, nms_v5_scalar_inputs_check)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{1, 2, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{1, 2, 2});
+
+    const auto scalar = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto non_scalar = make_shared<op::Parameter>(element::f32, Shape{1});
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, non_scalar, scalar, scalar);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "Expected a scalar for the 'max_output_boxes_per_class' input");
+    }
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, non_scalar, scalar);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'iou_threshold' input");
+    }
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, non_scalar);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'score_threshold' input");
+    }
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, scalar, non_scalar);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'soft_nms_sigma' input");
+    }
+}
+
+TEST(type_prop, nms_v5_output_shape)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{5, 2, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{5, 3, 2});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+    const auto nms_indices_out_ps = nms->get_output_partial_shape(0);
+
+    EXPECT_TRUE(nms_indices_out_ps.rank().is_static());
+    EXPECT_EQ(nms_indices_out_ps.rank().get_length(), 2);
+    EXPECT_EQ(nms->get_shape(), (Shape{0, 3}));
+}

From 0eb5f40ab57ca119cabbe06c2169d461e6272da1 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 17:11:32 +0300
Subject: [PATCH 035/173] Fixes in the test type_prop.nms_v5_output_shape.

---
 ngraph/test/type_prop/non_max_suppression.cpp | 76 ++++++++++++++++++-
 1 file changed, 75 insertions(+), 1 deletion(-)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 01462a0659697a..6ae27470c7232f 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -669,5 +669,79 @@ TEST(type_prop, nms_v5_output_shape)
 
     EXPECT_TRUE(nms_indices_out_ps.rank().is_static());
     EXPECT_EQ(nms_indices_out_ps.rank().get_length(), 2);
-    EXPECT_EQ(nms->get_shape(), (Shape{0, 3}));
+    EXPECT_EQ(nms->get_output_shape(0), (Shape{0, 3}));
+
+    const auto nms_scores_out_ps = nms->get_output_partial_shape(1);
+
+    EXPECT_TRUE(nms_scores_out_ps.rank().is_static());
+    EXPECT_EQ(nms_scores_out_ps.rank().get_length(), 2);
+    EXPECT_EQ(nms->get_output_shape(1), (Shape{0, 3}));
 }
+
+// TEST(type_prop, nms_v5_output_shape_2)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+//
+//     ASSERT_EQ(nms->get_element_type(), element::i64);
+//     ASSERT_EQ(nms->get_shape(), (Shape{2 * 5 * 3, 3}));
+// }
+//
+// TEST(type_prop, nms_v5_output_shape_3)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+//
+//     ASSERT_EQ(nms->get_element_type(), element::i64);
+//     ASSERT_EQ(nms->get_shape(), (Shape{2 * 5 * 7, 3}));
+// }
+//
+// TEST(type_prop, nms_v5_output_shape_i32)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms =
+//         make_shared<op::v5::NonMaxSuppression>(boxes,
+//                                                scores,
+//                                                max_output_boxes_per_class,
+//                                                iou_threshold,
+//                                                score_threshold,
+//                                                op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
+//                                                true,
+//                                                element::i32);
+//
+//     ASSERT_EQ(nms->get_element_type(), element::i32);
+//     ASSERT_EQ(nms->get_shape(), (Shape{30, 3}));
+// }
+//
+// TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+//     const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+//
+//     ASSERT_EQ(nms->get_element_type(), element::i64);
+//     ASSERT_TRUE(
+//         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+// }

From b98be7fac0e8aab066a28bdd6529499e4c88b4d8 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 2 Oct 2020 17:57:05 +0300
Subject: [PATCH 036/173] Fixes in tests nms_v5_output_shape_2 and
 nms_v5_output_shape.

---
 ngraph/core/src/op/non_max_suppression.cpp    |  2 +-
 ngraph/test/type_prop/non_max_suppression.cpp | 36 +++++++++++--------
 2 files changed, 22 insertions(+), 16 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index fcc7b697779848..6950ec5f957781 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -848,7 +848,7 @@ void op::v5::NonMaxSuppression::validate_and_infer_types()
     }
     set_output_type(0, m_output_type, out_shape);
     set_output_type(1, element::f32, out_shape);
-    set_output_type(2, element::f32, Shape{});
+    set_output_type(2, m_output_type, Shape{});
 }
 
 namespace ngraph
diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 6ae27470c7232f..551b0728b62202 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -676,23 +676,29 @@ TEST(type_prop, nms_v5_output_shape)
     EXPECT_TRUE(nms_scores_out_ps.rank().is_static());
     EXPECT_EQ(nms_scores_out_ps.rank().get_length(), 2);
     EXPECT_EQ(nms->get_output_shape(1), (Shape{0, 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+TEST(type_prop, nms_v5_output_shape_2)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+    const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 3, 3}));
+    ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 3, 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
 }
 
-// TEST(type_prop, nms_v5_output_shape_2)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-//
-//     ASSERT_EQ(nms->get_element_type(), element::i64);
-//     ASSERT_EQ(nms->get_shape(), (Shape{2 * 5 * 3, 3}));
-// }
-//
 // TEST(type_prop, nms_v5_output_shape_3)
 // {
 //     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});

From 1eda5aba460d2a9fdcac9802c87242be53ed8175 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 5 Oct 2020 12:18:11 +0300
Subject: [PATCH 037/173] Written tests for validate_and_infer_types() of
 NMS-5.

---
 ngraph/test/type_prop/non_max_suppression.cpp | 121 ++++++++++--------
 1 file changed, 69 insertions(+), 52 deletions(-)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 551b0728b62202..5bdc6a90856129 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -693,61 +693,78 @@ TEST(type_prop, nms_v5_output_shape_2)
 
     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 3, 3}));
     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 3, 3}));
 
     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
 }
 
-// TEST(type_prop, nms_v5_output_shape_3)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-//
-//     ASSERT_EQ(nms->get_element_type(), element::i64);
-//     ASSERT_EQ(nms->get_shape(), (Shape{2 * 5 * 7, 3}));
-// }
-//
-// TEST(type_prop, nms_v5_output_shape_i32)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms =
-//         make_shared<op::v5::NonMaxSuppression>(boxes,
-//                                                scores,
-//                                                max_output_boxes_per_class,
-//                                                iou_threshold,
-//                                                score_threshold,
-//                                                op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
-//                                                true,
-//                                                element::i32);
-//
-//     ASSERT_EQ(nms->get_element_type(), element::i32);
-//     ASSERT_EQ(nms->get_shape(), (Shape{30, 3}));
-// }
-//
-// TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-//     const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-//
-//     ASSERT_EQ(nms->get_element_type(), element::i64);
-//     ASSERT_TRUE(
-//         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-// }
+ TEST(type_prop, nms_v5_output_shape_3)
+ {
+     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
+     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 7, 3}));
+     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 7, 3}));
+
+     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+ }
+
+ TEST(type_prop, nms_v5_output_shape_i32)
+ {
+     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+     const auto nms =
+         make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                scores,
+                                                max_output_boxes_per_class,
+                                                iou_threshold,
+                                                score_threshold,
+                                                op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
+                                                true,
+                                                element::i32);
+
+     ASSERT_EQ(nms->get_output_element_type(0), element::i32);
+     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+     ASSERT_EQ(nms->get_output_element_type(2), element::i32);
+     ASSERT_EQ(nms->get_output_shape(0), (Shape{30, 3}));
+     ASSERT_EQ(nms->get_output_shape(1), (Shape{30, 3}));
+
+     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+ TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
+ {
+     const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+     const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+     ASSERT_TRUE(
+         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+     ASSERT_TRUE(
+         nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+ }

From b3a64931a571d1ea83cca52daa06830b1767c831 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 5 Oct 2020 14:06:15 +0300
Subject: [PATCH 038/173] Code style fixes.

---
 ngraph/test/type_prop/non_max_suppression.cpp | 136 +++++++++---------
 1 file changed, 68 insertions(+), 68 deletions(-)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 5bdc6a90856129..1c9c682131fe20 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -700,71 +700,71 @@ TEST(type_prop, nms_v5_output_shape_2)
     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
 }
 
- TEST(type_prop, nms_v5_output_shape_3)
- {
-     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
-     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-
-     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 7, 3}));
-     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 7, 3}));
-
-     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
- }
-
- TEST(type_prop, nms_v5_output_shape_i32)
- {
-     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-     const auto nms =
-         make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                scores,
-                                                max_output_boxes_per_class,
-                                                iou_threshold,
-                                                score_threshold,
-                                                op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
-                                                true,
-                                                element::i32);
-
-     ASSERT_EQ(nms->get_output_element_type(0), element::i32);
-     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-     ASSERT_EQ(nms->get_output_element_type(2), element::i32);
-     ASSERT_EQ(nms->get_output_shape(0), (Shape{30, 3}));
-     ASSERT_EQ(nms->get_output_shape(1), (Shape{30, 3}));
-
-     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-}
-
- TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
- {
-     const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-     const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-
-     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-     ASSERT_TRUE(
-         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-     ASSERT_TRUE(
-         nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-
-     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
- }
+TEST(type_prop, nms_v5_output_shape_3)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+    ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 7, 3}));
+    ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 7, 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+TEST(type_prop, nms_v5_output_shape_i32)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms =
+        make_shared<op::v5::NonMaxSuppression>(boxes,
+                                               scores,
+                                               max_output_boxes_per_class,
+                                               iou_threshold,
+                                               score_threshold,
+                                               op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
+                                               true,
+                                               element::i32);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i32);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i32);
+    ASSERT_EQ(nms->get_output_shape(0), (Shape{30, 3}));
+    ASSERT_EQ(nms->get_output_shape(1), (Shape{30, 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+
+
+TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+    const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}

From 2ad006645776e893947173cb9478c3ccb06e7a6e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 5 Oct 2020 14:42:37 +0300
Subject: [PATCH 039/173] Now NMS-5 evaluate() can have outputs with calculated
 shapes.

---
 ngraph/core/src/op/non_max_suppression.cpp    |  59 ++---
 ngraph/test/type_prop/non_max_suppression.cpp | 218 +++++++++---------
 2 files changed, 131 insertions(+), 146 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 6950ec5f957781..69d59a24723f96 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -831,21 +831,6 @@ void op::v5::NonMaxSuppression::validate_and_infer_types()
 
     validate();
 
-    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
-    {
-        const auto num_boxes_boxes = boxes_ps[1];
-        const auto max_output_boxes_per_class_node = input_value(2).get_node_shared_ptr();
-        if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static() &&
-            op::is_constant(max_output_boxes_per_class_node))
-        {
-            const auto num_boxes = num_boxes_boxes.get_length();
-            const auto num_classes = scores_ps[1].get_length();
-            const auto max_output_boxes_per_class = max_boxes_output_from_input();
-
-            out_shape[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
-                           scores_ps[0].get_length();
-        }
-    }
     set_output_type(0, m_output_type, out_shape);
     set_output_type(1, element::f32, out_shape);
     set_output_type(2, m_output_type, Shape{});
@@ -1037,19 +1022,19 @@ static void evaluate_postprocessing(const HostTensorVector& outputs,
                                     int64_t valid_outputs)
 {
     size_t num_of_outputs = outputs.size();
+    size_t selected_size = valid_outputs * 3;
 
     if (output_type == ngraph::element::i64)
     {
         int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
-        memcpy(indices_ptr, selected_indices.data(), selected_indices.size() * sizeof(int64_t));
+        memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
     }
     else
     {
         int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
-        for (int64_t idx : selected_indices)
+        for (size_t i = 0; i < selected_size; ++i)
         {
-            *indices_ptr = static_cast<int32_t>(idx);
-            indices_ptr++;
+            indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
         }
     }
 
@@ -1059,7 +1044,7 @@ static void evaluate_postprocessing(const HostTensorVector& outputs,
     }
 
     float* scores_ptr = outputs[1]->get_data_ptr<float>();
-    memcpy(scores_ptr, selected_scores.data(), selected_scores.size() * sizeof(float));
+    memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
 
     if (num_of_outputs < 3)
     {
@@ -1089,23 +1074,6 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
     auto selected_indices_shape = infer_selected_indices_shape(inputs, max_output_boxes_per_class);
     Shape out_shape = selected_indices_shape.to_shape();
 
-    outputs[0]->set_element_type(m_output_type);
-    outputs[0]->set_shape(out_shape);
-
-    size_t num_of_outputs = outputs.size();
-
-    if (num_of_outputs >= 2)
-    {
-        outputs[1]->set_element_type(element::f32);
-        outputs[1]->set_shape(out_shape);
-    }
-
-    if (num_of_outputs >= 3)
-    {
-        outputs[2]->set_element_type(element::f32);
-        outputs[2]->set_shape(Shape{});
-    }
-
     Shape boxes_shape = inputs[boxes_port]->get_shape();
     Shape scores_shape = inputs[scores_port]->get_shape();
 
@@ -1133,6 +1101,23 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                             &valid_outputs,
                                             m_sort_result_descending);
 
+    outputs[0]->set_element_type(m_output_type);
+    outputs[0]->set_shape(Shape{valid_outputs, 3});
+
+    size_t num_of_outputs = outputs.size();
+
+    if (num_of_outputs >= 2)
+    {
+        outputs[1]->set_element_type(element::f32);
+        outputs[1]->set_shape(Shape{valid_outputs, 3});
+    }
+
+    if (num_of_outputs >= 3)
+    {
+        outputs[2]->set_element_type(m_output_type);
+        outputs[2]->set_shape(Shape{});
+    }
+
     evaluate_postprocessing(
         outputs, m_output_type, selected_indices, selected_scores, valid_outputs);
 
diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 1c9c682131fe20..6a355d133a7628 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -659,112 +659,112 @@ TEST(type_prop, nms_v5_scalar_inputs_check)
     }
 }
 
-TEST(type_prop, nms_v5_output_shape)
-{
-    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{5, 2, 4});
-    const auto scores = make_shared<op::Parameter>(element::f32, Shape{5, 3, 2});
-
-    const auto nms = make_shared<op::v5::NonMaxSuppression>(boxes, scores);
-    const auto nms_indices_out_ps = nms->get_output_partial_shape(0);
-
-    EXPECT_TRUE(nms_indices_out_ps.rank().is_static());
-    EXPECT_EQ(nms_indices_out_ps.rank().get_length(), 2);
-    EXPECT_EQ(nms->get_output_shape(0), (Shape{0, 3}));
-
-    const auto nms_scores_out_ps = nms->get_output_partial_shape(1);
-
-    EXPECT_TRUE(nms_scores_out_ps.rank().is_static());
-    EXPECT_EQ(nms_scores_out_ps.rank().get_length(), 2);
-    EXPECT_EQ(nms->get_output_shape(1), (Shape{0, 3}));
-
-    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-}
-
-TEST(type_prop, nms_v5_output_shape_2)
-{
-    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-    const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
-    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-    const auto nms = make_shared<op::v5::NonMaxSuppression>(
-        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-
-    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-    ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 3, 3}));
-    ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 3, 3}));
-
-    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-}
-
-TEST(type_prop, nms_v5_output_shape_3)
-{
-    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
-    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-    const auto nms = make_shared<op::v5::NonMaxSuppression>(
-        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-
-    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-    ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 7, 3}));
-    ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 7, 3}));
-
-    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-}
-
-TEST(type_prop, nms_v5_output_shape_i32)
-{
-    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-    const auto nms =
-        make_shared<op::v5::NonMaxSuppression>(boxes,
-                                               scores,
-                                               max_output_boxes_per_class,
-                                               iou_threshold,
-                                               score_threshold,
-                                               op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
-                                               true,
-                                               element::i32);
-
-    ASSERT_EQ(nms->get_output_element_type(0), element::i32);
-    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-    ASSERT_EQ(nms->get_output_element_type(2), element::i32);
-    ASSERT_EQ(nms->get_output_shape(0), (Shape{30, 3}));
-    ASSERT_EQ(nms->get_output_shape(1), (Shape{30, 3}));
-
-    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-
-
-TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
-{
-    const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-    const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-
-    const auto nms = make_shared<op::v5::NonMaxSuppression>(
-        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-
-    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-    ASSERT_TRUE(
-        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-    ASSERT_TRUE(
-        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-
-    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-}
+// TEST(type_prop, nms_v5_output_shape)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{5, 2, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{5, 3, 2});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+//     const auto nms_indices_out_ps = nms->get_output_partial_shape(0);
+//
+//     EXPECT_TRUE(nms_indices_out_ps.rank().is_static());
+//     EXPECT_EQ(nms_indices_out_ps.rank().get_length(), 2);
+//     EXPECT_EQ(nms->get_output_shape(0), (Shape{0, 3}));
+//
+//     const auto nms_scores_out_ps = nms->get_output_partial_shape(1);
+//
+//     EXPECT_TRUE(nms_scores_out_ps.rank().is_static());
+//     EXPECT_EQ(nms_scores_out_ps.rank().get_length(), 2);
+//     EXPECT_EQ(nms->get_output_shape(1), (Shape{0, 3}));
+//
+//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+// }
+//
+// TEST(type_prop, nms_v5_output_shape_2)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+//
+//     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+//     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+//     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 3, 3}));
+//     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 3, 3}));
+//
+//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+// }
+//
+// TEST(type_prop, nms_v5_output_shape_3)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+//
+//     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+//     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+//     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 7, 3}));
+//     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 7, 3}));
+//
+//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+// }
+//
+// TEST(type_prop, nms_v5_output_shape_i32)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms =
+//         make_shared<op::v5::NonMaxSuppression>(boxes,
+//                                                scores,
+//                                                max_output_boxes_per_class,
+//                                                iou_threshold,
+//                                                score_threshold,
+//                                                op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
+//                                                true,
+//                                                element::i32);
+//
+//     ASSERT_EQ(nms->get_output_element_type(0), element::i32);
+//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+//     ASSERT_EQ(nms->get_output_element_type(2), element::i32);
+//     ASSERT_EQ(nms->get_output_shape(0), (Shape{30, 3}));
+//     ASSERT_EQ(nms->get_output_shape(1), (Shape{30, 3}));
+//
+//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+//
+//
+// TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
+// {
+//     const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+//     const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+//
+//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
+//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+//
+//     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+//     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+//     ASSERT_TRUE(
+//         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+//     ASSERT_TRUE(
+//         nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+//
+//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+// }

From 136b6b770175e208707f12e76cbc3df30beb4821 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 5 Oct 2020 14:55:32 +0300
Subject: [PATCH 040/173] Small fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 69d59a24723f96..4efcf87798878f 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -1102,14 +1102,14 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                             m_sort_result_descending);
 
     outputs[0]->set_element_type(m_output_type);
-    outputs[0]->set_shape(Shape{valid_outputs, 3});
+    outputs[0]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
 
     size_t num_of_outputs = outputs.size();
 
     if (num_of_outputs >= 2)
     {
         outputs[1]->set_element_type(element::f32);
-        outputs[1]->set_shape(Shape{valid_outputs, 3});
+        outputs[1]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
     }
 
     if (num_of_outputs >= 3)

From fe5a2546907d3b383e0ec9abb12f7e9344574919 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 5 Oct 2020 17:48:05 +0300
Subject: [PATCH 041/173] Corrected tests for NMS-5 validate_and_infer_type().

---
 ngraph/test/type_prop/non_max_suppression.cpp | 218 +++++++++---------
 1 file changed, 109 insertions(+), 109 deletions(-)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 6a355d133a7628..a0db6b606efc43 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -659,112 +659,112 @@ TEST(type_prop, nms_v5_scalar_inputs_check)
     }
 }
 
-// TEST(type_prop, nms_v5_output_shape)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{5, 2, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{5, 3, 2});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(boxes, scores);
-//     const auto nms_indices_out_ps = nms->get_output_partial_shape(0);
-//
-//     EXPECT_TRUE(nms_indices_out_ps.rank().is_static());
-//     EXPECT_EQ(nms_indices_out_ps.rank().get_length(), 2);
-//     EXPECT_EQ(nms->get_output_shape(0), (Shape{0, 3}));
-//
-//     const auto nms_scores_out_ps = nms->get_output_partial_shape(1);
-//
-//     EXPECT_TRUE(nms_scores_out_ps.rank().is_static());
-//     EXPECT_EQ(nms_scores_out_ps.rank().get_length(), 2);
-//     EXPECT_EQ(nms->get_output_shape(1), (Shape{0, 3}));
-//
-//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-// }
-//
-// TEST(type_prop, nms_v5_output_shape_2)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-//
-//     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-//     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-//     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 3, 3}));
-//     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 3, 3}));
-//
-//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-// }
-//
-// TEST(type_prop, nms_v5_output_shape_3)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-//
-//     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-//     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-//     ASSERT_EQ(nms->get_output_shape(0), (Shape{2 * 5 * 7, 3}));
-//     ASSERT_EQ(nms->get_output_shape(1), (Shape{2 * 5 * 7, 3}));
-//
-//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-// }
-//
-// TEST(type_prop, nms_v5_output_shape_i32)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
-//     const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms =
-//         make_shared<op::v5::NonMaxSuppression>(boxes,
-//                                                scores,
-//                                                max_output_boxes_per_class,
-//                                                iou_threshold,
-//                                                score_threshold,
-//                                                op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
-//                                                true,
-//                                                element::i32);
-//
-//     ASSERT_EQ(nms->get_output_element_type(0), element::i32);
-//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-//     ASSERT_EQ(nms->get_output_element_type(2), element::i32);
-//     ASSERT_EQ(nms->get_output_shape(0), (Shape{30, 3}));
-//     ASSERT_EQ(nms->get_output_shape(1), (Shape{30, 3}));
-//
-//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-//
-//
-// TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
-// {
-//     const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-//     const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
-//     const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
-//     const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//     const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
-//
-//     const auto nms = make_shared<op::v5::NonMaxSuppression>(
-//         boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
-//
-//     ASSERT_EQ(nms->get_output_element_type(0), element::i64);
-//     ASSERT_EQ(nms->get_output_element_type(1), element::f32);
-//     ASSERT_EQ(nms->get_output_element_type(2), element::i64);
-//     ASSERT_TRUE(
-//         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-//     ASSERT_TRUE(
-//         nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-//
-//     EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
-// }
+TEST(type_prop, nms_v5_output_shape)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{5, 2, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{5, 3, 2});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(boxes, scores);
+    
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+TEST(type_prop, nms_v5_output_shape_2)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+    const auto max_output_boxes_per_class = op::Constant::create(element::i32, Shape{}, {3});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+TEST(type_prop, nms_v5_output_shape_3)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {1000});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+TEST(type_prop, nms_v5_output_shape_i32)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, Shape{2, 7, 4});
+    const auto scores = make_shared<op::Parameter>(element::f32, Shape{2, 5, 7});
+    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms =
+        make_shared<op::v5::NonMaxSuppression>(boxes,
+                                               scores,
+                                               max_output_boxes_per_class,
+                                               iou_threshold,
+                                               score_threshold,
+                                               op::v5::NonMaxSuppression::BoxEncodingType::CORNER,
+                                               true,
+                                               element::i32);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i32);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i32);
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}
+
+TEST(type_prop, nms_v5_dynamic_boxes_and_scores)
+{
+    const auto boxes = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+    const auto scores = make_shared<op::Parameter>(element::f32, PartialShape::dynamic());
+    const auto max_output_boxes_per_class = op::Constant::create(element::i16, Shape{}, {3});
+    const auto iou_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+    const auto score_threshold = make_shared<op::Parameter>(element::f32, Shape{});
+
+    const auto nms = make_shared<op::v5::NonMaxSuppression>(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold);
+
+    ASSERT_EQ(nms->get_output_element_type(0), element::i64);
+    ASSERT_EQ(nms->get_output_element_type(1), element::f32);
+    ASSERT_EQ(nms->get_output_element_type(2), element::i64);
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+    ASSERT_TRUE(
+        nms->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+
+    EXPECT_EQ(nms->get_output_shape(2), (Shape{}));
+}

From 1d065669d53452e7b7b84d916503fff467074e9e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 09:36:06 +0300
Subject: [PATCH 042/173] Code style fixes.

---
 ngraph/test/type_prop/non_max_suppression.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index a0db6b606efc43..5bcb86de3ca57f 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -665,7 +665,7 @@ TEST(type_prop, nms_v5_output_shape)
     const auto scores = make_shared<op::Parameter>(element::f32, Shape{5, 3, 2});
 
     const auto nms = make_shared<op::v5::NonMaxSuppression>(boxes, scores);
-    
+
     ASSERT_TRUE(
         nms->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
     ASSERT_TRUE(

From fed50cc9cb68fb011a10c9fa189321841bdfea45 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 12:32:23 +0300
Subject: [PATCH 043/173] Started to write inner version of NMS-5 with static
 output shapes.

---
 .../include/ngraph_ops/nms_ie.hpp             | 26 +++++++++++++++++++
 .../transformations/src/ngraph_ops/nms_ie.cpp | 23 ++++++++++++++++
 2 files changed, 49 insertions(+)

diff --git a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
index 79a31dedcc1a83..130b21c14772bc 100644
--- a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
+++ b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
@@ -58,5 +58,31 @@ class TRANSFORMATIONS_API NonMaxSuppressionIE2 : public NonMaxSuppressionIE {
     std::shared_ptr<Node> clone_with_new_inputs(const OutputVector & new_args) const override;
 };
 
+class TRANSFORMATIONS_API NonMaxSuppressionIE3 : public Op {
+public:
+    static constexpr NodeTypeInfo type_info{"NonMaxSuppressionIE", 3};
+    const NodeTypeInfo& get_type_info() const override { return type_info; }
+
+    NonMaxSuppressionIE3(const Output<Node>& boxes,
+                         const Output<Node>& scores,
+                         const Output<Node>& max_output_boxes_per_class,
+                         const Output<Node>& iou_threshold,
+                         const Output<Node>& score_threshold,
+                         const Output<Node>& soft_nms_sigma,
+                         int center_point_box,
+                         bool sort_result_descending,
+                         const ngraph::element::Type& output_type = ngraph::element::i64);
+
+    void validate_and_infer_types() override;
+
+    bool visit_attributes(AttributeVisitor& visitor) override;
+
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector & new_args) const override;
+
+    int m_center_point_box;
+    bool m_sort_result_descending = true;
+    element::Type m_output_type;
+};
+
 }  // namespace op
 }  // namespace ngraph
diff --git a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
index 02473ad09131b4..7c65aad9cec9e0 100644
--- a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
@@ -101,3 +101,26 @@ void op::NonMaxSuppressionIE2::validate_and_infer_types() {
                                                            m_output_type);
     set_output_type(0, nms->output(0).get_element_type(), nms->output(0).get_partial_shape());
 }
+
+constexpr NodeTypeInfo op::NonMaxSuppressionIE3::type_info;
+
+op::NonMaxSuppressionIE3::NonMaxSuppressionIE3(const Output<Node>& boxes,
+                                               const Output<Node>& scores,
+                                               const Output<Node>& max_output_boxes_per_class,
+                                               const Output<Node>& iou_threshold,
+                                               const Output<Node>& score_threshold,
+                                               const Output<Node>& soft_nms_sigma,
+                                               int center_point_box,
+                                               bool sort_result_descending,
+                                               const ngraph::element::Type& output_type)
+        : Op({boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, soft_nms_sigma}),
+          m_center_point_box(center_point_box), m_sort_result_descending(sort_result_descending), m_output_type(output_type) {
+    constructor_validate_and_infer_types();
+}
+
+std::shared_ptr<Node> op::NonMaxSuppressionIE3::clone_with_new_inputs(const ngraph::OutputVector &new_args) const {
+    check_new_args_count(this, new_args);
+    return make_shared<NonMaxSuppressionIE2>(new_args.at(0), new_args.at(1), new_args.at(2), new_args.at(3),
+                                             new_args.at(4), new_args.at(5), m_center_point_box, m_sort_result_descending,
+                                             m_output_type);
+}

From 924061b37c20786d4fb21696fe0e1dc45b5efd0e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 13:58:55 +0300
Subject: [PATCH 044/173] Written draft of the inner operation
 NonMaxSuppressionIE3.

---
 .../include/ngraph_ops/nms_ie.hpp             |  2 +
 .../transformations/src/ngraph_ops/nms_ie.cpp | 51 +++++++++++++++++++
 ngraph/core/src/op/non_max_suppression.cpp    |  2 +-
 3 files changed, 54 insertions(+), 1 deletion(-)

diff --git a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
index 130b21c14772bc..9f37d1c083235d 100644
--- a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
+++ b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
@@ -82,6 +82,8 @@ class TRANSFORMATIONS_API NonMaxSuppressionIE3 : public Op {
     int m_center_point_box;
     bool m_sort_result_descending = true;
     element::Type m_output_type;
+private:
+    int64_t max_boxes_output_from_input() const;
 };
 
 }  // namespace op
diff --git a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
index 7c65aad9cec9e0..d54f46d50335f9 100644
--- a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
@@ -124,3 +124,54 @@ std::shared_ptr<Node> op::NonMaxSuppressionIE3::clone_with_new_inputs(const ngra
                                              new_args.at(4), new_args.at(5), m_center_point_box, m_sort_result_descending,
                                              m_output_type);
 }
+
+bool op::NonMaxSuppressionIE3::visit_attributes(AttributeVisitor& visitor) {
+    visitor.on_attribute("center_point_box", m_center_point_box);
+    visitor.on_attribute("sort_result_descending", m_sort_result_descending);
+    visitor.on_attribute("output_type", m_output_type);
+    return true;
+}
+
+static constexpr size_t boxes_port = 0;
+static constexpr size_t scores_port = 1;
+static constexpr size_t max_output_boxes_per_class_port = 2;
+
+int64_t op::NonMaxSuppressionIE3::max_boxes_output_from_input() const
+{
+    int64_t max_output_boxes{0};
+
+    const auto max_output_boxes_input =
+        as_type_ptr<op::Constant>(input_value(2).get_node_shared_ptr());
+    max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
+
+    return max_output_boxes;
+}
+
+void op::NonMaxSuppressionIE3::validate_and_infer_types() {
+    const auto boxes_ps = get_input_partial_shape(boxes_port);
+    const auto scores_ps = get_input_partial_shape(scores_port);
+
+    // NonMaxSuppression produces triplets
+    // that have the following format: [batch_index, class_index, box_index]
+    PartialShape out_shape = {Dimension::dynamic(), 3};
+
+    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
+    {
+        const auto num_boxes_boxes = boxes_ps[1];
+        const auto max_output_boxes_per_class_node = input_value(max_output_boxes_per_class_port).get_node_shared_ptr();
+        if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static() &&
+            op::is_constant(max_output_boxes_per_class_node))
+        {
+            const auto num_boxes = num_boxes_boxes.get_length();
+            const auto num_classes = scores_ps[1].get_length();
+            const auto max_output_boxes_per_class = max_boxes_output_from_input();
+
+            out_shape[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
+                           scores_ps[0].get_length();
+        }
+    }
+
+    set_output_type(0, m_output_type, out_shape);
+    set_output_type(1, element::f32, out_shape);
+    set_output_type(2, m_output_type, Shape{});
+}
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 4efcf87798878f..3c4c406780d054 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -863,7 +863,7 @@ static PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
                                                  int64_t max_output_boxes_per_class)
 {
     const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
-    const auto scores_ps = inputs[1]->get_partial_shape();
+    const auto scores_ps = inputs[scores_port]->get_partial_shape();
 
     // NonMaxSuppression produces triplets
     // that have the following format: [batch_index, class_index, box_index]

From 8a57122d1f51ac4a69cd6072552d270c246ad2f7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 16:28:54 +0300
Subject: [PATCH 045/173] Started to write conversion of
 op::v5::NonMaxSuppression into NonMaxSuppressionIE3.

---
 .../convert_nms_5_to_legacy.hpp               |  33 +++++
 .../convert_nms_5_to_legacy.cpp               | 113 ++++++++++++++++++
 2 files changed, 146 insertions(+)
 create mode 100644 inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
 create mode 100644 inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp

diff --git a/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp b/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
new file mode 100644
index 00000000000000..11fd03444f126c
--- /dev/null
+++ b/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
@@ -0,0 +1,33 @@
+// Copyright (C) 2018-2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include <transformations_visibility.hpp>
+
+#include <ngraph/pass/graph_rewrite.hpp>
+
+namespace ngraph {
+namespace pass {
+
+    class TRANSFORMATIONS_API ConvertNMS4ToLegacyMatcher;
+
+}  // namespace pass
+}  // namespace ngraph
+
+/*
+ * Description:
+ *      Convert NMS-5 directly to inner NMS.
+ */
+
+
+class ngraph::pass::ConvertNMS5ToLegacyMatcher: public ngraph::pass::MatcherPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+    ConvertNMS4ToLegacyMatcher();
+};
+
diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
new file mode 100644
index 00000000000000..b6ecb00fb7eac8
--- /dev/null
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -0,0 +1,113 @@
+// Copyright (C) 2018-2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <memory>
+#include <vector>
+
+#include <ngraph/graph_util.hpp>
+#include <ngraph/opsets/opset1.hpp>
+#include <ngraph/opsets/opset5.hpp>
+#include <ngraph_ops/nms_ie.hpp>
+#include <ngraph/rt_info.hpp>
+#include <transformations/utils/utils.hpp>
+
+#include "transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp"
+
+NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS5ToLegacyMatcher, "ConvertNMS4ToLegacyMatcher", 0);
+
+ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
+//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+//     auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
+//     auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
+//     auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
+//     auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+//                                                                    iou_threshold, score_threshold);
+//
+//     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
+//         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
+//         if (!nms_4) {
+//             return false;
+//         }
+//
+//         const auto new_args = nms_4->input_values();
+//         const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset4::Constant::create(element::i32, Shape{}, {0});
+//         const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
+//         const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
+//
+//         const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
+//         const auto iou_threshold_rank = arg3.get_partial_shape().rank();
+//         const auto score_threshold_rank = arg4.get_partial_shape().rank();
+//
+//         // Check that required ranks are not dynamic
+//         if (max_output_boxes_per_class_rank.is_dynamic() ||
+//             iou_threshold_rank.is_dynamic() ||
+//             score_threshold_rank.is_dynamic()) {
+//             return false;
+//         }
+//
+//         if (max_output_boxes_per_class_rank.get_length() == 1 &&
+//             iou_threshold_rank.get_length() == 1 &&
+//             score_threshold_rank.get_length() == 1) {
+//             return false;
+//         }
+//
+//         // vector of new nGraph operations
+//         NodeVector new_ops;
+//
+//         auto new_max_per_class = arg2;
+//         if (max_output_boxes_per_class_rank.get_length() == 0) {
+//             // WA: we need to create Constant manually because it requires by NMS shape inference
+//             //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
+//             //     if CF will be executed right after transformation and before Validate pass.
+//             if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
+//                 new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
+//             } else {
+//                 new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
+//                 new_ops.push_back(new_max_per_class.get_node_shared_ptr());
+//             }
+//         }
+//         auto new_iou_threshold = arg3;
+//         if (iou_threshold_rank.get_length() == 0) {
+//             new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
+//             new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
+//         }
+//         auto new_score_threshold = arg4;
+//         if (score_threshold_rank.get_length() == 0) {
+//             new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
+//             new_ops.push_back(new_score_threshold.get_node_shared_ptr());
+//         }
+//
+//         int center_point_box = 0;
+//         switch (nms_4->get_box_encoding()) {
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
+//                 center_point_box = 1;
+//                 break;
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
+//                 center_point_box = 0;
+//                 break;
+//             default:
+//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+//                                    " has unsupported box encoding");
+//         }
+//         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE2>(
+//                 new_args.at(0),
+//                 new_args.at(1),
+//                 new_max_per_class,
+//                 new_iou_threshold,
+//                 new_score_threshold,
+//                 center_point_box,
+//                 nms_4->get_sort_result_descending(),
+//                 nms_4->get_output_type());
+//         new_ops.push_back(nms_legacy);
+//
+//         nms_legacy->set_friendly_name(nms_4->get_friendly_name());
+//         ngraph::copy_runtime_info(nms_4, new_ops);
+//         ngraph::replace_node(nms_4, nms_legacy);
+//         return true;
+//     };
+//
+//     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMSLegacy");
+//     this->register_matcher(m, callback);
+}

From 44da98e6b7336b7d1fb12558a9e546dd09ae501d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 16:32:22 +0300
Subject: [PATCH 046/173] Small changes.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index b6ecb00fb7eac8..61f8ebd659d2ac 100644
--- a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -14,11 +14,11 @@
 
 #include "transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp"
 
-NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS5ToLegacyMatcher, "ConvertNMS4ToLegacyMatcher", 0);
+NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS5ToLegacyMatcher, "ConvertNMS5ToLegacyMatcher", 0);
 
 ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
-//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
 //     auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
 //     auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
 //     auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});

From b112a2fa3892c4bf961055dd047e9c0d899eb847 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 17:00:54 +0300
Subject: [PATCH 047/173] Some additions.

---
 .../transformations/src/ngraph_ops/nms_ie.cpp |  9 ++---
 .../convert_nms_5_to_legacy.cpp               | 37 ++++++++++---------
 2 files changed, 22 insertions(+), 24 deletions(-)

diff --git a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
index d54f46d50335f9..2cf573304f7385 100644
--- a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
@@ -136,8 +136,7 @@ static constexpr size_t boxes_port = 0;
 static constexpr size_t scores_port = 1;
 static constexpr size_t max_output_boxes_per_class_port = 2;
 
-int64_t op::NonMaxSuppressionIE3::max_boxes_output_from_input() const
-{
+int64_t op::NonMaxSuppressionIE3::max_boxes_output_from_input() const {
     int64_t max_output_boxes{0};
 
     const auto max_output_boxes_input =
@@ -155,13 +154,11 @@ void op::NonMaxSuppressionIE3::validate_and_infer_types() {
     // that have the following format: [batch_index, class_index, box_index]
     PartialShape out_shape = {Dimension::dynamic(), 3};
 
-    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
-    {
+    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static()) {
         const auto num_boxes_boxes = boxes_ps[1];
         const auto max_output_boxes_per_class_node = input_value(max_output_boxes_per_class_port).get_node_shared_ptr();
         if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static() &&
-            op::is_constant(max_output_boxes_per_class_node))
-        {
+            op::is_constant(max_output_boxes_per_class_node)) {
             const auto num_boxes = num_boxes_boxes.get_length();
             const auto num_classes = scores_ps[1].get_length();
             const auto max_output_boxes_per_class = max_boxes_output_from_input();
diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 61f8ebd659d2ac..bee33707a10427 100644
--- a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -19,19 +19,20 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS5ToLegacyMatcher, "ConvertNMS5ToL
 ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-//     auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
-//     auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
-//     auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
-//     auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-//                                                                    iou_threshold, score_threshold);
-//
-//     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
-//         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
-//         if (!nms_4) {
-//             return false;
-//         }
-//
-//         const auto new_args = nms_4->input_values();
+    auto max_output_boxes_per_class = ngraph::opset5::Constant::create(element::i64, Shape{}, {10});
+    auto iou_threshold = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.75});
+    auto score_threshold = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.7});
+    auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+    auto nms = std::make_shared<ngraph::opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                   iou_threshold, score_threshold, soft_nms_sigma);
+
+    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
+        auto nms_5 = std::dynamic_pointer_cast<ngraph::opset5::NonMaxSuppression>(m.get_match_root());
+        if (!nms_5) {
+            return false;
+        }
+
+        const auto new_args = nms_5->input_values();
 //         const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset4::Constant::create(element::i32, Shape{}, {0});
 //         const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
 //         const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
@@ -105,9 +106,9 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
 //         nms_legacy->set_friendly_name(nms_4->get_friendly_name());
 //         ngraph::copy_runtime_info(nms_4, new_ops);
 //         ngraph::replace_node(nms_4, nms_legacy);
-//         return true;
-//     };
-//
-//     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMSLegacy");
-//     this->register_matcher(m, callback);
+        return true;
+    };
+
+    auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS5ToNMSLegacy");
+    this->register_matcher(m, callback);
 }

From 285ce509417ad80a8a9ce965c7caa6e324fb0489 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 6 Oct 2020 17:33:57 +0300
Subject: [PATCH 048/173] Small fixes.

---
 .../include/ngraph_ops/nms_ie.hpp             |  1 +
 .../convert_nms_5_to_legacy.cpp               | 54 ++++++++++---------
 .../convert_opset1_to_legacy.cpp              |  2 +
 3 files changed, 32 insertions(+), 25 deletions(-)

diff --git a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
index 9f37d1c083235d..721392f8a42902 100644
--- a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
+++ b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
@@ -82,6 +82,7 @@ class TRANSFORMATIONS_API NonMaxSuppressionIE3 : public Op {
     int m_center_point_box;
     bool m_sort_result_descending = true;
     element::Type m_output_type;
+
 private:
     int64_t max_boxes_output_from_input() const;
 };
diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index bee33707a10427..26cde5f87fe788 100644
--- a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -33,31 +33,35 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         }
 
         const auto new_args = nms_5->input_values();
-//         const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset4::Constant::create(element::i32, Shape{}, {0});
-//         const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
-//         const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
-//
-//         const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
-//         const auto iou_threshold_rank = arg3.get_partial_shape().rank();
-//         const auto score_threshold_rank = arg4.get_partial_shape().rank();
-//
-//         // Check that required ranks are not dynamic
-//         if (max_output_boxes_per_class_rank.is_dynamic() ||
-//             iou_threshold_rank.is_dynamic() ||
-//             score_threshold_rank.is_dynamic()) {
-//             return false;
-//         }
-//
-//         if (max_output_boxes_per_class_rank.get_length() == 1 &&
-//             iou_threshold_rank.get_length() == 1 &&
-//             score_threshold_rank.get_length() == 1) {
-//             return false;
-//         }
-//
-//         // vector of new nGraph operations
-//         NodeVector new_ops;
-//
-//         auto new_max_per_class = arg2;
+        const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg5 = new_args.size() > 5 ? new_args.at(5) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+
+        const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
+        const auto iou_threshold_rank = arg3.get_partial_shape().rank();
+        const auto score_threshold_rank = arg4.get_partial_shape().rank();
+        const auto soft_nms_sigma_rank = arg5.get_partial_shape().rank();
+
+        // Check that required ranks are not dynamic
+        if (max_output_boxes_per_class_rank.is_dynamic() ||
+            iou_threshold_rank.is_dynamic() ||
+            score_threshold_rank.is_dynamic() ||
+            soft_nms_sigma_rank.is_dynamic()) {
+            return false;
+        }
+
+        if (max_output_boxes_per_class_rank.get_length() == 1 &&
+            iou_threshold_rank.get_length() == 1 &&
+            score_threshold_rank.get_length() == 1 &&
+            soft_nms_sigma_rank.get_length() == 1) {
+            return false;
+        }
+
+        // vector of new nGraph operations
+        NodeVector new_ops;
+
+        auto new_max_per_class = arg2;
 //         if (max_output_boxes_per_class_rank.get_length() == 0) {
 //             // WA: we need to create Constant manually because it requires by NMS shape inference
 //             //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp
index b6784268879769..204c8c017fe1b0 100644
--- a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp
@@ -21,6 +21,7 @@
 #include <transformations/convert_negative.hpp>
 #include <transformations/convert_opset1_to_legacy/convert_nms_to_nms_ie.hpp>
 #include <transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp>
+#include <transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp>
 #include <transformations/convert_opset1_to_legacy/convert_normalizel2_to_normalize_ie.hpp>
 #include <transformations/convert_opset1_to_legacy/convert_one_hot_to_one_hot_ie.hpp>
 #include <transformations/convert_opset1_to_legacy/convert_pad_to_pad_ie.hpp>
@@ -140,6 +141,7 @@ bool ngraph::pass::ConvertOpSet1ToLegacy::run_on_function(std::shared_ptr<ngraph
     anchor->add_matcher<ngraph::pass::ConvertTopKToTopKIEMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertNMSToNMSIEMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertNMS4ToLegacyMatcher>();
+    anchor->add_matcher<ngraph::pass::ConvertNMS5ToLegacyMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertGRUSequenceMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertRNNSequenceMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertLSTMSequenceMatcher>();

From 8063e01d6019fc20a70bdfc8b84d3b4663c7b179 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 09:21:16 +0300
Subject: [PATCH 049/173] Fixed typo.

---
 inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
index 2cf573304f7385..ce039843d9d701 100644
--- a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
@@ -120,7 +120,7 @@ op::NonMaxSuppressionIE3::NonMaxSuppressionIE3(const Output<Node>& boxes,
 
 std::shared_ptr<Node> op::NonMaxSuppressionIE3::clone_with_new_inputs(const ngraph::OutputVector &new_args) const {
     check_new_args_count(this, new_args);
-    return make_shared<NonMaxSuppressionIE2>(new_args.at(0), new_args.at(1), new_args.at(2), new_args.at(3),
+    return make_shared<NonMaxSuppressionIE3>(new_args.at(0), new_args.at(1), new_args.at(2), new_args.at(3),
                                              new_args.at(4), new_args.at(5), m_center_point_box, m_sort_result_descending,
                                              m_output_type);
 }

From c2db7479d7a9ecce7961180b415713c7021b4526 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 09:44:54 +0300
Subject: [PATCH 050/173] Fixed typos.

---
 .../convert_nms_5_to_legacy.hpp               |  2 +-
 .../convert_nms_5_to_legacy.cpp               | 72 ++++++++++---------
 2 files changed, 39 insertions(+), 35 deletions(-)

diff --git a/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp b/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
index 11fd03444f126c..5fd6b3ade95f60 100644
--- a/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
+++ b/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
@@ -14,7 +14,7 @@
 namespace ngraph {
 namespace pass {
 
-    class TRANSFORMATIONS_API ConvertNMS4ToLegacyMatcher;
+    class TRANSFORMATIONS_API ConvertNMS5ToLegacyMatcher;
 
 }  // namespace pass
 }  // namespace ngraph
diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 26cde5f87fe788..01c139fff0d41c 100644
--- a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -62,40 +62,44 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         NodeVector new_ops;
 
         auto new_max_per_class = arg2;
-//         if (max_output_boxes_per_class_rank.get_length() == 0) {
-//             // WA: we need to create Constant manually because it requires by NMS shape inference
-//             //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
-//             //     if CF will be executed right after transformation and before Validate pass.
-//             if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
-//                 new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
-//             } else {
-//                 new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
-//                 new_ops.push_back(new_max_per_class.get_node_shared_ptr());
-//             }
-//         }
-//         auto new_iou_threshold = arg3;
-//         if (iou_threshold_rank.get_length() == 0) {
-//             new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
-//             new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
-//         }
-//         auto new_score_threshold = arg4;
-//         if (score_threshold_rank.get_length() == 0) {
-//             new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
-//             new_ops.push_back(new_score_threshold.get_node_shared_ptr());
-//         }
-//
-//         int center_point_box = 0;
-//         switch (nms_4->get_box_encoding()) {
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-//                 center_point_box = 1;
-//                 break;
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-//                 center_point_box = 0;
-//                 break;
-//             default:
-//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
-//                                    " has unsupported box encoding");
-//         }
+        if (max_output_boxes_per_class_rank.get_length() == 0) {
+            // WA: we need to create Constant manually because it requires by NMS shape inference
+            //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
+            //     if CF will be executed right after transformation and before Validate pass.
+            if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
+                new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
+            } else {
+                new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
+                new_ops.push_back(new_max_per_class.get_node_shared_ptr());
+            }
+        }
+        auto new_iou_threshold = arg3;
+        if (iou_threshold_rank.get_length() == 0) {
+            new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::f32, Shape{1}, {0.0f}));
+            new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
+        }
+        auto new_score_threshold = arg4;
+        if (score_threshold_rank.get_length() == 0) {
+            new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::f32, Shape{1}, {0.0f}));
+            new_ops.push_back(new_score_threshold.get_node_shared_ptr());
+        }
+        auto new_soft_nms_sigma = arg5;
+        if (soft_nms_sigma_rank.get_length() == 0) {
+            new_soft_nms_sigma = std::make_shared<ngraph::op::Unsqueeze>(arg5, opset1::Constant::create(element::f32, Shape{1}, {0.0f}));
+            new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
+        }
+        int center_point_box = 0;
+        switch (nms_5->get_box_encoding()) {
+            case ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER:
+                center_point_box = 1;
+                break;
+            case ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER:
+                center_point_box = 0;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms_5->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
 //         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE2>(
 //                 new_args.at(0),
 //                 new_args.at(1),

From eb861a9e9c0ac7236ca604b092687d77714b5204 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 11:00:18 +0300
Subject: [PATCH 051/173] Written draft of the transformation
 ConvertNMS5ToLegacyMatcher that converts ngraph::opset5::NonMaxSuppression
 into op::NonMaxSuppressionIE3.

---
 .../convert_nms_5_to_legacy.hpp               |  2 +-
 .../convert_nms_5_to_legacy.cpp               | 29 ++++++++++---------
 2 files changed, 16 insertions(+), 15 deletions(-)

diff --git a/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp b/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
index 5fd6b3ade95f60..1dcfcbe1e2c4a8 100644
--- a/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
+++ b/inference-engine/src/transformations/include/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
@@ -28,6 +28,6 @@ namespace pass {
 class ngraph::pass::ConvertNMS5ToLegacyMatcher: public ngraph::pass::MatcherPass {
 public:
     NGRAPH_RTTI_DECLARATION;
-    ConvertNMS4ToLegacyMatcher();
+    ConvertNMS5ToLegacyMatcher();
 };
 
diff --git a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 01c139fff0d41c..a633b103501639 100644
--- a/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -100,20 +100,21 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
                 throw ngraph_error("NonMaxSuppression layer " + nms_5->get_friendly_name() +
                                    " has unsupported box encoding");
         }
-//         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE2>(
-//                 new_args.at(0),
-//                 new_args.at(1),
-//                 new_max_per_class,
-//                 new_iou_threshold,
-//                 new_score_threshold,
-//                 center_point_box,
-//                 nms_4->get_sort_result_descending(),
-//                 nms_4->get_output_type());
-//         new_ops.push_back(nms_legacy);
-//
-//         nms_legacy->set_friendly_name(nms_4->get_friendly_name());
-//         ngraph::copy_runtime_info(nms_4, new_ops);
-//         ngraph::replace_node(nms_4, nms_legacy);
+         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
+                 new_args.at(0),
+                 new_args.at(1),
+                 new_max_per_class,
+                 new_iou_threshold,
+                 new_score_threshold,
+                 new_soft_nms_sigma,
+                 center_point_box,
+                 nms_5->get_sort_result_descending(),
+                 nms_5->get_output_type());
+         new_ops.push_back(nms_legacy);
+
+         nms_legacy->set_friendly_name(nms_5->get_friendly_name());
+         ngraph::copy_runtime_info(nms_5, new_ops);
+         ngraph::replace_node(nms_5, nms_legacy);
         return true;
     };
 

From 88e17964a1bff052ebf991c65d1766e777119888 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 11:46:49 +0300
Subject: [PATCH 052/173] Written header file for the transformations from
 NMS-1, NMS-3, NMS-4 to NMS-5.

---
 .../convert_previous_nms_to_nms_5.hpp         | 39 +++++++++++++++++++
 1 file changed, 39 insertions(+)
 create mode 100644 inference-engine/src/transformations/include/transformations/convert_previous_nms_to_nms_5.hpp

diff --git a/inference-engine/src/transformations/include/transformations/convert_previous_nms_to_nms_5.hpp b/inference-engine/src/transformations/include/transformations/convert_previous_nms_to_nms_5.hpp
new file mode 100644
index 00000000000000..9b7d610dc73d07
--- /dev/null
+++ b/inference-engine/src/transformations/include/transformations/convert_previous_nms_to_nms_5.hpp
@@ -0,0 +1,39 @@
+// Copyright (C) 2018-2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <memory>
+#include <utility>
+
+#include <transformations_visibility.hpp>
+#include <ngraph/pass/graph_rewrite.hpp>
+
+namespace ngraph {
+namespace pass {
+
+class TRANSFORMATIONS_API ConvertNMS1ToNMS5;
+class TRANSFORMATIONS_API ConvertNMS3ToNMS5;
+class TRANSFORMATIONS_API ConvertNMS4ToNMS5;
+
+}  // namespace pass
+}  // namespace ngraph
+
+class ngraph::pass::ConvertNMS1ToNMS5: public ngraph::pass::MatcherPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+    ConvertNMS1ToNMS5();
+};
+
+class ngraph::pass::ConvertNMS3ToNMS5: public ngraph::pass::MatcherPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+    ConvertNMS3ToNMS5();
+};
+
+class ngraph::pass::ConvertNMS4ToNMS5: public ngraph::pass::MatcherPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+    ConvertNMS4ToNMS5();
+};

From e662ec4efef1da387ccbeec52eb68be06d7f44e5 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 12:54:13 +0300
Subject: [PATCH 053/173] Started to write conversion of NMS-4 to NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp         | 112 ++++++++++++++++++
 1 file changed, 112 insertions(+)
 create mode 100644 inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
new file mode 100644
index 00000000000000..4bbb1e492d42f2
--- /dev/null
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -0,0 +1,112 @@
+// Copyright (C) 2018-2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/convert_previous_nms_to_nms_5.hpp"
+
+#include <memory>
+#include <vector>
+
+#include <ngraph/opsets/opset1.hpp>
+#include <ngraph/opsets/opset3.hpp>
+#include <ngraph/opsets/opset5.hpp>
+#include <ngraph/rt_info.hpp>
+#include <ngraph/pattern/op/wrap_type.hpp>
+
+NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
+
+ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
+    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+    auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
+    auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
+    auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
+    auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                   iou_threshold, score_threshold);
+
+    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
+        auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
+        if (!nms_4) {
+            return false;
+        }
+
+//         const auto new_args = nms_4->input_values();
+//         const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset4::Constant::create(element::i32, Shape{}, {0});
+//         const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
+//         const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
+//
+//         const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
+//         const auto iou_threshold_rank = arg3.get_partial_shape().rank();
+//         const auto score_threshold_rank = arg4.get_partial_shape().rank();
+//
+//         // Check that required ranks are not dynamic
+//         if (max_output_boxes_per_class_rank.is_dynamic() ||
+//             iou_threshold_rank.is_dynamic() ||
+//             score_threshold_rank.is_dynamic()) {
+//             return false;
+//         }
+//
+//         if (max_output_boxes_per_class_rank.get_length() == 1 &&
+//             iou_threshold_rank.get_length() == 1 &&
+//             score_threshold_rank.get_length() == 1) {
+//             return false;
+//         }
+//
+//         // vector of new nGraph operations
+//         NodeVector new_ops;
+//
+//         auto new_max_per_class = arg2;
+//         if (max_output_boxes_per_class_rank.get_length() == 0) {
+//             // WA: we need to create Constant manually because it requires by NMS shape inference
+//             //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
+//             //     if CF will be executed right after transformation and before Validate pass.
+//             if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
+//                 new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
+//             } else {
+//                 new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
+//                 new_ops.push_back(new_max_per_class.get_node_shared_ptr());
+//             }
+//         }
+//         auto new_iou_threshold = arg3;
+//         if (iou_threshold_rank.get_length() == 0) {
+//             new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
+//             new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
+//         }
+//         auto new_score_threshold = arg4;
+//         if (score_threshold_rank.get_length() == 0) {
+//             new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
+//             new_ops.push_back(new_score_threshold.get_node_shared_ptr());
+//         }
+//
+//         int center_point_box = 0;
+//         switch (nms_4->get_box_encoding()) {
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
+//                 center_point_box = 1;
+//                 break;
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
+//                 center_point_box = 0;
+//                 break;
+//             default:
+//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+//                                    " has unsupported box encoding");
+//         }
+//         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE2>(
+//                 new_args.at(0),
+//                 new_args.at(1),
+//                 new_max_per_class,
+//                 new_iou_threshold,
+//                 new_score_threshold,
+//                 center_point_box,
+//                 nms_4->get_sort_result_descending(),
+//                 nms_4->get_output_type());
+//         new_ops.push_back(nms_legacy);
+//
+//         nms_legacy->set_friendly_name(nms_4->get_friendly_name());
+//         ngraph::copy_runtime_info(nms_4, new_ops);
+//         ngraph::replace_node(nms_4, nms_legacy);
+        return true;
+    };
+
+//     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMSLegacy");
+//     this->register_matcher(m, callback);
+}

From d8d0a191b81ab3b50c67b18e6a0fe71b1a32262f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 13:18:40 +0300
Subject: [PATCH 054/173] Added include for ngraph/opsets/opset4.hpp.

---
 .../src/transformations/convert_previous_nms_to_nms_5.cpp        | 1 +
 1 file changed, 1 insertion(+)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 4bbb1e492d42f2..01d0a0434fd26a 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -9,6 +9,7 @@
 
 #include <ngraph/opsets/opset1.hpp>
 #include <ngraph/opsets/opset3.hpp>
+#include <ngraph/opsets/opset4.hpp>
 #include <ngraph/opsets/opset5.hpp>
 #include <ngraph/rt_info.hpp>
 #include <ngraph/pattern/op/wrap_type.hpp>

From f6d5ea08945966d77ca278bfd87c44c31011bdd4 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 14:48:49 +0300
Subject: [PATCH 055/173] Started to write conversion of NMS-3 to NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp         | 147 +++++++++---------
 .../include/ngraph/op/non_max_suppression.hpp |   2 +-
 2 files changed, 72 insertions(+), 77 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 01d0a0434fd26a..6e7d2e91bcbd8c 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -4,6 +4,7 @@
 
 #include "transformations/convert_previous_nms_to_nms_5.hpp"
 
+#include <list>
 #include <memory>
 #include <vector>
 
@@ -31,83 +32,77 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
             return false;
         }
 
-//         const auto new_args = nms_4->input_values();
-//         const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset4::Constant::create(element::i32, Shape{}, {0});
-//         const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
-//         const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
-//
-//         const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
-//         const auto iou_threshold_rank = arg3.get_partial_shape().rank();
-//         const auto score_threshold_rank = arg4.get_partial_shape().rank();
-//
-//         // Check that required ranks are not dynamic
-//         if (max_output_boxes_per_class_rank.is_dynamic() ||
-//             iou_threshold_rank.is_dynamic() ||
-//             score_threshold_rank.is_dynamic()) {
-//             return false;
-//         }
-//
-//         if (max_output_boxes_per_class_rank.get_length() == 1 &&
-//             iou_threshold_rank.get_length() == 1 &&
-//             score_threshold_rank.get_length() == 1) {
-//             return false;
-//         }
-//
-//         // vector of new nGraph operations
-//         NodeVector new_ops;
-//
-//         auto new_max_per_class = arg2;
-//         if (max_output_boxes_per_class_rank.get_length() == 0) {
-//             // WA: we need to create Constant manually because it requires by NMS shape inference
-//             //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
-//             //     if CF will be executed right after transformation and before Validate pass.
-//             if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
-//                 new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
-//             } else {
-//                 new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
-//                 new_ops.push_back(new_max_per_class.get_node_shared_ptr());
-//             }
-//         }
-//         auto new_iou_threshold = arg3;
-//         if (iou_threshold_rank.get_length() == 0) {
-//             new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
-//             new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
-//         }
-//         auto new_score_threshold = arg4;
-//         if (score_threshold_rank.get_length() == 0) {
-//             new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
-//             new_ops.push_back(new_score_threshold.get_node_shared_ptr());
-//         }
-//
-//         int center_point_box = 0;
-//         switch (nms_4->get_box_encoding()) {
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-//                 center_point_box = 1;
-//                 break;
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-//                 center_point_box = 0;
-//                 break;
-//             default:
-//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
-//                                    " has unsupported box encoding");
-//         }
-//         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE2>(
-//                 new_args.at(0),
-//                 new_args.at(1),
-//                 new_max_per_class,
-//                 new_iou_threshold,
-//                 new_score_threshold,
-//                 center_point_box,
-//                 nms_4->get_sort_result_descending(),
-//                 nms_4->get_output_type());
-//         new_ops.push_back(nms_legacy);
-//
-//         nms_legacy->set_friendly_name(nms_4->get_friendly_name());
-//         ngraph::copy_runtime_info(nms_4, new_ops);
-//         ngraph::replace_node(nms_4, nms_legacy);
+        const auto new_args = nms_4->input_values();
+
+        size_t num_of_args = new_args.size();
+
+        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+
+        auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+        switch (nms_4->get_box_encoding()) {
+            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
+                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
+                break;
+            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
+                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
+
+        // list of new nGraph operations
+        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
+
+        new_ops_list.push_front(arg5);
+        if (num_of_args <= 4) {
+            new_ops_list.push_front(arg4);
+        }
+        if (num_of_args <= 3) {
+            new_ops_list.push_front(arg3);
+        }
+        if (num_of_args <= 2) {
+            new_ops_list.push_front(arg2);
+        }
+
+        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
+                new_args.at(0),
+                new_args.at(1),
+                arg2,
+                arg3,
+                arg4,
+                arg5,
+                box_encoding,
+                nms_4->get_sort_result_descending(),
+                nms_4->get_output_type());
+
+        new_ops_list.push_back(nms_5);
+
+        // vector of new nGraph operations
+        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
+
+        nms_5->set_friendly_name(nms_4->get_friendly_name());
+        ngraph::copy_runtime_info(nms_4, new_ops);
+        ngraph::replace_node(nms_4, nms_5);
         return true;
     };
 
-//     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMSLegacy");
-//     this->register_matcher(m, callback);
+    auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMS5");
+    this->register_matcher(m, callback);
+}
+
+
+NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS3ToNMS5, "ConvertNMS3ToNMS5", 0);
+
+ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
+    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+    auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
+    auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
+    auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
+    auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                   iou_threshold, score_threshold);
 }
diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 5ba5b63b52a8a1..a3828e9f65f704 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -253,7 +253,7 @@ namespace ngraph
                 const NodeTypeInfo& get_type_info() const override { return type_info; }
                 NonMaxSuppression() = default;
 
-                /// \brief Constructs a NonMaxSuppression operation with default values in the last.
+                /// \brief Constructs a NonMaxSuppression operation with default values in the last
                 ///        4 inputs.
                 ///
                 /// \param boxes Node producing the box coordinates

From f2adb311e397c0c3220d4c85fd25d00ade65a41c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 15:19:07 +0300
Subject: [PATCH 056/173] Small fixes.

---
 .../convert_previous_nms_to_nms_5.cpp         | 77 +++++++++++++++++--
 1 file changed, 72 insertions(+), 5 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 6e7d2e91bcbd8c..145436e8874f23 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -44,10 +44,10 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
         switch (nms_4->get_box_encoding()) {
             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
                 break;
             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
                 break;
             default:
                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
@@ -59,13 +59,13 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
 
         new_ops_list.push_front(arg5);
         if (num_of_args <= 4) {
-            new_ops_list.push_front(arg4);
+            new_ops_list.push_front(arg4.get_node_shared_ptr());
         }
         if (num_of_args <= 3) {
-            new_ops_list.push_front(arg3);
+            new_ops_list.push_front(arg3.get_node_shared_ptr());
         }
         if (num_of_args <= 2) {
-            new_ops_list.push_front(arg2);
+            new_ops_list.push_front(arg2.get_node_shared_ptr());
         }
 
         const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
@@ -105,4 +105,71 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
     auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
     auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
                                                                    iou_threshold, score_threshold);
+
+    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
+        auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(m.get_match_root());
+        if (!nms_3) {
+            return false;
+        }
+
+        const auto new_args = nms_3->input_values();
+
+        size_t num_of_args = new_args.size();
+
+        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+
+//         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+//         switch (nms_4->get_box_encoding()) {
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
+//                 center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
+//                 break;
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
+//                 center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+//                 break;
+//             default:
+//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+//                                    " has unsupported box encoding");
+//         }
+//
+//         // list of new nGraph operations
+//         std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
+//
+//         new_ops_list.push_front(arg5);
+//         if (num_of_args <= 4) {
+//             new_ops_list.push_front(arg4);
+//         }
+//         if (num_of_args <= 3) {
+//             new_ops_list.push_front(arg3);
+//         }
+//         if (num_of_args <= 2) {
+//             new_ops_list.push_front(arg2);
+//         }
+//
+//         const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
+//                 new_args.at(0),
+//                 new_args.at(1),
+//                 arg2,
+//                 arg3,
+//                 arg4,
+//                 arg5,
+//                 box_encoding,
+//                 nms_4->get_sort_result_descending(),
+//                 nms_4->get_output_type());
+//
+//         new_ops_list.push_back(nms_5);
+//
+//         // vector of new nGraph operations
+//         NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
+//
+//         nms_5->set_friendly_name(nms_4->get_friendly_name());
+//         ngraph::copy_runtime_info(nms_4, new_ops);
+//         ngraph::replace_node(nms_4, nms_5);
+//         return true;
+//     };
+//
+//     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMS5");
+//     this->register_matcher(m, callback);
 }

From 8fa5661b424fe461e26bfa8773ca6b92271061f3 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 15:48:26 +0300
Subject: [PATCH 057/173] Written draft of the conversion of NMS-3 into NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp         | 102 +++++++++---------
 1 file changed, 51 insertions(+), 51 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 145436e8874f23..8975aeca432a5e 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -121,55 +121,55 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
         const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
 
-//         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-//         switch (nms_4->get_box_encoding()) {
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-//                 center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
-//                 break;
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-//                 center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
-//                 break;
-//             default:
-//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
-//                                    " has unsupported box encoding");
-//         }
-//
-//         // list of new nGraph operations
-//         std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
-//
-//         new_ops_list.push_front(arg5);
-//         if (num_of_args <= 4) {
-//             new_ops_list.push_front(arg4);
-//         }
-//         if (num_of_args <= 3) {
-//             new_ops_list.push_front(arg3);
-//         }
-//         if (num_of_args <= 2) {
-//             new_ops_list.push_front(arg2);
-//         }
-//
-//         const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
-//                 new_args.at(0),
-//                 new_args.at(1),
-//                 arg2,
-//                 arg3,
-//                 arg4,
-//                 arg5,
-//                 box_encoding,
-//                 nms_4->get_sort_result_descending(),
-//                 nms_4->get_output_type());
-//
-//         new_ops_list.push_back(nms_5);
-//
-//         // vector of new nGraph operations
-//         NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
-//
-//         nms_5->set_friendly_name(nms_4->get_friendly_name());
-//         ngraph::copy_runtime_info(nms_4, new_ops);
-//         ngraph::replace_node(nms_4, nms_5);
-//         return true;
-//     };
-//
-//     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMS5");
-//     this->register_matcher(m, callback);
+        auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+        switch (nms_3->get_box_encoding()) {
+            case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CENTER:
+                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
+                break;
+            case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CORNER:
+                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
+
+        // list of new nGraph operations
+        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
+
+        new_ops_list.push_front(arg5);
+        if (num_of_args <= 4) {
+            new_ops_list.push_front(arg4.get_node_shared_ptr());
+        }
+        if (num_of_args <= 3) {
+            new_ops_list.push_front(arg3.get_node_shared_ptr());
+        }
+        if (num_of_args <= 2) {
+            new_ops_list.push_front(arg2.get_node_shared_ptr());
+        }
+
+        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
+                new_args.at(0),
+                new_args.at(1),
+                arg2,
+                arg3,
+                arg4,
+                arg5,
+                box_encoding,
+                nms_3->get_sort_result_descending(),
+                nms_3->get_output_type());
+
+        new_ops_list.push_back(nms_5);
+
+        // vector of new nGraph operations
+        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
+
+        nms_5->set_friendly_name(nms_3->get_friendly_name());
+        ngraph::copy_runtime_info(nms_3, new_ops);
+        ngraph::replace_node(nms_3, nms_5);
+        return true;
+    };
+
+    auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS3ToNMS5");
+    this->register_matcher(m, callback);
 }

From 57e4042f962b0f574835ab6043bea1226b111a01 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 15:49:53 +0300
Subject: [PATCH 058/173] Fixed typo.

---
 .../src/transformations/convert_previous_nms_to_nms_5.cpp       | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 8975aeca432a5e..dfee87a053be49 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -130,7 +130,7 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
                 center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
                 break;
             default:
-                throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+                throw ngraph_error("NonMaxSuppression layer " + nms_3->get_friendly_name() +
                                    " has unsupported box encoding");
         }
 

From 515dbb7149ddbad89758aeda96b401cc9ccdd76a Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 16:15:45 +0300
Subject: [PATCH 059/173] Started to write conversion of NMS-1 to NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp          | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index dfee87a053be49..123269ea3d3a14 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -44,7 +44,7 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
         switch (nms_4->get_box_encoding()) {
             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
                 break;
             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
                 box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
@@ -124,10 +124,10 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
         switch (nms_3->get_box_encoding()) {
             case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CENTER:
-                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;;
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
                 break;
             case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CORNER:
-                center_point_box = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
                 break;
             default:
                 throw ngraph_error("NonMaxSuppression layer " + nms_3->get_friendly_name() +
@@ -173,3 +173,15 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS3ToNMS5");
     this->register_matcher(m, callback);
 }
+
+NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS1ToNMS5, "ConvertNMS1ToNMS5", 0);
+
+ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
+    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+    auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
+    auto iou_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.75});
+    auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
+    auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                   iou_threshold, score_threshold);
+}

From a0b0a3fd2af4cef547429b44cfe06ca88195074c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 16:50:49 +0300
Subject: [PATCH 060/173] Written draft of the conversion NMS-1 to NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp         | 67 +++++++++++++++++++
 1 file changed, 67 insertions(+)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 123269ea3d3a14..5c7dbd8dfb7c24 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -184,4 +184,71 @@ ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
     auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
     auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
                                                                    iou_threshold, score_threshold);
+
+    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
+        auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());
+        if (!nms_1) {
+            return false;
+        }
+
+        const auto new_args = nms_1->input_values();
+
+        size_t num_of_args = new_args.size();
+
+        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+
+        auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+        switch (nms_1->get_box_encoding()) {
+            case ::ngraph::opset1::NonMaxSuppression::BoxEncodingType::CENTER:
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+                break;
+            case ::ngraph::opset1::NonMaxSuppression::BoxEncodingType::CORNER:
+                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms_1->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
+
+        // list of new nGraph operations
+        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
+
+        new_ops_list.push_front(arg5);
+        if (num_of_args <= 4) {
+            new_ops_list.push_front(arg4.get_node_shared_ptr());
+        }
+        if (num_of_args <= 3) {
+            new_ops_list.push_front(arg3.get_node_shared_ptr());
+        }
+        if (num_of_args <= 2) {
+            new_ops_list.push_front(arg2.get_node_shared_ptr());
+        }
+
+        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
+                new_args.at(0),
+                new_args.at(1),
+                arg2,
+                arg3,
+                arg4,
+                arg5,
+                box_encoding,
+                nms_1->get_sort_result_descending(),
+                ::ngraph::element::i64);
+
+        new_ops_list.push_back(nms_5);
+
+        // vector of new nGraph operations
+        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
+
+        nms_5->set_friendly_name(nms_1->get_friendly_name());
+        ngraph::copy_runtime_info(nms_1, new_ops);
+        ngraph::replace_node(nms_1, nms_5);
+        return true;
+    };
+
+    auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS1ToNMS5");
+    this->register_matcher(m, callback);
 }

From 56b52ddb0c7554ed48f3dae7fe6b2126329e5517 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 7 Oct 2020 17:37:35 +0300
Subject: [PATCH 061/173] Started to write tests for the conversion nGraph
 NMS-5 to inner NMS.

---
 .../common_optimizations.cpp                  |   5 +
 .../transformations/convert_nms5_test.cpp     | 148 ++++++++++++++++++
 2 files changed, 153 insertions(+)
 create mode 100644 inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp

diff --git a/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp b/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
index fae8de5cfcbae4..2c2b85c8f15bf8 100644
--- a/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
+++ b/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
@@ -41,6 +41,7 @@
 #include "transformations/remove_filtering_boxes_by_size.hpp"
 #include "transformations/hswish_decomposition.hpp"
 #include "transformations/hswish_fusion.hpp"
+#include "transformations/convert_previous_nms_to_nms_5.hpp"
 
 #include <ngraph/pass/manager.hpp>
 #include <ngraph/pass/constant_folding.hpp>
@@ -105,6 +106,10 @@ bool ngraph::pass::CommonOptimizations::run_on_function(std::shared_ptr<ngraph::
     manager.register_pass<ngraph::pass::GroupConvolutionBackpropDataMultiplyFusion>();
     manager.register_pass<ngraph::pass::ConstantFolding>();
 
+    manager.register_pass<ngraph::pass::ConvertNMS1ToNMS5>();
+    manager.register_pass<ngraph::pass::ConvertNMS3ToNMS5>();
+    manager.register_pass<ngraph::pass::ConvertNMS4ToNMS5>();
+
     auto fq_fusions = manager.register_pass<ngraph::pass::GraphRewrite>();
     fq_fusions->add_matcher<ngraph::pass::FakeQuantizeMulFusion>();
     fq_fusions->add_matcher<ngraph::pass::FakeQuantizeReshapeFusion>();
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
new file mode 100644
index 00000000000000..73d519d7e2f2e9
--- /dev/null
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -0,0 +1,148 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <gtest/gtest.h>
+
+#include <string>
+#include <memory>
+
+#include <ngraph/function.hpp>
+#include <ngraph/opsets/opset4.hpp>
+#include <ngraph_ops/nms_ie.hpp>
+#include <transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp>
+#include <transformations/init_node_info.hpp>
+#include <transformations/utils/utils.hpp>
+#include <ngraph/pass/manager.hpp>
+
+#include "common_test_utils/ngraph_test_utils.hpp"
+
+using namespace testing;
+using namespace ngraph;
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+//     {
+//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
+//         auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
+//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
+//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+//         auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+//                                                                opset4::NonMaxSuppression::BoxEncodingType::CORNER, true);
+//
+//         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+//
+//         const auto &orig_shape = f->get_output_partial_shape(0);
+//         pass::Manager manager;
+//         manager.register_pass<pass::InitNodeInfo>();
+//         manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
+//         manager.run_passes(f);
+//         ASSERT_NO_THROW(check_rt_info(f));
+//         ASSERT_TRUE(f->get_output_partial_shape(0).is_static()) << "Shape " << f->get_output_partial_shape(0) << " should be static";
+//     }
+//
+//     {
+//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
+//         auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
+//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
+//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+//         auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
+//                                                               std::make_shared<opset4::Unsqueeze>(iou_threshold,
+//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
+//                                                               std::make_shared<opset4::Unsqueeze>(score_threshold,
+//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
+//                                                               0, true);
+//         nms->set_friendly_name("nms");
+//
+//         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+//         ASSERT_TRUE(f_ref->get_output_partial_shape(0).is_static()) << "Shape " << f_ref->get_output_partial_shape(0) << " should be static";
+//     }
+//
+//     auto res = compare_functions(f, f_ref);
+//     ASSERT_TRUE(res.first) << res.second;
+}
+
+// TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic1) {
+//     std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+//     {
+//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
+//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
+//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
+//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+//         auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+//                                                                opset4::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
+//
+//         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+//
+//         pass::Manager manager;
+//         manager.register_pass<pass::InitNodeInfo>();
+//         manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
+//         manager.run_passes(f);
+//         f->validate_nodes_and_infer_types();
+//         ASSERT_NO_THROW(check_rt_info(f));
+//     }
+//
+//     {
+//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
+//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
+//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
+//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+//         auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
+//                                                               std::make_shared<opset4::Unsqueeze>(iou_threshold,
+//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
+//                                                               std::make_shared<opset4::Unsqueeze>(score_threshold,
+//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
+//                                                               0, true);
+//
+//         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+//     }
+//
+//     auto res = compare_functions(f, f_ref);
+//     ASSERT_TRUE(res.first) << res.second;
+// }
+//
+// TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic2) {
+//     std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+//     {
+//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
+//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+//         auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+//                                                                opset4::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
+//
+//         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+//
+//         pass::Manager manager;
+//         manager.register_pass<pass::InitNodeInfo>();
+//         manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
+//         manager.run_passes(f);
+//
+//         f->validate_nodes_and_infer_types();
+//         ASSERT_NO_THROW(check_rt_info(f));
+//     }
+//
+//     {
+//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
+//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+//         auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
+//                                                               std::make_shared<opset4::Unsqueeze>(iou_threshold,
+//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
+//                                                               std::make_shared<opset4::Unsqueeze>(score_threshold,
+//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
+//                                                               0, true);
+//
+//         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+//     }
+//
+//     auto res = compare_functions(f, f_ref);
+//     ASSERT_TRUE(res.first) << res.second;
+// }

From 00e5c1707dd74af5ca8d9bd9e4a87eaeb233dd2f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 8 Oct 2020 09:16:01 +0300
Subject: [PATCH 062/173] Started to write the test ConvertNMS5ToNMSIEStatic.

---
 .../transformations/convert_nms5_test.cpp     | 43 ++++++++++---------
 1 file changed, 23 insertions(+), 20 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 73d519d7e2f2e9..073e4d4240aabe 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -8,7 +8,7 @@
 #include <memory>
 
 #include <ngraph/function.hpp>
-#include <ngraph/opsets/opset4.hpp>
+#include <ngraph/opsets/opset5.hpp>
 #include <ngraph_ops/nms_ie.hpp>
 #include <transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp>
 #include <transformations/init_node_info.hpp>
@@ -22,26 +22,28 @@ using namespace ngraph;
 
 TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
     std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-//     {
-//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
-//         auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
-//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
-//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-//         auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-//                                                                opset4::NonMaxSuppression::BoxEncodingType::CORNER, true);
-//
-//         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-//
-//         const auto &orig_shape = f->get_output_partial_shape(0);
-//         pass::Manager manager;
-//         manager.register_pass<pass::InitNodeInfo>();
-//         manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
-//         manager.run_passes(f);
-//         ASSERT_NO_THROW(check_rt_info(f));
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
+        const auto &orig_selected_scores_shape = f->get_output_partial_shape(1);
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
 //         ASSERT_TRUE(f->get_output_partial_shape(0).is_static()) << "Shape " << f->get_output_partial_shape(0) << " should be static";
-//     }
-//
+    }
+
 //     {
 //         auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
 //         auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
@@ -62,6 +64,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
 //
 //     auto res = compare_functions(f, f_ref);
 //     ASSERT_TRUE(res.first) << res.second;
+    ASSERT_TRUE(true) << "";
 }
 
 // TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic1) {

From 97e6809e8b1d2f8ef2a7ceefa708b6370a802a77 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 8 Oct 2020 15:26:33 +0300
Subject: [PATCH 063/173] Written tests for conversion of nGraph NMS-5 to inner
 NMSIE3.

---
 .../transformations/convert_nms5_test.cpp     | 219 +++++++++---------
 1 file changed, 114 insertions(+), 105 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 073e4d4240aabe..e9adb14669b1d5 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -35,117 +35,126 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
-        const auto &orig_selected_scores_shape = f->get_output_partial_shape(1);
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
         manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
-//         ASSERT_TRUE(f->get_output_partial_shape(0).is_static()) << "Shape " << f->get_output_partial_shape(0) << " should be static";
+        ASSERT_TRUE(f->get_output_partial_shape(0).is_static()) << "Shape " << f->get_output_partial_shape(0) << " should be static";
     }
 
-//     {
-//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
-//         auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
-//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
-//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-//         auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
-//                                                               std::make_shared<opset4::Unsqueeze>(iou_threshold,
-//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
-//                                                               std::make_shared<opset4::Unsqueeze>(score_threshold,
-//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
-//                                                               0, true);
-//         nms->set_friendly_name("nms");
-//
-//         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-//         ASSERT_TRUE(f_ref->get_output_partial_shape(0).is_static()) << "Shape " << f_ref->get_output_partial_shape(0) << " should be static";
-//     }
-//
-//     auto res = compare_functions(f, f_ref);
-//     ASSERT_TRUE(res.first) << res.second;
-    ASSERT_TRUE(true) << "";
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{1}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, max_output_boxes_per_class,
+                                                              std::make_shared<opset5::Unsqueeze>(iou_threshold,
+                                                                                                  opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                              std::make_shared<opset5::Unsqueeze>(score_threshold,
+                                                                                                  opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                              std::make_shared<opset5::Unsqueeze>(soft_nms_sigma,
+                                                                                                  opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                              0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+        ASSERT_TRUE(f_ref->get_output_partial_shape(0).is_static()) << "Shape " << f_ref->get_output_partial_shape(0) << " should be static";
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
 }
 
-// TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic1) {
-//     std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-//     {
-//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
-//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-//         auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-//                                                                opset4::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
-//
-//         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-//
-//         pass::Manager manager;
-//         manager.register_pass<pass::InitNodeInfo>();
-//         manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
-//         manager.run_passes(f);
-//         f->validate_nodes_and_infer_types();
-//         ASSERT_NO_THROW(check_rt_info(f));
-//     }
-//
-//     {
-//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
-//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-//         auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
-//                                                               std::make_shared<opset4::Unsqueeze>(iou_threshold,
-//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
-//                                                               std::make_shared<opset4::Unsqueeze>(score_threshold,
-//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
-//                                                               0, true);
-//
-//         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-//     }
-//
-//     auto res = compare_functions(f, f_ref);
-//     ASSERT_TRUE(res.first) << res.second;
-// }
-//
-// TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic2) {
-//     std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-//     {
-//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
-//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
-//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-//         auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-//                                                                opset4::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
-//
-//         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-//
-//         pass::Manager manager;
-//         manager.register_pass<pass::InitNodeInfo>();
-//         manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
-//         manager.run_passes(f);
-//
-//         f->validate_nodes_and_infer_types();
-//         ASSERT_NO_THROW(check_rt_info(f));
-//     }
-//
-//     {
-//         auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
-//         auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-//         auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
-//         auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-//         auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-//         auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
-//                                                               std::make_shared<opset4::Unsqueeze>(iou_threshold,
-//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
-//                                                               std::make_shared<opset4::Unsqueeze>(score_threshold,
-//                                                                                                   opset4::Constant::create(element::i64, Shape{1}, {0})),
-//                                                               0, true);
-//
-//         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-//     }
-//
-//     auto res = compare_functions(f, f_ref);
-//     ASSERT_TRUE(res.first) << res.second;
-// }
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
+
+         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+         pass::Manager manager;
+         manager.register_pass<pass::InitNodeInfo>();
+         manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+         manager.run_passes(f);
+         f->validate_nodes_and_infer_types();
+         ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+     {
+         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{1}, {10});
+         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, max_output_boxes_per_class,
+                                                               std::make_shared<opset5::Unsqueeze>(iou_threshold,
+                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                               std::make_shared<opset5::Unsqueeze>(score_threshold,
+                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                               std::make_shared<opset5::Unsqueeze>(soft_nms_sigma,
+                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                               0, true, element::i32);
+
+         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+     }
+
+     auto res = compare_functions(f, f_ref);
+     ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
+
+         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+         pass::Manager manager;
+         manager.register_pass<pass::InitNodeInfo>();
+         manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+         manager.run_passes(f);
+
+         f->validate_nodes_and_infer_types();
+         ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{1}, {10});
+         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, max_output_boxes_per_class,
+                                                               std::make_shared<opset5::Unsqueeze>(iou_threshold,
+                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                               std::make_shared<opset5::Unsqueeze>(score_threshold,
+                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                               std::make_shared<opset5::Unsqueeze>(soft_nms_sigma,
+                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
+                                                               0, true, element::i32);
+
+         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}

From ca757963ab4a005c7b1eb81427ea9a8703faf41d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 8 Oct 2020 16:15:39 +0300
Subject: [PATCH 064/173] Started to write tests for conversion of previous NMS
 to nGraph NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp         | 72 +++++++++++++++++++
 1 file changed, 72 insertions(+)
 create mode 100644 inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
new file mode 100644
index 00000000000000..fded694e396142
--- /dev/null
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
@@ -0,0 +1,72 @@
+// Copyright (C) 2020 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include <gtest/gtest.h>
+
+#include <string>
+#include <memory>
+#include <queue>
+
+#include <ngraph/function.hpp>
+#include <ngraph/opsets/opset1.hpp>
+#include <ngraph/opsets/opset3.hpp>
+#include <ngraph/opsets/opset4.hpp>
+#include <ngraph/opsets/opset5.hpp>
+#include <transformations/convert_previous_nms_to_nms_5.hpp>
+#include <transformations/init_node_info.hpp>
+#include <transformations/utils/utils.hpp>
+#include <ngraph/pass/manager.hpp>
+
+#include "common_test_utils/ngraph_test_utils.hpp"
+
+using namespace testing;
+using namespace ngraph;
+
+TEST(TransformationTests, ConvertNMS4ToNMS5) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               opset4::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        pass::Manager m;
+        m.register_pass<pass::InitNodeInfo>();
+        m.register_pass<pass::ConvertNMS4ToNMS5>();
+        m.run_passes();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+//     {
+//         auto data = std::make_shared<ngraph::opset1::Parameter>(ngraph::element::f32, ngraph::Shape{3, 1, 2});
+//         auto divide_constant = ngraph::opset1::Constant::create(ngraph::element::f32, ngraph::Shape{1}, {1.5});
+//         auto divide = std::make_shared<ngraph::opset1::Divide>(data, divide_constant);
+//
+//         f = std::make_shared<ngraph::Function>(ngraph::NodeVector{divide}, ngraph::ParameterVector{data});
+//
+//         ngraph::pass::Manager m;
+//         m.register_pass<ngraph::pass::InitNodeInfo>();
+//         m.register_pass<ngraph::pass::ConvertDivide>();
+//         m.run_passes(f);
+//         ASSERT_NO_THROW(check_rt_info(f));
+//     }
+//
+//     {
+//         auto data = std::make_shared<ngraph::opset1::Parameter>(ngraph::element::f32, ngraph::Shape{3, 1, 2});
+//         auto divide_constant = ngraph::opset1::Constant::create(ngraph::element::f32, ngraph::Shape{1}, {1.5});
+//         auto pow = std::make_shared<ngraph::opset1::Power>(divide_constant,
+//                                                            ngraph::opset1::Constant::create(ngraph::element::f32, ngraph::Shape{1}, {-1}));
+//         auto mul = std::make_shared<ngraph::opset1::Multiply>(data, pow);
+//
+//         f_ref = std::make_shared<ngraph::Function>(ngraph::NodeVector{mul}, ngraph::ParameterVector{data});
+//     }
+//
+//     auto res = compare_functions(f, f_ref);
+//     ASSERT_TRUE(res.first) << res.second;
+    ASSERT_TRUE(true) << "";
+}

From deb6ff04914e4f6c90042a0cb59f2492eea1a08c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 9 Oct 2020 11:25:42 +0300
Subject: [PATCH 065/173] Written tests for conversion of old nGraph NMS to
 NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp         |  10 +-
 .../convert_previous_nms_to_nms_5.cpp         | 220 +++++++++++++++---
 2 files changed, 196 insertions(+), 34 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
index 5c7dbd8dfb7c24..ee4ab6a6f8dc99 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
@@ -36,7 +36,7 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
 
         size_t num_of_args = new_args.size();
 
-        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
         const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
@@ -86,7 +86,7 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
 
         nms_5->set_friendly_name(nms_4->get_friendly_name());
         ngraph::copy_runtime_info(nms_4, new_ops);
-        ngraph::replace_node(nms_4, nms_5);
+        nms_4->output(0).replace(nms_5->output(0));
         return true;
     };
 
@@ -116,7 +116,7 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
 
         size_t num_of_args = new_args.size();
 
-        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
         const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
@@ -166,7 +166,7 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
 
         nms_5->set_friendly_name(nms_3->get_friendly_name());
         ngraph::copy_runtime_info(nms_3, new_ops);
-        ngraph::replace_node(nms_3, nms_5);
+        nms_3->output(0).replace(nms_5->output(0));
         return true;
     };
 
@@ -245,7 +245,7 @@ ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
 
         nms_5->set_friendly_name(nms_1->get_friendly_name());
         ngraph::copy_runtime_info(nms_1, new_ops);
-        ngraph::replace_node(nms_1, nms_5);
+        nms_1->output(0).replace(nms_5->output(0));
         return true;
     };
 
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
index fded694e396142..e965d8d3d23e4f 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
@@ -23,7 +23,7 @@
 using namespace testing;
 using namespace ngraph;
 
-TEST(TransformationTests, ConvertNMS4ToNMS5) {
+TEST(TransformationTests, ConvertNMS4FiveInputsToNMS5) {
     std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
     {
         auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
@@ -39,34 +39,196 @@ TEST(TransformationTests, ConvertNMS4ToNMS5) {
         pass::Manager m;
         m.register_pass<pass::InitNodeInfo>();
         m.register_pass<pass::ConvertNMS4ToNMS5>();
-        m.run_passes();
+        m.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
     }
-//     {
-//         auto data = std::make_shared<ngraph::opset1::Parameter>(ngraph::element::f32, ngraph::Shape{3, 1, 2});
-//         auto divide_constant = ngraph::opset1::Constant::create(ngraph::element::f32, ngraph::Shape{1}, {1.5});
-//         auto divide = std::make_shared<ngraph::opset1::Divide>(data, divide_constant);
-//
-//         f = std::make_shared<ngraph::Function>(ngraph::NodeVector{divide}, ngraph::ParameterVector{data});
-//
-//         ngraph::pass::Manager m;
-//         m.register_pass<ngraph::pass::InitNodeInfo>();
-//         m.register_pass<ngraph::pass::ConvertDivide>();
-//         m.run_passes(f);
-//         ASSERT_NO_THROW(check_rt_info(f));
-//     }
-//
-//     {
-//         auto data = std::make_shared<ngraph::opset1::Parameter>(ngraph::element::f32, ngraph::Shape{3, 1, 2});
-//         auto divide_constant = ngraph::opset1::Constant::create(ngraph::element::f32, ngraph::Shape{1}, {1.5});
-//         auto pow = std::make_shared<ngraph::opset1::Power>(divide_constant,
-//                                                            ngraph::opset1::Constant::create(ngraph::element::f32, ngraph::Shape{1}, {-1}));
-//         auto mul = std::make_shared<ngraph::opset1::Multiply>(data, pow);
-//
-//         f_ref = std::make_shared<ngraph::Function>(ngraph::NodeVector{mul}, ngraph::ParameterVector{data});
-//     }
-//
-//     auto res = compare_functions(f, f_ref);
-//     ASSERT_TRUE(res.first) << res.second;
-    ASSERT_TRUE(true) << "";
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {.0f});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS4TwoInputsToNMS5) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, opset4::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        pass::Manager m;
+        m.register_pass<pass::InitNodeInfo>();
+        m.register_pass<pass::ConvertNMS4ToNMS5>();
+        m.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS3FiveInputsToNMS5) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset3::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset3::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset3::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               opset3::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        pass::Manager m;
+        m.register_pass<pass::InitNodeInfo>();
+        m.register_pass<pass::ConvertNMS3ToNMS5>();
+        m.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {.0f});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS3TwoInputsToNMS5) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto nms = std::make_shared<opset3::NonMaxSuppression>(boxes, scores, opset3::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        pass::Manager m;
+        m.register_pass<pass::InitNodeInfo>();
+        m.register_pass<pass::ConvertNMS3ToNMS5>();
+        m.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS1FiveInputsToNMS5) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset1::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset1::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset1::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               opset1::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        pass::Manager m;
+        m.register_pass<pass::InitNodeInfo>();
+        m.register_pass<pass::ConvertNMS1ToNMS5>();
+        m.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {.0f});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS1TwoInputsToNMS5) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset1::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto nms = std::make_shared<opset1::NonMaxSuppression>(boxes, scores, opset1::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+
+        pass::Manager m;
+        m.register_pass<pass::InitNodeInfo>();
+        m.register_pass<pass::ConvertNMS1ToNMS5>();
+        m.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
 }

From 2a41688a9e1a3523f4c3c37b9988715205cb1958 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 9 Oct 2020 12:51:04 +0300
Subject: [PATCH 066/173] Started to write tests for
 opset5::NonMaxSuppression::evaluate().

---
 ngraph/test/CMakeLists.txt                    |  1 +
 ngraph/test/op_eval/non_max_suppression_5.cpp | 33 +++++++++++++++++++
 2 files changed, 34 insertions(+)
 create mode 100644 ngraph/test/op_eval/non_max_suppression_5.cpp

diff --git a/ngraph/test/CMakeLists.txt b/ngraph/test/CMakeLists.txt
index 253c1e9737058a..bfe837a47e174f 100644
--- a/ngraph/test/CMakeLists.txt
+++ b/ngraph/test/CMakeLists.txt
@@ -74,6 +74,7 @@ set(SRC
     op_eval/interpolate.cpp
     op_eval/matmul.cpp
     op_eval/mish.cpp
+    op_eval/non_max_suppression_5.cpp
     op_eval/non_zero.cpp
     op_eval/reduce_l1.cpp
     op_eval/reduce_l2.cpp
diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
new file mode 100644
index 00000000000000..73f431ac5fa443
--- /dev/null
+++ b/ngraph/test/op_eval/non_max_suppression_5.cpp
@@ -0,0 +1,33 @@
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#include <string>
+#include <vector>
+
+#include "gtest/gtest.h"
+
+#include "ngraph/op/non_max_suppression.hpp"
+#include "ngraph/runtime/host_tensor.hpp"
+#include "ngraph/validation_util.hpp"
+#include "runtime/backend.hpp"
+#include "util/test_tools.hpp"
+#include "util/type_prop.hpp"
+
+using namespace std;
+using namespace ngraph;
+
+TEST(op_eval, nonmaxsuppression_center_point_box_format)
+{
+    std::vector<float> boxes = {0.5, 0.5, 1.0, 1.0, 0.5, 0.6, 1.0, 1.0, 0.5, 0.4, 1.0, 1.0, 0.5,
+                                10.5, 1.0, 1.0, 0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
+    std::vector<float> scores = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+    const int64_t max_output_boxes_per_class = 3;
+    const float iou_threshold = 0.0f;
+    const float score_threshold = 0.0f;
+    ASSERT_TRUE(true);
+}
\ No newline at end of file

From c0aea2d26fad5a34893c498b31fd29e4f10a9abb Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 9 Oct 2020 13:06:41 +0300
Subject: [PATCH 067/173] Some additions.

---
 ngraph/test/op_eval/non_max_suppression_5.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
index 73f431ac5fa443..c7d36e7bd80ce5 100644
--- a/ngraph/test/op_eval/non_max_suppression_5.cpp
+++ b/ngraph/test/op_eval/non_max_suppression_5.cpp
@@ -29,5 +29,9 @@ TEST(op_eval, nonmaxsuppression_center_point_box_format)
     const int64_t max_output_boxes_per_class = 3;
     const float iou_threshold = 0.0f;
     const float score_threshold = 0.0f;
+    const auto box_encoding = op::v5::BoxEncodingType::CENTER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
     ASSERT_TRUE(true);
 }
\ No newline at end of file

From 6944c1a500a5636c6b8ada79083fddacfe2bd4a6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 9 Oct 2020 13:08:50 +0300
Subject: [PATCH 068/173] Small fix.

---
 ngraph/test/op_eval/non_max_suppression_5.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
index c7d36e7bd80ce5..fbcc3dc00455cf 100644
--- a/ngraph/test/op_eval/non_max_suppression_5.cpp
+++ b/ngraph/test/op_eval/non_max_suppression_5.cpp
@@ -29,7 +29,7 @@ TEST(op_eval, nonmaxsuppression_center_point_box_format)
     const int64_t max_output_boxes_per_class = 3;
     const float iou_threshold = 0.0f;
     const float score_threshold = 0.0f;
-    const auto box_encoding = op::v5::BoxEncodingType::CENTER;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CENTER;
     const auto boxes_shape = Shape{1, 6, 4};
     const auto scores_shape = Shape{1, 1, 6};
 

From 9494d73665d6cc1c67f1fd75ea07102191081276 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 13 Oct 2020 15:48:46 +0300
Subject: [PATCH 069/173] Written tests for
 op::v5::NonMaxSuppression::evaluate().

---
 ngraph/core/src/op/non_max_suppression.cpp    |  87 ++-
 ngraph/test/op_eval/non_max_suppression_5.cpp | 526 +++++++++++++++++-
 2 files changed, 554 insertions(+), 59 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 3c4c406780d054..791d1e7c969606 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -889,63 +889,49 @@ using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
 
 static void normalize_corner(float* boxes, const Shape& boxes_shape)
 {
-    size_t num_batches = boxes_shape[0];
-    size_t num_boxes = boxes_shape[1];
-
-    float* box_ptr = boxes;
-
-    for (size_t batch = 0; batch < num_batches; ++batch)
+    size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+    for (size_t i = 0; i < total_num_of_boxes; ++i)
     {
-        for (size_t box = 0; box < num_boxes; ++num_boxes)
-        {
-            float y1 = box_ptr[0];
-            float x1 = box_ptr[1];
-            float y2 = box_ptr[2];
-            float x2 = box_ptr[3];
-
-            float ymin = std::min(y1, y2);
-            float ymax = std::max(y1, y2);
-            float xmin = std::min(x1, x2);
-            float xmax = std::max(x1, x2);
-
-            box_ptr[0] = ymin;
-            box_ptr[1] = xmin;
-            box_ptr[2] = ymax;
-            box_ptr[3] = xmax;
-
-            box_ptr += 4;
-        }
+        float* current_box = boxes + 4 * i;
+
+        float y1 = current_box[0];
+        float x1 = current_box[1];
+        float y2 = current_box[2];
+        float x2 = current_box[3];
+
+        float ymin = std::min(y1, y2);
+        float ymax = std::max(y1, y2);
+        float xmin = std::min(x1, x2);
+        float xmax = std::max(x1, x2);
+
+        current_box[0] = ymin;
+        current_box[1] = xmin;
+        current_box[2] = ymax;
+        current_box[3] = xmax;
     }
 }
 
 static void normalize_center(float* boxes, const Shape& boxes_shape)
 {
-    size_t num_batches = boxes_shape[0];
-    size_t num_boxes = boxes_shape[1];
-
-    float* box_ptr = boxes;
-
-    for (size_t batch = 0; batch < num_batches; ++batch)
+    size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+    for (size_t i = 0; i < total_num_of_boxes; ++i)
     {
-        for (size_t box = 0; box < num_boxes; ++num_boxes)
-        {
-            float x_center = box_ptr[0];
-            float y_center = box_ptr[1];
-            float width = box_ptr[2];
-            float height = box_ptr[3];
-
-            float y1 = y_center - height / 2.0;
-            float x1 = x_center - width / 2.0;
-            float y2 = y_center + height / 2.0;
-            float x2 = x_center + width / 2.0;
-
-            box_ptr[0] = y1;
-            box_ptr[1] = x1;
-            box_ptr[2] = y2;
-            box_ptr[3] = x2;
-
-            box_ptr += 4;
-        }
+        float* current_box = boxes + 4 * i;
+
+        float x_center = current_box[0];
+        float y_center = current_box[1];
+        float width = current_box[2];
+        float height = current_box[3];
+
+        float y1 = y_center - height / 2.0;
+        float x1 = x_center - width / 2.0;
+        float y2 = y_center + height / 2.0;
+        float x2 = x_center + width / 2.0;
+
+        current_box[0] = y1;
+        current_box[1] = x1;
+        current_box[2] = y2;
+        current_box[3] = x2;
     }
 }
 
@@ -986,7 +972,6 @@ static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
     }
 
     normalize_box_encoding(result.data(), boxes_shape, box_encoding);
-
     return result;
 }
 
diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
index fbcc3dc00455cf..cd8e496ae7fc44 100644
--- a/ngraph/test/op_eval/non_max_suppression_5.cpp
+++ b/ngraph/test/op_eval/non_max_suppression_5.cpp
@@ -11,6 +11,7 @@
 
 #include "gtest/gtest.h"
 
+#include "ngraph/op/constant.hpp"
 #include "ngraph/op/non_max_suppression.hpp"
 #include "ngraph/runtime/host_tensor.hpp"
 #include "ngraph/validation_util.hpp"
@@ -23,15 +24,524 @@ using namespace ngraph;
 
 TEST(op_eval, nonmaxsuppression_center_point_box_format)
 {
-    std::vector<float> boxes = {0.5, 0.5, 1.0, 1.0, 0.5, 0.6, 1.0, 1.0, 0.5, 0.4, 1.0, 1.0, 0.5,
-                                10.5, 1.0, 1.0, 0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
-    std::vector<float> scores = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-    const int64_t max_output_boxes_per_class = 3;
-    const float iou_threshold = 0.0f;
-    const float score_threshold = 0.0f;
+    std::vector<float> boxes_data = {
+        0.5, 0.5, 1.0, 1.0, 0.5, 0.6, 1.0, 1.0, 0.5, 0.4, 1.0, 1.0, 0.5, 10.5, 1.0, 1.0, 0.5, 10.6,
+        1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
     const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CENTER;
     const auto boxes_shape = Shape{1, 6, 4};
     const auto scores_shape = Shape{1, 1, 6};
 
-    ASSERT_TRUE(true);
-}
\ No newline at end of file
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
+    std::vector<int64_t> expected_valid_outputs = {3};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{3, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{3, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_flipped_coordinates)
+{
+    std::vector<float> boxes_data = {
+        1.0, 1.0, 0.0, 0.0, 0.0, 0.1, 1.0, 1.1, 0.0, 0.9, 1.0, -0.1, 0.0, 10.0, 1.0, 11.0,
+        1.0, 10.1, 0.0, 11.1, 1.0, 101.0, 0.0, 100.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
+    std::vector<int64_t> expected_valid_outputs = {3};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{3, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{3, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_identical_boxes)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0};
+
+    std::vector<float> scores_data = {0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 10, 4};
+    const auto scores_shape = Shape{1, 1, 10};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {1};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{1, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{1, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_limit_output_size)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
+        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 2;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {2};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{2, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{2, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_single_box)
+{
+    std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0};
+
+    std::vector<float> scores_data = {0.9};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 1, 4};
+    const auto scores_shape = Shape{1, 1, 1};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {1};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{1, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{1, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_suppress_by_IOU)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
+        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
+    std::vector<int64_t> expected_valid_outputs = {3};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{3, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{3, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_suppress_by_IOU_and_scores)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
+        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.4f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {2};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{2, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{2, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_two_batches)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
+        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1,
+        0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0, 0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {
+        0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 2;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{2, 6, 4};
+    const auto scores_shape = Shape{2, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 1, 0, 3, 1, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 1.0, 0.0, 0.95, 1.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {4};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{4, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{4, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}
+
+TEST(op_eval, nonmaxsuppression_two_classes)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9,
+        0.0, 10.0, 1.0, 11.0, 0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {
+        0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 2;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 2, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    HostTensorVector results(3);
+    for (auto& result : results)
+    {
+        result = make_shared<HostTensor>();
+    }
+    ASSERT_TRUE(
+        f->evaluate(results,
+                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 1, 3, 0, 1, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 1.0, 0.95, 0.0, 1.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {4};
+
+    EXPECT_EQ(results[0]->get_element_type(), element::i64);
+    EXPECT_EQ(results[1]->get_element_type(), element::f32);
+    EXPECT_EQ(results[2]->get_element_type(), element::i64);
+    EXPECT_EQ(results[0]->get_shape(), (Shape{4, 3}));
+    EXPECT_EQ(results[1]->get_shape(), (Shape{4, 3}));
+    EXPECT_EQ(results[2]->get_shape(), (Shape{}));
+    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
+    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
+    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
+}

From 54e89f51618734953304880652d626b76c2a7fc5 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 10:18:44 +0300
Subject: [PATCH 070/173] Used NGRAPH_RTTI_DECLARATION for
 NonMaxSuppressionIE3.

---
 .../src/transformations/include/ngraph_ops/nms_ie.hpp          | 3 +--
 inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp | 2 +-
 2 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
index 721392f8a42902..aace940cf0a87e 100644
--- a/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
+++ b/inference-engine/src/transformations/include/ngraph_ops/nms_ie.hpp
@@ -60,8 +60,7 @@ class TRANSFORMATIONS_API NonMaxSuppressionIE2 : public NonMaxSuppressionIE {
 
 class TRANSFORMATIONS_API NonMaxSuppressionIE3 : public Op {
 public:
-    static constexpr NodeTypeInfo type_info{"NonMaxSuppressionIE", 3};
-    const NodeTypeInfo& get_type_info() const override { return type_info; }
+    NGRAPH_RTTI_DECLARATION;
 
     NonMaxSuppressionIE3(const Output<Node>& boxes,
                          const Output<Node>& scores,
diff --git a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
index ce039843d9d701..965ecafb00c56b 100644
--- a/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/transformations/src/ngraph_ops/nms_ie.cpp
@@ -102,7 +102,7 @@ void op::NonMaxSuppressionIE2::validate_and_infer_types() {
     set_output_type(0, nms->output(0).get_element_type(), nms->output(0).get_partial_shape());
 }
 
-constexpr NodeTypeInfo op::NonMaxSuppressionIE3::type_info;
+NGRAPH_RTTI_DEFINITION(op::NonMaxSuppressionIE3, "NonMaxSuppressionIE", 3);
 
 op::NonMaxSuppressionIE3::NonMaxSuppressionIE3(const Output<Node>& boxes,
                                                const Output<Node>& scores,

From 61051a7b83c374604926ce748fed6895ed866368 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 10:25:33 +0300
Subject: [PATCH 071/173] Used NGRAPH_RTTI_DECLARATION for NMS-5.

---
 ngraph/core/include/ngraph/op/non_max_suppression.hpp | 3 +--
 ngraph/core/src/op/non_max_suppression.cpp            | 2 +-
 2 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index a3828e9f65f704..817e8128242c63 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -243,14 +243,13 @@ namespace ngraph
             class NGRAPH_API NonMaxSuppression : public Op
             {
             public:
+                NGRAPH_RTTI_DECLARATION;
                 enum class BoxEncodingType
                 {
                     CORNER,
                     CENTER
                 };
 
-                static constexpr NodeTypeInfo type_info{"NonMaxSuppression", 5};
-                const NodeTypeInfo& get_type_info() const override { return type_info; }
                 NonMaxSuppression() = default;
 
                 /// \brief Constructs a NonMaxSuppression operation with default values in the last
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 791d1e7c969606..cd6bebe2f84df2 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -536,7 +536,7 @@ void op::v4::NonMaxSuppression::validate_and_infer_types()
 
 // ------------------------------ V5 ------------------------------
 
-constexpr NodeTypeInfo op::v5::NonMaxSuppression::type_info;
+NGRAPH_RTTI_DEFINITION(op::v5::NonMaxSuppression, "NonMaxSuppression", 5);
 
 op::v5::NonMaxSuppression::NonMaxSuppression(
     const Output<Node>& boxes,

From 3e97da488b5564d294c05bfcd2f91afa228577f6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 10:42:05 +0300
Subject: [PATCH 072/173] All static local constants and functions for NMS-5
 were moved into non-named namespace.

---
 ngraph/core/src/op/non_max_suppression.cpp | 391 +++++++++++----------
 1 file changed, 197 insertions(+), 194 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index cd6bebe2f84df2..318d98f0dda858 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -773,11 +773,203 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
     return max_output_boxes;
 }
 
-static constexpr size_t boxes_port = 0;
-static constexpr size_t scores_port = 1;
-static constexpr size_t iou_threshold_port = 3;
-static constexpr size_t score_threshold_port = 4;
-static constexpr size_t soft_nms_sigma_port = 5;
+using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
+
+namespace {
+    constexpr size_t boxes_port = 0;
+    constexpr size_t scores_port = 1;
+    constexpr size_t iou_threshold_port = 3;
+    constexpr size_t score_threshold_port = 4;
+    constexpr size_t soft_nms_sigma_port = 5;
+
+    PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
+                                              int64_t max_output_boxes_per_class)
+    {
+        const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
+        const auto scores_ps = inputs[scores_port]->get_partial_shape();
+
+        // NonMaxSuppression produces triplets
+        // that have the following format: [batch_index, class_index, box_index]
+        PartialShape result = {Dimension::dynamic(), 3};
+
+        if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
+        {
+            const auto num_boxes_boxes = boxes_ps[1];
+            if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
+            {
+                const auto num_boxes = num_boxes_boxes.get_length();
+                const auto num_classes = scores_ps[1].get_length();
+
+                result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
+                            scores_ps[0].get_length();
+            }
+        }
+
+        return result;
+    }
+
+    void normalize_corner(float* boxes, const Shape& boxes_shape)
+    {
+        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+        for (size_t i = 0; i < total_num_of_boxes; ++i)
+        {
+            float* current_box = boxes + 4 * i;
+
+            float y1 = current_box[0];
+            float x1 = current_box[1];
+            float y2 = current_box[2];
+            float x2 = current_box[3];
+
+            float ymin = std::min(y1, y2);
+            float ymax = std::max(y1, y2);
+            float xmin = std::min(x1, x2);
+            float xmax = std::max(x1, x2);
+
+            current_box[0] = ymin;
+            current_box[1] = xmin;
+            current_box[2] = ymax;
+            current_box[3] = xmax;
+        }
+    }
+
+    void normalize_center(float* boxes, const Shape& boxes_shape)
+    {
+        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+        for (size_t i = 0; i < total_num_of_boxes; ++i)
+        {
+            float* current_box = boxes + 4 * i;
+
+            float x_center = current_box[0];
+            float y_center = current_box[1];
+            float width = current_box[2];
+            float height = current_box[3];
+
+            float y1 = y_center - height / 2.0;
+            float x1 = x_center - width / 2.0;
+            float y2 = y_center + height / 2.0;
+            float x2 = x_center + width / 2.0;
+
+            current_box[0] = y1;
+            current_box[1] = x1;
+            current_box[2] = y2;
+            current_box[3] = x2;
+        }
+    }
+
+    void normalize_box_encoding(float* boxes,
+                                const Shape& boxes_shape,
+                                const V5BoxEncoding box_encoding)
+    {
+        if (box_encoding == V5BoxEncoding::CORNER)
+        {
+            normalize_corner(boxes, boxes_shape);
+        }
+        else
+        {
+            normalize_center(boxes, boxes_shape);
+        }
+    }
+
+    std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
+                                          const Shape& boxes_shape,
+                                          const V5BoxEncoding box_encoding)
+    {
+        element::Type boxes_input_et = boxes->get_element_type();
+
+        size_t boxes_size = shape_size(boxes_shape);
+        std::vector<float> result(boxes_size);
+
+        if (boxes_input_et == ngraph::element::f32)
+        {
+            float* boxes_ptr = boxes->get_data_ptr<float>();
+            memcpy(result.data(), boxes_ptr, boxes_size * sizeof(float));
+        }
+        else
+        {
+            float16* boxes_ptr = boxes->get_data_ptr<float16>();
+            for (size_t i = 0; i < boxes_size; ++i)
+            {
+                result[i] = float(boxes_ptr[i]);
+            }
+        }
+
+        normalize_box_encoding(result.data(), boxes_shape, box_encoding);
+        return result;
+    }
+
+    std::vector<float> prepare_scores_data(const HostTensorPtr& scores,
+                                           const Shape& scores_shape)
+    {
+        element::Type scores_input_et = scores->get_element_type();
+
+        size_t scores_size = shape_size(scores_shape);
+        std::vector<float> result(scores_size);
+
+        if (scores_input_et == ngraph::element::f32)
+        {
+            float* scores_ptr = scores->get_data_ptr<float>();
+            memcpy(result.data(), scores_ptr, scores_size * sizeof(float));
+        }
+        else
+        {
+            float16* scores_ptr = scores->get_data_ptr<float16>();
+            for (size_t i = 0; i < scores_size; ++i)
+            {
+                result[i] = float(scores_ptr[i]);
+            }
+        }
+
+        return result;
+    }
+
+    void evaluate_postprocessing(const HostTensorVector& outputs,
+                                 const ngraph::element::Type output_type,
+                                 const std::vector<int64_t>& selected_indices,
+                                 const std::vector<float>& selected_scores,
+                                 int64_t valid_outputs)
+    {
+        size_t num_of_outputs = outputs.size();
+        size_t selected_size = valid_outputs * 3;
+
+        if (output_type == ngraph::element::i64)
+        {
+            int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
+            memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
+        }
+        else
+        {
+            int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
+            for (size_t i = 0; i < selected_size; ++i)
+            {
+                indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
+            }
+        }
+
+        if (num_of_outputs < 2)
+        {
+            return;
+        }
+
+        float* scores_ptr = outputs[1]->get_data_ptr<float>();
+        memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
+
+        if (num_of_outputs < 3)
+        {
+            return;
+        }
+
+        if (output_type == ngraph::element::i64)
+        {
+            int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
+            *valid_outputs_ptr = valid_outputs;
+        }
+        else
+        {
+            int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
+            *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
+        }
+    }
+}
 
 float op::v5::NonMaxSuppression::iou_threshold_from_input() const
 {
@@ -859,195 +1051,6 @@ namespace ngraph
     }
 } // namespace ngraph
 
-static PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
-                                                 int64_t max_output_boxes_per_class)
-{
-    const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
-    const auto scores_ps = inputs[scores_port]->get_partial_shape();
-
-    // NonMaxSuppression produces triplets
-    // that have the following format: [batch_index, class_index, box_index]
-    PartialShape result = {Dimension::dynamic(), 3};
-
-    if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
-    {
-        const auto num_boxes_boxes = boxes_ps[1];
-        if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
-        {
-            const auto num_boxes = num_boxes_boxes.get_length();
-            const auto num_classes = scores_ps[1].get_length();
-
-            result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
-                        scores_ps[0].get_length();
-        }
-    }
-
-    return result;
-}
-
-using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
-
-static void normalize_corner(float* boxes, const Shape& boxes_shape)
-{
-    size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
-    for (size_t i = 0; i < total_num_of_boxes; ++i)
-    {
-        float* current_box = boxes + 4 * i;
-
-        float y1 = current_box[0];
-        float x1 = current_box[1];
-        float y2 = current_box[2];
-        float x2 = current_box[3];
-
-        float ymin = std::min(y1, y2);
-        float ymax = std::max(y1, y2);
-        float xmin = std::min(x1, x2);
-        float xmax = std::max(x1, x2);
-
-        current_box[0] = ymin;
-        current_box[1] = xmin;
-        current_box[2] = ymax;
-        current_box[3] = xmax;
-    }
-}
-
-static void normalize_center(float* boxes, const Shape& boxes_shape)
-{
-    size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
-    for (size_t i = 0; i < total_num_of_boxes; ++i)
-    {
-        float* current_box = boxes + 4 * i;
-
-        float x_center = current_box[0];
-        float y_center = current_box[1];
-        float width = current_box[2];
-        float height = current_box[3];
-
-        float y1 = y_center - height / 2.0;
-        float x1 = x_center - width / 2.0;
-        float y2 = y_center + height / 2.0;
-        float x2 = x_center + width / 2.0;
-
-        current_box[0] = y1;
-        current_box[1] = x1;
-        current_box[2] = y2;
-        current_box[3] = x2;
-    }
-}
-
-static void
-    normalize_box_encoding(float* boxes, const Shape& boxes_shape, const V5BoxEncoding box_encoding)
-{
-    if (box_encoding == V5BoxEncoding::CORNER)
-    {
-        normalize_corner(boxes, boxes_shape);
-    }
-    else
-    {
-        normalize_center(boxes, boxes_shape);
-    }
-}
-
-static std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
-                                             const Shape& boxes_shape,
-                                             const V5BoxEncoding box_encoding)
-{
-    element::Type boxes_input_et = boxes->get_element_type();
-
-    size_t boxes_size = shape_size(boxes_shape);
-    std::vector<float> result(boxes_size);
-
-    if (boxes_input_et == ngraph::element::f32)
-    {
-        float* boxes_ptr = boxes->get_data_ptr<float>();
-        memcpy(result.data(), boxes_ptr, boxes_size * sizeof(float));
-    }
-    else
-    {
-        float16* boxes_ptr = boxes->get_data_ptr<float16>();
-        for (size_t i = 0; i < boxes_size; ++i)
-        {
-            result[i] = float(boxes_ptr[i]);
-        }
-    }
-
-    normalize_box_encoding(result.data(), boxes_shape, box_encoding);
-    return result;
-}
-
-static std::vector<float> prepare_scores_data(const HostTensorPtr& scores,
-                                              const Shape& scores_shape)
-{
-    element::Type scores_input_et = scores->get_element_type();
-
-    size_t scores_size = shape_size(scores_shape);
-    std::vector<float> result(scores_size);
-
-    if (scores_input_et == ngraph::element::f32)
-    {
-        float* scores_ptr = scores->get_data_ptr<float>();
-        memcpy(result.data(), scores_ptr, scores_size * sizeof(float));
-    }
-    else
-    {
-        float16* scores_ptr = scores->get_data_ptr<float16>();
-        for (size_t i = 0; i < scores_size; ++i)
-        {
-            result[i] = float(scores_ptr[i]);
-        }
-    }
-
-    return result;
-}
-
-static void evaluate_postprocessing(const HostTensorVector& outputs,
-                                    const ngraph::element::Type output_type,
-                                    const std::vector<int64_t>& selected_indices,
-                                    const std::vector<float>& selected_scores,
-                                    int64_t valid_outputs)
-{
-    size_t num_of_outputs = outputs.size();
-    size_t selected_size = valid_outputs * 3;
-
-    if (output_type == ngraph::element::i64)
-    {
-        int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
-        memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
-    }
-    else
-    {
-        int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
-        for (size_t i = 0; i < selected_size; ++i)
-        {
-            indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
-        }
-    }
-
-    if (num_of_outputs < 2)
-    {
-        return;
-    }
-
-    float* scores_ptr = outputs[1]->get_data_ptr<float>();
-    memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
-
-    if (num_of_outputs < 3)
-    {
-        return;
-    }
-
-    if (output_type == ngraph::element::i64)
-    {
-        int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
-        *valid_outputs_ptr = valid_outputs;
-    }
-    else
-    {
-        int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
-        *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
-    }
-}
-
 bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
                                          const HostTensorVector& inputs) const
 {

From 898eaf56634ad8f0d44198a2649777844319d5cf Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 11:15:10 +0300
Subject: [PATCH 073/173] Some code style fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp    |   6 +-
 ngraph/test/op_eval/non_max_suppression_5.cpp | 146 +++++++++---------
 2 files changed, 77 insertions(+), 75 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 318d98f0dda858..f6f96f85d013d4 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -775,7 +775,8 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
 
 using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
 
-namespace {
+namespace
+{
     constexpr size_t boxes_port = 0;
     constexpr size_t scores_port = 1;
     constexpr size_t iou_threshold_port = 3;
@@ -897,8 +898,7 @@ namespace {
         return result;
     }
 
-    std::vector<float> prepare_scores_data(const HostTensorPtr& scores,
-                                           const Shape& scores_shape)
+    std::vector<float> prepare_scores_data(const HostTensorPtr& scores, const Shape& scores_shape)
     {
         element::Type scores_input_et = scores->get_element_type();
 
diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
index cd8e496ae7fc44..fd70e33525386b 100644
--- a/ngraph/test/op_eval/non_max_suppression_5.cpp
+++ b/ngraph/test/op_eval/non_max_suppression_5.cpp
@@ -24,9 +24,9 @@ using namespace ngraph;
 
 TEST(op_eval, nonmaxsuppression_center_point_box_format)
 {
-    std::vector<float> boxes_data = {
-        0.5, 0.5, 1.0, 1.0, 0.5, 0.6, 1.0, 1.0, 0.5, 0.4, 1.0, 1.0, 0.5, 10.5, 1.0, 1.0, 0.5, 10.6,
-        1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
+    std::vector<float> boxes_data = {0.5, 0.5,  1.0, 1.0, 0.5, 0.6,   1.0, 1.0,
+                                     0.5, 0.4,  1.0, 1.0, 0.5, 10.5,  1.0, 1.0,
+                                     0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
 
     std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
 
@@ -42,7 +42,8 @@ TEST(op_eval, nonmaxsuppression_center_point_box_format)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -60,10 +61,9 @@ TEST(op_eval, nonmaxsuppression_center_point_box_format)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                           {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                            make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
@@ -82,9 +82,9 @@ TEST(op_eval, nonmaxsuppression_center_point_box_format)
 
 TEST(op_eval, nonmaxsuppression_flipped_coordinates)
 {
-    std::vector<float> boxes_data = {
-        1.0, 1.0, 0.0, 0.0, 0.0, 0.1, 1.0, 1.1, 0.0, 0.9, 1.0, -0.1, 0.0, 10.0, 1.0, 11.0,
-        1.0, 10.1, 0.0, 11.1, 1.0, 101.0, 0.0, 100.0};
+    std::vector<float> boxes_data = {1.0, 1.0,  0.0, 0.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, 0.9,  1.0, -0.1, 0.0, 10.0,  1.0, 11.0,
+                                     1.0, 10.1, 0.0, 11.1, 1.0, 101.0, 0.0, 100.0};
 
     std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
 
@@ -100,7 +100,8 @@ TEST(op_eval, nonmaxsuppression_flipped_coordinates)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -118,10 +119,9 @@ TEST(op_eval, nonmaxsuppression_flipped_coordinates)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
@@ -140,10 +140,10 @@ TEST(op_eval, nonmaxsuppression_flipped_coordinates)
 
 TEST(op_eval, nonmaxsuppression_identical_boxes)
 {
-    std::vector<float> boxes_data = {
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0};
+    std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
+                                     1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
+                                     0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
+                                     1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0};
 
     std::vector<float> scores_data = {0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9};
 
@@ -159,7 +159,8 @@ TEST(op_eval, nonmaxsuppression_identical_boxes)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -177,10 +178,9 @@ TEST(op_eval, nonmaxsuppression_identical_boxes)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 0};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
@@ -199,9 +199,9 @@ TEST(op_eval, nonmaxsuppression_identical_boxes)
 
 TEST(op_eval, nonmaxsuppression_limit_output_size)
 {
-    std::vector<float> boxes_data = {
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
-        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
 
     std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
 
@@ -217,7 +217,8 @@ TEST(op_eval, nonmaxsuppression_limit_output_size)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -235,10 +236,9 @@ TEST(op_eval, nonmaxsuppression_limit_output_size)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
@@ -273,7 +273,8 @@ TEST(op_eval, nonmaxsuppression_single_box)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -291,10 +292,9 @@ TEST(op_eval, nonmaxsuppression_single_box)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 0};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
@@ -313,9 +313,9 @@ TEST(op_eval, nonmaxsuppression_single_box)
 
 TEST(op_eval, nonmaxsuppression_suppress_by_IOU)
 {
-    std::vector<float> boxes_data = {
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
-        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
 
     std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
 
@@ -331,7 +331,8 @@ TEST(op_eval, nonmaxsuppression_suppress_by_IOU)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -349,10 +350,9 @@ TEST(op_eval, nonmaxsuppression_suppress_by_IOU)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
@@ -371,9 +371,9 @@ TEST(op_eval, nonmaxsuppression_suppress_by_IOU)
 
 TEST(op_eval, nonmaxsuppression_suppress_by_IOU_and_scores)
 {
-    std::vector<float> boxes_data = {
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
-        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
 
     std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
 
@@ -389,7 +389,8 @@ TEST(op_eval, nonmaxsuppression_suppress_by_IOU_and_scores)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -407,10 +408,9 @@ TEST(op_eval, nonmaxsuppression_suppress_by_IOU_and_scores)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
@@ -430,9 +430,9 @@ TEST(op_eval, nonmaxsuppression_suppress_by_IOU_and_scores)
 TEST(op_eval, nonmaxsuppression_two_batches)
 {
     std::vector<float> boxes_data = {
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0,
-        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1,
-        0.0, -0.1, 1.0, 0.9, 0.0, 10.0, 1.0, 11.0, 0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+        0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,   0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0, 0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+        0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,  0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
 
     std::vector<float> scores_data = {
         0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
@@ -449,7 +449,8 @@ TEST(op_eval, nonmaxsuppression_two_batches)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -467,13 +468,13 @@ TEST(op_eval, nonmaxsuppression_two_batches)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 1, 0, 3, 1, 0, 0};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 1.0, 0.0, 0.95, 1.0, 0.0, 0.9};
+    std::vector<float> expected_selected_scores = {
+        0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 1.0, 0.0, 0.95, 1.0, 0.0, 0.9};
     std::vector<int64_t> expected_valid_outputs = {4};
 
     EXPECT_EQ(results[0]->get_element_type(), element::i64);
@@ -489,9 +490,9 @@ TEST(op_eval, nonmaxsuppression_two_batches)
 
 TEST(op_eval, nonmaxsuppression_two_classes)
 {
-    std::vector<float> boxes_data = {
-        0.0, 0.0, 1.0, 1.0, 0.0, 0.1, 1.0, 1.1, 0.0, -0.1, 1.0, 0.9,
-        0.0, 10.0, 1.0, 11.0, 0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
 
     std::vector<float> scores_data = {
         0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
@@ -508,7 +509,8 @@ TEST(op_eval, nonmaxsuppression_two_classes)
     auto max_output_boxes_per_class =
         op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
     auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold = op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
     auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
     auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
                                                       scores,
@@ -526,13 +528,13 @@ TEST(op_eval, nonmaxsuppression_two_classes)
     {
         result = make_shared<HostTensor>();
     }
-    ASSERT_TRUE(
-        f->evaluate(results,
-                    {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                     make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+    ASSERT_TRUE(f->evaluate(results,
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 1, 3, 0, 1, 0};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 1.0, 0.95, 0.0, 1.0, 0.9};
+    std::vector<float> expected_selected_scores = {
+        0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 1.0, 0.95, 0.0, 1.0, 0.9};
     std::vector<int64_t> expected_valid_outputs = {4};
 
     EXPECT_EQ(results[0]->get_element_type(), element::i64);

From 5530ad03ce916e929622bdb8c1c7ad1ca66f75d7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 12:04:18 +0300
Subject: [PATCH 074/173] Moved some file.

---
 .../convert_opset1_to_legacy}/convert_nms_5_to_legacy.cpp         | 0
 1 file changed, 0 insertions(+), 0 deletions(-)
 rename inference-engine/src/{transformations/src/transformations/op_conversions => legacy_api/src/transformations/convert_opset1_to_legacy}/convert_nms_5_to_legacy.cpp (100%)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
similarity index 100%
rename from inference-engine/src/transformations/src/transformations/op_conversions/convert_nms_5_to_legacy.cpp
rename to inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp

From 5e412e558755591c4d3d71c16ead407a91ea2fa1 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 13:57:00 +0300
Subject: [PATCH 075/173] Small fix.

---
 .../src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp         | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp b/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp
index 7912993200ffc2..49e689e9bd4727 100644
--- a/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp
+++ b/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp
@@ -58,7 +58,7 @@ class INFERENCE_ENGINE_API_CLASS(NonMaxSuppressionIE2) : public NonMaxSuppressio
     std::shared_ptr<Node> clone_with_new_inputs(const OutputVector & new_args) const override;
 };
 
-class TRANSFORMATIONS_API NonMaxSuppressionIE3 : public Op {
+class INFERENCE_ENGINE_API_CLASS(NonMaxSuppressionIE3) : public Op {
 public:
     NGRAPH_RTTI_DECLARATION;
 

From 8ec8adbf58cd03ed715b7c6007375b4f9b59de52 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 14:11:26 +0300
Subject: [PATCH 076/173] Code style fix.

---
 ngraph/test/op_eval/non_max_suppression_5.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
index fd70e33525386b..78ee61ad2f14ff 100644
--- a/ngraph/test/op_eval/non_max_suppression_5.cpp
+++ b/ngraph/test/op_eval/non_max_suppression_5.cpp
@@ -62,8 +62,8 @@ TEST(op_eval, nonmaxsuppression_center_point_box_format)
         result = make_shared<HostTensor>();
     }
     ASSERT_TRUE(f->evaluate(results,
-                           {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                            make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
+                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
+                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};

From d9d761dce2e6c9027d6fe8867368f5cf066ff48b Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 14:43:33 +0300
Subject: [PATCH 077/173] Now NMS-5 supports all floating types in inputs 0 and
 1.

---
 ngraph/core/src/op/non_max_suppression.cpp | 117 +++++++++++++++------
 1 file changed, 82 insertions(+), 35 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index f6f96f85d013d4..c146c264bf80ef 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -20,6 +20,7 @@
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/util/op_types.hpp"
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
+#include "ngraph/type/bfloat16.hpp"
 #include "ngraph/type/float16.hpp"
 
 using namespace std;
@@ -871,54 +872,100 @@ namespace
         }
     }
 
-    std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
-                                          const Shape& boxes_shape,
-                                          const V5BoxEncoding box_encoding)
+    std::vector<float> get_floats(const HostTensorPtr& input, const Shape& shape)
     {
-        element::Type boxes_input_et = boxes->get_element_type();
-
-        size_t boxes_size = shape_size(boxes_shape);
-        std::vector<float> result(boxes_size);
+        size_t input_size = shape_size(shape);
+        std::vector<float> result(input_size);
 
-        if (boxes_input_et == ngraph::element::f32)
+        switch(input->get_element_type())
         {
-            float* boxes_ptr = boxes->get_data_ptr<float>();
-            memcpy(result.data(), boxes_ptr, boxes_size * sizeof(float));
-        }
-        else
-        {
-            float16* boxes_ptr = boxes->get_data_ptr<float16>();
-            for (size_t i = 0; i < boxes_size; ++i)
+        case element::Type_t::bf16:
             {
-                result[i] = float(boxes_ptr[i]);
+                bfloat16* p = input->get_data_ptr<bfloat16>();
+                for (size_t i = 0; i < input_size; ++i)
+                {
+                    result[i] = float(p[i]);
+                }
             }
+            break;
+        case element::Type_t::f16:
+            {
+                float16* p = input->get_data_ptr<float16>();
+                for (size_t i = 0; i < input_size; ++i)
+                {
+                    result[i] = float(p[i]);
+                }
+            }
+            break;
+        case element::Type_t::f32:
+            {
+                float* p = input->get_data_ptr<float>();
+                memcpy(result.data(), p, input_size * sizeof(float));
+            }
+            break;
+        case element::Type_t::f64:
+            {
+                double* p = input->get_data_ptr<double>();
+                for (size_t i = 0; i < input_size; ++i)
+                {
+                    result[i] = float(p[i]);
+                }
+            }
+            break;
+        default:;
         }
 
+        return result;
+    }
+
+    std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
+                                          const Shape& boxes_shape,
+                                          const V5BoxEncoding box_encoding)
+    {
+        //         element::Type boxes_input_et = boxes->get_element_type();
+        //
+        //         size_t boxes_size = shape_size(boxes_shape);
+        //         std::vector<float> result(boxes_size);
+        //
+        //         if (boxes_input_et == ngraph::element::f32)
+        //         {
+        //             float* boxes_ptr = boxes->get_data_ptr<float>();
+        //             memcpy(result.data(), boxes_ptr, boxes_size * sizeof(float));
+        //         }
+        //         else
+        //         {
+        //             float16* boxes_ptr = boxes->get_data_ptr<float16>();
+        //             for (size_t i = 0; i < boxes_size; ++i)
+        //             {
+        //                 result[i] = float(boxes_ptr[i]);
+        //             }
+        //         }
+        auto result = get_floats(boxes, boxes_shape);
         normalize_box_encoding(result.data(), boxes_shape, box_encoding);
         return result;
     }
 
     std::vector<float> prepare_scores_data(const HostTensorPtr& scores, const Shape& scores_shape)
     {
-        element::Type scores_input_et = scores->get_element_type();
-
-        size_t scores_size = shape_size(scores_shape);
-        std::vector<float> result(scores_size);
-
-        if (scores_input_et == ngraph::element::f32)
-        {
-            float* scores_ptr = scores->get_data_ptr<float>();
-            memcpy(result.data(), scores_ptr, scores_size * sizeof(float));
-        }
-        else
-        {
-            float16* scores_ptr = scores->get_data_ptr<float16>();
-            for (size_t i = 0; i < scores_size; ++i)
-            {
-                result[i] = float(scores_ptr[i]);
-            }
-        }
-
+        //         element::Type scores_input_et = scores->get_element_type();
+        //
+        //         size_t scores_size = shape_size(scores_shape);
+        //         std::vector<float> result(scores_size);
+        //
+        //         if (scores_input_et == ngraph::element::f32)
+        //         {
+        //             float* scores_ptr = scores->get_data_ptr<float>();
+        //             memcpy(result.data(), scores_ptr, scores_size * sizeof(float));
+        //         }
+        //         else
+        //         {
+        //             float16* scores_ptr = scores->get_data_ptr<float16>();
+        //             for (size_t i = 0; i < scores_size; ++i)
+        //             {
+        //                 result[i] = float(scores_ptr[i]);
+        //             }
+        //         }
+        auto result = get_floats(scores, scores_shape);
         return result;
     }
 

From 83a0eea1970e3936456e7299c1116db8f418eaa6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 15:13:29 +0300
Subject: [PATCH 078/173] Moved some files.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp        | 2 +-
 .../{ => op_conversions}/convert_previous_nms_to_nms_5.hpp      | 0
 .../common_optimizations/common_optimizations.cpp               | 1 +
 .../{ => op_conversions}/convert_previous_nms_to_nms_5.cpp      | 2 +-
 4 files changed, 3 insertions(+), 2 deletions(-)
 rename inference-engine/src/transformations/include/transformations/{ => op_conversions}/convert_previous_nms_to_nms_5.hpp (100%)
 rename inference-engine/src/transformations/src/transformations/{ => op_conversions}/convert_previous_nms_to_nms_5.cpp (99%)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index a633b103501639..9d15a172b6d03c 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -8,7 +8,7 @@
 #include <ngraph/graph_util.hpp>
 #include <ngraph/opsets/opset1.hpp>
 #include <ngraph/opsets/opset5.hpp>
-#include <ngraph_ops/nms_ie.hpp>
+#include <legacy/ngraph_ops/nms_ie.hpp>
 #include <ngraph/rt_info.hpp>
 #include <transformations/utils/utils.hpp>
 
diff --git a/inference-engine/src/transformations/include/transformations/convert_previous_nms_to_nms_5.hpp b/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
similarity index 100%
rename from inference-engine/src/transformations/include/transformations/convert_previous_nms_to_nms_5.hpp
rename to inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
diff --git a/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp b/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
index 23f121fa97f4db..6e09435a61650e 100644
--- a/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
+++ b/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
@@ -41,6 +41,7 @@
 #include "transformations/op_conversions/reduce_l1_decomposition.hpp"
 #include "transformations/op_conversions/reduce_l2_decomposition.hpp"
 #include "transformations/op_conversions/hswish_decomposition.hpp"
+#include "transformations/op_conversions/convert_previous_nms_to_nms_5.hpp"
 
 #include <ngraph/pass/manager.hpp>
 #include <ngraph/pass/constant_folding.hpp>
diff --git a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
similarity index 99%
rename from inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
rename to inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index ee4ab6a6f8dc99..6e0b78ae190f43 100644
--- a/inference-engine/src/transformations/src/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
-#include "transformations/convert_previous_nms_to_nms_5.hpp"
+#include "transformations/op_conversions/convert_previous_nms_to_nms_5.hpp"
 
 #include <list>
 #include <memory>

From 8428e209e9cf4e06f08ddfe43393ac314efeaab8 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 15:40:43 +0300
Subject: [PATCH 079/173] Fixed include directive in the file
 convert_nms_5_to_legacy.cpp with transformations NMS-1, 3, 4 -> NMS-5.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp        | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 9d15a172b6d03c..aceecea837ff94 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -12,7 +12,7 @@
 #include <ngraph/rt_info.hpp>
 #include <transformations/utils/utils.hpp>
 
-#include "transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp"
+#include "legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp"
 
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS5ToLegacyMatcher, "ConvertNMS5ToLegacyMatcher", 0);
 

From 571a7fde913762c9feeb140e881883441ad8c23c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 16:30:42 +0300
Subject: [PATCH 080/173] Small changes.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp      | 2 +-
 .../inference_engine/transformations/convert_nms5_test.cpp    | 4 ++--
 .../transformations/convert_previous_nms_to_nms_5.cpp         | 2 +-
 3 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp b/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
index 1dcfcbe1e2c4a8..630ef3f027bf99 100644
--- a/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
+++ b/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
@@ -14,7 +14,7 @@
 namespace ngraph {
 namespace pass {
 
-    class TRANSFORMATIONS_API ConvertNMS5ToLegacyMatcher;
+    class INFERENCE_ENGINE_API_CLASS(ConvertNMS5ToLegacyMatcher);
 
 }  // namespace pass
 }  // namespace ngraph
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index e9adb14669b1d5..6567330464f31c 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -9,8 +9,8 @@
 
 #include <ngraph/function.hpp>
 #include <ngraph/opsets/opset5.hpp>
-#include <ngraph_ops/nms_ie.hpp>
-#include <transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp>
+#include <legacy/ngraph_ops/nms_ie.hpp>
+#include <legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp>
 #include <transformations/init_node_info.hpp>
 #include <transformations/utils/utils.hpp>
 #include <ngraph/pass/manager.hpp>
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
index e965d8d3d23e4f..0f96d32b690c49 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
@@ -13,7 +13,7 @@
 #include <ngraph/opsets/opset3.hpp>
 #include <ngraph/opsets/opset4.hpp>
 #include <ngraph/opsets/opset5.hpp>
-#include <transformations/convert_previous_nms_to_nms_5.hpp>
+#include <transformations/op_conversions/convert_previous_nms_to_nms_5.hpp>
 #include <transformations/init_node_info.hpp>
 #include <transformations/utils/utils.hpp>
 #include <ngraph/pass/manager.hpp>

From 730cd6df057b4487b91a7ff60535952febe42ae0 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 17:39:50 +0300
Subject: [PATCH 081/173] Deleted conversion NMS-3 -> legacy.

---
 .../op_conversions/convert_nms3.hpp           | 31 ----------
 .../op_conversions/convert_nms3.cpp           | 45 --------------
 .../convert_opset3_to_opset2.cpp              |  1 -
 .../transformations/convert_nms3_test.cpp     | 62 -------------------
 4 files changed, 139 deletions(-)
 delete mode 100644 inference-engine/src/transformations/include/transformations/op_conversions/convert_nms3.hpp
 delete mode 100644 inference-engine/src/transformations/src/transformations/op_conversions/convert_nms3.cpp
 delete mode 100644 inference-engine/tests/functional/inference_engine/transformations/convert_nms3_test.cpp

diff --git a/inference-engine/src/transformations/include/transformations/op_conversions/convert_nms3.hpp b/inference-engine/src/transformations/include/transformations/op_conversions/convert_nms3.hpp
deleted file mode 100644
index 724566fdd2391d..00000000000000
--- a/inference-engine/src/transformations/include/transformations/op_conversions/convert_nms3.hpp
+++ /dev/null
@@ -1,31 +0,0 @@
-// Copyright (C) 2018-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-//
-
-#pragma once
-
-#include <vector>
-#include <memory>
-
-#include <transformations_visibility.hpp>
-
-#include <ngraph/pass/graph_rewrite.hpp>
-
-namespace ngraph {
-namespace pass {
-
-class TRANSFORMATIONS_API ConvertNMS1ToNMS3;
-
-}  // namespace pass
-}  // namespace ngraph
-
-class ngraph::pass::ConvertNMS1ToNMS3: public ngraph::pass::GraphRewrite {
-public:
-    NGRAPH_RTTI_DECLARATION;
-    ConvertNMS1ToNMS3() : GraphRewrite() {
-        convert_nms1_to_nms3();
-    }
-
-private:
-    void convert_nms1_to_nms3();
-};
diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_nms3.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_nms3.cpp
deleted file mode 100644
index fa90efa9c58c81..00000000000000
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_nms3.cpp
+++ /dev/null
@@ -1,45 +0,0 @@
-// Copyright (C) 2018-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-//
-
-#include "transformations/op_conversions/convert_nms3.hpp"
-
-#include <memory>
-#include <vector>
-
-#include <ngraph/opsets/opset1.hpp>
-#include <ngraph/opsets/opset2.hpp>
-#include <ngraph/opsets/opset3.hpp>
-#include <ngraph/rt_info.hpp>
-
-NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS1ToNMS3, "ConvertNMS1ToNMS3", 0);
-
-void ngraph::pass::ConvertNMS1ToNMS3::convert_nms1_to_nms3() {
-    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-    auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
-    auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
-    auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
-    auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                   iou_threshold, score_threshold);
-
-    ngraph::graph_rewrite_callback callback = [](pattern::Matcher &m) {
-        auto nms = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());
-        if (!nms) {
-            return false;
-        }
-
-        auto nms3 = std::make_shared<ngraph::opset3::NonMaxSuppression>(nms->input_value(0), nms->input_value(1),
-                nms->input_value(2), nms->input_value(3), nms->input_value(4),
-                static_cast<const op::v3::NonMaxSuppression::BoxEncodingType>(nms->get_box_encoding()),
-                nms->get_sort_result_descending());
-
-        nms3->set_friendly_name(nms->get_friendly_name());
-        ngraph::copy_runtime_info(nms, nms3);
-        ngraph::replace_node(nms, nms3);
-        return true;
-    };
-
-    auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS1ToNMS3");
-    this->add_matcher(m, callback, PassProperty::CHANGE_DYNAMIC_STATE);
-}
diff --git a/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp b/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp
index 02c45bd0ee4a3d..c9fbecc4f13cbe 100644
--- a/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp
+++ b/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp
@@ -5,7 +5,6 @@
 #include "transformations/opset_conversions/convert_opset3_to_opset2.hpp"
 
 #include "transformations/op_conversions/convert_broadcast3.hpp"
-#include "transformations/op_conversions/convert_nms3.hpp"
 #include "transformations/op_conversions/convert_shapeof3.hpp"
 #include "transformations/op_conversions/convert_shuffle_channels3.hpp"
 #include "transformations/op_conversions/convert_topk3.hpp"
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms3_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms3_test.cpp
deleted file mode 100644
index ae4ee04c26b7c6..00000000000000
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms3_test.cpp
+++ /dev/null
@@ -1,62 +0,0 @@
-// Copyright (C) 2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-//
-
-#include <gtest/gtest.h>
-
-#include <string>
-#include <memory>
-#include <queue>
-
-#include <ngraph/function.hpp>
-#include <ngraph/opsets/opset1.hpp>
-#include <ngraph/opsets/opset2.hpp>
-#include <ngraph/opsets/opset3.hpp>
-#include <transformations/op_conversions/convert_nms3.hpp>
-#include <transformations/init_node_info.hpp>
-#include <transformations/utils/utils.hpp>
-
-#include "common_test_utils/ngraph_test_utils.hpp"
-
-using namespace testing;
-using namespace ngraph;
-
-TEST(TransformationTests, ConvertNMS3I32Output) {
-    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-    {
-        auto boxes = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = opset3::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = opset3::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset3::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                iou_threshold, score_threshold, opset1::NonMaxSuppression::BoxEncodingType::CORNER, true);
-        nms->set_friendly_name("nms");
-
-        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-
-        pass::InitNodeInfo().run_on_function(f);
-        pass::ConvertNMS1ToNMS3().run_on_function(f);
-        ASSERT_NO_THROW(check_rt_info(f));
-    }
-
-    {
-        auto boxes = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<opset3::Parameter>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = opset3::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = opset3::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset3::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                iou_threshold, score_threshold, opset3::NonMaxSuppression::BoxEncodingType::CORNER, true);
-        nms->set_friendly_name("nms");
-
-        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-    }
-
-    auto res = compare_functions(f, f_ref);
-    ASSERT_TRUE(res.first) << res.second;
-
-    auto result_node_of_converted_f = f->get_output_op(0);
-    auto nms_node = result_node_of_converted_f->input(0).get_source_output().get_node_shared_ptr();
-    ASSERT_TRUE(nms_node->get_friendly_name() == "nms") << "Transformation ConvertNMS1ToNMS3 should keep output names.\n";
-}

From 32337fc6966acafcc0af1c40a10510e5e60db80f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 17:48:27 +0300
Subject: [PATCH 082/173] Small changes.

---
 .../opset_conversions/convert_opset3_to_opset2.cpp               | 1 -
 1 file changed, 1 deletion(-)

diff --git a/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp b/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp
index c9fbecc4f13cbe..666d6e80d3eb05 100644
--- a/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp
+++ b/inference-engine/src/transformations/src/transformations/opset_conversions/convert_opset3_to_opset2.cpp
@@ -25,7 +25,6 @@ bool ngraph::pass::ConvertOpSet3ToOpSet2::run_on_function(std::shared_ptr<ngraph
     ngraph::pass::Manager manager;
 
     manager.register_pass<ngraph::pass::ConvertBroadcast3>();
-    manager.register_pass<ngraph::pass::ConvertNMS1ToNMS3>();
     manager.register_pass<ngraph::pass::ConvertShapeOf3>();
     manager.register_pass<ngraph::pass::ConvertShuffleChannels3>();
     manager.register_pass<ngraph::pass::ConvertTopK3>();

From 24a6a2f3dcdcec3b936d8e12bab286b540c5826e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 14 Oct 2020 17:54:20 +0300
Subject: [PATCH 083/173] Fix in op::v5::NonMaxSuppression::evaluate: output
 shape [1] instead of [] in the output port 2.

---
 ngraph/core/src/op/non_max_suppression.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 81e8efcbf78419..e9755309f89b36 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -1114,7 +1114,7 @@ bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
     if (num_of_outputs >= 3)
     {
         outputs[2]->set_element_type(m_output_type);
-        outputs[2]->set_shape(Shape{});
+        outputs[2]->set_shape(Shape{1});
     }
 
     evaluate_postprocessing(

From c0f9697e59c188796ed9524798968806c71df1b8 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 15 Oct 2020 09:25:36 +0300
Subject: [PATCH 084/173] Code style fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp | 48 +++++++++++-----------
 1 file changed, 24 insertions(+), 24 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index e9755309f89b36..26f9f7878b6817 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -877,41 +877,41 @@ namespace
         size_t input_size = shape_size(shape);
         std::vector<float> result(input_size);
 
-        switch(input->get_element_type())
+        switch (input->get_element_type())
         {
         case element::Type_t::bf16:
+        {
+            bfloat16* p = input->get_data_ptr<bfloat16>();
+            for (size_t i = 0; i < input_size; ++i)
             {
-                bfloat16* p = input->get_data_ptr<bfloat16>();
-                for (size_t i = 0; i < input_size; ++i)
-                {
-                    result[i] = float(p[i]);
-                }
+                result[i] = float(p[i]);
             }
-            break;
+        }
+        break;
         case element::Type_t::f16:
+        {
+            float16* p = input->get_data_ptr<float16>();
+            for (size_t i = 0; i < input_size; ++i)
             {
-                float16* p = input->get_data_ptr<float16>();
-                for (size_t i = 0; i < input_size; ++i)
-                {
-                    result[i] = float(p[i]);
-                }
+                result[i] = float(p[i]);
             }
-            break;
+        }
+        break;
         case element::Type_t::f32:
-            {
-                float* p = input->get_data_ptr<float>();
-                memcpy(result.data(), p, input_size * sizeof(float));
-            }
-            break;
+        {
+            float* p = input->get_data_ptr<float>();
+            memcpy(result.data(), p, input_size * sizeof(float));
+        }
+        break;
         case element::Type_t::f64:
+        {
+            double* p = input->get_data_ptr<double>();
+            for (size_t i = 0; i < input_size; ++i)
             {
-                double* p = input->get_data_ptr<double>();
-                for (size_t i = 0; i < input_size; ++i)
-                {
-                    result[i] = float(p[i]);
-                }
+                result[i] = float(p[i]);
             }
-            break;
+        }
+        break;
         default:;
         }
 

From ad6ea78b96fd301623f61d28e87e8690565ae898 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 15 Oct 2020 10:26:37 +0300
Subject: [PATCH 085/173] Deleted conversion of NMS-4 into legacy NMS.

---
 .../convert_nms_4_to_legacy.hpp               |  33 ----
 .../convert_nms_4_to_legacy.cpp               | 113 -------------
 .../convert_opset1_to_legacy.cpp              |   2 -
 .../transformations/convert_nms4_test.cpp     | 148 ------------------
 4 files changed, 296 deletions(-)
 delete mode 100644 inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp
 delete mode 100644 inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.cpp
 delete mode 100644 inference-engine/tests/functional/inference_engine/transformations/convert_nms4_test.cpp

diff --git a/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp b/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp
deleted file mode 100644
index affd092ce56b0a..00000000000000
--- a/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp
+++ /dev/null
@@ -1,33 +0,0 @@
-// Copyright (C) 2018-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-//
-
-#pragma once
-
-#include <vector>
-#include <memory>
-
-#include <ie_api.h>
-
-#include <ngraph/pass/graph_rewrite.hpp>
-
-namespace ngraph {
-namespace pass {
-
-    class INFERENCE_ENGINE_API_CLASS(ConvertNMS4ToLegacyMatcher);
-
-}  // namespace pass
-}  // namespace ngraph
-
-/*
- * Description:
- *      Convert NMS-4 directly to legacy NMS because NMS-3 and NMS-1 have different shape infer function
- */
-
-
-class ngraph::pass::ConvertNMS4ToLegacyMatcher: public ngraph::pass::MatcherPass {
-public:
-    NGRAPH_RTTI_DECLARATION;
-    ConvertNMS4ToLegacyMatcher();
-};
-
diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.cpp
deleted file mode 100644
index ca6a7459b63952..00000000000000
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.cpp
+++ /dev/null
@@ -1,113 +0,0 @@
-// Copyright (C) 2018-2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-//
-
-#include <memory>
-#include <vector>
-
-#include <ngraph/graph_util.hpp>
-#include <ngraph/opsets/opset1.hpp>
-#include <ngraph/opsets/opset4.hpp>
-#include <legacy/ngraph_ops/nms_ie.hpp>
-#include <ngraph/rt_info.hpp>
-#include <transformations/utils/utils.hpp>
-
-#include "legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp"
-
-NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToLegacyMatcher, "ConvertNMS4ToLegacyMatcher", 0);
-
-ngraph::pass::ConvertNMS4ToLegacyMatcher::ConvertNMS4ToLegacyMatcher() {
-    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-    auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
-    auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
-    auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
-    auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                   iou_threshold, score_threshold);
-
-    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
-        auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
-        if (!nms_4) {
-            return false;
-        }
-
-        const auto new_args = nms_4->input_values();
-        const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset4::Constant::create(element::i32, Shape{}, {0});
-        const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset4::Constant::create(element::f32, Shape{}, {.0f});
-
-        const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
-        const auto iou_threshold_rank = arg3.get_partial_shape().rank();
-        const auto score_threshold_rank = arg4.get_partial_shape().rank();
-
-        // Check that required ranks are not dynamic
-        if (max_output_boxes_per_class_rank.is_dynamic() ||
-            iou_threshold_rank.is_dynamic() ||
-            score_threshold_rank.is_dynamic()) {
-            return false;
-        }
-
-        if (max_output_boxes_per_class_rank.get_length() == 1 &&
-            iou_threshold_rank.get_length() == 1 &&
-            score_threshold_rank.get_length() == 1) {
-            return false;
-        }
-
-        // vector of new nGraph operations
-        NodeVector new_ops;
-
-        auto new_max_per_class = arg2;
-        if (max_output_boxes_per_class_rank.get_length() == 0) {
-            // WA: we need to create Constant manually because it requires by NMS shape inference
-            //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
-            //     if CF will be executed right after transformation and before Validate pass.
-            if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
-                new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
-            } else {
-                new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
-                new_ops.push_back(new_max_per_class.get_node_shared_ptr());
-            }
-        }
-        auto new_iou_threshold = arg3;
-        if (iou_threshold_rank.get_length() == 0) {
-            new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
-            new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
-        }
-        auto new_score_threshold = arg4;
-        if (score_threshold_rank.get_length() == 0) {
-            new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
-            new_ops.push_back(new_score_threshold.get_node_shared_ptr());
-        }
-
-        int center_point_box = 0;
-        switch (nms_4->get_box_encoding()) {
-            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-                center_point_box = 1;
-                break;
-            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-                center_point_box = 0;
-                break;
-            default:
-                throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
-                                   " has unsupported box encoding");
-        }
-        const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE2>(
-                new_args.at(0),
-                new_args.at(1),
-                new_max_per_class,
-                new_iou_threshold,
-                new_score_threshold,
-                center_point_box,
-                nms_4->get_sort_result_descending(),
-                nms_4->get_output_type());
-        new_ops.push_back(nms_legacy);
-
-        nms_legacy->set_friendly_name(nms_4->get_friendly_name());
-        ngraph::copy_runtime_info(nms_4, new_ops);
-        ngraph::replace_node(nms_4, nms_legacy);
-        return true;
-    };
-
-    auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMSLegacy");
-    this->register_matcher(m, callback);
-}
diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp
index cef550019877d6..5f9a657c06be5c 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.cpp
@@ -13,7 +13,6 @@
 #include "legacy/transformations/convert_opset1_to_legacy/convert_mul_add_to_scaleshift_or_power.hpp"
 #include "legacy/transformations/convert_opset1_to_legacy/convert_mul_or_add_finally.hpp"
 #include "legacy/transformations/convert_opset1_to_legacy/convert_nms_to_nms_ie.hpp"
-#include "legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp"
 #include "legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp"
 #include "legacy/transformations/convert_opset1_to_legacy/convert_normalizel2_to_normalize_ie.hpp"
 #include "legacy/transformations/convert_opset1_to_legacy/convert_one_hot_to_one_hot_ie.hpp"
@@ -134,7 +133,6 @@ bool ngraph::pass::ConvertOpSet1ToLegacy::run_on_function(std::shared_ptr<ngraph
     anchor->add_matcher<ngraph::pass::ConvertGatherTreeToGatherTreeIEMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertTopKToTopKIEMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertNMSToNMSIEMatcher>();
-    anchor->add_matcher<ngraph::pass::ConvertNMS4ToLegacyMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertNMS5ToLegacyMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertGRUSequenceMatcher>();
     anchor->add_matcher<ngraph::pass::ConvertRNNSequenceMatcher>();
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms4_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms4_test.cpp
deleted file mode 100644
index 6b62756b247b84..00000000000000
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms4_test.cpp
+++ /dev/null
@@ -1,148 +0,0 @@
-// Copyright (C) 2020 Intel Corporation
-// SPDX-License-Identifier: Apache-2.0
-//
-
-#include <gtest/gtest.h>
-
-#include <string>
-#include <memory>
-
-#include <ngraph/function.hpp>
-#include <ngraph/opsets/opset4.hpp>
-#include <legacy/ngraph_ops/nms_ie.hpp>
-#include <legacy/transformations/convert_opset1_to_legacy/convert_nms_4_to_legacy.hpp>
-#include <transformations/init_node_info.hpp>
-#include <transformations/utils/utils.hpp>
-#include <ngraph/pass/manager.hpp>
-
-#include "common_test_utils/ngraph_test_utils.hpp"
-
-using namespace testing;
-using namespace ngraph;
-
-TEST(TransformationTests, ConvertNMS4ToNMSIEStatic) {
-    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-    {
-        auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-                                                               opset4::NonMaxSuppression::BoxEncodingType::CORNER, true);
-
-        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-
-        const auto &orig_shape = f->get_output_partial_shape(0);
-        pass::Manager manager;
-        manager.register_pass<pass::InitNodeInfo>();
-        manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
-        manager.run_passes(f);
-        ASSERT_NO_THROW(check_rt_info(f));
-        ASSERT_TRUE(f->get_output_partial_shape(0).is_static()) << "Shape " << f->get_output_partial_shape(0) << " should be static";
-    }
-
-    {
-        auto boxes = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<opset4::Parameter>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
-        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
-                                                              std::make_shared<opset4::Unsqueeze>(iou_threshold,
-                                                                                                  opset4::Constant::create(element::i64, Shape{1}, {0})),
-                                                              std::make_shared<opset4::Unsqueeze>(score_threshold,
-                                                                                                  opset4::Constant::create(element::i64, Shape{1}, {0})),
-                                                              0, true);
-        nms->set_friendly_name("nms");
-
-        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-        ASSERT_TRUE(f_ref->get_output_partial_shape(0).is_static()) << "Shape " << f_ref->get_output_partial_shape(0) << " should be static";
-    }
-
-    auto res = compare_functions(f, f_ref);
-    ASSERT_TRUE(res.first) << res.second;
-}
-
-TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic1) {
-    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-    {
-        auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-        auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-                                                               opset4::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
-
-        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-
-        pass::Manager manager;
-        manager.register_pass<pass::InitNodeInfo>();
-        manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
-        manager.run_passes(f);
-        f->validate_nodes_and_infer_types();
-        ASSERT_NO_THROW(check_rt_info(f));
-    }
-
-    {
-        auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-        auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape::dynamic());
-        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
-        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
-                                                              std::make_shared<opset4::Unsqueeze>(iou_threshold,
-                                                                                                  opset4::Constant::create(element::i64, Shape{1}, {0})),
-                                                              std::make_shared<opset4::Unsqueeze>(score_threshold,
-                                                                                                  opset4::Constant::create(element::i64, Shape{1}, {0})),
-                                                              0, true);
-
-        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-    }
-
-    auto res = compare_functions(f, f_ref);
-    ASSERT_TRUE(res.first) << res.second;
-}
-
-TEST(TransformationTests, ConvertNMS4ToNMSIEDynamic2) {
-    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-    {
-        auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
-        auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-                                                               opset4::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
-
-        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-
-        pass::Manager manager;
-        manager.register_pass<pass::InitNodeInfo>();
-        manager.register_pass<pass::ConvertNMS4ToLegacyMatcher>();
-        manager.run_passes(f);
-
-        f->validate_nodes_and_infer_types();
-        ASSERT_NO_THROW(check_rt_info(f));
-    }
-
-    {
-        auto boxes = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
-        auto scores = std::make_shared<opset4::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-        auto max_output_boxes_per_class = opset4::Constant::create(element::i64, Shape{1}, {10});
-        auto iou_threshold = opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<op::NonMaxSuppressionIE2>(boxes, scores, max_output_boxes_per_class,
-                                                              std::make_shared<opset4::Unsqueeze>(iou_threshold,
-                                                                                                  opset4::Constant::create(element::i64, Shape{1}, {0})),
-                                                              std::make_shared<opset4::Unsqueeze>(score_threshold,
-                                                                                                  opset4::Constant::create(element::i64, Shape{1}, {0})),
-                                                              0, true);
-
-        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-    }
-
-    auto res = compare_functions(f, f_ref);
-    ASSERT_TRUE(res.first) << res.second;
-}

From 0952f96e96edc947c19dc09b6d31cf86faa4e57c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 15 Oct 2020 10:55:18 +0300
Subject: [PATCH 086/173] Deleted redundant ifs.

---
 ngraph/core/src/op/non_max_suppression.cpp | 56 +++++++++-------------
 1 file changed, 22 insertions(+), 34 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 26f9f7878b6817..c27f75d270ebf1 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -700,43 +700,31 @@ void op::v5::NonMaxSuppression::validate()
                           "Expected a 3D tensor for the 'scores' input. Got: ",
                           scores_ps);
 
-    if (inputs().size() >= 3)
-    {
-        const auto max_boxes_ps = get_input_partial_shape(2);
-        NODE_VALIDATION_CHECK(this,
-                              max_boxes_ps.is_dynamic() || is_scalar(max_boxes_ps.to_shape()),
-                              "Expected a scalar for the 'max_output_boxes_per_class' input. Got: ",
-                              max_boxes_ps);
-    }
+    const auto max_boxes_ps = get_input_partial_shape(2);
+    NODE_VALIDATION_CHECK(this,
+                          max_boxes_ps.is_dynamic() || is_scalar(max_boxes_ps.to_shape()),
+                          "Expected a scalar for the 'max_output_boxes_per_class' input. Got: ",
+                          max_boxes_ps);
 
-    if (inputs().size() >= 4)
-    {
-        const auto iou_threshold_ps = get_input_partial_shape(3);
-        NODE_VALIDATION_CHECK(this,
-                              iou_threshold_ps.is_dynamic() ||
-                                  is_scalar(iou_threshold_ps.to_shape()),
-                              "Expected a scalar for the 'iou_threshold' input. Got: ",
-                              iou_threshold_ps);
-    }
+    const auto iou_threshold_ps = get_input_partial_shape(3);
+    NODE_VALIDATION_CHECK(this,
+                          iou_threshold_ps.is_dynamic() ||
+                              is_scalar(iou_threshold_ps.to_shape()),
+                          "Expected a scalar for the 'iou_threshold' input. Got: ",
+                          iou_threshold_ps);
 
-    if (inputs().size() >= 5)
-    {
-        const auto score_threshold_ps = get_input_partial_shape(4);
-        NODE_VALIDATION_CHECK(this,
-                              score_threshold_ps.is_dynamic() ||
-                                  is_scalar(score_threshold_ps.to_shape()),
-                              "Expected a scalar for the 'score_threshold' input. Got: ",
-                              score_threshold_ps);
-    }
+    const auto score_threshold_ps = get_input_partial_shape(4);
+    NODE_VALIDATION_CHECK(this,
+                          score_threshold_ps.is_dynamic() ||
+                              is_scalar(score_threshold_ps.to_shape()),
+                          "Expected a scalar for the 'score_threshold' input. Got: ",
+                          score_threshold_ps);
 
-    if (inputs().size() >= 6)
-    {
-        const auto soft_nms_sigma = get_input_partial_shape(5);
-        NODE_VALIDATION_CHECK(this,
-                              soft_nms_sigma.is_dynamic() || is_scalar(soft_nms_sigma.to_shape()),
-                              "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
-                              soft_nms_sigma);
-    }
+    const auto soft_nms_sigma = get_input_partial_shape(5);
+    NODE_VALIDATION_CHECK(this,
+                          soft_nms_sigma.is_dynamic() || is_scalar(soft_nms_sigma.to_shape()),
+                          "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
+                          soft_nms_sigma);
 
     const auto num_batches_boxes = boxes_ps[0];
     const auto num_batches_scores = scores_ps[0];

From 44b6d09d707b11cd47829fbe05a8a2cf3486ced5 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 15 Oct 2020 13:16:51 +0300
Subject: [PATCH 087/173] Added NMS-5 to Python API.

---
 ngraph/python/src/ngraph/opset5/__init__.py |  2 +-
 ngraph/python/src/ngraph/opset5/ops.py      | 47 +++++++++++++++++++++
 2 files changed, 48 insertions(+), 1 deletion(-)

diff --git a/ngraph/python/src/ngraph/opset5/__init__.py b/ngraph/python/src/ngraph/opset5/__init__.py
index e2b4a83c488ec3..16663ef7521d0f 100644
--- a/ngraph/python/src/ngraph/opset5/__init__.py
+++ b/ngraph/python/src/ngraph/opset5/__init__.py
@@ -95,7 +95,7 @@
 from ngraph.opset1.ops import multiply
 from ngraph.opset2.ops import mvn
 from ngraph.opset1.ops import negative
-from ngraph.opset4.ops import non_max_suppression
+from ngraph.opset5.ops import non_max_suppression
 from ngraph.opset3.ops import non_zero
 from ngraph.opset1.ops import normalize_l2
 from ngraph.opset1.ops import not_equal
diff --git a/ngraph/python/src/ngraph/opset5/ops.py b/ngraph/python/src/ngraph/opset5/ops.py
index 8c1950e93bede2..193133963f5008 100644
--- a/ngraph/python/src/ngraph/opset5/ops.py
+++ b/ngraph/python/src/ngraph/opset5/ops.py
@@ -90,3 +90,50 @@ def log_softmax(data: NodeInput, axis: int, name: Optional[str] = None) -> Node:
     :return: The new node with LogSoftmax operation applied on each element.
     """
     return _get_node_factory_opset5().create("LogSoftmax", [as_node(data)], {"axis": axis})
+
+
+@nameable_op
+def non_max_suppression(
+    boxes: NodeInput,
+    scores: NodeInput,
+    max_output_boxes_per_class: Optional[NodeInput] = None,
+    iou_threshold: Optional[NodeInput] = None,
+    score_threshold: Optional[NodeInput] = None,
+    soft_nms_sigma: Optional[NodeInput] = None,
+    box_encoding: str = "corner",
+    sort_result_descending: bool = True,
+    output_type: str = "i64",
+    name: Optional[str] = None,
+) -> Node:
+    """Return a node which performs NonMaxSuppression.
+
+    :param boxes: Tensor with box coordinates.
+    :param scores: Tensor with box scores.
+    :param max_output_boxes_per_class: Tensor Specifying maximum number of boxes
+                                        to be selected per class.
+    :param iou_threshold: Tensor specifying intersection over union threshold
+    :param score_threshold: Tensor specifying minimum score to consider box for the processing.
+    :param soft_nms_sigma: Tensor specifying the sigma parameter for Soft-NMS.
+    :param box_encoding: Format of boxes data encoding.
+    :param sort_result_descending: Flag that specifies whenever it is necessary to sort selected
+                                   boxes across batches or not.
+    :param output_type: Output element type.
+    :return: The new node which performs NonMaxSuppression
+    """
+    if max_output_boxes_per_class is None:
+        max_output_boxes_per_class = make_constant_node(0, np.int64)
+    if iou_threshold is None:
+        iou_threshold = make_constant_node(0, np.float32)
+    if score_threshold is None:
+        score_threshold = make_constant_node(0, np.float32)
+    if soft_nms_sigma is None:
+        soft_nms_sigma = make_constant_node(0, np.float32)
+
+    inputs = as_nodes(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, soft_nms_sigma)
+    attributes = {
+        "box_encoding": box_encoding,
+        "sort_result_descending": sort_result_descending,
+        "output_type": output_type,
+    }
+
+    return _get_node_factory_opset5().create("NonMaxSuppression", inputs, attributes)

From cca2675b68a408e6caa36ec60dcc999340276bbe Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 15 Oct 2020 15:39:04 +0300
Subject: [PATCH 088/173] Code style fix.

---
 ngraph/core/src/op/non_max_suppression.cpp | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index c27f75d270ebf1..909f1c5e9216b9 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -708,8 +708,7 @@ void op::v5::NonMaxSuppression::validate()
 
     const auto iou_threshold_ps = get_input_partial_shape(3);
     NODE_VALIDATION_CHECK(this,
-                          iou_threshold_ps.is_dynamic() ||
-                              is_scalar(iou_threshold_ps.to_shape()),
+                          iou_threshold_ps.is_dynamic() || is_scalar(iou_threshold_ps.to_shape()),
                           "Expected a scalar for the 'iou_threshold' input. Got: ",
                           iou_threshold_ps);
 

From c2e181ac02037c4a068b021a769b702dcaddbf27 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 15 Oct 2020 17:18:54 +0300
Subject: [PATCH 089/173] Small change.

---
 ngraph/python/src/ngraph/opset5/ops.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/ngraph/python/src/ngraph/opset5/ops.py b/ngraph/python/src/ngraph/opset5/ops.py
index 193133963f5008..aa49942988143b 100644
--- a/ngraph/python/src/ngraph/opset5/ops.py
+++ b/ngraph/python/src/ngraph/opset5/ops.py
@@ -129,7 +129,9 @@ def non_max_suppression(
     if soft_nms_sigma is None:
         soft_nms_sigma = make_constant_node(0, np.float32)
 
-    inputs = as_nodes(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, soft_nms_sigma)
+    inputs = as_nodes(
+        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, soft_nms_sigma
+    )
     attributes = {
         "box_encoding": box_encoding,
         "sort_result_descending": sort_result_descending,

From 7597f3095699680f9c638cc8ca63b9e4dc6f767d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 09:44:49 +0300
Subject: [PATCH 090/173] Fixed element type for constants in the conversion of
 NMS-5 to NMSIE3.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp    | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index aceecea837ff94..7d89d43327fe05 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -75,17 +75,17 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         }
         auto new_iou_threshold = arg3;
         if (iou_threshold_rank.get_length() == 0) {
-            new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::f32, Shape{1}, {0.0f}));
+            new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
             new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
         }
         auto new_score_threshold = arg4;
         if (score_threshold_rank.get_length() == 0) {
-            new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::f32, Shape{1}, {0.0f}));
+            new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
             new_ops.push_back(new_score_threshold.get_node_shared_ptr());
         }
         auto new_soft_nms_sigma = arg5;
         if (soft_nms_sigma_rank.get_length() == 0) {
-            new_soft_nms_sigma = std::make_shared<ngraph::op::Unsqueeze>(arg5, opset1::Constant::create(element::f32, Shape{1}, {0.0f}));
+            new_soft_nms_sigma = std::make_shared<ngraph::op::Unsqueeze>(arg5, opset1::Constant::create(element::i64, Shape{1}, {0}));
             new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
         }
         int center_point_box = 0;

From debdf367b4166167cb7d024606d6293c156190f8 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 09:50:47 +0300
Subject: [PATCH 091/173] Deleted support of f64 in NMS-5.

---
 ngraph/core/src/op/non_max_suppression.cpp | 9 ---------
 1 file changed, 9 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 909f1c5e9216b9..4415565761e72b 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -890,15 +890,6 @@ namespace
             memcpy(result.data(), p, input_size * sizeof(float));
         }
         break;
-        case element::Type_t::f64:
-        {
-            double* p = input->get_data_ptr<double>();
-            for (size_t i = 0; i < input_size; ++i)
-            {
-                result[i] = float(p[i]);
-            }
-        }
-        break;
         default:;
         }
 

From 3d9d2e4ae33d9f95fe19d8da970374bc4c47060f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 10:46:01 +0300
Subject: [PATCH 092/173] Added checks for input element types for inputs #0,
 #1, #3, #4, #5.

---
 ngraph/core/src/op/non_max_suppression.cpp | 208 ++++++++++++---------
 1 file changed, 123 insertions(+), 85 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 4415565761e72b..c64598a63fad9b 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -676,91 +676,6 @@ shared_ptr<Node>
                                                        m_output_type);
 }
 
-void op::v5::NonMaxSuppression::validate()
-{
-    const auto boxes_ps = get_input_partial_shape(0);
-    const auto scores_ps = get_input_partial_shape(1);
-
-    NODE_VALIDATION_CHECK(this,
-                          m_output_type == element::i64 || m_output_type == element::i32,
-                          "Output type must be i32 or i64");
-
-    if (boxes_ps.is_dynamic() || scores_ps.is_dynamic())
-    {
-        return;
-    }
-
-    NODE_VALIDATION_CHECK(this,
-                          boxes_ps.rank().is_static() && boxes_ps.rank().get_length() == 3,
-                          "Expected a 3D tensor for the 'boxes' input. Got: ",
-                          boxes_ps);
-
-    NODE_VALIDATION_CHECK(this,
-                          scores_ps.rank().is_static() && scores_ps.rank().get_length() == 3,
-                          "Expected a 3D tensor for the 'scores' input. Got: ",
-                          scores_ps);
-
-    const auto max_boxes_ps = get_input_partial_shape(2);
-    NODE_VALIDATION_CHECK(this,
-                          max_boxes_ps.is_dynamic() || is_scalar(max_boxes_ps.to_shape()),
-                          "Expected a scalar for the 'max_output_boxes_per_class' input. Got: ",
-                          max_boxes_ps);
-
-    const auto iou_threshold_ps = get_input_partial_shape(3);
-    NODE_VALIDATION_CHECK(this,
-                          iou_threshold_ps.is_dynamic() || is_scalar(iou_threshold_ps.to_shape()),
-                          "Expected a scalar for the 'iou_threshold' input. Got: ",
-                          iou_threshold_ps);
-
-    const auto score_threshold_ps = get_input_partial_shape(4);
-    NODE_VALIDATION_CHECK(this,
-                          score_threshold_ps.is_dynamic() ||
-                              is_scalar(score_threshold_ps.to_shape()),
-                          "Expected a scalar for the 'score_threshold' input. Got: ",
-                          score_threshold_ps);
-
-    const auto soft_nms_sigma = get_input_partial_shape(5);
-    NODE_VALIDATION_CHECK(this,
-                          soft_nms_sigma.is_dynamic() || is_scalar(soft_nms_sigma.to_shape()),
-                          "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
-                          soft_nms_sigma);
-
-    const auto num_batches_boxes = boxes_ps[0];
-    const auto num_batches_scores = scores_ps[0];
-    NODE_VALIDATION_CHECK(this,
-                          num_batches_boxes.same_scheme(num_batches_scores),
-                          "The first dimension of both 'boxes' and 'scores' must match. Boxes: ",
-                          num_batches_boxes,
-                          "; Scores: ",
-                          num_batches_scores);
-
-    const auto num_boxes_boxes = boxes_ps[1];
-    const auto num_boxes_scores = scores_ps[2];
-    NODE_VALIDATION_CHECK(this,
-                          num_boxes_boxes.same_scheme(num_boxes_scores),
-                          "'boxes' and 'scores' input shapes must match at the second and third "
-                          "dimension respectively. Boxes: ",
-                          num_boxes_boxes,
-                          "; Scores: ",
-                          num_boxes_scores);
-
-    NODE_VALIDATION_CHECK(this,
-                          boxes_ps[2].is_static() && boxes_ps[2].get_length() == 4u,
-                          "The last dimension of the 'boxes' input must be equal to 4. Got:",
-                          boxes_ps[2]);
-}
-
-int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
-{
-    int64_t max_output_boxes{0};
-
-    const auto max_output_boxes_input =
-        as_type_ptr<op::Constant>(input_value(2).get_node_shared_ptr());
-    max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
-
-    return max_output_boxes;
-}
-
 using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
 
 namespace
@@ -958,6 +873,129 @@ namespace
             *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
         }
     }
+
+    inline bool is_float_type_admissible(const element::Type& t)
+    {
+        return t == element::f32 || t == element::f16 || t == element::bf16;
+    }
+}
+
+void op::v5::NonMaxSuppression::validate()
+{
+    const auto boxes_ps = get_input_partial_shape(0);
+    const auto scores_ps = get_input_partial_shape(1);
+
+    NODE_VALIDATION_CHECK(this,
+                          m_output_type == element::i64 || m_output_type == element::i32,
+                          "Output type must be i32 or i64");
+
+    if (boxes_ps.is_dynamic() || scores_ps.is_dynamic())
+    {
+        return;
+    }
+
+    NODE_VALIDATION_CHECK(this,
+                          is_float_type_admissible(get_input_element_type(0)),
+                          "Expected bf16, fp16 or fp32 as element type for the 'boxes' input.");
+
+    NODE_VALIDATION_CHECK(this,
+                          is_float_type_admissible(get_input_element_type(1)),
+                          "Expected bf16, fp16 or fp32 as element type for the 'scores' input.");
+
+    NODE_VALIDATION_CHECK(this,
+                          boxes_ps.rank().is_static() && boxes_ps.rank().get_length() == 3,
+                          "Expected a 3D tensor for the 'boxes' input. Got: ",
+                          boxes_ps);
+
+    NODE_VALIDATION_CHECK(this,
+                          scores_ps.rank().is_static() && scores_ps.rank().get_length() == 3,
+                          "Expected a 3D tensor for the 'scores' input. Got: ",
+                          scores_ps);
+
+    if (inputs().size() >= 3)
+    {
+        const auto max_boxes_ps = get_input_partial_shape(2);
+        NODE_VALIDATION_CHECK(this,
+                              max_boxes_ps.is_dynamic() || is_scalar(max_boxes_ps.to_shape()),
+                              "Expected a scalar for the 'max_output_boxes_per_class' input. Got: ",
+                              max_boxes_ps);
+    }
+
+    if (inputs().size() >= 4)
+    {
+        const auto iou_threshold_ps = get_input_partial_shape(3);
+        NODE_VALIDATION_CHECK(this,
+                              is_float_type_admissible(get_input_element_type(3)),
+                              "Expected bf16, fp16 or fp32 as element type for the "
+                              "'iou_threshold' input.");
+        NODE_VALIDATION_CHECK(this,
+                              iou_threshold_ps.is_dynamic() ||
+                                  is_scalar(iou_threshold_ps.to_shape()),
+                              "Expected a scalar for the 'iou_threshold' input. Got: ",
+                              iou_threshold_ps);
+    }
+
+    if (inputs().size() >= 5)
+    {
+        const auto score_threshold_ps = get_input_partial_shape(4);
+        NODE_VALIDATION_CHECK(this,
+                              is_float_type_admissible(get_input_element_type(4)),
+                              "Expected bf16, fp16 or fp32 as element type for the "
+                              "'score_threshold_ps' input.");
+        NODE_VALIDATION_CHECK(this,
+                              score_threshold_ps.is_dynamic() ||
+                                  is_scalar(score_threshold_ps.to_shape()),
+                              "Expected a scalar for the 'score_threshold' input. Got: ",
+                              score_threshold_ps);
+    }
+
+    if (inputs().size() >= 6)
+    {
+        const auto soft_nms_sigma = get_input_partial_shape(5);
+        NODE_VALIDATION_CHECK(this,
+                              is_float_type_admissible(get_input_element_type(5)),
+                              "Expected bf16, fp16 or fp32 as element type for the "
+                              "'soft_nms_sigma' input.");
+        NODE_VALIDATION_CHECK(this,
+                              soft_nms_sigma.is_dynamic() || is_scalar(soft_nms_sigma.to_shape()),
+                              "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
+                              soft_nms_sigma);
+    }
+
+    const auto num_batches_boxes = boxes_ps[0];
+    const auto num_batches_scores = scores_ps[0];
+    NODE_VALIDATION_CHECK(this,
+                          num_batches_boxes.same_scheme(num_batches_scores),
+                          "The first dimension of both 'boxes' and 'scores' must match. Boxes: ",
+                          num_batches_boxes,
+                          "; Scores: ",
+                          num_batches_scores);
+
+    const auto num_boxes_boxes = boxes_ps[1];
+    const auto num_boxes_scores = scores_ps[2];
+    NODE_VALIDATION_CHECK(this,
+                          num_boxes_boxes.same_scheme(num_boxes_scores),
+                          "'boxes' and 'scores' input shapes must match at the second and third "
+                          "dimension respectively. Boxes: ",
+                          num_boxes_boxes,
+                          "; Scores: ",
+                          num_boxes_scores);
+
+    NODE_VALIDATION_CHECK(this,
+                          boxes_ps[2].is_static() && boxes_ps[2].get_length() == 4u,
+                          "The last dimension of the 'boxes' input must be equal to 4. Got:",
+                          boxes_ps[2]);
+}
+
+int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
+{
+    int64_t max_output_boxes{0};
+
+    const auto max_output_boxes_input =
+        as_type_ptr<op::Constant>(input_value(2).get_node_shared_ptr());
+    max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
+
+    return max_output_boxes;
 }
 
 float op::v5::NonMaxSuppression::iou_threshold_from_input() const

From e6cbf479011a0265ed64a1e8853520e1bb7f47bd Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 11:02:37 +0300
Subject: [PATCH 093/173] Small change.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp        | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp b/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
index 630ef3f027bf99..375ec6cdc71ada 100644
--- a/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
+++ b/inference-engine/src/legacy_api/include/legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp
@@ -14,7 +14,7 @@
 namespace ngraph {
 namespace pass {
 
-    class INFERENCE_ENGINE_API_CLASS(ConvertNMS5ToLegacyMatcher);
+class INFERENCE_ENGINE_API_CLASS(ConvertNMS5ToLegacyMatcher);
 
 }  // namespace pass
 }  // namespace ngraph

From e004a33fb54e8535d8988ac13c48870226064c56 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 11:17:54 +0300
Subject: [PATCH 094/173] Now get_floats throws an exception for unsupported
 types.

---
 ngraph/core/src/op/non_max_suppression.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index c64598a63fad9b..2d1ab63bfbc521 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -22,6 +22,7 @@
 #include "ngraph/runtime/reference/non_max_suppression.hpp"
 #include "ngraph/type/bfloat16.hpp"
 #include "ngraph/type/float16.hpp"
+#include "ngraph/util.hpp"
 
 using namespace std;
 using namespace ngraph;
@@ -805,7 +806,7 @@ namespace
             memcpy(result.data(), p, input_size * sizeof(float));
         }
         break;
-        default:;
+        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
         }
 
         return result;

From ddea6e26eee10272646d06d507b790b3e1d5cf6d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 11:37:25 +0300
Subject: [PATCH 095/173] Now nGraph NMS-5 supports 0D and 1D tensors in inputs
 #2, #3, #4, #5.

---
 ngraph/core/src/op/non_max_suppression.cpp    | 31 ++++++--
 ngraph/test/type_prop/non_max_suppression.cpp | 73 ++++++++++---------
 2 files changed, 60 insertions(+), 44 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 2d1ab63bfbc521..16dc12c55560e0 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -879,6 +879,18 @@ namespace
     {
         return t == element::f32 || t == element::f16 || t == element::bf16;
     }
+
+    inline bool is_scalar_or_1d_tensor_with_1_element(const PartialShape& p)
+    {
+        if (p.is_dynamic())
+        {
+            return false;
+        }
+
+        Shape shape = p.to_shape();
+
+        return is_scalar(shape) || is_vector(shape) && (shape[0] == 1);
+    }
 }
 
 void op::v5::NonMaxSuppression::validate()
@@ -917,8 +929,10 @@ void op::v5::NonMaxSuppression::validate()
     {
         const auto max_boxes_ps = get_input_partial_shape(2);
         NODE_VALIDATION_CHECK(this,
-                              max_boxes_ps.is_dynamic() || is_scalar(max_boxes_ps.to_shape()),
-                              "Expected a scalar for the 'max_output_boxes_per_class' input. Got: ",
+                              max_boxes_ps.is_dynamic() ||
+                                  is_scalar_or_1d_tensor_with_1_element(max_boxes_ps),
+                              "Expected 0D or 1D tensor for the 'max_output_boxes_per_class' input. "
+                              "Got: ",
                               max_boxes_ps);
     }
 
@@ -931,8 +945,8 @@ void op::v5::NonMaxSuppression::validate()
                               "'iou_threshold' input.");
         NODE_VALIDATION_CHECK(this,
                               iou_threshold_ps.is_dynamic() ||
-                                  is_scalar(iou_threshold_ps.to_shape()),
-                              "Expected a scalar for the 'iou_threshold' input. Got: ",
+                                  is_scalar_or_1d_tensor_with_1_element(iou_threshold_ps),
+                              "Expected 0D or 1D tensor for the 'iou_threshold' input. Got: ",
                               iou_threshold_ps);
     }
 
@@ -945,8 +959,8 @@ void op::v5::NonMaxSuppression::validate()
                               "'score_threshold_ps' input.");
         NODE_VALIDATION_CHECK(this,
                               score_threshold_ps.is_dynamic() ||
-                                  is_scalar(score_threshold_ps.to_shape()),
-                              "Expected a scalar for the 'score_threshold' input. Got: ",
+                                  is_scalar_or_1d_tensor_with_1_element(score_threshold_ps),
+                              "Expected 0D or 1D tensor for the 'score_threshold' input. Got: ",
                               score_threshold_ps);
     }
 
@@ -958,8 +972,9 @@ void op::v5::NonMaxSuppression::validate()
                               "Expected bf16, fp16 or fp32 as element type for the "
                               "'soft_nms_sigma' input.");
         NODE_VALIDATION_CHECK(this,
-                              soft_nms_sigma.is_dynamic() || is_scalar(soft_nms_sigma.to_shape()),
-                              "Expected a scalar for the 'soft_nms_sigma' input. Got: ",
+                              soft_nms_sigma.is_dynamic() ||
+                                is_scalar_or_1d_tensor_with_1_element(soft_nms_sigma),
+                              "Expected 0D or 1D tensor for the 'soft_nms_sigma' input. Got: ",
                               soft_nms_sigma);
     }
 
diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index d1ced839867ad3..7c6531629eca09 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -621,42 +621,43 @@ TEST(type_prop, nms_v5_scalar_inputs_check)
     const auto scalar = make_shared<op::Parameter>(element::f32, Shape{});
     const auto non_scalar = make_shared<op::Parameter>(element::f32, Shape{1});
 
-    try
-    {
-        make_shared<op::v5::NonMaxSuppression>(boxes, scores, non_scalar, scalar, scalar);
-    }
-    catch (const NodeValidationFailure& error)
-    {
-        EXPECT_HAS_SUBSTRING(error.what(),
-                             "Expected a scalar for the 'max_output_boxes_per_class' input");
-    }
-
-    try
-    {
-        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, non_scalar, scalar);
-    }
-    catch (const NodeValidationFailure& error)
-    {
-        EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'iou_threshold' input");
-    }
-
-    try
-    {
-        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, non_scalar);
-    }
-    catch (const NodeValidationFailure& error)
-    {
-        EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'score_threshold' input");
-    }
-
-    try
-    {
-        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, scalar, non_scalar);
-    }
-    catch (const NodeValidationFailure& error)
-    {
-        EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'soft_nms_sigma' input");
-    }
+//     try
+//     {
+//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, non_scalar, scalar, scalar);
+//     }
+//     catch (const NodeValidationFailure& error)
+//     {
+//         EXPECT_HAS_SUBSTRING(error.what(),
+//                              "Expected 0D or 1D tensor for the 'max_output_boxes_per_class' input");
+//     }
+//
+//     try
+//     {
+//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, non_scalar, scalar);
+//     }
+//     catch (const NodeValidationFailure& error)
+//     {
+//         EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'iou_threshold' input");
+//     }
+//
+//     try
+//     {
+//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, non_scalar);
+//     }
+//     catch (const NodeValidationFailure& error)
+//     {
+//         EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'score_threshold' input");
+//     }
+//
+//     try
+//     {
+//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, scalar, non_scalar);
+//     }
+//     catch (const NodeValidationFailure& error)
+//     {
+//         EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'soft_nms_sigma' input");
+//     }
+    ASSERT_TRUE(true);
 }
 
 TEST(type_prop, nms_v5_output_shape)

From c09b791270be247d7f531c46daf080769d65a39b Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 11:48:30 +0300
Subject: [PATCH 096/173] Small fix in test_non_max_suppression.

---
 ngraph/python/tests/test_ngraph/test_reduction.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/python/tests/test_ngraph/test_reduction.py b/ngraph/python/tests/test_ngraph/test_reduction.py
index 6aa94a7c52fa4b..b23e8b242b3720 100644
--- a/ngraph/python/tests/test_ngraph/test_reduction.py
+++ b/ngraph/python/tests/test_ngraph/test_reduction.py
@@ -112,7 +112,7 @@ def test_non_max_suppression():
     node = ng.non_max_suppression(boxes_parameter, scores_parameter)
 
     assert node.get_type_name() == "NonMaxSuppression"
-    assert node.get_output_size() == 1
+    assert node.get_output_size() == 3
     assert list(node.get_output_shape(0)) == expected_shape
 
 

From ab571576b39549172a64c85f3dfbf34f1ac111e7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 12:20:26 +0300
Subject: [PATCH 097/173] Deleted using namespace std

---
 ngraph/core/src/op/non_max_suppression.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 16dc12c55560e0..445917f5be4310 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -24,7 +24,6 @@
 #include "ngraph/type/float16.hpp"
 #include "ngraph/util.hpp"
 
-using namespace std;
 using namespace ngraph;
 
 // ------------------------------ V1 ------------------------------

From 53190d908384150e54876ed448a5e47f329c5dd0 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 12:29:22 +0300
Subject: [PATCH 098/173] Fixes in test_non_max_suppression().

---
 ngraph/python/tests/test_ngraph/test_reduction.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/ngraph/python/tests/test_ngraph/test_reduction.py b/ngraph/python/tests/test_ngraph/test_reduction.py
index b23e8b242b3720..6cf5f4a81e71ee 100644
--- a/ngraph/python/tests/test_ngraph/test_reduction.py
+++ b/ngraph/python/tests/test_ngraph/test_reduction.py
@@ -105,7 +105,6 @@ def test_non_max_suppression():
 
     boxes_shape = [1, 1000, 4]
     scores_shape = [1, 1, 1000]
-    expected_shape = [0, 3]
     boxes_parameter = ng.parameter(boxes_shape, name="Boxes", dtype=np.float32)
     scores_parameter = ng.parameter(scores_shape, name="Scores", dtype=np.float32)
 
@@ -113,7 +112,9 @@ def test_non_max_suppression():
 
     assert node.get_type_name() == "NonMaxSuppression"
     assert node.get_output_size() == 3
-    assert list(node.get_output_shape(0)) == expected_shape
+    assert node.get_output_partial_shape(0).same_scheme(PartialShape([-1, 3]))
+    assert node.get_output_partial_shape(1).same_scheme(PartialShape([-1, 3]))
+    assert list(node.get_output_shape(2)) == [1]
 
 
 def test_non_zero():

From 2e23b0aeadb9e74f37be4bd0ba3c013e3397826c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 12:33:33 +0300
Subject: [PATCH 099/173] Small fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 445917f5be4310..c115ad08e3ef96 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -61,7 +61,7 @@ op::v1::NonMaxSuppression::NonMaxSuppression(
     constructor_validate_and_infer_types();
 }
 
-shared_ptr<Node>
+std::shared_ptr<Node>
     op::v1::NonMaxSuppression::clone_with_new_inputs(const OutputVector& new_args) const
 {
     check_new_args_count(this, new_args);
@@ -259,7 +259,7 @@ op::v3::NonMaxSuppression::NonMaxSuppression(
     constructor_validate_and_infer_types();
 }
 
-shared_ptr<Node>
+std::shared_ptr<Node>
     op::v3::NonMaxSuppression::clone_with_new_inputs(const OutputVector& new_args) const
 {
     check_new_args_count(this, new_args);
@@ -478,7 +478,7 @@ op::v4::NonMaxSuppression::NonMaxSuppression(
     constructor_validate_and_infer_types();
 }
 
-shared_ptr<Node>
+std::shared_ptr<Node>
     op::v4::NonMaxSuppression::clone_with_new_inputs(const OutputVector& new_args) const
 {
     check_new_args_count(this, new_args);
@@ -644,7 +644,7 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     constructor_validate_and_infer_types();
 }
 
-shared_ptr<Node>
+std::shared_ptr<Node>
     op::v5::NonMaxSuppression::clone_with_new_inputs(const OutputVector& new_args) const
 {
     check_new_args_count(this, new_args);

From facc5083f6ab4c2ed1cc55c3a0e112b57812db6a Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 13:03:16 +0300
Subject: [PATCH 100/173] Added 'import PartialShape' in test_reduction.py.

---
 ngraph/python/tests/test_ngraph/test_reduction.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/ngraph/python/tests/test_ngraph/test_reduction.py b/ngraph/python/tests/test_ngraph/test_reduction.py
index 6cf5f4a81e71ee..4ca00f59dce8fb 100644
--- a/ngraph/python/tests/test_ngraph/test_reduction.py
+++ b/ngraph/python/tests/test_ngraph/test_reduction.py
@@ -15,6 +15,7 @@
 # ******************************************************************************
 import numpy as np
 import pytest
+from _pyngraph import PartialShape
 
 import ngraph as ng
 from tests.runtime import get_runtime

From 4e27e3589b81a3010e0410c09c221646b462ab3b Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 13:13:36 +0300
Subject: [PATCH 101/173] Deleted creating fake inputs in the ctor of
 op::v5::NonMaxSuppression.

---
 ngraph/core/src/op/non_max_suppression.cpp | 31 ++++------------------
 1 file changed, 5 insertions(+), 26 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index c115ad08e3ef96..f5f19fa5b7bd74 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -545,12 +545,7 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
     const bool sort_result_descending,
     const element::Type& output_type)
-    : Op({boxes,
-          scores,
-          op::Constant::create(element::i64, Shape{}, {0}),
-          op::Constant::create(element::f32, Shape{}, {.0f}),
-          op::Constant::create(element::f32, Shape{}, {.0f}),
-          op::Constant::create(element::f32, Shape{}, {.0f})})
+    : Op({boxes, scores})
     , m_box_encoding{box_encoding}
     , m_sort_result_descending{sort_result_descending}
     , m_output_type{output_type}
@@ -565,12 +560,7 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
     const bool sort_result_descending,
     const element::Type& output_type)
-    : Op({boxes,
-          scores,
-          max_output_boxes_per_class,
-          op::Constant::create(element::f32, Shape{}, {.0f}),
-          op::Constant::create(element::f32, Shape{}, {.0f}),
-          op::Constant::create(element::f32, Shape{}, {.0f})})
+    : Op({boxes, scores, max_output_boxes_per_class})
     , m_box_encoding{box_encoding}
     , m_sort_result_descending{sort_result_descending}
     , m_output_type{output_type}
@@ -586,12 +576,7 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
     const bool sort_result_descending,
     const element::Type& output_type)
-    : Op({boxes,
-          scores,
-          max_output_boxes_per_class,
-          iou_threshold,
-          op::Constant::create(element::f32, Shape{}, {.0f}),
-          op::Constant::create(element::f32, Shape{}, {.0f})})
+    : Op({boxes, scores, max_output_boxes_per_class, iou_threshold})
     , m_box_encoding{box_encoding}
     , m_sort_result_descending{sort_result_descending}
     , m_output_type{output_type}
@@ -608,12 +593,7 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
     const bool sort_result_descending,
     const element::Type& output_type)
-    : Op({boxes,
-          scores,
-          max_output_boxes_per_class,
-          iou_threshold,
-          score_threshold,
-          op::Constant::create(element::f32, Shape{}, {.0f})})
+    : Op({boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold})
     , m_box_encoding{box_encoding}
     , m_sort_result_descending{sort_result_descending}
     , m_output_type{output_type}
@@ -631,8 +611,7 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
     const bool sort_result_descending,
     const element::Type& output_type)
-    : Op({boxes,
-          scores,
+    : Op({boxes, scores,
           max_output_boxes_per_class,
           iou_threshold,
           score_threshold,

From 8249034a53d69b3e4b340dd2157ec2fe31d19acb Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 13:27:59 +0300
Subject: [PATCH 102/173] Deleted creating fake inputs in
 op::v5::NonMaxSuppression::clone_with_new_inputs.

---
 ngraph/core/src/op/non_max_suppression.cpp | 70 +++++++++++++++-------
 1 file changed, 48 insertions(+), 22 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index f5f19fa5b7bd74..acad31c1a930a4 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -631,28 +631,54 @@ std::shared_ptr<Node>
                           new_args.size() >= 2 && new_args.size() <= 6,
                           "Number of inputs must be 2, 3, 4, 5 or 6");
 
-    const auto& arg2 = new_args.size() > 2
-                           ? new_args.at(2)
-                           : ngraph::op::Constant::create(element::i64, Shape{}, {0});
-    const auto& arg3 = new_args.size() > 3
-                           ? new_args.at(3)
-                           : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
-    const auto& arg4 = new_args.size() > 4
-                           ? new_args.at(4)
-                           : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
-    const auto& arg5 = new_args.size() > 5
-                           ? new_args.at(5)
-                           : ngraph::op::Constant::create(element::f32, Shape{}, {.0f});
-
-    return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
-                                                       new_args.at(1),
-                                                       arg2,
-                                                       arg3,
-                                                       arg4,
-                                                       arg5,
-                                                       m_box_encoding,
-                                                       m_sort_result_descending,
-                                                       m_output_type);
+    switch (new_args.size())
+    {
+    case 2:
+        return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
+                                                           new_args.at(1),
+                                                           m_box_encoding,
+                                                           m_sort_result_descending,
+                                                           m_output_type)
+        break;
+    case 3:
+        return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
+                                                           new_args.at(1),
+                                                           new_args.at(2),
+                                                           m_box_encoding,
+                                                           m_sort_result_descending,
+                                                           m_output_type)
+        break;
+    case 4:
+        return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
+                                                           new_args.at(1),
+                                                           new_args.at(2),
+                                                           new_args.at(3),
+                                                           m_box_encoding,
+                                                           m_sort_result_descending,
+                                                           m_output_type)
+        break;
+    case 5:
+        return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
+                                                           new_args.at(1),
+                                                           new_args.at(2),
+                                                           new_args.at(3),
+                                                           new_args.at(4),
+                                                           m_box_encoding,
+                                                           m_sort_result_descending,
+                                                           m_output_type)
+        break;
+    default:
+        return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
+                                                           new_args.at(1),
+                                                           new_args.at(2),
+                                                           new_args.at(3),
+                                                           new_args.at(4),
+                                                           new_args.at(5),
+                                                           m_box_encoding,
+                                                           m_sort_result_descending,
+                                                           m_output_type)
+        break;
+    }
 }
 
 using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;

From fcd2d00ede2df686b12d0bd0632c5cdb243f1694 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 13:35:02 +0300
Subject: [PATCH 103/173] Corrections in int64_t
 op::v5::NonMaxSuppression::max_boxes_output_from_input() const.

---
 ngraph/core/src/op/non_max_suppression.cpp | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index acad31c1a930a4..b16565a5f738ac 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -687,6 +687,7 @@ namespace
 {
     constexpr size_t boxes_port = 0;
     constexpr size_t scores_port = 1;
+    constexpr size_t max_output_boxes_port = 2;
     constexpr size_t iou_threshold_port = 3;
     constexpr size_t score_threshold_port = 4;
     constexpr size_t soft_nms_sigma_port = 5;
@@ -1011,8 +1012,14 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
 {
     int64_t max_output_boxes{0};
 
+    size_t num_of_inputs = inputs().size();
+    if (num_of_inputs < 3)
+    {
+        return 0;
+    }
+
     const auto max_output_boxes_input =
-        as_type_ptr<op::Constant>(input_value(2).get_node_shared_ptr());
+        as_type_ptr<op::Constant>(input_value(max_output_boxes_port).get_node_shared_ptr());
     max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
 
     return max_output_boxes;

From 635babdb46c11c658341a27fd97393be24cefd26 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 14:13:05 +0300
Subject: [PATCH 104/173] Corrected functions float
 op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const, float
 op::v5::NonMaxSuppression::score_threshold_from_input() const,  float
 op::v5::NonMaxSuppression::iou_threshold_from_input() const.

---
 ngraph/core/src/op/non_max_suppression.cpp | 50 ++++++++++++++++++++--
 1 file changed, 47 insertions(+), 3 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index b16565a5f738ac..53e1ca2db8951b 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -896,6 +896,26 @@ namespace
 
         return is_scalar(shape) || is_vector(shape) && (shape[0] == 1);
     }
+
+    float float_from_constant_node(const std::shared_ptr<op::Constant>& constant,
+                                   const element::Type& et)
+    {
+        float result = 0.0f;
+        switch (et)
+        {
+        case element::Type_t::bf16:
+            result = static_cast<float>(constant->cast_vector<bfloat16>().at(0));
+            break;
+        case element::Type_t::f16:
+            result = static_cast<float>(constant->cast_vector<float16>().at(0));
+            break;
+        case element::Type_t::f32:
+            result = constant->cast_vector<float>().at(0);
+            break;
+        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
+        }
+        return result;
+    }
 }
 
 void op::v5::NonMaxSuppression::validate()
@@ -1029,9 +1049,17 @@ float op::v5::NonMaxSuppression::iou_threshold_from_input() const
 {
     float iou_threshold = 0.0f;
 
+    size_t num_of_inputs = inputs().size();
+    if (num_of_inputs < 4)
+    {
+        return iou_threshold;
+    }
+
     const auto iou_threshold_input =
         as_type_ptr<op::Constant>(input_value(iou_threshold_port).get_node_shared_ptr());
-    iou_threshold = iou_threshold_input->cast_vector<float>().at(0);
+    iou_threshold = float_from_constant_node(iou_threshold_input,
+                                             get_input_element_type(iou_threshold_port));
+    // iou_threshold = iou_threshold_input->cast_vector<float>().at(0);
 
     return iou_threshold;
 }
@@ -1040,9 +1068,17 @@ float op::v5::NonMaxSuppression::score_threshold_from_input() const
 {
     float score_threshold = 0.0f;
 
+    size_t num_of_inputs = inputs().size();
+    if (num_of_inputs < 5)
+    {
+        return score_threshold;
+    }
+
     const auto score_threshold_input =
         as_type_ptr<op::Constant>(input_value(score_threshold_port).get_node_shared_ptr());
-    score_threshold = score_threshold_input->cast_vector<float>().at(0);
+    score_threshold = float_from_constant_node(score_threshold_input,
+                                               get_input_element_type(score_threshold_port));
+    // score_threshold = score_threshold_input->cast_vector<float>().at(0);
 
     return score_threshold;
 }
@@ -1051,9 +1087,17 @@ float op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const
 {
     float soft_nms_sigma = 0.0f;
 
+    size_t num_of_inputs = inputs().size();
+    if (num_of_inputs < 6)
+    {
+        return soft_nms_sigma;
+    }
+
     const auto soft_nms_sigma_input =
         as_type_ptr<op::Constant>(input_value(soft_nms_sigma_port).get_node_shared_ptr());
-    soft_nms_sigma = soft_nms_sigma_input->cast_vector<float>().at(0);
+    soft_nms_sigma = float_from_constant_node(soft_nms_sigma_input,
+                                              get_input_element_type(soft_nms_sigma_port));
+    // soft_nms_sigma = soft_nms_sigma_input->cast_vector<float>().at(0);
 
     return soft_nms_sigma;
 }

From e1b4ee05785d5ef3b7a3ac73354b331799fccd2e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 14:40:29 +0300
Subject: [PATCH 105/173] Small fixes.

---
 ngraph/core/src/op/non_max_suppression.cpp | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 53e1ca2db8951b..09f98d5dfa8c34 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -638,7 +638,7 @@ std::shared_ptr<Node>
                                                            new_args.at(1),
                                                            m_box_encoding,
                                                            m_sort_result_descending,
-                                                           m_output_type)
+                                                           m_output_type);
         break;
     case 3:
         return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
@@ -646,7 +646,7 @@ std::shared_ptr<Node>
                                                            new_args.at(2),
                                                            m_box_encoding,
                                                            m_sort_result_descending,
-                                                           m_output_type)
+                                                           m_output_type);
         break;
     case 4:
         return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
@@ -655,7 +655,7 @@ std::shared_ptr<Node>
                                                            new_args.at(3),
                                                            m_box_encoding,
                                                            m_sort_result_descending,
-                                                           m_output_type)
+                                                           m_output_type);
         break;
     case 5:
         return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
@@ -665,7 +665,7 @@ std::shared_ptr<Node>
                                                            new_args.at(4),
                                                            m_box_encoding,
                                                            m_sort_result_descending,
-                                                           m_output_type)
+                                                           m_output_type);
         break;
     default:
         return std::make_shared<op::v5::NonMaxSuppression>(new_args.at(0),
@@ -676,7 +676,7 @@ std::shared_ptr<Node>
                                                            new_args.at(5),
                                                            m_box_encoding,
                                                            m_sort_result_descending,
-                                                           m_output_type)
+                                                           m_output_type);
         break;
     }
 }

From 3f14ed9c3b17d38d121bc792016eebb4651762c6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 15:38:19 +0300
Subject: [PATCH 106/173] Deleted commented code.

---
 ngraph/core/src/op/non_max_suppression.cpp | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 09f98d5dfa8c34..235781c2a63683 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -1059,7 +1059,6 @@ float op::v5::NonMaxSuppression::iou_threshold_from_input() const
         as_type_ptr<op::Constant>(input_value(iou_threshold_port).get_node_shared_ptr());
     iou_threshold = float_from_constant_node(iou_threshold_input,
                                              get_input_element_type(iou_threshold_port));
-    // iou_threshold = iou_threshold_input->cast_vector<float>().at(0);
 
     return iou_threshold;
 }
@@ -1078,7 +1077,6 @@ float op::v5::NonMaxSuppression::score_threshold_from_input() const
         as_type_ptr<op::Constant>(input_value(score_threshold_port).get_node_shared_ptr());
     score_threshold = float_from_constant_node(score_threshold_input,
                                                get_input_element_type(score_threshold_port));
-    // score_threshold = score_threshold_input->cast_vector<float>().at(0);
 
     return score_threshold;
 }
@@ -1097,7 +1095,6 @@ float op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const
         as_type_ptr<op::Constant>(input_value(soft_nms_sigma_port).get_node_shared_ptr());
     soft_nms_sigma = float_from_constant_node(soft_nms_sigma_input,
                                               get_input_element_type(soft_nms_sigma_port));
-    // soft_nms_sigma = soft_nms_sigma_input->cast_vector<float>().at(0);
 
     return soft_nms_sigma;
 }

From 6cc4fc9f18f30a62d48fe85b41c01409fa2491fb Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 16 Oct 2020 16:19:04 +0300
Subject: [PATCH 107/173] Fixes in nms_v5_scalar_inputs_check.

---
 ngraph/test/type_prop/non_max_suppression.cpp | 78 ++++++++++---------
 1 file changed, 40 insertions(+), 38 deletions(-)

diff --git a/ngraph/test/type_prop/non_max_suppression.cpp b/ngraph/test/type_prop/non_max_suppression.cpp
index 7c6531629eca09..ab70f7cb45709b 100644
--- a/ngraph/test/type_prop/non_max_suppression.cpp
+++ b/ngraph/test/type_prop/non_max_suppression.cpp
@@ -619,45 +619,47 @@ TEST(type_prop, nms_v5_scalar_inputs_check)
     const auto scores = make_shared<op::Parameter>(element::f32, Shape{1, 2, 2});
 
     const auto scalar = make_shared<op::Parameter>(element::f32, Shape{});
-    const auto non_scalar = make_shared<op::Parameter>(element::f32, Shape{1});
+    const auto non_0d_or_1d = make_shared<op::Parameter>(element::f32, Shape{2});
 
-//     try
-//     {
-//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, non_scalar, scalar, scalar);
-//     }
-//     catch (const NodeValidationFailure& error)
-//     {
-//         EXPECT_HAS_SUBSTRING(error.what(),
-//                              "Expected 0D or 1D tensor for the 'max_output_boxes_per_class' input");
-//     }
-//
-//     try
-//     {
-//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, non_scalar, scalar);
-//     }
-//     catch (const NodeValidationFailure& error)
-//     {
-//         EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'iou_threshold' input");
-//     }
-//
-//     try
-//     {
-//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, non_scalar);
-//     }
-//     catch (const NodeValidationFailure& error)
-//     {
-//         EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'score_threshold' input");
-//     }
-//
-//     try
-//     {
-//         make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, scalar, non_scalar);
-//     }
-//     catch (const NodeValidationFailure& error)
-//     {
-//         EXPECT_HAS_SUBSTRING(error.what(), "Expected a scalar for the 'soft_nms_sigma' input");
-//     }
-    ASSERT_TRUE(true);
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, non_0d_or_1d, scalar, scalar);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "Expected 0D or 1D tensor for the 'max_output_boxes_per_class' input");
+    }
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, non_0d_or_1d, scalar);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "Expected 0D or 1D tensor for the 'iou_threshold' input");
+    }
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, non_0d_or_1d);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "Expected 0D or 1D tensor for the 'score_threshold' input");
+    }
+
+    try
+    {
+        make_shared<op::v5::NonMaxSuppression>(boxes, scores, scalar, scalar, scalar, non_0d_or_1d);
+    }
+    catch (const NodeValidationFailure& error)
+    {
+        EXPECT_HAS_SUBSTRING(error.what(),
+                             "Expected 0D or 1D tensor for the 'soft_nms_sigma' input");
+    }
 }
 
 TEST(type_prop, nms_v5_output_shape)

From 7da307f4f6e5400b20842424efc4266bb8b4f59d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 10:25:18 +0300
Subject: [PATCH 108/173] Some changes.

---
 .../convert_previous_nms_to_nms_5.hpp         |   7 +
 .../convert_previous_nms_to_nms_5.cpp         | 257 ++++++++++++++++--
 ngraph/core/src/op/non_max_suppression.cpp    |   2 +-
 3 files changed, 244 insertions(+), 22 deletions(-)

diff --git a/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp b/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
index 9b7d610dc73d07..c3f5e1a9cfd6d0 100644
--- a/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
+++ b/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
@@ -16,6 +16,7 @@ namespace pass {
 class TRANSFORMATIONS_API ConvertNMS1ToNMS5;
 class TRANSFORMATIONS_API ConvertNMS3ToNMS5;
 class TRANSFORMATIONS_API ConvertNMS4ToNMS5;
+class TRANSFORMATIONS_API ConvertPreviousNMSToNMS5;
 
 }  // namespace pass
 }  // namespace ngraph
@@ -37,3 +38,9 @@ class ngraph::pass::ConvertNMS4ToNMS5: public ngraph::pass::MatcherPass {
     NGRAPH_RTTI_DECLARATION;
     ConvertNMS4ToNMS5();
 };
+
+class ngraph::pass::ConvertPreviousNMSToNMS5: public ngraph::pass::MatcherPass {
+public:
+    NGRAPH_RTTI_DECLARATION;
+    ConvertPreviousNMSToNMS5();
+};
diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 6e0b78ae190f43..6a4d3182e298f9 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -15,16 +15,229 @@
 #include <ngraph/rt_info.hpp>
 #include <ngraph/pattern/op/wrap_type.hpp>
 
+namespace
+{
+    std::shared_ptr<ngraph::opset4::NonMaxSuppression> nms4_pattern()
+    {
+        auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                       iou_threshold, score_threshold);
+        return nms;
+    }
+
+    std::shared_ptr<ngraph::opset3::NonMaxSuppression> nms3_pattern()
+    {
+        auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                       iou_threshold, score_threshold);
+        return nms;
+    }
+
+    std::shared_ptr<ngraph::opset1::NonMaxSuppression> nms1_pattern()
+    {
+        auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                                       iou_threshold, score_threshold);
+        return nms;
+    }
+
+    using NMS5BoxEncoding = ngraph::opset5::NonMaxSuppression::BoxEncodingType;
+
+    struct NMSAttributes
+    {
+        ngraph::element::Type output_type;
+        NMS5BoxEncoding box_encoding;
+        bool sort_result_descending;
+        bool is_supported_nms;
+    };
+
+    NMSAttributes get_nms4_attrs(const std::shared_ptr<ngraph::opset4::NonMaxSuppression>& nms4)
+    {
+        NMSAttributes attrs;
+
+        attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+        attrs.is_supported_nms = true;
+        attrs.sort_result_descending = true;
+        attrs.output_type = ::ngraph::element::i64;
+
+        switch (nms4->get_box_encoding()) {
+            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
+                attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+                break;
+            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
+                attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms4->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
+
+        attrs.sort_result_descending = nms4->get_sort_result_descending(),
+        attrs.output_type = nms4->get_output_type();
+
+        return attrs;
+    }
+
+    NMSAttributes get_nms3_attrs(const std::shared_ptr<ngraph::opset3::NonMaxSuppression>& nms3)
+    {
+        NMSAttributes attrs;
+
+        attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+        attrs.is_supported_nms = true;
+        attrs.sort_result_descending = true;
+        attrs.output_type = ::ngraph::element::i64;
+
+        switch (nms3->get_box_encoding()) {
+            case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CENTER:
+                attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+                break;
+            case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CORNER:
+                attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms3->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
+
+        attrs.sort_result_descending = nms3->get_sort_result_descending(),
+        attrs.output_type = nms3->get_output_type();
+
+        return attrs;
+    }
+
+    NMSAttributes get_nms1_attrs(const std::shared_ptr<ngraph::opset1::NonMaxSuppression>& nms1)
+    {
+        NMSAttributes attrs;
+
+        attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+        attrs.is_supported_nms = true;
+        attrs.sort_result_descending = true;
+        attrs.output_type = ::ngraph::element::i64;
+
+        switch (nms1->get_box_encoding()) {
+            case ::ngraph::opset1::NonMaxSuppression::BoxEncodingType::CENTER:
+                attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+                break;
+            case ::ngraph::opset1::NonMaxSuppression::BoxEncodingType::CORNER:
+                attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+                break;
+            default:
+                throw ngraph_error("NonMaxSuppression layer " + nms1->get_friendly_name() +
+                                   " has unsupported box encoding");
+        }
+
+        attrs.sort_result_descending = nms1->get_sort_result_descending(),
+
+        return attrs;
+    }
+
+    NMSAttributes get_nms_attrs(const std::shared_ptr<ngraph::Node>& root)
+    {
+        NMSAttributes attrs;
+        attrs.output_type = ::ngraph::element::i64;
+        attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+        attrs.sort_result_descending = false;
+        attrs.is_supported_nms = false;
+
+        auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
+        if (nms_4) {
+            return get_nms4_attrs(nms_4);
+        }
+        auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(m.get_match_root());
+        if (nms_4) {
+            return get_nms3_attrs(nms_3);
+        }
+        auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());
+        if (nms_1) {
+            return get_nms1_attrs(nms_3);
+        }
+
+        return attrs;
+    }
+
+    bool callback_func(pattern::Matcher &m)
+    {
+        auto root = m.get_match_root();
+
+        auto attrs = get_nms_attrs(root);
+        if (!attrs.is_supported_nms) {
+            return false;
+        }
+
+        const auto new_args = root->input_values();
+
+        size_t num_of_args = new_args.size();
+
+        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
+        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+
+        // list of new nGraph operations
+        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
+
+        new_ops_list.push_front(arg5);
+        if (num_of_args <= 4) {
+            new_ops_list.push_front(arg4.get_node_shared_ptr());
+        }
+        if (num_of_args <= 3) {
+            new_ops_list.push_front(arg3.get_node_shared_ptr());
+        }
+        if (num_of_args <= 2) {
+            new_ops_list.push_front(arg2.get_node_shared_ptr());
+        }
+
+        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
+                new_args.at(0),
+                new_args.at(1),
+                arg2,
+                arg3,
+                arg4,
+                arg5,
+                box_encoding,
+                attrs.sort_result_descending,
+                attrs.output_type);
+
+        new_ops_list.push_back(nms_5);
+
+        // vector of new nGraph operations
+        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
+
+        nms_5->set_friendly_name(root->get_friendly_name());
+        ngraph::copy_runtime_info(root, new_ops);
+        root->output(0).replace(nms_5->output(0));
+        return true;
+    }
+}
+
+NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertPreviousNMSToNMS5, "ConvertPreviousNMSToNMS5", 0);
+
+ngraph::pass::ConvertPreviousNMSToNMS5::ConvertPreviousNMSToNMS5() {
+}
+
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
 
 ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
-    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-    auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
-    auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
-    auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
-    auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                   iou_threshold, score_threshold);
+    auto nms = nms4_pattern();
+//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+//     auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
+//     auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
+//     auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
+//     auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+//                                                                    iou_threshold, score_threshold);
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
@@ -98,13 +311,14 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS3ToNMS5, "ConvertNMS3ToNMS5", 0);
 
 ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
-    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-    auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
-    auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
-    auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
-    auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                   iou_threshold, score_threshold);
+    auto nms = nms3_pattern();
+//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+//     auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
+//     auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
+//     auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
+//     auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+//                                                                    iou_threshold, score_threshold);
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(m.get_match_root());
@@ -177,13 +391,14 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS1ToNMS5, "ConvertNMS1ToNMS5", 0);
 
 ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
-    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-    auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
-    auto iou_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.75});
-    auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
-    auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                   iou_threshold, score_threshold);
+    auto nms = nms1_pattern();
+//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
+//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
+//     auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
+//     auto iou_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.75});
+//     auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
+//     auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+//                                                                    iou_threshold, score_threshold);
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 235781c2a63683..4259e2d4514482 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -894,7 +894,7 @@ namespace
 
         Shape shape = p.to_shape();
 
-        return is_scalar(shape) || is_vector(shape) && (shape[0] == 1);
+        return is_scalar(shape) || (is_vector(shape) && (shape[0] == 1));
     }
 
     float float_from_constant_node(const std::shared_ptr<op::Constant>& constant,

From 40adcbb0ff98b3af2ab5ca0bf147b7d7b39f87cc Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 10:57:33 +0300
Subject: [PATCH 109/173] Small fixes.

---
 .../convert_previous_nms_to_nms_5.cpp         | 32 +++++++------------
 1 file changed, 11 insertions(+), 21 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 6a4d3182e298f9..de2c905bbb0359 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -15,10 +15,8 @@
 #include <ngraph/rt_info.hpp>
 #include <ngraph/pattern/op/wrap_type.hpp>
 
-namespace
-{
-    std::shared_ptr<ngraph::opset4::NonMaxSuppression> nms4_pattern()
-    {
+namespace {
+    std::shared_ptr<ngraph::opset4::NonMaxSuppression> nms4_pattern() {
         auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
         auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
         auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
@@ -29,8 +27,7 @@ namespace
         return nms;
     }
 
-    std::shared_ptr<ngraph::opset3::NonMaxSuppression> nms3_pattern()
-    {
+    std::shared_ptr<ngraph::opset3::NonMaxSuppression> nms3_pattern() {
         auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
         auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
         auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
@@ -41,8 +38,7 @@ namespace
         return nms;
     }
 
-    std::shared_ptr<ngraph::opset1::NonMaxSuppression> nms1_pattern()
-    {
+    std::shared_ptr<ngraph::opset1::NonMaxSuppression> nms1_pattern() {
         auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
         auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
         auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
@@ -55,16 +51,14 @@ namespace
 
     using NMS5BoxEncoding = ngraph::opset5::NonMaxSuppression::BoxEncodingType;
 
-    struct NMSAttributes
-    {
+    struct NMSAttributes {
         ngraph::element::Type output_type;
         NMS5BoxEncoding box_encoding;
         bool sort_result_descending;
         bool is_supported_nms;
     };
 
-    NMSAttributes get_nms4_attrs(const std::shared_ptr<ngraph::opset4::NonMaxSuppression>& nms4)
-    {
+    NMSAttributes get_nms4_attrs(const std::shared_ptr<ngraph::opset4::NonMaxSuppression>& nms4) {
         NMSAttributes attrs;
 
         attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
@@ -90,8 +84,7 @@ namespace
         return attrs;
     }
 
-    NMSAttributes get_nms3_attrs(const std::shared_ptr<ngraph::opset3::NonMaxSuppression>& nms3)
-    {
+    NMSAttributes get_nms3_attrs(const std::shared_ptr<ngraph::opset3::NonMaxSuppression>& nms3) {
         NMSAttributes attrs;
 
         attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
@@ -117,8 +110,7 @@ namespace
         return attrs;
     }
 
-    NMSAttributes get_nms1_attrs(const std::shared_ptr<ngraph::opset1::NonMaxSuppression>& nms1)
-    {
+    NMSAttributes get_nms1_attrs(const std::shared_ptr<ngraph::opset1::NonMaxSuppression>& nms1) {
         NMSAttributes attrs;
 
         attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
@@ -143,8 +135,7 @@ namespace
         return attrs;
     }
 
-    NMSAttributes get_nms_attrs(const std::shared_ptr<ngraph::Node>& root)
-    {
+    NMSAttributes get_nms_attrs(const std::shared_ptr<ngraph::Node>& root) {
         NMSAttributes attrs;
         attrs.output_type = ::ngraph::element::i64;
         attrs.box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
@@ -167,8 +158,7 @@ namespace
         return attrs;
     }
 
-    bool callback_func(pattern::Matcher &m)
-    {
+    bool callback_func(pattern::Matcher &m) {
         auto root = m.get_match_root();
 
         auto attrs = get_nms_attrs(root);
@@ -220,7 +210,7 @@ namespace
         root->output(0).replace(nms_5->output(0));
         return true;
     }
-}
+} // namespace
 
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertPreviousNMSToNMS5, "ConvertPreviousNMSToNMS5", 0);
 

From 83b14a8c3882fc82b7dc1b7fd3d1210cd45fa199 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 11:36:07 +0300
Subject: [PATCH 110/173] Code style fixes.

---
 .../convert_previous_nms_to_nms_5.cpp         | 18 ++++++------
 ngraph/core/src/op/non_max_suppression.cpp    | 29 +++++++++----------
 2 files changed, 23 insertions(+), 24 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index de2c905bbb0359..46f40ab9bd1f46 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -15,6 +15,15 @@
 #include <ngraph/rt_info.hpp>
 #include <ngraph/pattern/op/wrap_type.hpp>
 
+using NMS5BoxEncoding = ngraph::opset5::NonMaxSuppression::BoxEncodingType;
+
+struct NMSAttributes {
+    ngraph::element::Type output_type;
+    NMS5BoxEncoding box_encoding;
+    bool sort_result_descending;
+    bool is_supported_nms;
+};
+
 namespace {
     std::shared_ptr<ngraph::opset4::NonMaxSuppression> nms4_pattern() {
         auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
@@ -49,15 +58,6 @@ namespace {
         return nms;
     }
 
-    using NMS5BoxEncoding = ngraph::opset5::NonMaxSuppression::BoxEncodingType;
-
-    struct NMSAttributes {
-        ngraph::element::Type output_type;
-        NMS5BoxEncoding box_encoding;
-        bool sort_result_descending;
-        bool is_supported_nms;
-    };
-
     NMSAttributes get_nms4_attrs(const std::shared_ptr<ngraph::opset4::NonMaxSuppression>& nms4) {
         NMSAttributes attrs;
 
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 4259e2d4514482..2d581b2f2c5f46 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -611,7 +611,8 @@ op::v5::NonMaxSuppression::NonMaxSuppression(
     const op::v5::NonMaxSuppression::BoxEncodingType box_encoding,
     const bool sort_result_descending,
     const element::Type& output_type)
-    : Op({boxes, scores,
+    : Op({boxes,
+          scores,
           max_output_boxes_per_class,
           iou_threshold,
           score_threshold,
@@ -909,9 +910,7 @@ namespace
         case element::Type_t::f16:
             result = static_cast<float>(constant->cast_vector<float16>().at(0));
             break;
-        case element::Type_t::f32:
-            result = constant->cast_vector<float>().at(0);
-            break;
+        case element::Type_t::f32: result = constant->cast_vector<float>().at(0); break;
         default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
         }
         return result;
@@ -953,12 +952,12 @@ void op::v5::NonMaxSuppression::validate()
     if (inputs().size() >= 3)
     {
         const auto max_boxes_ps = get_input_partial_shape(2);
-        NODE_VALIDATION_CHECK(this,
-                              max_boxes_ps.is_dynamic() ||
-                                  is_scalar_or_1d_tensor_with_1_element(max_boxes_ps),
-                              "Expected 0D or 1D tensor for the 'max_output_boxes_per_class' input. "
-                              "Got: ",
-                              max_boxes_ps);
+        NODE_VALIDATION_CHECK(
+            this,
+            max_boxes_ps.is_dynamic() || is_scalar_or_1d_tensor_with_1_element(max_boxes_ps),
+            "Expected 0D or 1D tensor for the 'max_output_boxes_per_class' input. "
+            "Got: ",
+            max_boxes_ps);
     }
 
     if (inputs().size() >= 4)
@@ -998,7 +997,7 @@ void op::v5::NonMaxSuppression::validate()
                               "'soft_nms_sigma' input.");
         NODE_VALIDATION_CHECK(this,
                               soft_nms_sigma.is_dynamic() ||
-                                is_scalar_or_1d_tensor_with_1_element(soft_nms_sigma),
+                                  is_scalar_or_1d_tensor_with_1_element(soft_nms_sigma),
                               "Expected 0D or 1D tensor for the 'soft_nms_sigma' input. Got: ",
                               soft_nms_sigma);
     }
@@ -1057,8 +1056,8 @@ float op::v5::NonMaxSuppression::iou_threshold_from_input() const
 
     const auto iou_threshold_input =
         as_type_ptr<op::Constant>(input_value(iou_threshold_port).get_node_shared_ptr());
-    iou_threshold = float_from_constant_node(iou_threshold_input,
-                                             get_input_element_type(iou_threshold_port));
+    iou_threshold =
+        float_from_constant_node(iou_threshold_input, get_input_element_type(iou_threshold_port));
 
     return iou_threshold;
 }
@@ -1093,8 +1092,8 @@ float op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const
 
     const auto soft_nms_sigma_input =
         as_type_ptr<op::Constant>(input_value(soft_nms_sigma_port).get_node_shared_ptr());
-    soft_nms_sigma = float_from_constant_node(soft_nms_sigma_input,
-                                              get_input_element_type(soft_nms_sigma_port));
+    soft_nms_sigma =
+        float_from_constant_node(soft_nms_sigma_input, get_input_element_type(soft_nms_sigma_port));
 
     return soft_nms_sigma;
 }

From f7da96dc6d11dcb37c13da1c01912a79e2751b81 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 12:00:50 +0300
Subject: [PATCH 111/173] Small changes.

---
 .../op_conversions/convert_previous_nms_to_nms_5.cpp            | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 46f40ab9bd1f46..a1474079700065 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -24,6 +24,8 @@ struct NMSAttributes {
     bool is_supported_nms;
 };
 
+using namespace ngraph;
+
 namespace {
     std::shared_ptr<ngraph::opset4::NonMaxSuppression> nms4_pattern() {
         auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});

From 743e87bc8d7c05ccb288571dd932dca17fa64b70 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 12:26:52 +0300
Subject: [PATCH 112/173] Small changes.

---
 .../convert_previous_nms_to_nms_5.cpp              | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index a1474079700065..fc9c968f2ea686 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -80,7 +80,7 @@ namespace {
                                    " has unsupported box encoding");
         }
 
-        attrs.sort_result_descending = nms4->get_sort_result_descending(),
+        attrs.sort_result_descending = nms4->get_sort_result_descending();
         attrs.output_type = nms4->get_output_type();
 
         return attrs;
@@ -106,7 +106,7 @@ namespace {
                                    " has unsupported box encoding");
         }
 
-        attrs.sort_result_descending = nms3->get_sort_result_descending(),
+        attrs.sort_result_descending = nms3->get_sort_result_descending();
         attrs.output_type = nms3->get_output_type();
 
         return attrs;
@@ -132,7 +132,7 @@ namespace {
                                    " has unsupported box encoding");
         }
 
-        attrs.sort_result_descending = nms1->get_sort_result_descending(),
+        attrs.sort_result_descending = nms1->get_sort_result_descending();
 
         return attrs;
     }
@@ -144,15 +144,15 @@ namespace {
         attrs.sort_result_descending = false;
         attrs.is_supported_nms = false;
 
-        auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
+        auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(root);
         if (nms_4) {
             return get_nms4_attrs(nms_4);
         }
-        auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(m.get_match_root());
+        auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(root);
         if (nms_4) {
             return get_nms3_attrs(nms_3);
         }
-        auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());
+        auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(root);
         if (nms_1) {
             return get_nms1_attrs(nms_3);
         }
@@ -198,7 +198,7 @@ namespace {
                 arg3,
                 arg4,
                 arg5,
-                box_encoding,
+                attrs.box_encoding,
                 attrs.sort_result_descending,
                 attrs.output_type);
 

From db12b8ad6973281aa0980d715a720c00234f2e1a Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 12:49:15 +0300
Subject: [PATCH 113/173] Small fix.

---
 .../op_conversions/convert_previous_nms_to_nms_5.cpp            | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index fc9c968f2ea686..3ae8672519e80e 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -154,7 +154,7 @@ namespace {
         }
         auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(root);
         if (nms_1) {
-            return get_nms1_attrs(nms_3);
+            return get_nms1_attrs(nms_1);
         }
 
         return attrs;

From b38e91657320ad4ddb8a538e4d8fb911fc353989 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 13:17:33 +0300
Subject: [PATCH 114/173] Deleted commented code.

---
 .../convert_previous_nms_to_nms_5.cpp         | 21 -------------------
 1 file changed, 21 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 3ae8672519e80e..ffb05fcce69b62 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -223,13 +223,6 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
 
 ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
     auto nms = nms4_pattern();
-//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-//     auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
-//     auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
-//     auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
-//     auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-//                                                                    iou_threshold, score_threshold);
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
@@ -304,13 +297,6 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS3ToNMS5, "ConvertNMS3ToNMS5", 0);
 
 ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
     auto nms = nms3_pattern();
-//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-//     auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
-//     auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
-//     auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
-//     auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-//                                                                    iou_threshold, score_threshold);
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(m.get_match_root());
@@ -384,13 +370,6 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS1ToNMS5, "ConvertNMS1ToNMS5", 0);
 
 ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
     auto nms = nms1_pattern();
-//     auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-//     auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-//     auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
-//     auto iou_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.75});
-//     auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
-//     auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-//                                                                    iou_threshold, score_threshold);
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());

From fdbf03c02fb8a433a5b6f5aaa5c7d33b61f94420 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 13:22:26 +0300
Subject: [PATCH 115/173] Some refactoring in ConvertNMS4ToNMS5 ctor.

---
 .../convert_previous_nms_to_nms_5.cpp         | 127 +++++++++---------
 1 file changed, 64 insertions(+), 63 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index ffb05fcce69b62..0d18e152677311 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -224,69 +224,70 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
 ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
     auto nms = nms4_pattern();
 
-    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
-        auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
-        if (!nms_4) {
-            return false;
-        }
-
-        const auto new_args = nms_4->input_values();
-
-        size_t num_of_args = new_args.size();
-
-        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
-        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-
-        auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-        switch (nms_4->get_box_encoding()) {
-            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-                break;
-            case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
-                break;
-            default:
-                throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
-                                   " has unsupported box encoding");
-        }
-
-        // list of new nGraph operations
-        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
-
-        new_ops_list.push_front(arg5);
-        if (num_of_args <= 4) {
-            new_ops_list.push_front(arg4.get_node_shared_ptr());
-        }
-        if (num_of_args <= 3) {
-            new_ops_list.push_front(arg3.get_node_shared_ptr());
-        }
-        if (num_of_args <= 2) {
-            new_ops_list.push_front(arg2.get_node_shared_ptr());
-        }
-
-        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
-                new_args.at(0),
-                new_args.at(1),
-                arg2,
-                arg3,
-                arg4,
-                arg5,
-                box_encoding,
-                nms_4->get_sort_result_descending(),
-                nms_4->get_output_type());
-
-        new_ops_list.push_back(nms_5);
-
-        // vector of new nGraph operations
-        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
-
-        nms_5->set_friendly_name(nms_4->get_friendly_name());
-        ngraph::copy_runtime_info(nms_4, new_ops);
-        nms_4->output(0).replace(nms_5->output(0));
-        return true;
-    };
+    ngraph::matcher_pass_callback = callback_func;
+//     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
+//         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
+//         if (!nms_4) {
+//             return false;
+//         }
+//
+//         const auto new_args = nms_4->input_values();
+//
+//         size_t num_of_args = new_args.size();
+//
+//         const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
+//         const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+//         const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+//         const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+//
+//         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+//         switch (nms_4->get_box_encoding()) {
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
+//                 box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
+//                 break;
+//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
+//                 box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
+//                 break;
+//             default:
+//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
+//                                    " has unsupported box encoding");
+//         }
+//
+//         // list of new nGraph operations
+//         std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
+//
+//         new_ops_list.push_front(arg5);
+//         if (num_of_args <= 4) {
+//             new_ops_list.push_front(arg4.get_node_shared_ptr());
+//         }
+//         if (num_of_args <= 3) {
+//             new_ops_list.push_front(arg3.get_node_shared_ptr());
+//         }
+//         if (num_of_args <= 2) {
+//             new_ops_list.push_front(arg2.get_node_shared_ptr());
+//         }
+//
+//         const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
+//                 new_args.at(0),
+//                 new_args.at(1),
+//                 arg2,
+//                 arg3,
+//                 arg4,
+//                 arg5,
+//                 box_encoding,
+//                 nms_4->get_sort_result_descending(),
+//                 nms_4->get_output_type());
+//
+//         new_ops_list.push_back(nms_5);
+//
+//         // vector of new nGraph operations
+//         NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
+//
+//         nms_5->set_friendly_name(nms_4->get_friendly_name());
+//         ngraph::copy_runtime_info(nms_4, new_ops);
+//         nms_4->output(0).replace(nms_5->output(0));
+//         return true;
+//     };
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMS5");
     this->register_matcher(m, callback);

From 032e98bc1c500909a7705eae7d2f155ecf00c914 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 13:59:45 +0300
Subject: [PATCH 116/173] Small fix.

---
 .../op_conversions/convert_previous_nms_to_nms_5.cpp            | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 0d18e152677311..c7bb62215060ab 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -224,7 +224,7 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
 ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
     auto nms = nms4_pattern();
 
-    ngraph::matcher_pass_callback = callback_func;
+    ngraph::matcher_pass_callback callback = callback_func;
 //     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
 //         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
 //         if (!nms_4) {

From 0d739a95d56e016e26590b05dec68a16e28159ac Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 19 Oct 2020 15:27:03 +0300
Subject: [PATCH 117/173] Common part of conversions NMS-1 -> NMS-5, NMS-3 ->
 NMS-5, NMS-4 -> NMS-5 was moved into the separate function.

---
 .../convert_previous_nms_to_nms_5.cpp         | 191 +-----------------
 1 file changed, 2 insertions(+), 189 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index c7bb62215060ab..8eff19f4672b68 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -225,69 +225,6 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
     auto nms = nms4_pattern();
 
     ngraph::matcher_pass_callback callback = callback_func;
-//     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
-//         auto nms_4 = std::dynamic_pointer_cast<ngraph::opset4::NonMaxSuppression>(m.get_match_root());
-//         if (!nms_4) {
-//             return false;
-//         }
-//
-//         const auto new_args = nms_4->input_values();
-//
-//         size_t num_of_args = new_args.size();
-//
-//         const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
-//         const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-//         const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-//         const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-//
-//         auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-//         switch (nms_4->get_box_encoding()) {
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CENTER:
-//                 box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-//                 break;
-//             case ::ngraph::opset4::NonMaxSuppression::BoxEncodingType::CORNER:
-//                 box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
-//                 break;
-//             default:
-//                 throw ngraph_error("NonMaxSuppression layer " + nms_4->get_friendly_name() +
-//                                    " has unsupported box encoding");
-//         }
-//
-//         // list of new nGraph operations
-//         std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
-//
-//         new_ops_list.push_front(arg5);
-//         if (num_of_args <= 4) {
-//             new_ops_list.push_front(arg4.get_node_shared_ptr());
-//         }
-//         if (num_of_args <= 3) {
-//             new_ops_list.push_front(arg3.get_node_shared_ptr());
-//         }
-//         if (num_of_args <= 2) {
-//             new_ops_list.push_front(arg2.get_node_shared_ptr());
-//         }
-//
-//         const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
-//                 new_args.at(0),
-//                 new_args.at(1),
-//                 arg2,
-//                 arg3,
-//                 arg4,
-//                 arg5,
-//                 box_encoding,
-//                 nms_4->get_sort_result_descending(),
-//                 nms_4->get_output_type());
-//
-//         new_ops_list.push_back(nms_5);
-//
-//         // vector of new nGraph operations
-//         NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
-//
-//         nms_5->set_friendly_name(nms_4->get_friendly_name());
-//         ngraph::copy_runtime_info(nms_4, new_ops);
-//         nms_4->output(0).replace(nms_5->output(0));
-//         return true;
-//     };
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMS5");
     this->register_matcher(m, callback);
@@ -299,69 +236,7 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS3ToNMS5, "ConvertNMS3ToNMS5", 0);
 ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
     auto nms = nms3_pattern();
 
-    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
-        auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(m.get_match_root());
-        if (!nms_3) {
-            return false;
-        }
-
-        const auto new_args = nms_3->input_values();
-
-        size_t num_of_args = new_args.size();
-
-        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
-        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-
-        auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-        switch (nms_3->get_box_encoding()) {
-            case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CENTER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-                break;
-            case ::ngraph::opset3::NonMaxSuppression::BoxEncodingType::CORNER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
-                break;
-            default:
-                throw ngraph_error("NonMaxSuppression layer " + nms_3->get_friendly_name() +
-                                   " has unsupported box encoding");
-        }
-
-        // list of new nGraph operations
-        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
-
-        new_ops_list.push_front(arg5);
-        if (num_of_args <= 4) {
-            new_ops_list.push_front(arg4.get_node_shared_ptr());
-        }
-        if (num_of_args <= 3) {
-            new_ops_list.push_front(arg3.get_node_shared_ptr());
-        }
-        if (num_of_args <= 2) {
-            new_ops_list.push_front(arg2.get_node_shared_ptr());
-        }
-
-        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
-                new_args.at(0),
-                new_args.at(1),
-                arg2,
-                arg3,
-                arg4,
-                arg5,
-                box_encoding,
-                nms_3->get_sort_result_descending(),
-                nms_3->get_output_type());
-
-        new_ops_list.push_back(nms_5);
-
-        // vector of new nGraph operations
-        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
-
-        nms_5->set_friendly_name(nms_3->get_friendly_name());
-        ngraph::copy_runtime_info(nms_3, new_ops);
-        nms_3->output(0).replace(nms_5->output(0));
-        return true;
-    };
+    ngraph::matcher_pass_callback callback = callback_func;
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS3ToNMS5");
     this->register_matcher(m, callback);
@@ -372,69 +247,7 @@ NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS1ToNMS5, "ConvertNMS1ToNMS5", 0);
 ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
     auto nms = nms1_pattern();
 
-    ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
-        auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(m.get_match_root());
-        if (!nms_1) {
-            return false;
-        }
-
-        const auto new_args = nms_1->input_values();
-
-        size_t num_of_args = new_args.size();
-
-        const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
-        const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-
-        auto box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-        switch (nms_1->get_box_encoding()) {
-            case ::ngraph::opset1::NonMaxSuppression::BoxEncodingType::CENTER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER;
-                break;
-            case ::ngraph::opset1::NonMaxSuppression::BoxEncodingType::CORNER:
-                box_encoding = ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CORNER;
-                break;
-            default:
-                throw ngraph_error("NonMaxSuppression layer " + nms_1->get_friendly_name() +
-                                   " has unsupported box encoding");
-        }
-
-        // list of new nGraph operations
-        std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
-
-        new_ops_list.push_front(arg5);
-        if (num_of_args <= 4) {
-            new_ops_list.push_front(arg4.get_node_shared_ptr());
-        }
-        if (num_of_args <= 3) {
-            new_ops_list.push_front(arg3.get_node_shared_ptr());
-        }
-        if (num_of_args <= 2) {
-            new_ops_list.push_front(arg2.get_node_shared_ptr());
-        }
-
-        const auto nms_5 = std::make_shared<ngraph::op::v5::NonMaxSuppression>(
-                new_args.at(0),
-                new_args.at(1),
-                arg2,
-                arg3,
-                arg4,
-                arg5,
-                box_encoding,
-                nms_1->get_sort_result_descending(),
-                ::ngraph::element::i64);
-
-        new_ops_list.push_back(nms_5);
-
-        // vector of new nGraph operations
-        NodeVector new_ops(new_ops_list.begin(), new_ops_list.end());
-
-        nms_5->set_friendly_name(nms_1->get_friendly_name());
-        ngraph::copy_runtime_info(nms_1, new_ops);
-        nms_1->output(0).replace(nms_5->output(0));
-        return true;
-    };
+    ngraph::matcher_pass_callback callback = callback_func;
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS1ToNMS5");
     this->register_matcher(m, callback);

From 525481ba19d8f3bfc414aefd9017ab60ac5ebd56 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 20 Oct 2020 10:45:43 +0300
Subject: [PATCH 118/173] Now conversions NMS-1 -> NMS-5, NMS-3 -> NMS-5, NMS-4
 -> NMS-5 support both 2 inputs, and 5 inputs.

---
 .../samples/speech_sample/main.cpp            |  2 +-
 .../convert_previous_nms_to_nms_5.cpp         | 44 ++-----------------
 2 files changed, 5 insertions(+), 41 deletions(-)

diff --git a/inference-engine/samples/speech_sample/main.cpp b/inference-engine/samples/speech_sample/main.cpp
index 5bae50b1d0bd5b..17ce5904ba61fa 100644
--- a/inference-engine/samples/speech_sample/main.cpp
+++ b/inference-engine/samples/speech_sample/main.cpp
@@ -702,7 +702,7 @@ int main(int argc, char *argv[]) {
                 outputs.push_back(outBlobName.substr(0, pos_layer));
                 try {
                     ports.push_back(std::stoi(outBlobName.substr(pos_layer + 1)));
-                } catch (std::exception) {
+                } catch (const std::exception&) {
                     throw std::logic_error("Ports should have integer type");
                 }
             }
diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 8eff19f4672b68..5fff024ee860d0 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -27,39 +27,6 @@ struct NMSAttributes {
 using namespace ngraph;
 
 namespace {
-    std::shared_ptr<ngraph::opset4::NonMaxSuppression> nms4_pattern() {
-        auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = ngraph::opset4::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = ngraph::opset4::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<ngraph::opset4::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                       iou_threshold, score_threshold);
-        return nms;
-    }
-
-    std::shared_ptr<ngraph::opset3::NonMaxSuppression> nms3_pattern() {
-        auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = ngraph::opset3::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = ngraph::opset3::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<ngraph::opset3::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                       iou_threshold, score_threshold);
-        return nms;
-    }
-
-    std::shared_ptr<ngraph::opset1::NonMaxSuppression> nms1_pattern() {
-        auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-        auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = ngraph::opset1::Constant::create(element::i64, Shape{}, {10});
-        auto iou_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.75});
-        auto score_threshold = ngraph::opset1::Constant::create(element::f32, Shape{}, {0.7});
-        auto nms = std::make_shared<ngraph::opset1::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                       iou_threshold, score_threshold);
-        return nms;
-    }
-
     NMSAttributes get_nms4_attrs(const std::shared_ptr<ngraph::opset4::NonMaxSuppression>& nms4) {
         NMSAttributes attrs;
 
@@ -149,7 +116,7 @@ namespace {
             return get_nms4_attrs(nms_4);
         }
         auto nms_3 = std::dynamic_pointer_cast<ngraph::opset3::NonMaxSuppression>(root);
-        if (nms_4) {
+        if (nms_3) {
             return get_nms3_attrs(nms_3);
         }
         auto nms_1 = std::dynamic_pointer_cast<ngraph::opset1::NonMaxSuppression>(root);
@@ -222,8 +189,7 @@ ngraph::pass::ConvertPreviousNMSToNMS5::ConvertPreviousNMSToNMS5() {
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
 
 ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
-    auto nms = nms4_pattern();
-
+    auto nms = ngraph::pattern::wrap_type<ngraph::opset4::NonMaxSuppression>();
     ngraph::matcher_pass_callback callback = callback_func;
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS4ToNMS5");
@@ -234,8 +200,7 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS3ToNMS5, "ConvertNMS3ToNMS5", 0);
 
 ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
-    auto nms = nms3_pattern();
-
+    auto nms = ngraph::pattern::wrap_type<ngraph::opset3::NonMaxSuppression>();
     ngraph::matcher_pass_callback callback = callback_func;
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS3ToNMS5");
@@ -245,8 +210,7 @@ ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS1ToNMS5, "ConvertNMS1ToNMS5", 0);
 
 ngraph::pass::ConvertNMS1ToNMS5::ConvertNMS1ToNMS5() {
-    auto nms = nms1_pattern();
-
+    auto nms = ngraph::pattern::wrap_type<ngraph::opset1::NonMaxSuppression>();
     ngraph::matcher_pass_callback callback = callback_func;
 
     auto m = std::make_shared<ngraph::pattern::Matcher>(nms, "ConvertNMS1ToNMS5");

From 398c2e0b562e1474db1a039e7d85c5069fc360ac Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 20 Oct 2020 11:04:28 +0300
Subject: [PATCH 119/173] Now transformations NMS-1 -> NMS-5, NMS-3 -> NMS-5,
 NMS-4 -> NMS-5 are called from 'umbrella' transformation.

---
 .../op_conversions/convert_previous_nms_to_nms_5.hpp      | 8 ++++++--
 .../common_optimizations/common_optimizations.cpp         | 4 +---
 .../op_conversions/convert_previous_nms_to_nms_5.cpp      | 4 ----
 3 files changed, 7 insertions(+), 9 deletions(-)

diff --git a/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp b/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
index c3f5e1a9cfd6d0..f0296717526cae 100644
--- a/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
+++ b/inference-engine/src/transformations/include/transformations/op_conversions/convert_previous_nms_to_nms_5.hpp
@@ -39,8 +39,12 @@ class ngraph::pass::ConvertNMS4ToNMS5: public ngraph::pass::MatcherPass {
     ConvertNMS4ToNMS5();
 };
 
-class ngraph::pass::ConvertPreviousNMSToNMS5: public ngraph::pass::MatcherPass {
+class ngraph::pass::ConvertPreviousNMSToNMS5: public ngraph::pass::GraphRewrite {
 public:
     NGRAPH_RTTI_DECLARATION;
-    ConvertPreviousNMSToNMS5();
+    ConvertPreviousNMSToNMS5() {
+        add_matcher<ngraph::pass::ConvertNMS1ToNMS5>();
+        add_matcher<ngraph::pass::ConvertNMS3ToNMS5>();
+        add_matcher<ngraph::pass::ConvertNMS4ToNMS5>();
+    }
 };
diff --git a/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp b/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
index 6e09435a61650e..36ad9e14a4ceb0 100644
--- a/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
+++ b/inference-engine/src/transformations/src/transformations/common_optimizations/common_optimizations.cpp
@@ -106,9 +106,7 @@ bool ngraph::pass::CommonOptimizations::run_on_function(std::shared_ptr<ngraph::
     manager.register_pass<ngraph::pass::GroupConvolutionBackpropDataMultiplyFusion>();
     manager.register_pass<ngraph::pass::ConstantFolding>();
 
-    manager.register_pass<ngraph::pass::ConvertNMS1ToNMS5>();
-    manager.register_pass<ngraph::pass::ConvertNMS3ToNMS5>();
-    manager.register_pass<ngraph::pass::ConvertNMS4ToNMS5>();
+    manager.register_pass<ngraph::pass::ConvertPreviousNMSToNMS5>();
 
     auto fq_fusions = manager.register_pass<ngraph::pass::GraphRewrite>();
     fq_fusions->add_matcher<ngraph::pass::FakeQuantizeMulFusion>();
diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 5fff024ee860d0..e16b62daf7d22a 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -183,9 +183,6 @@ namespace {
 
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertPreviousNMSToNMS5, "ConvertPreviousNMSToNMS5", 0);
 
-ngraph::pass::ConvertPreviousNMSToNMS5::ConvertPreviousNMSToNMS5() {
-}
-
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS4ToNMS5, "ConvertNMS4ToNMS5", 0);
 
 ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
@@ -196,7 +193,6 @@ ngraph::pass::ConvertNMS4ToNMS5::ConvertNMS4ToNMS5() {
     this->register_matcher(m, callback);
 }
 
-
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS3ToNMS5, "ConvertNMS3ToNMS5", 0);
 
 ngraph::pass::ConvertNMS3ToNMS5::ConvertNMS3ToNMS5() {

From d1225554ffe0db2a975f610fc6e46b8f29be69a1 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 20 Oct 2020 13:07:33 +0300
Subject: [PATCH 120/173] Now the transformation ConvertNMS5ToLegacyMatcher
 supports NMS-5 with 2, 3, 4, 5, or 6 inputs.

---
 .../convert_nms_5_to_legacy.cpp               | 38 ++++++++-----------
 1 file changed, 16 insertions(+), 22 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 7d89d43327fe05..54ee510421479f 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -10,6 +10,7 @@
 #include <ngraph/opsets/opset5.hpp>
 #include <legacy/ngraph_ops/nms_ie.hpp>
 #include <ngraph/rt_info.hpp>
+#include <ngraph/pattern/op/wrap_type.hpp>
 #include <transformations/utils/utils.hpp>
 
 #include "legacy/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.hpp"
@@ -17,14 +18,7 @@
 NGRAPH_RTTI_DEFINITION(ngraph::pass::ConvertNMS5ToLegacyMatcher, "ConvertNMS5ToLegacyMatcher", 0);
 
 ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
-    auto boxes = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1000, 4});
-    auto scores = std::make_shared<pattern::op::Label>(element::f32, Shape{1, 1, 1000});
-    auto max_output_boxes_per_class = ngraph::opset5::Constant::create(element::i64, Shape{}, {10});
-    auto iou_threshold = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.75});
-    auto score_threshold = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.7});
-    auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
-    auto nms = std::make_shared<ngraph::opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
-                                                                   iou_threshold, score_threshold, soft_nms_sigma);
+    auto nms = ngraph::pattern::wrap_type<ngraph::opset5::NonMaxSuppression>();
 
     ngraph::matcher_pass_callback callback = [](pattern::Matcher &m) {
         auto nms_5 = std::dynamic_pointer_cast<ngraph::opset5::NonMaxSuppression>(m.get_match_root());
@@ -100,21 +94,21 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
                 throw ngraph_error("NonMaxSuppression layer " + nms_5->get_friendly_name() +
                                    " has unsupported box encoding");
         }
-         const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
-                 new_args.at(0),
-                 new_args.at(1),
-                 new_max_per_class,
-                 new_iou_threshold,
-                 new_score_threshold,
-                 new_soft_nms_sigma,
-                 center_point_box,
-                 nms_5->get_sort_result_descending(),
-                 nms_5->get_output_type());
-         new_ops.push_back(nms_legacy);
+        const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
+                new_args.at(0),
+                new_args.at(1),
+                new_max_per_class,
+                new_iou_threshold,
+                new_score_threshold,
+                new_soft_nms_sigma,
+                center_point_box,
+                nms_5->get_sort_result_descending(),
+                nms_5->get_output_type());
+        new_ops.push_back(nms_legacy);
 
-         nms_legacy->set_friendly_name(nms_5->get_friendly_name());
-         ngraph::copy_runtime_info(nms_5, new_ops);
-         ngraph::replace_node(nms_5, nms_legacy);
+        nms_legacy->set_friendly_name(nms_5->get_friendly_name());
+        ngraph::copy_runtime_info(nms_5, new_ops);
+        ngraph::replace_node(nms_5, nms_legacy);
         return true;
     };
 

From a61629e25fc21648d76a7da6183cf0315eea44d2 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 21 Oct 2020 09:38:18 +0300
Subject: [PATCH 121/173] The transformation ConvertNMS5ToLegacyMatcher was
 rewritten using Reshape instead of Unsqueeze.

---
 .../convert_nms_5_to_legacy.cpp               | 76 +++++++------------
 1 file changed, 26 insertions(+), 50 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 54ee510421479f..d3344f567244dd 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -27,61 +27,37 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         }
 
         const auto new_args = nms_5->input_values();
-        const auto& arg2 = new_args.size() > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
-        const auto& arg3 = new_args.size() > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg4 = new_args.size() > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg5 = new_args.size() > 5 ? new_args.at(5) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-
-        const auto max_output_boxes_per_class_rank = arg2.get_partial_shape().rank();
-        const auto iou_threshold_rank = arg3.get_partial_shape().rank();
-        const auto score_threshold_rank = arg4.get_partial_shape().rank();
-        const auto soft_nms_sigma_rank = arg5.get_partial_shape().rank();
-
-        // Check that required ranks are not dynamic
-        if (max_output_boxes_per_class_rank.is_dynamic() ||
-            iou_threshold_rank.is_dynamic() ||
-            score_threshold_rank.is_dynamic() ||
-            soft_nms_sigma_rank.is_dynamic()) {
-            return false;
-        }
+        const std::size_t num_of_inputs = new_args.size();
 
-        if (max_output_boxes_per_class_rank.get_length() == 1 &&
-            iou_threshold_rank.get_length() == 1 &&
-            score_threshold_rank.get_length() == 1 &&
-            soft_nms_sigma_rank.get_length() == 1) {
-            return false;
-        }
+        const auto& arg2 = num_of_inputs > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
+        const auto& arg3 = num_of_inputs > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg4 = num_of_inputs > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+        const auto& arg5 = num_of_inputs > 5 ? new_args.at(5) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
 
         // vector of new nGraph operations
         NodeVector new_ops;
 
-        auto new_max_per_class = arg2;
-        if (max_output_boxes_per_class_rank.get_length() == 0) {
-            // WA: we need to create Constant manually because it requires by NMS shape inference
-            //     otherwise we will get dynamic shape until first CF is executed. It can be resolved
-            //     if CF will be executed right after transformation and before Validate pass.
-            if (auto new_max_per_class_const = std::dynamic_pointer_cast<opset1::Constant>(new_max_per_class.get_node_shared_ptr())) {
-                new_max_per_class = opset1::Constant::create(element::i64, Shape{1}, new_max_per_class_const->cast_vector<int64_t>());
-            } else {
-                new_max_per_class = std::make_shared<ngraph::op::Unsqueeze>(arg2, opset1::Constant::create(element::i64, Shape{1}, {0}));
-                new_ops.push_back(new_max_per_class.get_node_shared_ptr());
-            }
-        }
-        auto new_iou_threshold = arg3;
-        if (iou_threshold_rank.get_length() == 0) {
-            new_iou_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg3, opset1::Constant::create(element::i64, Shape{1}, {0}));
-            new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
-        }
-        auto new_score_threshold = arg4;
-        if (score_threshold_rank.get_length() == 0) {
-            new_score_threshold = std::make_shared<ngraph::op::Unsqueeze>(arg4, opset1::Constant::create(element::i64, Shape{1}, {0}));
-            new_ops.push_back(new_score_threshold.get_node_shared_ptr());
-        }
-        auto new_soft_nms_sigma = arg5;
-        if (soft_nms_sigma_rank.get_length() == 0) {
-            new_soft_nms_sigma = std::make_shared<ngraph::op::Unsqueeze>(arg5, opset1::Constant::create(element::i64, Shape{1}, {0}));
-            new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
-        }
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<ngraph::op::Reshape>(arg2,
+                                                                       opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                                one_dim_shape), true);
+        new_ops.push_back(new_max_per_class.get_node_shared_ptr());
+
+        auto new_iou_threshold = std::make_shared<ngraph::op::Reshape>(arg3,
+                                                                       opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                                one_dim_shape), true);
+        new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
+
+        auto new_score_threshold = std::make_shared<ngraph::op::Reshape>(arg4,
+                                                                         opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                                  one_dim_shape), true);
+        new_ops.push_back(new_score_threshold.get_node_shared_ptr());
+
+        auto new_soft_nms_sigma = std::make_shared<ngraph::op::Reshape>(arg5,
+                                                                        opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                                 one_dim_shape), true);
+        new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
+
         int center_point_box = 0;
         switch (nms_5->get_box_encoding()) {
             case ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER:

From 37b7b11f98e79635e12298180c0aaeabe15f16f7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 21 Oct 2020 10:23:50 +0300
Subject: [PATCH 122/173] Started to rewrite tests for the transformation
 ConvertNMS5ToLegacyMatcher.

---
 .../transformations/convert_nms5_test.cpp     | 117 +++---------------
 1 file changed, 17 insertions(+), 100 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 6567330464f31c..391d4953654fa4 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -46,18 +46,26 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
     {
         auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
         auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{1}, {10});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
-        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, max_output_boxes_per_class,
-                                                              std::make_shared<opset5::Unsqueeze>(iou_threshold,
-                                                                                                  opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                              std::make_shared<opset5::Unsqueeze>(score_threshold,
-                                                                                                  opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                              std::make_shared<opset5::Unsqueeze>(soft_nms_sigma,
-                                                                                                  opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                              0, true);
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                       opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                                one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -67,94 +75,3 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
     auto res = compare_functions(f, f_ref);
     ASSERT_TRUE(res.first) << res.second;
 }
-
-TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1) {
-    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-    {
-         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
-         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
-         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
-         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
-         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
-         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
-
-         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-
-         pass::Manager manager;
-         manager.register_pass<pass::InitNodeInfo>();
-         manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
-         manager.run_passes(f);
-         f->validate_nodes_and_infer_types();
-         ASSERT_NO_THROW(check_rt_info(f));
-    }
-
-     {
-         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
-         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
-         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{1}, {10});
-         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
-         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
-         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, max_output_boxes_per_class,
-                                                               std::make_shared<opset5::Unsqueeze>(iou_threshold,
-                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                               std::make_shared<opset5::Unsqueeze>(score_threshold,
-                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                               std::make_shared<opset5::Unsqueeze>(soft_nms_sigma,
-                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                               0, true, element::i32);
-
-         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-     }
-
-     auto res = compare_functions(f, f_ref);
-     ASSERT_TRUE(res.first) << res.second;
-}
-
-TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2) {
-    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
-    {
-         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
-         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
-         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
-         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
-         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
-                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true, element::i32);
-
-         f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-
-         pass::Manager manager;
-         manager.register_pass<pass::InitNodeInfo>();
-         manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
-         manager.run_passes(f);
-
-         f->validate_nodes_and_infer_types();
-         ASSERT_NO_THROW(check_rt_info(f));
-    }
-
-    {
-         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
-         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{1}, {10});
-         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
-         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
-         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, max_output_boxes_per_class,
-                                                               std::make_shared<opset5::Unsqueeze>(iou_threshold,
-                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                               std::make_shared<opset5::Unsqueeze>(score_threshold,
-                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                               std::make_shared<opset5::Unsqueeze>(soft_nms_sigma,
-                                                                                                   opset5::Constant::create(element::i64, Shape{1}, {0})),
-                                                               0, true, element::i32);
-
-         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-    }
-
-    auto res = compare_functions(f, f_ref);
-    ASSERT_TRUE(res.first) << res.second;
-}

From 8fcf0469c86d5e34abe1ed2241228bf41d50f770 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 21 Oct 2020 11:02:57 +0300
Subject: [PATCH 123/173] Some fixes.

---
 .../convert_nms_5_to_legacy.cpp               | 30 +++++++++++--------
 .../transformations/convert_nms5_test.cpp     |  4 +--
 2 files changed, 20 insertions(+), 14 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index d3344f567244dd..8efb347abe03e7 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -38,24 +38,30 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         NodeVector new_ops;
 
         auto one_dim_shape = Shape{1};
-        auto new_max_per_class = std::make_shared<ngraph::op::Reshape>(arg2,
-                                                                       opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                                one_dim_shape), true);
+
+        auto new_max_per_class = arg2;
+        auto new_iou_threshold = arg3;
+        auto new_score_threshold = arg4;
+        auto new_soft_nms_sigma = arg5;
+
+        new_max_per_class = std::make_shared<ngraph::op::Reshape>(arg2,
+                                                                  opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                           one_dim_shape), true);
         new_ops.push_back(new_max_per_class.get_node_shared_ptr());
 
-        auto new_iou_threshold = std::make_shared<ngraph::op::Reshape>(arg3,
-                                                                       opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                                one_dim_shape), true);
+        new_iou_threshold = std::make_shared<ngraph::op::Reshape>(arg3,
+                                                                  opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                           one_dim_shape), true);
         new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
 
-        auto new_score_threshold = std::make_shared<ngraph::op::Reshape>(arg4,
-                                                                         opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                                  one_dim_shape), true);
+        new_score_threshold = std::make_shared<ngraph::op::Reshape>(arg4,
+                                                                    opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
         new_ops.push_back(new_score_threshold.get_node_shared_ptr());
 
-        auto new_soft_nms_sigma = std::make_shared<ngraph::op::Reshape>(arg5,
-                                                                        opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                                 one_dim_shape), true);
+        new_soft_nms_sigma = std::make_shared<ngraph::op::Reshape>(arg5,
+                                                                   opset1::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
         new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
 
         int center_point_box = 0;
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 391d4953654fa4..b6b8883313546a 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -53,8 +53,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
-                                                                       opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                                one_dim_shape), true);
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
         auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                             one_dim_shape), true);

From 33fe2f0d4ed06c6ca9fd3164eb5839ec7a61d28f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 21 Oct 2020 12:02:14 +0300
Subject: [PATCH 124/173] Small fixes.

---
 .../convert_nms_5_to_legacy.cpp               | 31 +++++++++----------
 1 file changed, 14 insertions(+), 17 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index 8efb347abe03e7..dae4adaf0510ba 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -39,29 +39,26 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
 
         auto one_dim_shape = Shape{1};
 
-        auto new_max_per_class = arg2;
-        auto new_iou_threshold = arg3;
-        auto new_score_threshold = arg4;
-        auto new_soft_nms_sigma = arg5;
-
-        new_max_per_class = std::make_shared<ngraph::op::Reshape>(arg2,
-                                                                  opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                           one_dim_shape), true);
+        Output<Node> new_max_per_class;
+        Output<Node> new_iou_threshold;
+        Output<Node> new_score_threshold;
+        Output<Node> new_soft_nms_sigma;
+
+        Output<Node> new_shape_for_max_per_class = opset1::Constant::create(ngraph::element::i64, Shape{1}, {1});
+        Output<Node> new_shape_for_iou_threshold = opset1::Constant::create(ngraph::element::i64, Shape{1}, {1});
+        Output<Node> new_shape_for_score_threshold = opset1::Constant::create(ngraph::element::i64, Shape{1}, {1});
+        Output<Node> new_shape_for_soft_nms_sigma = opset1::Constant::create(ngraph::element::i64, Shape{1}, {1});
+
+        new_max_per_class = std::make_shared<opset1::Reshape>(arg2, new_shape_for_max_per_class, true);
         new_ops.push_back(new_max_per_class.get_node_shared_ptr());
 
-        new_iou_threshold = std::make_shared<ngraph::op::Reshape>(arg3,
-                                                                  opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                           one_dim_shape), true);
+        new_iou_threshold = std::make_shared<opset1::Reshape>(arg3, new_shape_for_iou_threshold, true);
         new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
 
-        new_score_threshold = std::make_shared<ngraph::op::Reshape>(arg4,
-                                                                    opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
+        new_score_threshold = std::make_shared<opset1::Reshape>(arg4, new_shape_for_score_threshold, true);
         new_ops.push_back(new_score_threshold.get_node_shared_ptr());
 
-        new_soft_nms_sigma = std::make_shared<ngraph::op::Reshape>(arg5,
-                                                                   opset1::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                            one_dim_shape), true);
+        new_soft_nms_sigma = std::make_shared<opset1::Reshape>(arg5, new_shape_for_soft_nms_sigma, true);
         new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
 
         int center_point_box = 0;

From 3fbcb28b8bca8bc30eb791892312c6b7189d2d6d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 21 Oct 2020 16:50:21 +0300
Subject: [PATCH 125/173] Corrected tests for the transformation NMS-5 ->
 NMSIE3.

---
 .../src/legacy_api/src/ngraph_ops/nms_ie.cpp  |   2 +-
 .../transformations/convert_nms5_test.cpp     | 766 +++++++++++++++++-
 2 files changed, 761 insertions(+), 7 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
index 4e838696811e9b..347e939505290b 100644
--- a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
@@ -140,7 +140,7 @@ int64_t op::NonMaxSuppressionIE3::max_boxes_output_from_input() const {
     int64_t max_output_boxes{0};
 
     const auto max_output_boxes_input =
-        as_type_ptr<op::Constant>(input_value(2).get_node_shared_ptr());
+        as_type_ptr<op::Constant>(input_value(max_output_boxes_per_class_port).get_node_shared_ptr());
     max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
 
     return max_output_boxes;
diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index b6b8883313546a..29cb46fe0fd28a 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -20,7 +20,7 @@
 using namespace testing;
 using namespace ngraph;
 
-TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
+TEST(TransformationTests, ConvertNMS5ToNMSIEStaticSixInputs) {
     std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
     {
         auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
@@ -32,20 +32,305 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+        ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+        ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+        ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        ASSERT_NO_THROW(check_rt_info(f));
+        ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
 
-        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
         manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
-        ASSERT_TRUE(f->get_output_partial_shape(0).is_static()) << "Shape " << f->get_output_partial_shape(0) << " should be static";
+        ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
     }
 
     {
         auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
         auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
+        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1SixInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
@@ -68,8 +353,477 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStatic) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
-        ASSERT_TRUE(f_ref->get_output_partial_shape(0).is_static()) << "Shape " << f_ref->get_output_partial_shape(0) << " should be static";
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FiveInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FourInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1ThreeInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1TwoInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2SixInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.25});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FiveInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FourInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2ThreeInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+    }
+
+    auto res = compare_functions(f, f_ref);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2TwoInputs) {
+    std::shared_ptr<Function> f(nullptr), f_ref(nullptr);
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores,
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+
+        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+        pass::Manager manager;
+        manager.register_pass<pass::InitNodeInfo>();
+        manager.register_pass<pass::ConvertNMS5ToLegacyMatcher>();
+        manager.run_passes(f);
+        f->validate_nodes_and_infer_types();
+        ASSERT_NO_THROW(check_rt_info(f));
+    }
+
+    {
+        auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
+        auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
+        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
+
+        auto one_dim_shape = Shape{1};
+        auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_iou_threshold = std::make_shared<opset5::Reshape>(iou_threshold,
+                                                                   opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                            one_dim_shape), true);
+        auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
+                                                                     opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                              one_dim_shape), true);
+        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
+                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
+                                                                                             one_dim_shape), true);
+        auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
+                                                              new_soft_nms_sigma, 0, true);
+        nms->set_friendly_name("nms");
+
+        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);

From 707666afc8b24ab3fe1fdadf972725fe5c3ef044 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 09:51:54 +0300
Subject: [PATCH 126/173] Small formatting fix.

---
 ngraph/python/src/ngraph/opset5/ops.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/ngraph/python/src/ngraph/opset5/ops.py b/ngraph/python/src/ngraph/opset5/ops.py
index 62a233755671ca..01a62111ff52f8 100644
--- a/ngraph/python/src/ngraph/opset5/ops.py
+++ b/ngraph/python/src/ngraph/opset5/ops.py
@@ -140,6 +140,7 @@ def non_max_suppression(
 
     return _get_node_factory_opset5().create("NonMaxSuppression", inputs, attributes)
 
+
 @nameable_op
 def round(data: NodeInput, mode: str = "half_to_even", name: Optional[str] = None) -> Node:
     """Apply Round operation on each element of input tensor.

From cdb77fa6344c61dc2b2d54fbeab018c8121b1857 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 10:29:01 +0300
Subject: [PATCH 127/173] Now methods max_boxes_output_from_input(),
 iou_threshold_from_input(), score_threshold_from_input(),
 soft_nms_sigma_from_input() of op::v5::NunMaxSuppression are public.

---
 ngraph/core/include/ngraph/op/non_max_suppression.hpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 817e8128242c63..ffab830ad19bb5 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -376,15 +376,15 @@ namespace ngraph
                 }
                 using Node::set_output_type;
 
+                int64_t max_boxes_output_from_input() const;
+                float iou_threshold_from_input() const;
+                float score_threshold_from_input() const;
+                float soft_nms_sigma_from_input() const;
             protected:
                 BoxEncodingType m_box_encoding = BoxEncodingType::CORNER;
                 bool m_sort_result_descending = true;
                 ngraph::element::Type m_output_type = ngraph::element::i64;
                 void validate();
-                int64_t max_boxes_output_from_input() const;
-                float iou_threshold_from_input() const;
-                float score_threshold_from_input() const;
-                float soft_nms_sigma_from_input() const;
             };
         } // namespace v5
     }     // namespace op

From 3ab28d3388767cfe06597457e49992a8a27ce59f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 10:31:15 +0300
Subject: [PATCH 128/173] Started to move op::v5::NonMaxSuppression::evaluate()
 into ngraph/test/runtime/interpreter.

---
 ngraph/test/runtime/interpreter/int_executable.hpp | 13 +++++++++++++
 ngraph/test/runtime/interpreter/opset_int_tbl.hpp  |  1 +
 2 files changed, 14 insertions(+)

diff --git a/ngraph/test/runtime/interpreter/int_executable.hpp b/ngraph/test/runtime/interpreter/int_executable.hpp
index be993f6c98033e..7a1afbca25c14c 100644
--- a/ngraph/test/runtime/interpreter/int_executable.hpp
+++ b/ngraph/test/runtime/interpreter/int_executable.hpp
@@ -70,6 +70,7 @@
 #include "ngraph/runtime/reference/max_pool.hpp"
 #include "ngraph/runtime/reference/min.hpp"
 #include "ngraph/runtime/reference/negate.hpp"
+#include "ngraph/runtime/reference/non_max_suppression.hpp"
 #include "ngraph/runtime/reference/normalize_l2.hpp"
 #include "ngraph/runtime/reference/not.hpp"
 #include "ngraph/runtime/reference/one_hot.hpp"
@@ -1444,6 +1445,18 @@ class INTERPRETER_BACKEND_API ngraph::runtime::interpreter::INTExecutable : publ
                                        norm->get_eps_mode());
             break;
         }
+        case OP_TYPEID::NonMaxSuppression_v5:
+        {
+            const op::NonMaxSuppression* nms =
+                static_cast<const op::v5::NonMaxSuppression*>(&node);
+
+            int64_t max_output_boxes_per_class = nms->max_boxes_output_from_input();
+            float iou_threshold = nms->iou_threshold_from_input();
+            float score_threshold = nms->score_threshold_from_input();
+            float soft_nms_sigma = nms->soft_nms_sigma_from_input();
+
+            break;
+        }
 
         // Fused Ops are not supported in interpreter. They need to be decomposed before execution
         case OP_TYPEID::DepthToSpace:
diff --git a/ngraph/test/runtime/interpreter/opset_int_tbl.hpp b/ngraph/test/runtime/interpreter/opset_int_tbl.hpp
index f9e1ee46126810..9db0f38a0f4d39 100644
--- a/ngraph/test/runtime/interpreter/opset_int_tbl.hpp
+++ b/ngraph/test/runtime/interpreter/opset_int_tbl.hpp
@@ -59,4 +59,5 @@ NGRAPH_OP(GRUSequence, op::v5)
 NGRAPH_OP(RNNSequence, op::v5)
 NGRAPH_OP(Round, op::v5)
 NGRAPH_OP(LogSoftmax, op::v5)
+NGRAPH_OP(NonMaxSuppression, op::v5)
 #undef ID_SUFFIX

From a0a562eef5261570e4a708ff959429a115579de8 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 12:34:09 +0300
Subject: [PATCH 129/173] Added NMS-5 to
 ngraph/test/runtime/interpreter/int_executable.

---
 .../runtime/interpreter/int_executable.cpp    | 307 ++++++++++++++++++
 .../runtime/interpreter/int_executable.hpp    |  13 +-
 2 files changed, 314 insertions(+), 6 deletions(-)

diff --git a/ngraph/test/runtime/interpreter/int_executable.cpp b/ngraph/test/runtime/interpreter/int_executable.cpp
index 9b54f73215f7c2..b44b4ad78b5117 100644
--- a/ngraph/test/runtime/interpreter/int_executable.cpp
+++ b/ngraph/test/runtime/interpreter/int_executable.cpp
@@ -21,12 +21,16 @@
 #include "ngraph/op/util/op_types.hpp"
 #include "ngraph/ops.hpp"
 #include "ngraph/pass/manager.hpp"
+#include "ngraph/type/bfloat16.hpp"
+#include "ngraph/type/float16.hpp"
 #include "ngraph/util.hpp"
 #include "pass/fused_op_decomposition.hpp"
 #include "pass/like_replacement.hpp"
 #include "pass/liveness.hpp"
 #include "pass/opset0_downgrade.hpp"
 #include "pass/opset1_downgrade.hpp"
+#include <algorithm>
+#include <cstring>
 
 using namespace std;
 using namespace ngraph;
@@ -55,6 +59,309 @@ runtime::interpreter::OP_TYPEID runtime::interpreter::INTExecutable::get_typeid(
     return rc;
 }
 
+using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
+
+namespace
+{
+    constexpr size_t boxes_port = 0;
+    constexpr size_t scores_port = 1;
+    constexpr size_t max_output_boxes_port = 2;
+    constexpr size_t iou_threshold_port = 3;
+    constexpr size_t score_threshold_port = 4;
+    constexpr size_t soft_nms_sigma_port = 5;
+
+    PartialShape infer_selected_indices_shape(
+        const std::vector<std::shared_ptr<HostTensor>>& inputs,
+        int64_t max_output_boxes_per_class)
+    {
+        const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
+        const auto scores_ps = inputs[scores_port]->get_partial_shape();
+
+        // NonMaxSuppression produces triplets
+        // that have the following format: [batch_index, class_index, box_index]
+        PartialShape result = {Dimension::dynamic(), 3};
+
+        if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
+        {
+            const auto num_boxes_boxes = boxes_ps[1];
+            if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
+            {
+                const auto num_boxes = num_boxes_boxes.get_length();
+                const auto num_classes = scores_ps[1].get_length();
+
+                result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
+                            scores_ps[0].get_length();
+            }
+        }
+
+        return result;
+    }
+
+    void normalize_corner(float* boxes, const Shape& boxes_shape)
+    {
+        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+        for (size_t i = 0; i < total_num_of_boxes; ++i)
+        {
+            float* current_box = boxes + 4 * i;
+
+            float y1 = current_box[0];
+            float x1 = current_box[1];
+            float y2 = current_box[2];
+            float x2 = current_box[3];
+
+            float ymin = std::min(y1, y2);
+            float ymax = std::max(y1, y2);
+            float xmin = std::min(x1, x2);
+            float xmax = std::max(x1, x2);
+
+            current_box[0] = ymin;
+            current_box[1] = xmin;
+            current_box[2] = ymax;
+            current_box[3] = xmax;
+        }
+    }
+
+    void normalize_center(float* boxes, const Shape& boxes_shape)
+    {
+        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+        for (size_t i = 0; i < total_num_of_boxes; ++i)
+        {
+            float* current_box = boxes + 4 * i;
+
+            float x_center = current_box[0];
+            float y_center = current_box[1];
+            float width = current_box[2];
+            float height = current_box[3];
+
+            float y1 = y_center - height / 2.0;
+            float x1 = x_center - width / 2.0;
+            float y2 = y_center + height / 2.0;
+            float x2 = x_center + width / 2.0;
+
+            current_box[0] = y1;
+            current_box[1] = x1;
+            current_box[2] = y2;
+            current_box[3] = x2;
+        }
+    }
+
+    void normalize_box_encoding(float* boxes,
+                                const Shape& boxes_shape,
+                                const V5BoxEncoding box_encoding)
+    {
+        if (box_encoding == V5BoxEncoding::CORNER)
+        {
+            normalize_corner(boxes, boxes_shape);
+        }
+        else
+        {
+            normalize_center(boxes, boxes_shape);
+        }
+    }
+
+    std::vector<float> get_floats(const std::shared_ptr<HostTensor>& input, const Shape& shape)
+    {
+        size_t input_size = shape_size(shape);
+        std::vector<float> result(input_size);
+
+        switch (input->get_element_type())
+        {
+        case element::Type_t::bf16:
+        {
+            bfloat16* p = input->get_data_ptr<bfloat16>();
+            for (size_t i = 0; i < input_size; ++i)
+            {
+                result[i] = float(p[i]);
+            }
+        }
+        break;
+        case element::Type_t::f16:
+        {
+            float16* p = input->get_data_ptr<float16>();
+            for (size_t i = 0; i < input_size; ++i)
+            {
+                result[i] = float(p[i]);
+            }
+        }
+        break;
+        case element::Type_t::f32:
+        {
+            float* p = input->get_data_ptr<float>();
+            memcpy(result.data(), p, input_size * sizeof(float));
+        }
+        break;
+        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
+        }
+
+        return result;
+    }
+
+    std::vector<float> prepare_boxes_data(const std::shared_ptr<HostTensor>& boxes,
+                                          const Shape& boxes_shape,
+                                          const V5BoxEncoding box_encoding)
+    {
+        auto result = get_floats(boxes, boxes_shape);
+        normalize_box_encoding(result.data(), boxes_shape, box_encoding);
+        return result;
+    }
+
+    std::vector<float> prepare_scores_data(const std::shared_ptr<HostTensor>& scores,
+                                           const Shape& scores_shape)
+    {
+        auto result = get_floats(scores, scores_shape);
+        return result;
+    }
+
+    void nms_postprocessing(const std::vector<std::shared_ptr<HostTensor>>& outputs,
+                            const ngraph::element::Type output_type,
+                            const std::vector<int64_t>& selected_indices,
+                            const std::vector<float>& selected_scores,
+                            int64_t valid_outputs,
+                            const ngraph::element::Type selected_scores_type)
+    {
+        size_t num_of_outputs = outputs.size();
+        size_t selected_size = valid_outputs * 3;
+
+        if (output_type == ngraph::element::i64)
+        {
+            int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
+            memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
+        }
+        else
+        {
+            int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
+            for (size_t i = 0; i < selected_size; ++i)
+            {
+                indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
+            }
+        }
+
+        if (num_of_outputs < 2)
+        {
+            return;
+        }
+
+        size_t selected_scores_size = selected_scores.size();
+
+        switch (selected_scores_type)
+        {
+        case element::Type_t::bf16:
+        {
+            bfloat16* scores_ptr = outputs[1]->get_data_ptr<bfloat16>();
+            for (size_t i = 0; i < selected_scores_size; ++i)
+            {
+                scores_ptr[i] = bfloat16(selected_scores[i]);
+            }
+        }
+        break;
+        case element::Type_t::f16:
+        {
+            float16* scores_ptr = outputs[1]->get_data_ptr<float16>();
+            for (size_t i = 0; i < selected_scores_size; ++i)
+            {
+                scores_ptr[i] = float16(selected_scores[i]);
+            }
+        }
+        break;
+        case element::Type_t::f32:
+        {
+            float* scores_ptr = outputs[1]->get_data_ptr<float>();
+            memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
+        }
+        break;
+        default:;
+        }
+
+        if (num_of_outputs < 3)
+        {
+            return;
+        }
+
+        if (output_type == ngraph::element::i64)
+        {
+            int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
+            *valid_outputs_ptr = valid_outputs;
+        }
+        else
+        {
+            int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
+            *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
+        }
+    }
+}
+
+void runtime::interpreter::run_nms5(
+    const op::v5::NonMaxSuppression* nms5,
+    const std::vector<std::shared_ptr<HostTensor>>& outputs,
+    const std::vector<std::shared_ptr<HostTensor>>& inputs)
+{
+    int64_t max_output_boxes_per_class = nms->max_boxes_output_from_input();
+    float iou_threshold = nms->iou_threshold_from_input();
+    float score_threshold = nms->score_threshold_from_input();
+    float soft_nms_sigma = nms->soft_nms_sigma_from_input();
+
+    auto selected_indices_shape = infer_selected_indices_shape(inputs,
+                                                               max_output_boxes_per_class);
+    Shape out_shape = selected_indices_shape.to_shape();
+
+    Shape boxes_shape = inputs[boxes_port]->get_shape();
+    Shape scores_shape = inputs[scores_port]->get_shape();
+
+    auto boxes_data = prepare_boxes_data(inputs[boxes_port], boxes_shape, nms5->get_box_encoding());
+    auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
+
+    size_t out_shape_size = shape_size(out_shape);
+
+    std::vector<int64_t> selected_indices(out_shape_size);
+    std::vector<float> selected_scores(out_shape_size);
+    int64_t valid_outputs = 0;
+
+    runtime::reference::non_max_suppression(boxes_data.data(),
+                                            boxes_shape,
+                                            scores_data.data(),
+                                            scores_shape,
+                                            max_output_boxes_per_class,
+                                            iou_threshold,
+                                            score_threshold,
+                                            soft_nms_sigma,
+                                            selected_indices.data(),
+                                            out_shape,
+                                            selected_scores.data(),
+                                            out_shape,
+                                            &valid_outputs,
+                                            nms5->get_sort_result_descending());
+
+    auto output_type = nms5->get_output_type();
+
+    outputs[0]->set_element_type(output_type);
+    outputs[0]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
+
+    size_t num_of_outputs = outputs.size();
+
+    size_t num_of_outputs = outputs.size();
+
+    auto selected_scores_type = (inputs.size() < 4) ? element::f32 : inputs[3]->get_element_type();
+
+    if (num_of_outputs >= 2)
+    {
+        outputs[1]->set_element_type(selected_scores_type);
+        outputs[1]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
+    }
+
+    if (num_of_outputs >= 3)
+    {
+        outputs[2]->set_element_type(output_type);
+        outputs[2]->set_shape(Shape{1});
+    }
+
+    nms_postprocessing(outputs,
+                       output_type,
+                       selected_indices,
+                       selected_scores,
+                       valid_outputs,
+                       selected_scores_type);
+}
+
 runtime::interpreter::INTExecutable::INTExecutable(const shared_ptr<Function>& function,
                                                    bool enable_performance_collection)
     : m_is_compiled{true}
diff --git a/ngraph/test/runtime/interpreter/int_executable.hpp b/ngraph/test/runtime/interpreter/int_executable.hpp
index 7a1afbca25c14c..f5571b2d894313 100644
--- a/ngraph/test/runtime/interpreter/int_executable.hpp
+++ b/ngraph/test/runtime/interpreter/int_executable.hpp
@@ -129,6 +129,11 @@ namespace ngraph
 #undef NGRAPH_OP
                 UnknownOp
             };
+
+            void run_nms5(const op::v5::NonMaxSuppression* nms5,
+                          const std::vector<std::shared_ptr<HostTensor>>& out,
+                          const std::vector<std::shared_ptr<HostTensor>>& args);
+
         } // namespace interpreter
     }     // namespace runtime
 } // namespace ngraph
@@ -1447,14 +1452,10 @@ class INTERPRETER_BACKEND_API ngraph::runtime::interpreter::INTExecutable : publ
         }
         case OP_TYPEID::NonMaxSuppression_v5:
         {
-            const op::NonMaxSuppression* nms =
+            const op::v5::NonMaxSuppression* nms =
                 static_cast<const op::v5::NonMaxSuppression*>(&node);
 
-            int64_t max_output_boxes_per_class = nms->max_boxes_output_from_input();
-            float iou_threshold = nms->iou_threshold_from_input();
-            float score_threshold = nms->score_threshold_from_input();
-            float soft_nms_sigma = nms->soft_nms_sigma_from_input();
-
+            run_nms5(nms, out, args);
             break;
         }
 

From 6761c749ca3e707ce8b73567c2e22388f0d0e009 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 12:47:43 +0300
Subject: [PATCH 130/173] Small fixes.

---
 ngraph/test/runtime/interpreter/int_executable.cpp | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/ngraph/test/runtime/interpreter/int_executable.cpp b/ngraph/test/runtime/interpreter/int_executable.cpp
index b44b4ad78b5117..034a06cf9ae56b 100644
--- a/ngraph/test/runtime/interpreter/int_executable.cpp
+++ b/ngraph/test/runtime/interpreter/int_executable.cpp
@@ -295,10 +295,10 @@ void runtime::interpreter::run_nms5(
     const std::vector<std::shared_ptr<HostTensor>>& outputs,
     const std::vector<std::shared_ptr<HostTensor>>& inputs)
 {
-    int64_t max_output_boxes_per_class = nms->max_boxes_output_from_input();
-    float iou_threshold = nms->iou_threshold_from_input();
-    float score_threshold = nms->score_threshold_from_input();
-    float soft_nms_sigma = nms->soft_nms_sigma_from_input();
+    int64_t max_output_boxes_per_class = nms5->max_boxes_output_from_input();
+    float iou_threshold = nms5->iou_threshold_from_input();
+    float score_threshold = nms5->score_threshold_from_input();
+    float soft_nms_sigma = nms5->soft_nms_sigma_from_input();
 
     auto selected_indices_shape = infer_selected_indices_shape(inputs,
                                                                max_output_boxes_per_class);
@@ -338,8 +338,6 @@ void runtime::interpreter::run_nms5(
 
     size_t num_of_outputs = outputs.size();
 
-    size_t num_of_outputs = outputs.size();
-
     auto selected_scores_type = (inputs.size() < 4) ? element::f32 : inputs[3]->get_element_type();
 
     if (num_of_outputs >= 2)

From 02fb546746da7013486da5e37d51d39c6bfb8c2d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 13:21:54 +0300
Subject: [PATCH 131/173] Code style fixes.

---
 .../include/ngraph/op/non_max_suppression.hpp |  1 +
 .../runtime/interpreter/int_executable.cpp    | 20 +++++++++----------
 2 files changed, 10 insertions(+), 11 deletions(-)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index ffab830ad19bb5..0c710efc889b79 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -380,6 +380,7 @@ namespace ngraph
                 float iou_threshold_from_input() const;
                 float score_threshold_from_input() const;
                 float soft_nms_sigma_from_input() const;
+
             protected:
                 BoxEncodingType m_box_encoding = BoxEncodingType::CORNER;
                 bool m_sort_result_descending = true;
diff --git a/ngraph/test/runtime/interpreter/int_executable.cpp b/ngraph/test/runtime/interpreter/int_executable.cpp
index 034a06cf9ae56b..035a9317a0aad6 100644
--- a/ngraph/test/runtime/interpreter/int_executable.cpp
+++ b/ngraph/test/runtime/interpreter/int_executable.cpp
@@ -15,6 +15,8 @@
 //*****************************************************************************
 
 #include "int_executable.hpp"
+#include <algorithm>
+#include <cstring>
 #include "backend_manager.hpp"
 #include "ngraph/chrome_trace.hpp"
 #include "ngraph/except.hpp"
@@ -29,8 +31,6 @@
 #include "pass/liveness.hpp"
 #include "pass/opset0_downgrade.hpp"
 #include "pass/opset1_downgrade.hpp"
-#include <algorithm>
-#include <cstring>
 
 using namespace std;
 using namespace ngraph;
@@ -70,9 +70,9 @@ namespace
     constexpr size_t score_threshold_port = 4;
     constexpr size_t soft_nms_sigma_port = 5;
 
-    PartialShape infer_selected_indices_shape(
-        const std::vector<std::shared_ptr<HostTensor>>& inputs,
-        int64_t max_output_boxes_per_class)
+    PartialShape
+        infer_selected_indices_shape(const std::vector<std::shared_ptr<HostTensor>>& inputs,
+                                     int64_t max_output_boxes_per_class)
     {
         const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
         const auto scores_ps = inputs[scores_port]->get_partial_shape();
@@ -290,18 +290,16 @@ namespace
     }
 }
 
-void runtime::interpreter::run_nms5(
-    const op::v5::NonMaxSuppression* nms5,
-    const std::vector<std::shared_ptr<HostTensor>>& outputs,
-    const std::vector<std::shared_ptr<HostTensor>>& inputs)
+void runtime::interpreter::run_nms5(const op::v5::NonMaxSuppression* nms5,
+                                    const std::vector<std::shared_ptr<HostTensor>>& outputs,
+                                    const std::vector<std::shared_ptr<HostTensor>>& inputs)
 {
     int64_t max_output_boxes_per_class = nms5->max_boxes_output_from_input();
     float iou_threshold = nms5->iou_threshold_from_input();
     float score_threshold = nms5->score_threshold_from_input();
     float soft_nms_sigma = nms5->soft_nms_sigma_from_input();
 
-    auto selected_indices_shape = infer_selected_indices_shape(inputs,
-                                                               max_output_boxes_per_class);
+    auto selected_indices_shape = infer_selected_indices_shape(inputs, max_output_boxes_per_class);
     Shape out_shape = selected_indices_shape.to_shape();
 
     Shape boxes_shape = inputs[boxes_port]->get_shape();

From 95a235ffe00b32db0758c698aaa003c2304947ef Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 16:14:50 +0300
Subject: [PATCH 132/173] Written draft test
 nonmaxsuppression_center_point_box_format_backend in ngraph/test/backend.

---
 ngraph/test/CMakeLists.txt                    |   1 +
 .../test/backend/non_max_suppression.in.cpp   | 106 ++++++++++++++++++
 2 files changed, 107 insertions(+)
 create mode 100644 ngraph/test/backend/non_max_suppression.in.cpp

diff --git a/ngraph/test/CMakeLists.txt b/ngraph/test/CMakeLists.txt
index a795f922ed1f43..56a7005034461d 100644
--- a/ngraph/test/CMakeLists.txt
+++ b/ngraph/test/CMakeLists.txt
@@ -313,6 +313,7 @@ set(MULTI_TEST_SRC
     backend/node_name.in.cpp
     backend/normalize_l2.in.cpp
     backend/not.in.cpp
+    backend/non_max_suppression.in.cpp
     backend/non_zero.in.cpp
     backend/numeric.in.cpp
     backend/one_hot.in.cpp
diff --git a/ngraph/test/backend/non_max_suppression.in.cpp b/ngraph/test/backend/non_max_suppression.in.cpp
new file mode 100644
index 00000000000000..35aaf36ab1fda4
--- /dev/null
+++ b/ngraph/test/backend/non_max_suppression.in.cpp
@@ -0,0 +1,106 @@
+//*****************************************************************************
+// Copyright 2020 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+// clang-format off
+#ifdef ${BACKEND_NAME}_FLOAT_TOLERANCE_BITS
+#define DEFAULT_FLOAT_TOLERANCE_BITS ${BACKEND_NAME}_FLOAT_TOLERANCE_BITS
+#endif
+
+#ifdef ${BACKEND_NAME}_DOUBLE_TOLERANCE_BITS
+#define DEFAULT_DOUBLE_TOLERANCE_BITS ${BACKEND_NAME}_DOUBLE_TOLERANCE_BITS
+#endif
+// clang-format on
+
+#include "gtest/gtest.h"
+#include "runtime/backend.hpp"
+#include "ngraph/runtime/tensor.hpp"
+#include "ngraph/ngraph.hpp"
+#include "util/all_close.hpp"
+#include "util/all_close_f.hpp"
+#include "util/known_element_types.hpp"
+#include "util/ndarray.hpp"
+#include "util/test_control.hpp"
+#include "util/test_tools.hpp"
+
+NGRAPH_SUPPRESS_DEPRECATED_START
+
+using namespace std;
+using namespace ngraph;
+
+static string s_manifest = "${MANIFEST}";
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format_backend)
+{
+    std::vector<float> boxes_data = {0.5, 0.5,  1.0, 1.0, 0.5, 0.6,   1.0, 1.0,
+                                     0.5, 0.4,  1.0, 1.0, 0.5, 10.5,  1.0, 1.0,
+                                     0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CENTER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    int64_t valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
+    std::vector<int64_t> expected_valid_outputs = {3};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}

From 13ee76ea915a07854a41f93b52463ce6724c9894 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 16:40:24 +0300
Subject: [PATCH 133/173] Small fix.

---
 ngraph/test/backend/non_max_suppression.in.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/test/backend/non_max_suppression.in.cpp b/ngraph/test/backend/non_max_suppression.in.cpp
index 35aaf36ab1fda4..8aa173791f8731 100644
--- a/ngraph/test/backend/non_max_suppression.in.cpp
+++ b/ngraph/test/backend/non_max_suppression.in.cpp
@@ -94,7 +94,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format_backend)
 
     auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
     auto selected_scores_value = read_vector<float>(selected_scores);
-    int64_t valid_outputs_value = read_vector<int64_t>(valid_outputs);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
 
     std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
     std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};

From c4f4edf97b46a8b7b22090c32067c2a1aa71f1c7 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 22 Oct 2020 17:20:25 +0300
Subject: [PATCH 134/173] Written draft tests of NonMaxSuppression in
 ngraph/test/runtime.

---
 .../test/backend/non_max_suppression.in.cpp   | 508 ++++++++++++++++++
 1 file changed, 508 insertions(+)

diff --git a/ngraph/test/backend/non_max_suppression.in.cpp b/ngraph/test/backend/non_max_suppression.in.cpp
index 8aa173791f8731..367e71612950f7 100644
--- a/ngraph/test/backend/non_max_suppression.in.cpp
+++ b/ngraph/test/backend/non_max_suppression.in.cpp
@@ -104,3 +104,511 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format_backend)
     EXPECT_EQ(expected_selected_scores, selected_scores_value);
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_flipped_coordinates_backend)
+{
+    std::vector<float> boxes_data = {1.0, 1.0,  0.0, 0.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, 0.9,  1.0, -0.1, 0.0, 10.0,  1.0, 11.0,
+                                     1.0, 10.1, 0.0, 11.1, 1.0, 101.0, 0.0, 100.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
+    std::vector<int64_t> expected_valid_outputs = {3};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_identical_boxes_backend)
+{
+    std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
+                                     1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
+                                     0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
+                                     1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0};
+
+    std::vector<float> scores_data = {0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 10, 4};
+    const auto scores_shape = Shape{1, 1, 10};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {1};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_limit_output_size_backend)
+{
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 2;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {2};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_single_box_backend)
+{
+    std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0};
+
+    std::vector<float> scores_data = {0.9};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 1, 4};
+    const auto scores_shape = Shape{1, 1, 1};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {1};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_backend)
+{
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
+    std::vector<int64_t> expected_valid_outputs = {3};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_and_scores_backend)
+{
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 3;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.4f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
+    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {2};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_batches_backend)
+{
+    std::vector<float> boxes_data = {
+        0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,   0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0, 0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+        0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,  0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {
+        0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 2;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{2, 6, 4};
+    const auto scores_shape = Shape{2, 1, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 1, 0, 3, 1, 0, 0};
+    std::vector<float> expected_selected_scores = {
+        0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 1.0, 0.0, 0.95, 1.0, 0.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {4};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}
+
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_classes_backend)
+{
+    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
+                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
+                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
+
+    std::vector<float> scores_data = {
+        0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
+
+    const int64_t max_output_boxes_per_class_data = 2;
+    const float iou_threshold_data = 0.5f;
+    const float score_threshold_data = 0.0f;
+    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
+    const auto boxes_shape = Shape{1, 6, 4};
+    const auto scores_shape = Shape{1, 2, 6};
+
+    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
+    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
+    auto max_output_boxes_per_class =
+        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
+    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
+    auto score_threshold =
+        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
+    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
+    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
+                                                      scores,
+                                                      max_output_boxes_per_class,
+                                                      iou_threshold,
+                                                      score_threshold,
+                                                      soft_nms_sigma,
+                                                      box_encoding,
+                                                      false);
+
+    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
+
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+
+    auto selected_indeces = backend->create_tensor();
+    auto selected_scores = backend->create_tensor();
+    auto valid_outputs = backend->create_tensor();
+
+    auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
+    auto backend_scores = backend->create_tensor(element::f32, scores_shape);
+    copy_data(backend_boxes, boxes_data);
+    copy_data(backend_scores, scores_data);
+
+    auto handle = backend->compile(f);
+
+    handle->call({selected_indeces, selected_scores, valid_outputs},
+                 {backend_boxes, backend_scores});
+
+    auto selected_indeces_value = read_vector<int64_t>(selected_indeces);
+    auto selected_scores_value = read_vector<float>(selected_scores);
+    auto valid_outputs_value = read_vector<int64_t>(valid_outputs);
+
+    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 1, 3, 0, 1, 0};
+    std::vector<float> expected_selected_scores = {
+        0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 1.0, 0.95, 0.0, 1.0, 0.9};
+    std::vector<int64_t> expected_valid_outputs = {4};
+
+    EXPECT_EQ(expected_selected_indices, selected_indeces_value);
+    EXPECT_EQ(expected_selected_scores, selected_scores_value);
+    EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
+}

From 33e15ce561d2413aa5c564f3be54f95e79348d74 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 09:42:30 +0300
Subject: [PATCH 135/173] Some changes.

---
 ngraph/test/backend/non_max_suppression.in.cpp |  6 +++---
 ngraph/test/runtime/ie/unit_test.manifest      | 11 +++++++++++
 2 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/ngraph/test/backend/non_max_suppression.in.cpp b/ngraph/test/backend/non_max_suppression.in.cpp
index 367e71612950f7..aec40302ee3c49 100644
--- a/ngraph/test/backend/non_max_suppression.in.cpp
+++ b/ngraph/test/backend/non_max_suppression.in.cpp
@@ -585,9 +585,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_classes_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{3, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{3, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
diff --git a/ngraph/test/runtime/ie/unit_test.manifest b/ngraph/test/runtime/ie/unit_test.manifest
index bbad53c317835d..f29c41fbe4de17 100644
--- a/ngraph/test/runtime/ie/unit_test.manifest
+++ b/ngraph/test/runtime/ie/unit_test.manifest
@@ -1139,6 +1139,17 @@ IE_CPU.log_softmax_2d_axis_neg2
 IE_CPU.log_softmax_3d_axis_0
 IE_CPU.log_softmax_3d_axis_neg3
 
+# Unsupported NMS-5
+# IE_CPU.nonmaxsuppression_center_point_box_format_backend
+IE_CPU.nonmaxsuppression_flipped_coordinates_backend
+IE_CPU.nonmaxsuppression_identical_boxes_backend
+IE_CPU.nonmaxsuppression_limit_output_size_backend
+IE_CPU.nonmaxsuppression_single_box_backend
+IE_CPU.nonmaxsuppression_suppress_by_IOU_backend
+IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores_backend
+IE_CPU.nonmaxsuppression_two_batches_backend
+IE_CPU.nonmaxsuppression_two_classes_backend
+
 #-------------------------------------------------------------------------------
 #
 #       Inference Engine GPU plugin excludes

From b98c59a77991aa818160289e02cd4b1acad1c714 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 10:27:40 +0300
Subject: [PATCH 136/173] Small changes.

---
 .../test/backend/non_max_suppression.in.cpp   | 52 +++++++++----------
 ngraph/test/runtime/ie/unit_test.manifest     | 16 +++---
 2 files changed, 34 insertions(+), 34 deletions(-)

diff --git a/ngraph/test/backend/non_max_suppression.in.cpp b/ngraph/test/backend/non_max_suppression.in.cpp
index aec40302ee3c49..2c9514d9538db2 100644
--- a/ngraph/test/backend/non_max_suppression.in.cpp
+++ b/ngraph/test/backend/non_max_suppression.in.cpp
@@ -78,9 +78,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{3, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{3, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -141,9 +141,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_flipped_coordinates_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{3, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{3, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -205,9 +205,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_identical_boxes_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{1, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{1, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -268,9 +268,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_limit_output_size_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{2, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{2, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -329,9 +329,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_single_box_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{1, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{1, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -392,9 +392,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{3, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{3, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -455,9 +455,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_and_scores_backen
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{2, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{2, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -520,9 +520,9 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_batches_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor();
-    auto selected_scores = backend->create_tensor();
-    auto valid_outputs = backend->create_tensor();
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{4, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{4, 3});
+    auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
     auto backend_scores = backend->create_tensor(element::f32, scores_shape);
@@ -585,8 +585,8 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_classes_backend)
 
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
-    auto selected_indeces = backend->create_tensor(element::i64, Shape{3, 3});
-    auto selected_scores = backend->create_tensor(element::f32, Shape{3, 3});
+    auto selected_indeces = backend->create_tensor(element::i64, Shape{4, 3});
+    auto selected_scores = backend->create_tensor(element::f32, Shape{4, 3});
     auto valid_outputs = backend->create_tensor(element::i64, Shape{1});
 
     auto backend_boxes = backend->create_tensor(element::f32, boxes_shape);
diff --git a/ngraph/test/runtime/ie/unit_test.manifest b/ngraph/test/runtime/ie/unit_test.manifest
index f29c41fbe4de17..83aa4a0781a38b 100644
--- a/ngraph/test/runtime/ie/unit_test.manifest
+++ b/ngraph/test/runtime/ie/unit_test.manifest
@@ -1141,14 +1141,14 @@ IE_CPU.log_softmax_3d_axis_neg3
 
 # Unsupported NMS-5
 # IE_CPU.nonmaxsuppression_center_point_box_format_backend
-IE_CPU.nonmaxsuppression_flipped_coordinates_backend
-IE_CPU.nonmaxsuppression_identical_boxes_backend
-IE_CPU.nonmaxsuppression_limit_output_size_backend
-IE_CPU.nonmaxsuppression_single_box_backend
-IE_CPU.nonmaxsuppression_suppress_by_IOU_backend
-IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores_backend
-IE_CPU.nonmaxsuppression_two_batches_backend
-IE_CPU.nonmaxsuppression_two_classes_backend
+# IE_CPU.nonmaxsuppression_flipped_coordinates_backend
+# IE_CPU.nonmaxsuppression_identical_boxes_backend
+# IE_CPU.nonmaxsuppression_limit_output_size_backend
+# IE_CPU.nonmaxsuppression_single_box_backend
+# IE_CPU.nonmaxsuppression_suppress_by_IOU_backend
+# IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores_backend
+# IE_CPU.nonmaxsuppression_two_batches_backend
+# IE_CPU.nonmaxsuppression_two_classes_backend
 
 #-------------------------------------------------------------------------------
 #

From 7cc7aac9c9acac600dee0844fbc90282d3d325e9 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 10:51:57 +0300
Subject: [PATCH 137/173] Disabled IE_CPU tests for NMS-5.

---
 ngraph/test/runtime/ie/unit_test.manifest | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/ngraph/test/runtime/ie/unit_test.manifest b/ngraph/test/runtime/ie/unit_test.manifest
index 83aa4a0781a38b..18de0871915153 100644
--- a/ngraph/test/runtime/ie/unit_test.manifest
+++ b/ngraph/test/runtime/ie/unit_test.manifest
@@ -1140,15 +1140,15 @@ IE_CPU.log_softmax_3d_axis_0
 IE_CPU.log_softmax_3d_axis_neg3
 
 # Unsupported NMS-5
-# IE_CPU.nonmaxsuppression_center_point_box_format_backend
-# IE_CPU.nonmaxsuppression_flipped_coordinates_backend
-# IE_CPU.nonmaxsuppression_identical_boxes_backend
-# IE_CPU.nonmaxsuppression_limit_output_size_backend
-# IE_CPU.nonmaxsuppression_single_box_backend
-# IE_CPU.nonmaxsuppression_suppress_by_IOU_backend
-# IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores_backend
-# IE_CPU.nonmaxsuppression_two_batches_backend
-# IE_CPU.nonmaxsuppression_two_classes_backend
+IE_CPU.nonmaxsuppression_center_point_box_format_backend
+IE_CPU.nonmaxsuppression_flipped_coordinates_backend
+IE_CPU.nonmaxsuppression_identical_boxes_backend
+IE_CPU.nonmaxsuppression_limit_output_size_backend
+IE_CPU.nonmaxsuppression_single_box_backend
+IE_CPU.nonmaxsuppression_suppress_by_IOU_backend
+IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores_backend
+IE_CPU.nonmaxsuppression_two_batches_backend
+IE_CPU.nonmaxsuppression_two_classes_backend
 
 #-------------------------------------------------------------------------------
 #

From 45d1ff730e37824c72d7b802c1d88d804f9aeaf3 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 10:56:55 +0300
Subject: [PATCH 138/173] Deleted op_eval tests for NMS-5.

---
 ngraph/test/CMakeLists.txt                    |   1 -
 ngraph/test/op_eval/non_max_suppression_5.cpp | 549 ------------------
 2 files changed, 550 deletions(-)
 delete mode 100644 ngraph/test/op_eval/non_max_suppression_5.cpp

diff --git a/ngraph/test/CMakeLists.txt b/ngraph/test/CMakeLists.txt
index 56a7005034461d..f58348147fb155 100644
--- a/ngraph/test/CMakeLists.txt
+++ b/ngraph/test/CMakeLists.txt
@@ -75,7 +75,6 @@ set(SRC
     op_eval/interpolate.cpp
     op_eval/matmul.cpp
     op_eval/mish.cpp
-    op_eval/non_max_suppression_5.cpp
     op_eval/non_zero.cpp
     op_eval/reduce_l1.cpp
     op_eval/reduce_l2.cpp
diff --git a/ngraph/test/op_eval/non_max_suppression_5.cpp b/ngraph/test/op_eval/non_max_suppression_5.cpp
deleted file mode 100644
index c47cc69b527321..00000000000000
--- a/ngraph/test/op_eval/non_max_suppression_5.cpp
+++ /dev/null
@@ -1,549 +0,0 @@
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-//*****************************************************************************
-
-#include <string>
-#include <vector>
-
-#include "gtest/gtest.h"
-
-#include "ngraph/op/constant.hpp"
-#include "ngraph/op/non_max_suppression.hpp"
-#include "ngraph/runtime/host_tensor.hpp"
-#include "ngraph/validation_util.hpp"
-#include "runtime/backend.hpp"
-#include "util/test_tools.hpp"
-#include "util/type_prop.hpp"
-
-using namespace std;
-using namespace ngraph;
-
-TEST(op_eval, nonmaxsuppression_center_point_box_format)
-{
-    std::vector<float> boxes_data = {0.5, 0.5,  1.0, 1.0, 0.5, 0.6,   1.0, 1.0,
-                                     0.5, 0.4,  1.0, 1.0, 0.5, 10.5,  1.0, 1.0,
-                                     0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
-
-    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 3;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CENTER;
-    const auto boxes_shape = Shape{1, 6, 4};
-    const auto scores_shape = Shape{1, 1, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
-    std::vector<int64_t> expected_valid_outputs = {3};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{3, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{3, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_flipped_coordinates)
-{
-    std::vector<float> boxes_data = {1.0, 1.0,  0.0, 0.0,  0.0, 0.1,   1.0, 1.1,
-                                     0.0, 0.9,  1.0, -0.1, 0.0, 10.0,  1.0, 11.0,
-                                     1.0, 10.1, 0.0, 11.1, 1.0, 101.0, 0.0, 100.0};
-
-    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 3;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 6, 4};
-    const auto scores_shape = Shape{1, 1, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
-    std::vector<int64_t> expected_valid_outputs = {3};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{3, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{3, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_identical_boxes)
-{
-    std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
-                                     1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
-                                     0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
-                                     1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0};
-
-    std::vector<float> scores_data = {0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9};
-
-    const int64_t max_output_boxes_per_class_data = 3;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 10, 4};
-    const auto scores_shape = Shape{1, 1, 10};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 0};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
-    std::vector<int64_t> expected_valid_outputs = {1};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{1, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{1, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_limit_output_size)
-{
-    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
-                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
-                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
-
-    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 2;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 6, 4};
-    const auto scores_shape = Shape{1, 1, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
-    std::vector<int64_t> expected_valid_outputs = {2};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{2, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{2, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_single_box)
-{
-    std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0};
-
-    std::vector<float> scores_data = {0.9};
-
-    const int64_t max_output_boxes_per_class_data = 3;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 1, 4};
-    const auto scores_shape = Shape{1, 1, 1};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 0};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.9};
-    std::vector<int64_t> expected_valid_outputs = {1};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{1, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{1, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_suppress_by_IOU)
-{
-    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
-                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
-                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
-
-    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 3;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 6, 4};
-    const auto scores_shape = Shape{1, 1, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 0, 5};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 0.0, 0.3};
-    std::vector<int64_t> expected_valid_outputs = {3};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{3, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{3, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_suppress_by_IOU_and_scores)
-{
-    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
-                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
-                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
-
-    std::vector<float> scores_data = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 3;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.4f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 6, 4};
-    const auto scores_shape = Shape{1, 1, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0};
-    std::vector<float> expected_selected_scores = {0.0, 0.0, 0.95, 0.0, 0.0, 0.9};
-    std::vector<int64_t> expected_valid_outputs = {2};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{2, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{2, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_two_batches)
-{
-    std::vector<float> boxes_data = {
-        0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,   0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
-        0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0, 0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
-        0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,  0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
-
-    std::vector<float> scores_data = {
-        0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 2;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{2, 6, 4};
-    const auto scores_shape = Shape{2, 1, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 1, 0, 3, 1, 0, 0};
-    std::vector<float> expected_selected_scores = {
-        0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 1.0, 0.0, 0.95, 1.0, 0.0, 0.9};
-    std::vector<int64_t> expected_valid_outputs = {4};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{4, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{4, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}
-
-TEST(op_eval, nonmaxsuppression_two_classes)
-{
-    std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
-                                     0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
-                                     0.0, 10.1, 1.0, 11.1, 0.0, 100.0, 1.0, 101.0};
-
-    std::vector<float> scores_data = {
-        0.9, 0.75, 0.6, 0.95, 0.5, 0.3, 0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
-
-    const int64_t max_output_boxes_per_class_data = 2;
-    const float iou_threshold_data = 0.5f;
-    const float score_threshold_data = 0.0f;
-    const auto box_encoding = op::v5::NonMaxSuppression::BoxEncodingType::CORNER;
-    const auto boxes_shape = Shape{1, 6, 4};
-    const auto scores_shape = Shape{1, 2, 6};
-
-    const auto boxes = make_shared<op::Parameter>(element::f32, boxes_shape);
-    const auto scores = make_shared<op::Parameter>(element::f32, scores_shape);
-    auto max_output_boxes_per_class =
-        op::Constant::create<int64_t>(element::i64, Shape{}, {max_output_boxes_per_class_data});
-    auto iou_threshold = op::Constant::create<float>(element::f32, Shape{}, {iou_threshold_data});
-    auto score_threshold =
-        op::Constant::create<float>(element::f32, Shape{}, {score_threshold_data});
-    auto soft_nms_sigma = op::Constant::create<float>(element::f32, Shape{}, {0.0f});
-    auto nms = make_shared<op::v5::NonMaxSuppression>(boxes,
-                                                      scores,
-                                                      max_output_boxes_per_class,
-                                                      iou_threshold,
-                                                      score_threshold,
-                                                      soft_nms_sigma,
-                                                      box_encoding,
-                                                      false);
-
-    auto f = make_shared<Function>(nms, ParameterVector{boxes, scores});
-
-    HostTensorVector results(3);
-    for (auto& result : results)
-    {
-        result = make_shared<HostTensor>();
-    }
-    ASSERT_TRUE(f->evaluate(results,
-                            {make_host_tensor<element::Type_t::f32>(boxes_shape, boxes_data),
-                             make_host_tensor<element::Type_t::f32>(scores_shape, scores_data)}));
-
-    std::vector<int64_t> expected_selected_indices = {0, 0, 3, 0, 0, 0, 0, 1, 3, 0, 1, 0};
-    std::vector<float> expected_selected_scores = {
-        0.0, 0.0, 0.95, 0.0, 0.0, 0.9, 0.0, 1.0, 0.95, 0.0, 1.0, 0.9};
-    std::vector<int64_t> expected_valid_outputs = {4};
-
-    EXPECT_EQ(results[0]->get_element_type(), element::i64);
-    EXPECT_EQ(results[1]->get_element_type(), element::f32);
-    EXPECT_EQ(results[2]->get_element_type(), element::i64);
-    EXPECT_EQ(results[0]->get_shape(), (Shape{4, 3}));
-    EXPECT_EQ(results[1]->get_shape(), (Shape{4, 3}));
-    EXPECT_EQ(results[2]->get_shape(), (Shape{1}));
-    EXPECT_EQ(read_vector<int64_t>(results[0]), expected_selected_indices);
-    EXPECT_EQ(read_vector<float>(results[1]), expected_selected_scores);
-    EXPECT_EQ(read_vector<int64_t>(results[2]), expected_valid_outputs);
-}

From 304b54c3b51396e30aecf0a9b2ae2e794873ad50 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 11:09:23 +0300
Subject: [PATCH 139/173] Deleted evaluate() method of NMS-5.

---
 .../include/ngraph/op/non_max_suppression.hpp |   2 -
 ngraph/core/src/op/non_max_suppression.cpp    | 249 ------------------
 2 files changed, 251 deletions(-)

diff --git a/ngraph/core/include/ngraph/op/non_max_suppression.hpp b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
index 0c710efc889b79..b6a93610f62a3d 100644
--- a/ngraph/core/include/ngraph/op/non_max_suppression.hpp
+++ b/ngraph/core/include/ngraph/op/non_max_suppression.hpp
@@ -355,8 +355,6 @@ namespace ngraph
 
                 std::shared_ptr<Node>
                     clone_with_new_inputs(const OutputVector& new_args) const override;
-                bool evaluate(const HostTensorVector& outputs,
-                              const HostTensorVector& inputs) const override;
 
                 BoxEncodingType get_box_encoding() const { return m_box_encoding; }
                 void set_box_encoding(const BoxEncodingType box_encoding)
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 2d581b2f2c5f46..8adc25f0e16770 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -693,194 +693,6 @@ namespace
     constexpr size_t score_threshold_port = 4;
     constexpr size_t soft_nms_sigma_port = 5;
 
-    PartialShape infer_selected_indices_shape(const HostTensorVector& inputs,
-                                              int64_t max_output_boxes_per_class)
-    {
-        const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
-        const auto scores_ps = inputs[scores_port]->get_partial_shape();
-
-        // NonMaxSuppression produces triplets
-        // that have the following format: [batch_index, class_index, box_index]
-        PartialShape result = {Dimension::dynamic(), 3};
-
-        if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
-        {
-            const auto num_boxes_boxes = boxes_ps[1];
-            if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
-            {
-                const auto num_boxes = num_boxes_boxes.get_length();
-                const auto num_classes = scores_ps[1].get_length();
-
-                result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
-                            scores_ps[0].get_length();
-            }
-        }
-
-        return result;
-    }
-
-    void normalize_corner(float* boxes, const Shape& boxes_shape)
-    {
-        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
-        for (size_t i = 0; i < total_num_of_boxes; ++i)
-        {
-            float* current_box = boxes + 4 * i;
-
-            float y1 = current_box[0];
-            float x1 = current_box[1];
-            float y2 = current_box[2];
-            float x2 = current_box[3];
-
-            float ymin = std::min(y1, y2);
-            float ymax = std::max(y1, y2);
-            float xmin = std::min(x1, x2);
-            float xmax = std::max(x1, x2);
-
-            current_box[0] = ymin;
-            current_box[1] = xmin;
-            current_box[2] = ymax;
-            current_box[3] = xmax;
-        }
-    }
-
-    void normalize_center(float* boxes, const Shape& boxes_shape)
-    {
-        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
-        for (size_t i = 0; i < total_num_of_boxes; ++i)
-        {
-            float* current_box = boxes + 4 * i;
-
-            float x_center = current_box[0];
-            float y_center = current_box[1];
-            float width = current_box[2];
-            float height = current_box[3];
-
-            float y1 = y_center - height / 2.0;
-            float x1 = x_center - width / 2.0;
-            float y2 = y_center + height / 2.0;
-            float x2 = x_center + width / 2.0;
-
-            current_box[0] = y1;
-            current_box[1] = x1;
-            current_box[2] = y2;
-            current_box[3] = x2;
-        }
-    }
-
-    void normalize_box_encoding(float* boxes,
-                                const Shape& boxes_shape,
-                                const V5BoxEncoding box_encoding)
-    {
-        if (box_encoding == V5BoxEncoding::CORNER)
-        {
-            normalize_corner(boxes, boxes_shape);
-        }
-        else
-        {
-            normalize_center(boxes, boxes_shape);
-        }
-    }
-
-    std::vector<float> get_floats(const HostTensorPtr& input, const Shape& shape)
-    {
-        size_t input_size = shape_size(shape);
-        std::vector<float> result(input_size);
-
-        switch (input->get_element_type())
-        {
-        case element::Type_t::bf16:
-        {
-            bfloat16* p = input->get_data_ptr<bfloat16>();
-            for (size_t i = 0; i < input_size; ++i)
-            {
-                result[i] = float(p[i]);
-            }
-        }
-        break;
-        case element::Type_t::f16:
-        {
-            float16* p = input->get_data_ptr<float16>();
-            for (size_t i = 0; i < input_size; ++i)
-            {
-                result[i] = float(p[i]);
-            }
-        }
-        break;
-        case element::Type_t::f32:
-        {
-            float* p = input->get_data_ptr<float>();
-            memcpy(result.data(), p, input_size * sizeof(float));
-        }
-        break;
-        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
-        }
-
-        return result;
-    }
-
-    std::vector<float> prepare_boxes_data(const HostTensorPtr& boxes,
-                                          const Shape& boxes_shape,
-                                          const V5BoxEncoding box_encoding)
-    {
-        auto result = get_floats(boxes, boxes_shape);
-        normalize_box_encoding(result.data(), boxes_shape, box_encoding);
-        return result;
-    }
-
-    std::vector<float> prepare_scores_data(const HostTensorPtr& scores, const Shape& scores_shape)
-    {
-        auto result = get_floats(scores, scores_shape);
-        return result;
-    }
-
-    void evaluate_postprocessing(const HostTensorVector& outputs,
-                                 const ngraph::element::Type output_type,
-                                 const std::vector<int64_t>& selected_indices,
-                                 const std::vector<float>& selected_scores,
-                                 int64_t valid_outputs)
-    {
-        size_t num_of_outputs = outputs.size();
-        size_t selected_size = valid_outputs * 3;
-
-        if (output_type == ngraph::element::i64)
-        {
-            int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
-            memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
-        }
-        else
-        {
-            int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
-            for (size_t i = 0; i < selected_size; ++i)
-            {
-                indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
-            }
-        }
-
-        if (num_of_outputs < 2)
-        {
-            return;
-        }
-
-        float* scores_ptr = outputs[1]->get_data_ptr<float>();
-        memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
-
-        if (num_of_outputs < 3)
-        {
-            return;
-        }
-
-        if (output_type == ngraph::element::i64)
-        {
-            int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
-            *valid_outputs_ptr = valid_outputs;
-        }
-        else
-        {
-            int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
-            *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
-        }
-    }
-
     inline bool is_float_type_admissible(const element::Type& t)
     {
         return t == element::f32 || t == element::f16 || t == element::bf16;
@@ -1144,64 +956,3 @@ namespace ngraph
         return s << as_string(type);
     }
 } // namespace ngraph
-
-bool op::v5::NonMaxSuppression::evaluate(const HostTensorVector& outputs,
-                                         const HostTensorVector& inputs) const
-{
-    int64_t max_output_boxes_per_class = max_boxes_output_from_input();
-    float iou_threshold = iou_threshold_from_input();
-    float score_threshold = score_threshold_from_input();
-    float soft_nms_sigma = soft_nms_sigma_from_input();
-
-    auto selected_indices_shape = infer_selected_indices_shape(inputs, max_output_boxes_per_class);
-    Shape out_shape = selected_indices_shape.to_shape();
-
-    Shape boxes_shape = inputs[boxes_port]->get_shape();
-    Shape scores_shape = inputs[scores_port]->get_shape();
-
-    auto boxes_data = prepare_boxes_data(inputs[boxes_port], boxes_shape, m_box_encoding);
-    auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
-
-    size_t out_shape_size = shape_size(out_shape);
-
-    std::vector<int64_t> selected_indices(out_shape_size);
-    std::vector<float> selected_scores(out_shape_size);
-    int64_t valid_outputs = 0;
-
-    runtime::reference::non_max_suppression(boxes_data.data(),
-                                            boxes_shape,
-                                            scores_data.data(),
-                                            scores_shape,
-                                            max_output_boxes_per_class,
-                                            iou_threshold,
-                                            score_threshold,
-                                            soft_nms_sigma,
-                                            selected_indices.data(),
-                                            out_shape,
-                                            selected_scores.data(),
-                                            out_shape,
-                                            &valid_outputs,
-                                            m_sort_result_descending);
-
-    outputs[0]->set_element_type(m_output_type);
-    outputs[0]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
-
-    size_t num_of_outputs = outputs.size();
-
-    if (num_of_outputs >= 2)
-    {
-        outputs[1]->set_element_type(element::f32);
-        outputs[1]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
-    }
-
-    if (num_of_outputs >= 3)
-    {
-        outputs[2]->set_element_type(m_output_type);
-        outputs[2]->set_shape(Shape{1});
-    }
-
-    evaluate_postprocessing(
-        outputs, m_output_type, selected_indices, selected_scores, valid_outputs);
-
-    return true;
-}

From 977e7a9b946c4a51c9a436f9f8034dc2e10f3764 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 15:30:40 +0300
Subject: [PATCH 140/173] Now all nGraph functions in tests of the
 transformation NMS-5 -> NMSIE3 have one output.

---
 .../transformations/convert_nms5_test.cpp     | 60 +++++++++----------
 1 file changed, 30 insertions(+), 30 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 29cb46fe0fd28a..038598101dd1bd 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -32,7 +32,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticSixInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -69,7 +69,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticSixInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
@@ -90,7 +90,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -127,7 +127,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
@@ -147,7 +147,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -184,7 +184,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
@@ -203,7 +203,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -240,7 +240,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
@@ -258,7 +258,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -295,7 +295,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
         EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
@@ -317,7 +317,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1SixInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
         pass::Manager manager;
@@ -353,7 +353,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1SixInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -371,7 +371,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FiveInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
         pass::Manager manager;
@@ -407,7 +407,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FiveInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -424,7 +424,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FourInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
         pass::Manager manager;
@@ -460,7 +460,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FourInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -476,7 +476,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1ThreeInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
         pass::Manager manager;
@@ -512,7 +512,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1ThreeInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -527,7 +527,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1TwoInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         const auto &orig_selected_indices_shape = f->get_output_partial_shape(0);
         pass::Manager manager;
@@ -563,7 +563,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1TwoInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -582,7 +582,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2SixInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -617,7 +617,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2SixInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -635,7 +635,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FiveInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -670,7 +670,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FiveInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -687,7 +687,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FourInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class, iou_threshold,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -722,7 +722,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FourInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -738,7 +738,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2ThreeInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -773,7 +773,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2ThreeInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);
@@ -788,7 +788,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2TwoInputs) {
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores,
                                                                opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
-        f = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
 
         pass::Manager manager;
         manager.register_pass<pass::InitNodeInfo>();
@@ -823,7 +823,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2TwoInputs) {
                                                               new_soft_nms_sigma, 0, true);
         nms->set_friendly_name("nms");
 
-        f_ref = std::make_shared<Function>(nms, ParameterVector{boxes, scores});
+        f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
 
     auto res = compare_functions(f, f_ref);

From cdc5d68bf042e758395b7e75ea27c6a4f678f762 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 16:07:26 +0300
Subject: [PATCH 141/173] Now preprocessing and postprocessing of the
 calculation of NMS-5 in reference implementation.

---
 .../runtime/reference/non_max_suppression.hpp |  19 +-
 .../runtime/reference/non_max_suppression.cpp | 555 ++++++++++++++----
 .../test/backend/non_max_suppression.in.cpp   |  18 +-
 ngraph/test/runtime/ie/unit_test.manifest     |  18 +-
 .../runtime/interpreter/int_executable.cpp    | 299 ----------
 .../runtime/interpreter/int_executable.hpp    |   2 +-
 6 files changed, 453 insertions(+), 458 deletions(-)

diff --git a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
index aa5725650f5da2..a0e370c8b7d950 100644
--- a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
+++ b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
@@ -23,8 +23,10 @@
 #include <cstddef>
 #include <functional>
 #include <map>
+#include <ngraph/runtime/host_tensor.hpp>
 #include "ngraph/coordinate_transform.hpp"
 #include "ngraph/op/util/op_types.hpp"
+#include "ngraph/ops.hpp"
 #include "ngraph/shape_util.hpp"
 
 namespace ngraph
@@ -33,20 +35,9 @@ namespace ngraph
     {
         namespace reference
         {
-            void non_max_suppression(const float* boxes_data,
-                                     const Shape& boxes_data_shape,
-                                     const float* scores_data,
-                                     const Shape& scores_data_shape,
-                                     int64_t max_output_boxes_per_class,
-                                     float iou_threshold,
-                                     float score_threshold,
-                                     float soft_nms_sigma,
-                                     int64_t* selected_indices,
-                                     const Shape& selected_indices_shape,
-                                     float* selected_scores,
-                                     const Shape& selected_scores_shape,
-                                     int64_t* valid_outputs,
-                                     const bool sort_result_descending);
+            void non_max_suppression(const op::v5::NonMaxSuppression* nms5,
+                                     const std::vector<std::shared_ptr<HostTensor>>& out,
+                                     const std::vector<std::shared_ptr<HostTensor>>& args);
         }
     }
 }
diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 2f7e81b4ba543e..9b2b50db684ce8 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -128,169 +128,472 @@ struct BoxInfo
     float score = 0.0f;
 };
 
-namespace ngraph
+using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
+
+namespace
 {
-    namespace runtime
+    constexpr size_t boxes_port = 0;
+    constexpr size_t scores_port = 1;
+    constexpr size_t max_output_boxes_port = 2;
+    constexpr size_t iou_threshold_port = 3;
+    constexpr size_t score_threshold_port = 4;
+    constexpr size_t soft_nms_sigma_port = 5;
+
+    PartialShape
+        infer_selected_indices_shape(const std::vector<std::shared_ptr<HostTensor>>& inputs,
+                                     int64_t max_output_boxes_per_class)
     {
-        namespace reference
+        const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
+        const auto scores_ps = inputs[scores_port]->get_partial_shape();
+
+        // NonMaxSuppression produces triplets
+        // that have the following format: [batch_index, class_index, box_index]
+        PartialShape result = {Dimension::dynamic(), 3};
+
+        if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
         {
-            void non_max_suppression(const float* boxes_data,
-                                     const Shape& boxes_data_shape,
-                                     const float* scores_data,
-                                     const Shape& scores_data_shape,
-                                     int64_t max_output_boxes_per_class,
-                                     float iou_threshold,
-                                     float score_threshold,
-                                     float soft_nms_sigma,
-                                     int64_t* selected_indices,
-                                     const Shape& selected_indices_shape,
-                                     float* selected_scores,
-                                     const Shape& selected_scores_shape,
-                                     int64_t* valid_outputs,
-                                     const bool sort_result_descending)
+            const auto num_boxes_boxes = boxes_ps[1];
+            if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
             {
-                float scale = 0.0f;
-                if (soft_nms_sigma > 0.0f)
-                {
-                    scale = -0.5f / soft_nms_sigma;
-                }
+                const auto num_boxes = num_boxes_boxes.get_length();
+                const auto num_classes = scores_ps[1].get_length();
 
-                auto func = [iou_threshold, scale](float iou) {
-                    const float weight = std::exp(scale * iou * iou);
-                    return iou <= iou_threshold ? weight : 0.0f;
-                };
+                result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
+                            scores_ps[0].get_length();
+            }
+        }
 
-                // boxes shape: {num_batches, num_boxes, 4}
-                // scores shape: {num_batches, num_classes, num_boxes}
-                int64_t num_batches = static_cast<int64_t>(scores_data_shape[0]);
-                int64_t num_classes = static_cast<int64_t>(scores_data_shape[1]);
-                int64_t num_boxes = static_cast<int64_t>(boxes_data_shape[1]);
+        return result;
+    }
 
-                SelectedIndex* selected_indices_ptr =
-                    reinterpret_cast<SelectedIndex*>(selected_indices);
-                SelectedScore* selected_scores_ptr =
-                    reinterpret_cast<SelectedScore*>(selected_scores);
+    void normalize_corner(float* boxes, const Shape& boxes_shape)
+    {
+        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+        for (size_t i = 0; i < total_num_of_boxes; ++i)
+        {
+            float* current_box = boxes + 4 * i;
+
+            float y1 = current_box[0];
+            float x1 = current_box[1];
+            float y2 = current_box[2];
+            float x2 = current_box[3];
+
+            float ymin = std::min(y1, y2);
+            float ymax = std::max(y1, y2);
+            float xmin = std::min(x1, x2);
+            float xmax = std::max(x1, x2);
+
+            current_box[0] = ymin;
+            current_box[1] = xmin;
+            current_box[2] = ymax;
+            current_box[3] = xmax;
+        }
+    }
 
-                size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
+    void normalize_center(float* boxes, const Shape& boxes_shape)
+    {
+        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
+        for (size_t i = 0; i < total_num_of_boxes; ++i)
+        {
+            float* current_box = boxes + 4 * i;
+
+            float x_center = current_box[0];
+            float y_center = current_box[1];
+            float width = current_box[2];
+            float height = current_box[3];
+
+            float y1 = y_center - height / 2.0;
+            float x1 = x_center - width / 2.0;
+            float y2 = y_center + height / 2.0;
+            float x2 = x_center + width / 2.0;
+
+            current_box[0] = y1;
+            current_box[1] = x1;
+            current_box[2] = y2;
+            current_box[3] = x2;
+        }
+    }
 
-                int64_t num_of_valid_boxes = 0;
+    void normalize_box_encoding(float* boxes,
+                                const Shape& boxes_shape,
+                                const V5BoxEncoding box_encoding)
+    {
+        if (box_encoding == V5BoxEncoding::CORNER)
+        {
+            normalize_corner(boxes, boxes_shape);
+        }
+        else
+        {
+            normalize_center(boxes, boxes_shape);
+        }
+    }
 
-                std::vector<BoxInfo> filteredBoxes;
+    std::vector<float> get_floats(const std::shared_ptr<HostTensor>& input, const Shape& shape)
+    {
+        size_t input_size = shape_size(shape);
+        std::vector<float> result(input_size);
 
-                for (int64_t batch = 0; batch < num_batches; batch++)
-                {
-                    const float* boxesPtr = boxes_data + batch * num_boxes * 4;
-                    Rectangle* r = reinterpret_cast<Rectangle*>(const_cast<float*>(boxesPtr));
+        switch (input->get_element_type())
+        {
+        case element::Type_t::bf16:
+        {
+            bfloat16* p = input->get_data_ptr<bfloat16>();
+            for (size_t i = 0; i < input_size; ++i)
+            {
+                result[i] = float(p[i]);
+            }
+        }
+        break;
+        case element::Type_t::f16:
+        {
+            float16* p = input->get_data_ptr<float16>();
+            for (size_t i = 0; i < input_size; ++i)
+            {
+                result[i] = float(p[i]);
+            }
+        }
+        break;
+        case element::Type_t::f32:
+        {
+            float* p = input->get_data_ptr<float>();
+            memcpy(result.data(), p, input_size * sizeof(float));
+        }
+        break;
+        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
+        }
 
-                    for (int64_t class_idx = 0; class_idx < num_classes; class_idx++)
+        return result;
+    }
+
+    std::vector<float> prepare_boxes_data(const std::shared_ptr<HostTensor>& boxes,
+                                          const Shape& boxes_shape,
+                                          const V5BoxEncoding box_encoding)
+    {
+        auto result = get_floats(boxes, boxes_shape);
+        normalize_box_encoding(result.data(), boxes_shape, box_encoding);
+        return result;
+    }
+
+    std::vector<float> prepare_scores_data(const std::shared_ptr<HostTensor>& scores,
+                                           const Shape& scores_shape)
+    {
+        auto result = get_floats(scores, scores_shape);
+        return result;
+    }
+
+    void nms_postprocessing(const std::vector<std::shared_ptr<HostTensor>>& outputs,
+                            const ngraph::element::Type output_type,
+                            const std::vector<int64_t>& selected_indices,
+                            const std::vector<float>& selected_scores,
+                            int64_t valid_outputs,
+                            const ngraph::element::Type selected_scores_type)
+    {
+        size_t num_of_outputs = outputs.size();
+        size_t selected_size = valid_outputs * 3;
+
+        if (output_type == ngraph::element::i64)
+        {
+            int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
+            memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
+        }
+        else
+        {
+            int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
+            for (size_t i = 0; i < selected_size; ++i)
+            {
+                indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
+            }
+        }
+
+        if (num_of_outputs < 2)
+        {
+            return;
+        }
+
+        size_t selected_scores_size = selected_scores.size();
+
+        switch (selected_scores_type)
+        {
+        case element::Type_t::bf16:
+        {
+            bfloat16* scores_ptr = outputs[1]->get_data_ptr<bfloat16>();
+            for (size_t i = 0; i < selected_scores_size; ++i)
+            {
+                scores_ptr[i] = bfloat16(selected_scores[i]);
+            }
+        }
+        break;
+        case element::Type_t::f16:
+        {
+            float16* scores_ptr = outputs[1]->get_data_ptr<float16>();
+            for (size_t i = 0; i < selected_scores_size; ++i)
+            {
+                scores_ptr[i] = float16(selected_scores[i]);
+            }
+        }
+        break;
+        case element::Type_t::f32:
+        {
+            float* scores_ptr = outputs[1]->get_data_ptr<float>();
+            memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
+        }
+        break;
+        default:;
+        }
+
+        if (num_of_outputs < 3)
+        {
+            return;
+        }
+
+        if (output_type == ngraph::element::i64)
+        {
+            int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
+            *valid_outputs_ptr = valid_outputs;
+        }
+        else
+        {
+            int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
+            *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
+        }
+    }
+
+    void eval(const float* boxes_data,
+              const Shape& boxes_data_shape,
+              const float* scores_data,
+              const Shape& scores_data_shape,
+              int64_t max_output_boxes_per_class,
+              float iou_threshold,
+              float score_threshold,
+              float soft_nms_sigma,
+              int64_t* selected_indices,
+              const Shape& selected_indices_shape,
+              float* selected_scores,
+              const Shape& selected_scores_shape,
+              int64_t* valid_outputs,
+              const bool sort_result_descending)
+    {
+        float scale = 0.0f;
+        if (soft_nms_sigma > 0.0f)
+        {
+            scale = -0.5f / soft_nms_sigma;
+        }
+
+        auto func = [iou_threshold, scale](float iou) {
+            const float weight = std::exp(scale * iou * iou);
+            return iou <= iou_threshold ? weight : 0.0f;
+        };
+
+        // boxes shape: {num_batches, num_boxes, 4}
+        // scores shape: {num_batches, num_classes, num_boxes}
+        int64_t num_batches = static_cast<int64_t>(scores_data_shape[0]);
+        int64_t num_classes = static_cast<int64_t>(scores_data_shape[1]);
+        int64_t num_boxes = static_cast<int64_t>(boxes_data_shape[1]);
+
+        SelectedIndex* selected_indices_ptr =
+            reinterpret_cast<SelectedIndex*>(selected_indices);
+        SelectedScore* selected_scores_ptr =
+            reinterpret_cast<SelectedScore*>(selected_scores);
+
+        size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
+
+        int64_t num_of_valid_boxes = 0;
+
+        std::vector<BoxInfo> filteredBoxes;
+
+        for (int64_t batch = 0; batch < num_batches; batch++)
+        {
+            const float* boxesPtr = boxes_data + batch * num_boxes * 4;
+            Rectangle* r = reinterpret_cast<Rectangle*>(const_cast<float*>(boxesPtr));
+
+            for (int64_t class_idx = 0; class_idx < num_classes; class_idx++)
+            {
+                const float* scoresPtr =
+                    scores_data + batch * (num_classes * num_boxes) + class_idx * num_boxes;
+
+                std::vector<BoxInfo> candidate_boxes;
+                candidate_boxes.reserve(num_boxes);
+
+                for (size_t box_idx = 0; box_idx < num_boxes; box_idx++)
+                {
+                    if (scoresPtr[box_idx] > score_threshold)
                     {
-                        const float* scoresPtr =
-                            scores_data + batch * (num_classes * num_boxes) + class_idx * num_boxes;
+                        candidate_boxes.emplace_back(
+                            r[box_idx], box_idx, scoresPtr[box_idx], 0, batch, class_idx);
+                    }
+                }
 
-                        std::vector<BoxInfo> candidate_boxes;
-                        candidate_boxes.reserve(num_boxes);
+                std::priority_queue<BoxInfo> sorted_boxes(std::less<BoxInfo>(),
+                                                          std::move(candidate_boxes));
 
-                        for (size_t box_idx = 0; box_idx < num_boxes; box_idx++)
-                        {
-                            if (scoresPtr[box_idx] > score_threshold)
-                            {
-                                candidate_boxes.emplace_back(
-                                    r[box_idx], box_idx, scoresPtr[box_idx], 0, batch, class_idx);
-                            }
-                        }
+                std::vector<BoxInfo> selected;
+                // Get the next box with top score, filter by iou_threshold
 
-                        std::priority_queue<BoxInfo> sorted_boxes(std::less<BoxInfo>(),
-                                                                  std::move(candidate_boxes));
+                BoxInfo next_candidate;
+                float original_score;
 
-                        std::vector<BoxInfo> selected;
-                        // Get the next box with top score, filter by iou_threshold
+                while (!sorted_boxes.empty() && selected.size() < boxes_per_class)
+                {
+                    next_candidate = sorted_boxes.top();
+                    original_score = next_candidate.score;
+                    sorted_boxes.pop();
+
+                    bool should_hard_suppress = false;
+                    for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
+                         j >= next_candidate.suppress_begin_index;
+                         --j)
+                    {
+                        float iou =
+                            intersectionOverUnion(next_candidate.box, selected[j].box);
+                        next_candidate.score *= func(iou);
 
-                        BoxInfo next_candidate;
-                        float original_score;
+                        if (iou >= iou_threshold)
+                        {
+                            should_hard_suppress = true;
+                            break;
+                        }
 
-                        while (!sorted_boxes.empty() && selected.size() < boxes_per_class)
+                        if (next_candidate.score <= score_threshold)
                         {
-                            next_candidate = sorted_boxes.top();
-                            original_score = next_candidate.score;
-                            sorted_boxes.pop();
-
-                            bool should_hard_suppress = false;
-                            for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
-                                 j >= next_candidate.suppress_begin_index;
-                                 --j)
-                            {
-                                float iou =
-                                    intersectionOverUnion(next_candidate.box, selected[j].box);
-                                next_candidate.score *= func(iou);
-
-                                if (iou >= iou_threshold)
-                                {
-                                    should_hard_suppress = true;
-                                    break;
-                                }
-
-                                if (next_candidate.score <= score_threshold)
-                                {
-                                    break;
-                                }
-                            }
-
-                            next_candidate.suppress_begin_index = selected.size();
-
-                            if (!should_hard_suppress)
-                            {
-                                if (next_candidate.score == original_score)
-                                {
-                                    selected.push_back(next_candidate);
-                                    continue;
-                                }
-                                if (next_candidate.score > score_threshold)
-                                {
-                                    sorted_boxes.push(next_candidate);
-                                }
-                            }
+                            break;
                         }
+                    }
 
-                        for (const auto& box_info : selected)
+                    next_candidate.suppress_begin_index = selected.size();
+
+                    if (!should_hard_suppress)
+                    {
+                        if (next_candidate.score == original_score)
                         {
-                            filteredBoxes.push_back(box_info);
+                            selected.push_back(next_candidate);
+                            continue;
+                        }
+                        if (next_candidate.score > score_threshold)
+                        {
+                            sorted_boxes.push(next_candidate);
                         }
                     }
                 }
 
-                if (sort_result_descending)
+                for (const auto& box_info : selected)
                 {
-                    std::sort(filteredBoxes.begin(),
-                              filteredBoxes.end(),
-                              [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
+                    filteredBoxes.push_back(box_info);
                 }
+            }
+        }
 
-                size_t max_num_of_selected_indices = selected_indices_shape[0];
-                size_t output_size = std::min(filteredBoxes.size(), max_num_of_selected_indices);
+        if (sort_result_descending)
+        {
+            std::sort(filteredBoxes.begin(),
+                      filteredBoxes.end(),
+                      [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
+        }
 
-                *valid_outputs = output_size;
+        size_t max_num_of_selected_indices = selected_indices_shape[0];
+        size_t output_size = std::min(filteredBoxes.size(), max_num_of_selected_indices);
 
-                size_t idx;
-                for (idx = 0; idx < output_size; idx++)
+        *valid_outputs = output_size;
+
+        size_t idx;
+        for (idx = 0; idx < output_size; idx++)
+        {
+            const auto& box_info = filteredBoxes[idx];
+            SelectedIndex selected_index{
+                box_info.batch_index, box_info.class_index, box_info.index};
+            SelectedScore selected_score{static_cast<float>(box_info.batch_index),
+                                         static_cast<float>(box_info.class_index),
+                                         box_info.score};
+
+            selected_indices_ptr[idx] = selected_index;
+            selected_scores_ptr[idx] = selected_score;
+        }
+
+        SelectedIndex selected_index_filler{0, 0, 0};
+        SelectedScore selected_score_filler{0.0f, 0.0f, 0.0f};
+        for (; idx < max_num_of_selected_indices; idx++)
+        {
+            selected_indices_ptr[idx] = selected_index_filler;
+            selected_scores_ptr[idx] = selected_score_filler;
+        }
+    }
+}
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace reference
+        {
+
+            void non_max_suppression(const op::v5::NonMaxSuppression* nms5,
+                                     const std::vector<std::shared_ptr<HostTensor>>& outputs,
+                                     const std::vector<std::shared_ptr<HostTensor>>& inputs)
+            {
+                int64_t max_output_boxes_per_class = nms5->max_boxes_output_from_input();
+                float iou_threshold = nms5->iou_threshold_from_input();
+                float score_threshold = nms5->score_threshold_from_input();
+                float soft_nms_sigma = nms5->soft_nms_sigma_from_input();
+
+                auto selected_indices_shape =
+                    infer_selected_indices_shape(inputs, max_output_boxes_per_class);
+                Shape out_shape = selected_indices_shape.to_shape();
+
+                Shape boxes_shape = inputs[boxes_port]->get_shape();
+                Shape scores_shape = inputs[scores_port]->get_shape();
+
+                auto boxes_data =
+                    prepare_boxes_data(inputs[boxes_port], boxes_shape, nms5->get_box_encoding());
+                auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
+
+                size_t out_shape_size = shape_size(out_shape);
+
+                std::vector<int64_t> selected_indices(out_shape_size);
+                std::vector<float> selected_scores(out_shape_size);
+                int64_t valid_outputs = 0;
+
+                eval(boxes_data.data(),
+                     boxes_shape,
+                     scores_data.data(),
+                     scores_shape,
+                     max_output_boxes_per_class,
+                     iou_threshold,
+                     score_threshold,
+                     soft_nms_sigma,
+                     selected_indices.data(),
+                     out_shape,
+                     selected_scores.data(),
+                     out_shape,
+                     &valid_outputs,
+                     nms5->get_sort_result_descending());
+
+                auto output_type = nms5->get_output_type();
+
+                outputs[0]->set_element_type(output_type);
+                outputs[0]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
+
+                size_t num_of_outputs = outputs.size();
+
+                auto selected_scores_type =
+                    (inputs.size() < 4) ? element::f32 : inputs[3]->get_element_type();
+
+                if (num_of_outputs >= 2)
                 {
-                    const auto& box_info = filteredBoxes[idx];
-                    SelectedIndex selected_index{
-                        box_info.batch_index, box_info.class_index, box_info.index};
-                    SelectedScore selected_score{static_cast<float>(box_info.batch_index),
-                                                 static_cast<float>(box_info.class_index),
-                                                 box_info.score};
-
-                    selected_indices_ptr[idx] = selected_index;
-                    selected_scores_ptr[idx] = selected_score;
+                    outputs[1]->set_element_type(selected_scores_type);
+                    outputs[1]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
                 }
 
-                SelectedIndex selected_index_filler{0, 0, 0};
-                SelectedScore selected_score_filler{0.0f, 0.0f, 0.0f};
-                for (; idx < max_num_of_selected_indices; idx++)
+                if (num_of_outputs >= 3)
                 {
-                    selected_indices_ptr[idx] = selected_index_filler;
-                    selected_scores_ptr[idx] = selected_score_filler;
+                    outputs[2]->set_element_type(output_type);
+                    outputs[2]->set_shape(Shape{1});
                 }
+
+                nms_postprocessing(outputs,
+                                output_type,
+                                selected_indices,
+                                selected_scores,
+                                valid_outputs,
+                                selected_scores_type);
             }
         }
     }
diff --git a/ngraph/test/backend/non_max_suppression.in.cpp b/ngraph/test/backend/non_max_suppression.in.cpp
index 2c9514d9538db2..e258d272e41dcd 100644
--- a/ngraph/test/backend/non_max_suppression.in.cpp
+++ b/ngraph/test/backend/non_max_suppression.in.cpp
@@ -42,7 +42,7 @@ using namespace ngraph;
 
 static string s_manifest = "${MANIFEST}";
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format)
 {
     std::vector<float> boxes_data = {0.5, 0.5,  1.0, 1.0, 0.5, 0.6,   1.0, 1.0,
                                      0.5, 0.4,  1.0, 1.0, 0.5, 10.5,  1.0, 1.0,
@@ -105,7 +105,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_center_point_box_format_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_flipped_coordinates_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_flipped_coordinates)
 {
     std::vector<float> boxes_data = {1.0, 1.0,  0.0, 0.0,  0.0, 0.1,   1.0, 1.1,
                                      0.0, 0.9,  1.0, -0.1, 0.0, 10.0,  1.0, 11.0,
@@ -168,7 +168,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_flipped_coordinates_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_identical_boxes_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_identical_boxes)
 {
     std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0,
                                      1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0,
@@ -232,7 +232,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_identical_boxes_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_limit_output_size_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_limit_output_size)
 {
     std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
                                      0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
@@ -295,7 +295,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_limit_output_size_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_single_box_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_single_box)
 {
     std::vector<float> boxes_data = {0.0, 0.0, 1.0, 1.0};
 
@@ -356,7 +356,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_single_box_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU)
 {
     std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
                                      0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
@@ -419,7 +419,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_and_scores_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_and_scores)
 {
     std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
                                      0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
@@ -482,7 +482,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_suppress_by_IOU_and_scores_backen
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_batches_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_batches)
 {
     std::vector<float> boxes_data = {
         0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,   0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
@@ -548,7 +548,7 @@ NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_batches_backend)
     EXPECT_EQ(expected_valid_outputs, valid_outputs_value);
 }
 
-NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_classes_backend)
+NGRAPH_TEST(${BACKEND_NAME}, nonmaxsuppression_two_classes)
 {
     std::vector<float> boxes_data = {0.0, 0.0,  1.0, 1.0,  0.0, 0.1,   1.0, 1.1,
                                      0.0, -0.1, 1.0, 0.9,  0.0, 10.0,  1.0, 11.0,
diff --git a/ngraph/test/runtime/ie/unit_test.manifest b/ngraph/test/runtime/ie/unit_test.manifest
index 18de0871915153..97bb32c0f03730 100644
--- a/ngraph/test/runtime/ie/unit_test.manifest
+++ b/ngraph/test/runtime/ie/unit_test.manifest
@@ -1140,15 +1140,15 @@ IE_CPU.log_softmax_3d_axis_0
 IE_CPU.log_softmax_3d_axis_neg3
 
 # Unsupported NMS-5
-IE_CPU.nonmaxsuppression_center_point_box_format_backend
-IE_CPU.nonmaxsuppression_flipped_coordinates_backend
-IE_CPU.nonmaxsuppression_identical_boxes_backend
-IE_CPU.nonmaxsuppression_limit_output_size_backend
-IE_CPU.nonmaxsuppression_single_box_backend
-IE_CPU.nonmaxsuppression_suppress_by_IOU_backend
-IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores_backend
-IE_CPU.nonmaxsuppression_two_batches_backend
-IE_CPU.nonmaxsuppression_two_classes_backend
+IE_CPU.nonmaxsuppression_center_point_box_format
+IE_CPU.nonmaxsuppression_flipped_coordinates
+IE_CPU.nonmaxsuppression_identical_boxes
+IE_CPU.nonmaxsuppression_limit_output_size
+IE_CPU.nonmaxsuppression_single_box
+IE_CPU.nonmaxsuppression_suppress_by_IOU
+IE_CPU.nonmaxsuppression_suppress_by_IOU_and_scores
+IE_CPU.nonmaxsuppression_two_batches
+IE_CPU.nonmaxsuppression_two_classes
 
 #-------------------------------------------------------------------------------
 #
diff --git a/ngraph/test/runtime/interpreter/int_executable.cpp b/ngraph/test/runtime/interpreter/int_executable.cpp
index d72ad211f535ff..a0bfff3463a001 100644
--- a/ngraph/test/runtime/interpreter/int_executable.cpp
+++ b/ngraph/test/runtime/interpreter/int_executable.cpp
@@ -59,305 +59,6 @@ runtime::interpreter::OP_TYPEID runtime::interpreter::INTExecutable::get_typeid(
     return rc;
 }
 
-using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
-
-namespace
-{
-    constexpr size_t boxes_port = 0;
-    constexpr size_t scores_port = 1;
-    constexpr size_t max_output_boxes_port = 2;
-    constexpr size_t iou_threshold_port = 3;
-    constexpr size_t score_threshold_port = 4;
-    constexpr size_t soft_nms_sigma_port = 5;
-
-    PartialShape
-        infer_selected_indices_shape(const std::vector<std::shared_ptr<HostTensor>>& inputs,
-                                     int64_t max_output_boxes_per_class)
-    {
-        const auto boxes_ps = inputs[boxes_port]->get_partial_shape();
-        const auto scores_ps = inputs[scores_port]->get_partial_shape();
-
-        // NonMaxSuppression produces triplets
-        // that have the following format: [batch_index, class_index, box_index]
-        PartialShape result = {Dimension::dynamic(), 3};
-
-        if (boxes_ps.rank().is_static() && scores_ps.rank().is_static())
-        {
-            const auto num_boxes_boxes = boxes_ps[1];
-            if (num_boxes_boxes.is_static() && scores_ps[0].is_static() && scores_ps[1].is_static())
-            {
-                const auto num_boxes = num_boxes_boxes.get_length();
-                const auto num_classes = scores_ps[1].get_length();
-
-                result[0] = std::min(num_boxes, max_output_boxes_per_class) * num_classes *
-                            scores_ps[0].get_length();
-            }
-        }
-
-        return result;
-    }
-
-    void normalize_corner(float* boxes, const Shape& boxes_shape)
-    {
-        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
-        for (size_t i = 0; i < total_num_of_boxes; ++i)
-        {
-            float* current_box = boxes + 4 * i;
-
-            float y1 = current_box[0];
-            float x1 = current_box[1];
-            float y2 = current_box[2];
-            float x2 = current_box[3];
-
-            float ymin = std::min(y1, y2);
-            float ymax = std::max(y1, y2);
-            float xmin = std::min(x1, x2);
-            float xmax = std::max(x1, x2);
-
-            current_box[0] = ymin;
-            current_box[1] = xmin;
-            current_box[2] = ymax;
-            current_box[3] = xmax;
-        }
-    }
-
-    void normalize_center(float* boxes, const Shape& boxes_shape)
-    {
-        size_t total_num_of_boxes = shape_size(boxes_shape) / 4;
-        for (size_t i = 0; i < total_num_of_boxes; ++i)
-        {
-            float* current_box = boxes + 4 * i;
-
-            float x_center = current_box[0];
-            float y_center = current_box[1];
-            float width = current_box[2];
-            float height = current_box[3];
-
-            float y1 = y_center - height / 2.0;
-            float x1 = x_center - width / 2.0;
-            float y2 = y_center + height / 2.0;
-            float x2 = x_center + width / 2.0;
-
-            current_box[0] = y1;
-            current_box[1] = x1;
-            current_box[2] = y2;
-            current_box[3] = x2;
-        }
-    }
-
-    void normalize_box_encoding(float* boxes,
-                                const Shape& boxes_shape,
-                                const V5BoxEncoding box_encoding)
-    {
-        if (box_encoding == V5BoxEncoding::CORNER)
-        {
-            normalize_corner(boxes, boxes_shape);
-        }
-        else
-        {
-            normalize_center(boxes, boxes_shape);
-        }
-    }
-
-    std::vector<float> get_floats(const std::shared_ptr<HostTensor>& input, const Shape& shape)
-    {
-        size_t input_size = shape_size(shape);
-        std::vector<float> result(input_size);
-
-        switch (input->get_element_type())
-        {
-        case element::Type_t::bf16:
-        {
-            bfloat16* p = input->get_data_ptr<bfloat16>();
-            for (size_t i = 0; i < input_size; ++i)
-            {
-                result[i] = float(p[i]);
-            }
-        }
-        break;
-        case element::Type_t::f16:
-        {
-            float16* p = input->get_data_ptr<float16>();
-            for (size_t i = 0; i < input_size; ++i)
-            {
-                result[i] = float(p[i]);
-            }
-        }
-        break;
-        case element::Type_t::f32:
-        {
-            float* p = input->get_data_ptr<float>();
-            memcpy(result.data(), p, input_size * sizeof(float));
-        }
-        break;
-        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
-        }
-
-        return result;
-    }
-
-    std::vector<float> prepare_boxes_data(const std::shared_ptr<HostTensor>& boxes,
-                                          const Shape& boxes_shape,
-                                          const V5BoxEncoding box_encoding)
-    {
-        auto result = get_floats(boxes, boxes_shape);
-        normalize_box_encoding(result.data(), boxes_shape, box_encoding);
-        return result;
-    }
-
-    std::vector<float> prepare_scores_data(const std::shared_ptr<HostTensor>& scores,
-                                           const Shape& scores_shape)
-    {
-        auto result = get_floats(scores, scores_shape);
-        return result;
-    }
-
-    void nms_postprocessing(const std::vector<std::shared_ptr<HostTensor>>& outputs,
-                            const ngraph::element::Type output_type,
-                            const std::vector<int64_t>& selected_indices,
-                            const std::vector<float>& selected_scores,
-                            int64_t valid_outputs,
-                            const ngraph::element::Type selected_scores_type)
-    {
-        size_t num_of_outputs = outputs.size();
-        size_t selected_size = valid_outputs * 3;
-
-        if (output_type == ngraph::element::i64)
-        {
-            int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
-            memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
-        }
-        else
-        {
-            int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
-            for (size_t i = 0; i < selected_size; ++i)
-            {
-                indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
-            }
-        }
-
-        if (num_of_outputs < 2)
-        {
-            return;
-        }
-
-        size_t selected_scores_size = selected_scores.size();
-
-        switch (selected_scores_type)
-        {
-        case element::Type_t::bf16:
-        {
-            bfloat16* scores_ptr = outputs[1]->get_data_ptr<bfloat16>();
-            for (size_t i = 0; i < selected_scores_size; ++i)
-            {
-                scores_ptr[i] = bfloat16(selected_scores[i]);
-            }
-        }
-        break;
-        case element::Type_t::f16:
-        {
-            float16* scores_ptr = outputs[1]->get_data_ptr<float16>();
-            for (size_t i = 0; i < selected_scores_size; ++i)
-            {
-                scores_ptr[i] = float16(selected_scores[i]);
-            }
-        }
-        break;
-        case element::Type_t::f32:
-        {
-            float* scores_ptr = outputs[1]->get_data_ptr<float>();
-            memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
-        }
-        break;
-        default:;
-        }
-
-        if (num_of_outputs < 3)
-        {
-            return;
-        }
-
-        if (output_type == ngraph::element::i64)
-        {
-            int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
-            *valid_outputs_ptr = valid_outputs;
-        }
-        else
-        {
-            int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
-            *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
-        }
-    }
-}
-
-void runtime::interpreter::run_nms5(const op::v5::NonMaxSuppression* nms5,
-                                    const std::vector<std::shared_ptr<HostTensor>>& outputs,
-                                    const std::vector<std::shared_ptr<HostTensor>>& inputs)
-{
-    int64_t max_output_boxes_per_class = nms5->max_boxes_output_from_input();
-    float iou_threshold = nms5->iou_threshold_from_input();
-    float score_threshold = nms5->score_threshold_from_input();
-    float soft_nms_sigma = nms5->soft_nms_sigma_from_input();
-
-    auto selected_indices_shape = infer_selected_indices_shape(inputs, max_output_boxes_per_class);
-    Shape out_shape = selected_indices_shape.to_shape();
-
-    Shape boxes_shape = inputs[boxes_port]->get_shape();
-    Shape scores_shape = inputs[scores_port]->get_shape();
-
-    auto boxes_data = prepare_boxes_data(inputs[boxes_port], boxes_shape, nms5->get_box_encoding());
-    auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
-
-    size_t out_shape_size = shape_size(out_shape);
-
-    std::vector<int64_t> selected_indices(out_shape_size);
-    std::vector<float> selected_scores(out_shape_size);
-    int64_t valid_outputs = 0;
-
-    runtime::reference::non_max_suppression(boxes_data.data(),
-                                            boxes_shape,
-                                            scores_data.data(),
-                                            scores_shape,
-                                            max_output_boxes_per_class,
-                                            iou_threshold,
-                                            score_threshold,
-                                            soft_nms_sigma,
-                                            selected_indices.data(),
-                                            out_shape,
-                                            selected_scores.data(),
-                                            out_shape,
-                                            &valid_outputs,
-                                            nms5->get_sort_result_descending());
-
-    auto output_type = nms5->get_output_type();
-
-    outputs[0]->set_element_type(output_type);
-    outputs[0]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
-
-    size_t num_of_outputs = outputs.size();
-
-    auto selected_scores_type = (inputs.size() < 4) ? element::f32 : inputs[3]->get_element_type();
-
-    if (num_of_outputs >= 2)
-    {
-        outputs[1]->set_element_type(selected_scores_type);
-        outputs[1]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
-    }
-
-    if (num_of_outputs >= 3)
-    {
-        outputs[2]->set_element_type(output_type);
-        outputs[2]->set_shape(Shape{1});
-    }
-
-    nms_postprocessing(outputs,
-                       output_type,
-                       selected_indices,
-                       selected_scores,
-                       valid_outputs,
-                       selected_scores_type);
-}
-
 runtime::interpreter::INTExecutable::INTExecutable(const shared_ptr<Function>& function,
                                                    bool enable_performance_collection)
     : m_is_compiled{true}
diff --git a/ngraph/test/runtime/interpreter/int_executable.hpp b/ngraph/test/runtime/interpreter/int_executable.hpp
index 1107cfc4a1ac6b..f86bc9574c818a 100644
--- a/ngraph/test/runtime/interpreter/int_executable.hpp
+++ b/ngraph/test/runtime/interpreter/int_executable.hpp
@@ -1434,7 +1434,7 @@ class INTERPRETER_BACKEND_API ngraph::runtime::interpreter::INTExecutable : publ
             const op::v5::NonMaxSuppression* nms =
                 static_cast<const op::v5::NonMaxSuppression*>(&node);
 
-            run_nms5(nms, out, args);
+            reference::non_max_suppression(nms, out, args);
             break;
         }
 

From ab5a374d564aaf8d0afea46cd21bdc2c4bd56141 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 16:20:47 +0300
Subject: [PATCH 142/173] Code style fixes.

---
 .../runtime/reference/non_max_suppression.cpp | 20 ++++++++-----------
 1 file changed, 8 insertions(+), 12 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 9b2b50db684ce8..a6a113c7ca04f4 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -390,10 +390,8 @@ namespace
         int64_t num_classes = static_cast<int64_t>(scores_data_shape[1]);
         int64_t num_boxes = static_cast<int64_t>(boxes_data_shape[1]);
 
-        SelectedIndex* selected_indices_ptr =
-            reinterpret_cast<SelectedIndex*>(selected_indices);
-        SelectedScore* selected_scores_ptr =
-            reinterpret_cast<SelectedScore*>(selected_scores);
+        SelectedIndex* selected_indices_ptr = reinterpret_cast<SelectedIndex*>(selected_indices);
+        SelectedScore* selected_scores_ptr = reinterpret_cast<SelectedScore*>(selected_scores);
 
         size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
 
@@ -443,8 +441,7 @@ namespace
                          j >= next_candidate.suppress_begin_index;
                          --j)
                     {
-                        float iou =
-                            intersectionOverUnion(next_candidate.box, selected[j].box);
+                        float iou = intersectionOverUnion(next_candidate.box, selected[j].box);
                         next_candidate.score *= func(iou);
 
                         if (iou >= iou_threshold)
@@ -524,7 +521,6 @@ namespace ngraph
     {
         namespace reference
         {
-
             void non_max_suppression(const op::v5::NonMaxSuppression* nms5,
                                      const std::vector<std::shared_ptr<HostTensor>>& outputs,
                                      const std::vector<std::shared_ptr<HostTensor>>& inputs)
@@ -589,11 +585,11 @@ namespace ngraph
                 }
 
                 nms_postprocessing(outputs,
-                                output_type,
-                                selected_indices,
-                                selected_scores,
-                                valid_outputs,
-                                selected_scores_type);
+                                   output_type,
+                                   selected_indices,
+                                   selected_scores,
+                                   valid_outputs,
+                                   selected_scores_type);
             }
         }
     }

From 0b690a46091be69dc8b4d33f91c61fcdeb7baa62 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 16:57:53 +0300
Subject: [PATCH 143/173] Some fixes in tests for the transformation NMS-5 ->
 NMSIE3.

---
 .../transformations/convert_nms5_test.cpp     | 20 -------------------
 1 file changed, 20 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 038598101dd1bd..9f1e538637324c 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -40,8 +40,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticSixInputs) {
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
         ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
     }
 
     {
@@ -71,8 +69,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticSixInputs) {
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
     }
 
     auto res = compare_functions(f, f_ref);
@@ -98,8 +94,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
         ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
     }
 
     {
@@ -129,8 +123,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
     }
 
     auto res = compare_functions(f, f_ref);
@@ -155,8 +147,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
         ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
     }
 
     {
@@ -186,8 +176,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
     }
 
     auto res = compare_functions(f, f_ref);
@@ -211,8 +199,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
         ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
     }
 
     {
@@ -242,8 +228,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
     }
 
     auto res = compare_functions(f, f_ref);
@@ -266,8 +250,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
         manager.run_passes(f);
         ASSERT_NO_THROW(check_rt_info(f));
         ASSERT_TRUE(f->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f->get_output_shape(2), (Shape{1}));
     }
 
     {
@@ -297,8 +279,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
         ASSERT_TRUE(f_ref->get_output_partial_shape(0).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        ASSERT_TRUE(f_ref->get_output_partial_shape(1).same_scheme(PartialShape{Dimension::dynamic(), 3}));
-        EXPECT_EQ(f_ref->get_output_shape(2), (Shape{1}));
     }
 
     auto res = compare_functions(f, f_ref);

From 211963362f867e1063606acc1be00f7fa48a8856 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 23 Oct 2020 17:58:35 +0300
Subject: [PATCH 144/173] Replaced precision i64 -> i32 for some constants in
 tests for the transformation NMS-5 -> NMSIE3.

---
 .../inference_engine/transformations/convert_nms5_test.cpp  | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 9f1e538637324c..c3aba0880fe847 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -255,7 +255,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
     {
         auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1000, 4});
         auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 1, 1000});
-        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i32, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
@@ -521,7 +521,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1TwoInputs) {
     {
         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape::dynamic());
-        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i32, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
@@ -781,7 +781,7 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2TwoInputs) {
     {
         auto boxes = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1000, 4});
         auto scores = std::make_shared<opset5::Parameter>(element::f32, PartialShape{DYN, 1, 1000});
-        auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
+        auto max_output_boxes_per_class = opset5::Constant::create(element::i32, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});

From 5cbdf1f430df26b6686b61dbfe884feb2ef2c0bb Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Sat, 24 Oct 2020 17:32:31 +0300
Subject: [PATCH 145/173] Written creating CNNLayer for NMS-5.

---
 .../src/convert_function_to_cnn_network.cpp   | 10 ++
 .../src/ie_cnn_layer_builder_ngraph.cpp       | 42 ++++++++
 .../src/legacy_api/src/ngraph_ops/nms_ie.cpp  |  2 +-
 .../non_max_suppression_tests.cpp             | 99 +++++++++++--------
 4 files changed, 110 insertions(+), 43 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp b/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
index 0cf25b94f7df74..8797f651479182 100644
--- a/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
+++ b/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
@@ -535,6 +535,15 @@ InferenceEngine::details::CNNLayerCreator::CNNLayerCreator(const std::shared_ptr
 
     });
 
+    addSpecificCreator({"NonMaxSuppressionIE3"}, [](const std::shared_ptr<::ngraph::Node>& node,
+        const std::map<std::string, std::string>& params) -> CNNLayerPtr {
+        LayerParams attrs = {node->get_friendly_name(), "NonMaxSuppressionIE3",
+            details::convertPrecision(node->get_output_element_type(0))};
+        auto res = std::make_shared<InferenceEngine::CNNLayer>(attrs);
+        res->params = params;
+        return res;
+    });
+
     addSpecificCreator({"NonMaxSuppressionIE"}, [](const std::shared_ptr<::ngraph::Node>& node,
                                                  const std::map<std::string, std::string>& params) -> CNNLayerPtr {
         LayerParams attrs = {node->get_friendly_name(), "NonMaxSuppression", details::convertPrecision(node->get_output_element_type(0))};
@@ -814,6 +823,7 @@ void convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function
                 std::make_shared<Builder::NodeConverter<::ngraph::op::v1::LogicalNot>>(),
                 std::make_shared<Builder::NodeConverter<::ngraph::op::ShuffleChannels>>(),
                 std::make_shared<Builder::NodeConverter<::ngraph::op::v4::Interpolate>>(),
+                std::make_shared<Builder::NodeConverter<::ngraph::op::v5::NonMaxSuppression>>(),
                 std::make_shared<Builder::NodeConverter<::ExecGraphInfoSerialization::ExecutionNode>>(),
         };
         CNNLayerPtr result;
diff --git a/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp b/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
index 07f74fe52ba829..838468b1f76428 100644
--- a/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
+++ b/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
@@ -1726,6 +1726,48 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v0::Interpolate>::createLayer(const std:
     THROW_IE_EXCEPTION << "Interpolate operation should be converted to Interp";
 }
 
+template <>
+CNNLayer::Ptr NodeConverter<ngraph::op::v5::NonMaxSuppression>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
+    LayerParams params = {layer->get_friendly_name(), "NonMaxSuppressionIE3",
+                          details::convertPrecision(layer->get_output_element_type(0))};
+    auto castedLayer = ngraph::as_type_ptr<ngraph::op::v5::NonMaxSuppression>(layer);
+    if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
+
+    auto box_encoding = castedLayer->get_box_encoding();
+
+    auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
+
+    switch (box_encoding) {
+        case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CORNER:
+            res->params["box_encoding"] = "corner";
+            break;
+        case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CENTER:
+            res->params["box_encoding"] = "center";
+            break;
+        default:
+            THROW_IE_EXCEPTION << "Unsupported box encoding for NonMaxSuppression op";
+            break;
+    }
+
+    auto output_type = details::convertPrecision(castedLayer->get_output_type());
+    std::string output_type_str;
+    switch (output_type) {
+    case Precision::I32:
+        output_type_str = "I32";
+        break;
+    case Precision::I64:
+        output_type_str = "I64";
+        break;
+    default:
+        THROW_IE_EXCEPTION << "Unsupported output type";
+    }
+    res->params["output_type"] = output_type_str;
+
+    bool sort_result_descending = castedLayer->get_sort_result_descending();
+    res->params["sort_result_descending"] = sort_result_descending ? "1" : "0";
+    return res;
+}
+
 template <>
 CNNLayer::Ptr NodeConverter<ngraph::op::v4::Interpolate>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
     LayerParams params = {layer->get_friendly_name(), "Interpolate",
diff --git a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
index 347e939505290b..0bee3d569db920 100644
--- a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
@@ -102,7 +102,7 @@ void op::NonMaxSuppressionIE2::validate_and_infer_types() {
     set_output_type(0, nms->output(0).get_element_type(), nms->output(0).get_partial_shape());
 }
 
-NGRAPH_RTTI_DEFINITION(op::NonMaxSuppressionIE3, "NonMaxSuppressionIE", 3);
+NGRAPH_RTTI_DEFINITION(op::NonMaxSuppressionIE3, "NonMaxSuppressionIE3", 3);
 
 op::NonMaxSuppressionIE3::NonMaxSuppressionIE3(const Output<Node>& boxes,
                                                const Output<Node>& scores,
diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index bc2f3aa651eee3..ab6e69281097dc 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -4,12 +4,12 @@
 
 #include <string>
 #include "ngraph_reader_tests.hpp"
-/*
-TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
+
+TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     std::string model = R"V0G0N(
 <net name="Network" version="10">
     <layers>
-        <layer id="0" name="in1" type="Parameter" >
+        <layer id="0" name="in1" type="Parameter" version="opset1">
             <data element_type="f32" shape="1,15130,4"/>
             <output>
                 <port id="0" precision="FP32">
@@ -19,7 +19,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
                 </port>
             </output>
         </layer>
-        <layer id="1" name="in2" type="Parameter" >
+        <layer id="1" name="in2" type="Parameter" version="opset1">
             <data element_type="f32" shape="1,80,15130"/>
             <output>
                 <port id="0" precision="FP32">
@@ -29,55 +29,62 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
                 </port>
             </output>
         </layer>
-        <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const">
+        <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const" version="opset1">
             <data offset="0" size="8"/>
             <output>
-                <port id="0"/>
+                <port id="0" precision="I64"/>
             </output>
         </layer>
-        <layer id="3" name="iou_threshold" precision="FP32" type="Const">
+        <layer id="3" name="iou_threshold" precision="FP32" type="Const" version="opset1">
             <data offset="8" size="4"/>
             <output>
-                <port id="0"/>
+                <port id="0" precision="FP32"/>
             </output>
         </layer>
-        <layer id="4" name="score_threshold" precision="FP32" type="Const">
+        <layer id="4" name="score_threshold" precision="FP32" type="Const" version="opset1">
             <data offset="12" size="4"/>
             <output>
-                <port id="0"/>
+                <port id="0" precision="FP32"/>
+            </output>
+        </layer>
+        <layer id="5" name="soft_nms_sigma" precision="FP32" type="Const" version="opset1">
+            <data offset="16" size="4"/>
+            <output>
+                <port id="0" precision="FP32"/>
             </output>
         </layer>
-        <layer id="5" name="nms" type="NonMaxSuppression">
+        <layer id="6" name="nms" type="NonMaxSuppression" version="opset5">
             <data box_encoding="corner" sort_result_descending="0"/>
             <input>
-                <port id="0">
+                <port id="0" precision="FP32">
                     <dim>1</dim>
                     <dim>15130</dim>
                     <dim>4</dim>
                 </port>
-                <port id="1">
+                <port id="1" precision="FP32">
                     <dim>1</dim>
                     <dim>80</dim>
                     <dim>15130</dim>
                 </port>
-                <port id="2"/>
-                <port id="3"/>
-                <port id="4"/>
+                <port id="2" precision="I64"/>
+                <port id="3" precision="FP32"/>
+                <port id="4" precision="FP32"/>
+                <port id="5" precision="FP32"/>
             </input>
             <output>
-                <port id="5" precision="I64">
+                <port id="6" precision="I64">
                     <dim>15130</dim>
                     <dim>3</dim>
                 </port>
             </output>
         </layer>
-        <layer id="6" name="mul" type="Multiply">
+        <layer id="7" name="mul" type="Multiply" version="opset1">
             <input>
-                <port id="0">
+                <port id="0" precision="I64">
                     <dim>15130</dim>
                     <dim>3</dim>
                 </port>
-                <port id="1">
+                <port id="1" precision="I64">
                     <dim>15130</dim>
                     <dim>3</dim>
                 </port>
@@ -89,7 +96,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
                 </port>
             </output>
         </layer>
-        <layer id="7" name="output" type="Result" >
+        <layer id="8" name="output" type="Result" version="opset1">
             <input>
                 <port id="0" precision="I64">
                     <dim>15130</dim>
@@ -99,14 +106,15 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
         </layer>
     </layers>
     <edges>
-        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
-        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
-        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
-        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
-        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
-        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
-        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
-        <edge from-layer="6" from-port="2" to-layer="7" to-port="0"/>
+        <edge from-layer="0" from-port="0" to-layer="6" to-port="0"/>
+        <edge from-layer="1" from-port="0" to-layer="6" to-port="1"/>
+        <edge from-layer="2" from-port="0" to-layer="6" to-port="2"/>
+        <edge from-layer="3" from-port="0" to-layer="6" to-port="3"/>
+        <edge from-layer="4" from-port="0" to-layer="6" to-port="4"/>
+        <edge from-layer="5" from-port="0" to-layer="6" to-port="5"/>
+        <edge from-layer="6" from-port="6" to-layer="7" to-port="0"/>
+        <edge from-layer="6" from-port="6" to-layer="7" to-port="1"/>
+        <edge from-layer="7" from-port="2" to-layer="8" to-port="0"/>
     </edges>
 </net>
 )V0G0N";
@@ -151,7 +159,13 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
                 <port id="0"/>
             </output>
         </layer>
-        <layer id="5" name="nms" type="NonMaxSuppression" precision="I32">
+        <layer id="5" name="soft_nms_sigma" precision="FP32" type="Const">
+            <data offset="16" size="4"/>
+            <output>
+                <port id="0"/>
+            </output>
+        </layer>
+        <layer id="6" name="nms" type="NonMaxSuppressionIE3" precision="I32">
             <data box_encoding="corner" sort_result_descending="0"/>
             <input>
                 <port id="0">
@@ -167,15 +181,16 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
                 <port id="2"/>
                 <port id="3"/>
                 <port id="4"/>
+                <port id="5"/>
             </input>
             <output>
-                <port id="5">
+                <port id="6">
                     <dim>15130</dim>
                     <dim>3</dim>
                 </port>
             </output>
         </layer>
-        <layer id="6" name="mul" type="Multiply" precision="I32">
+        <layer id="7" name="mul" type="Multiply" precision="I32">
             <input>
                 <port id="0">
                     <dim>15130</dim>
@@ -195,25 +210,25 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression) {
         </layer>
     </layers>
     <edges>
-        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
-        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
-        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
-        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
-        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
-        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
-        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
+        <edge from-layer="0" from-port="0" to-layer="6" to-port="0"/>
+        <edge from-layer="1" from-port="0" to-layer="6" to-port="1"/>
+        <edge from-layer="2" from-port="0" to-layer="6" to-port="2"/>
+        <edge from-layer="3" from-port="0" to-layer="6" to-port="3"/>
+        <edge from-layer="4" from-port="0" to-layer="6" to-port="4"/>
+        <edge from-layer="5" from-port="0" to-layer="6" to-port="5"/>
+        <edge from-layer="6" from-port="6" to-layer="7" to-port="0"/>
+        <edge from-layer="6" from-port="6" to-layer="7" to-port="1"/>
     </edges>
 </net>
 )V0G0N";
 
-    compareIRs(model, modelV5, 16, [](Blob::Ptr& weights) {
+    compareIRs(model, modelV5, 20, [](Blob::Ptr& weights) {
         auto * i64w = weights->buffer().as<int64_t*>();
         i64w[0] = 200;
 
         auto * fp32w = weights->buffer().as<float*>();
         fp32w[2] = 0.5;
         fp32w[3] = 0.05;
+        fp32w[4] = 0.0;
     });
 }
-
- */

From c19e54f0e08da62b69349bf1deb95996394fd4a3 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Sun, 25 Oct 2020 15:41:42 +0300
Subject: [PATCH 146/173] Added creating CNNLayer for NonMaxSuppressionIE3.

---
 .../src/ie_cnn_layer_builder_ngraph.cpp       | 32 ++++++++-
 .../non_max_suppression_tests.cpp             | 69 ++++++++++++++-----
 .../functional_test_utils/network_utils.cpp   | 17 +++++
 3 files changed, 97 insertions(+), 21 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp b/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
index 838468b1f76428..c066bfa00e5caa 100644
--- a/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
+++ b/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
@@ -1726,6 +1726,34 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v0::Interpolate>::createLayer(const std:
     THROW_IE_EXCEPTION << "Interpolate operation should be converted to Interp";
 }
 
+template <>
+CNNLayer::Ptr NodeConverter<ngraph::op::NonMaxSuppressionIE3>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
+    LayerParams params = {layer->get_friendly_name(), "NonMaxSuppressionIE3",
+                          details::convertPrecision(layer->get_output_element_type(0))};
+    auto castedLayer = ngraph::as_type_ptr<ngraph::op::NonMaxSuppressionIE3>(layer);
+    if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
+
+    auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
+    res->params["center_point_box"] = asString(castedLayer->m_center_point_box);
+    res->params["sort_result_descending"] = castedLayer->m_sort_result_descending ? "1" : "0";
+
+    auto output_type = details::convertPrecision(castedLayer->m_output_type);
+    std::string output_type_str;
+    switch (output_type) {
+    case Precision::I32:
+        output_type_str = "I32";
+        break;
+    case Precision::I64:
+        output_type_str = "I64";
+        break;
+    default:
+        THROW_IE_EXCEPTION << "Unsupported output type";
+    }
+    res->params["output_type"] = output_type_str;
+
+    return res;
+}
+
 template <>
 CNNLayer::Ptr NodeConverter<ngraph::op::v5::NonMaxSuppression>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
     LayerParams params = {layer->get_friendly_name(), "NonMaxSuppressionIE3",
@@ -1739,10 +1767,10 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v5::NonMaxSuppression>::createLayer(cons
 
     switch (box_encoding) {
         case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CORNER:
-            res->params["box_encoding"] = "corner";
+            res->params["center_point_box"] = "0";
             break;
         case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CENTER:
-            res->params["box_encoding"] = "center";
+            res->params["center_point_box"] = "1";
             break;
         default:
             THROW_IE_EXCEPTION << "Unsupported box encoding for NonMaxSuppression op";
diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index ab6e69281097dc..60f5bafdfd9dde 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -121,7 +121,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     std::string modelV5 = R"V0G0N(
 <net name="Network" version="5">
     <layers>
-        <layer id="0" name="in1" type="Input" >
+        <layer id="0" name="in1" precision="FP32" type="Input" >
             <data precision="I32"/>
             <output>
                 <port id="0">
@@ -131,7 +131,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </output>
         </layer>
-        <layer id="1" name="in2" type="Input" >
+        <layer id="1" name="in2" precision="FP32" type="Input" >
             <data precision="I32"/>
             <output>
                 <port id="0">
@@ -142,30 +142,46 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
             </output>
         </layer>
         <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const">
-            <data offset="0" size="8"/>
             <output>
-                <port id="0"/>
+                <port id="0">
+                    <dim>1</dim>
+                </port>
             </output>
+            <blobs>
+                <custom offset="0" size="8"/>
+            </blobs>
         </layer>
         <layer id="3" name="iou_threshold" precision="FP32" type="Const">
-            <data offset="8" size="4"/>
             <output>
-                <port id="0"/>
+                <port id="0">
+                    <dim>1</dim>
+                </port>
             </output>
+            <blobs>
+                <custom offset="8" size="4"/>
+            </blobs>
         </layer>
         <layer id="4" name="score_threshold" precision="FP32" type="Const">
-            <data offset="12" size="4"/>
             <output>
-                <port id="0"/>
+                <port id="0">
+                    <dim>1</dim>
+                </port>
             </output>
+            <blobs>
+                <custom offset="12" size="4"/>
+            </blobs>
         </layer>
         <layer id="5" name="soft_nms_sigma" precision="FP32" type="Const">
-            <data offset="16" size="4"/>
             <output>
-                <port id="0"/>
+                <port id="0">
+                    <dim>1</dim>
+                </port>
             </output>
+            <blobs>
+                <custom offset="16" size="4"/>
+            </blobs>
         </layer>
-        <layer id="6" name="nms" type="NonMaxSuppressionIE3" precision="I32">
+        <layer id="6" name="nms" type="NonMaxSuppressionIE3" precision="I64">
             <data box_encoding="corner" sort_result_descending="0"/>
             <input>
                 <port id="0">
@@ -178,32 +194,47 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                     <dim>80</dim>
                     <dim>15130</dim>
                 </port>
-                <port id="2"/>
-                <port id="3"/>
-                <port id="4"/>
-                <port id="5"/>
+                <port id="2">
+                    <dim>1</dim>
+                </port>
+                <port id="3">
+                    <dim>1</dim>
+                </port>
+                <port id="4">
+                    <dim>1</dim>
+                </port>
+                <port id="5">
+                    <dim>1</dim>
+                </port>
             </input>
             <output>
                 <port id="6">
-                    <dim>15130</dim>
+                    <dim>16000</dim>
+                    <dim>3</dim>
+                </port>
+                <port id="7">
+                    <dim>16000</dim>
                     <dim>3</dim>
                 </port>
+                <port id="8">
+                    <dim>1</dim>
+                </port>
             </output>
         </layer>
         <layer id="7" name="mul" type="Multiply" precision="I32">
             <input>
                 <port id="0">
-                    <dim>15130</dim>
+                    <dim>16000</dim>
                     <dim>3</dim>
                 </port>
                 <port id="1">
-                    <dim>15130</dim>
+                    <dim>16000</dim>
                     <dim>3</dim>
                 </port>
             </input>
             <output>
                 <port id="2">
-                    <dim>15130</dim>
+                    <dim>16000</dim>
                     <dim>3</dim>
                 </port>
             </output>
diff --git a/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp b/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
index 8f4a507eec94ed..83709dac917a5a 100644
--- a/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
+++ b/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
@@ -119,6 +119,23 @@ namespace FuncTestUtils {
             if (refLayer->outData[i]->getTensorDesc() != layer->outData[i]->getTensorDesc()) {
                 err_log.push_back("Layer " + layer->name + " and ref layer " + refLayer->name +
                                   " have different tensor desc for out Data");
+                std::cout << "For output " << i << "\n";
+                auto refDescriptor = refLayer->outData[i]->getTensorDesc();
+                auto layerDescriptor = layer->outData[i]->getTensorDesc();
+                std::cout << "ref precision:   " << refDescriptor.getPrecision() << "\n";
+                std::cout << "layer precision: " << layerDescriptor.getPrecision() << "\n";
+                auto refDims = refDescriptor.getDims();
+                auto layerDims = layerDescriptor.getDims();
+                std::cout <<"ref dims:    ";
+                for (size_t d : refDims) {
+                    std::cout << d << " ";
+                }
+                std::cout << "\n";
+                std::cout <<"layer dims:  ";
+                for (size_t d : layerDims) {
+                    std::cout << d << " ";
+                }
+                std::cout << "\n";
             }
             success = success && refLayer->outData[i]->getTensorDesc() == layer->outData[i]->getTensorDesc();
         }

From d9232f08e78c7931d115fcb23b3ebd4fd77da3ab Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Mon, 26 Oct 2020 15:21:24 +0300
Subject: [PATCH 147/173] some changes.

---
 .../non_max_suppression_tests.cpp             | 41 +++++++++++--------
 .../runtime/interpreter/int_executable.cpp    |  1 -
 .../runtime/interpreter/int_executable.hpp    |  5 ---
 3 files changed, 23 insertions(+), 24 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index 60f5bafdfd9dde..236858d1073a37 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -76,6 +76,13 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                     <dim>15130</dim>
                     <dim>3</dim>
                 </port>
+                <port id="7" precision="FP32">
+                    <dim>15130</dim>
+                    <dim>3</dim>
+                </port>
+                <port id="8" precision="I64">
+                    <dim>1</dim>
+                </port>
             </output>
         </layer>
         <layer id="7" name="mul" type="Multiply" version="opset1">
@@ -119,12 +126,11 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
 </net>
 )V0G0N";
     std::string modelV5 = R"V0G0N(
-<net name="Network" version="5">
+<net name="Network" version="7">
     <layers>
         <layer id="0" name="in1" precision="FP32" type="Input" >
-            <data precision="I32"/>
             <output>
-                <port id="0">
+                <port id="0" precision="FP32">
                     <dim>1</dim>
                     <dim>15130</dim>
                     <dim>4</dim>
@@ -132,9 +138,8 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
             </output>
         </layer>
         <layer id="1" name="in2" precision="FP32" type="Input" >
-            <data precision="I32"/>
             <output>
-                <port id="0">
+                <port id="0" precision="FP32">
                     <dim>1</dim>
                     <dim>80</dim>
                     <dim>15130</dim>
@@ -143,42 +148,42 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
         </layer>
         <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const">
             <output>
-                <port id="0">
+                <port id="0" precision="I64">
                     <dim>1</dim>
                 </port>
             </output>
             <blobs>
-                <custom offset="0" size="8"/>
+                <custom offset="0" precision="I64" size="8"/>
             </blobs>
         </layer>
         <layer id="3" name="iou_threshold" precision="FP32" type="Const">
             <output>
-                <port id="0">
+                <port id="0" precision="FP32">
                     <dim>1</dim>
                 </port>
             </output>
             <blobs>
-                <custom offset="8" size="4"/>
+                <custom offset="8" precision="FP32" size="4"/>
             </blobs>
         </layer>
         <layer id="4" name="score_threshold" precision="FP32" type="Const">
             <output>
-                <port id="0">
+                <port id="0" precision="FP32">
                     <dim>1</dim>
                 </port>
             </output>
             <blobs>
-                <custom offset="12" size="4"/>
+                <custom offset="12" precision="FP32" size="4"/>
             </blobs>
         </layer>
         <layer id="5" name="soft_nms_sigma" precision="FP32" type="Const">
             <output>
-                <port id="0">
+                <port id="0" precision="FP32">
                     <dim>1</dim>
                 </port>
             </output>
             <blobs>
-                <custom offset="16" size="4"/>
+                <custom offset="16" precision="FP32" size="4"/>
             </blobs>
         </layer>
         <layer id="6" name="nms" type="NonMaxSuppressionIE3" precision="I64">
@@ -208,20 +213,20 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </input>
             <output>
-                <port id="6">
+                <port id="6" precision="I64">
                     <dim>16000</dim>
                     <dim>3</dim>
                 </port>
-                <port id="7">
+                <port id="7" precision="FP32">
                     <dim>16000</dim>
                     <dim>3</dim>
                 </port>
-                <port id="8">
+                <port id="8" precision="I64">
                     <dim>1</dim>
                 </port>
             </output>
         </layer>
-        <layer id="7" name="mul" type="Multiply" precision="I32">
+        <layer id="7" name="mul" type="Multiply" precision="I64">
             <input>
                 <port id="0">
                     <dim>16000</dim>
@@ -233,7 +238,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </input>
             <output>
-                <port id="2">
+                <port id="2" precision="I64">
                     <dim>16000</dim>
                     <dim>3</dim>
                 </port>
diff --git a/ngraph/test/runtime/interpreter/int_executable.cpp b/ngraph/test/runtime/interpreter/int_executable.cpp
index a0bfff3463a001..186c5de48fd7a7 100644
--- a/ngraph/test/runtime/interpreter/int_executable.cpp
+++ b/ngraph/test/runtime/interpreter/int_executable.cpp
@@ -15,7 +15,6 @@
 //*****************************************************************************
 
 #include "int_executable.hpp"
-#include <algorithm>
 #include <cstring>
 #include "backend_manager.hpp"
 #include "ngraph/chrome_trace.hpp"
diff --git a/ngraph/test/runtime/interpreter/int_executable.hpp b/ngraph/test/runtime/interpreter/int_executable.hpp
index f86bc9574c818a..f11b798e3f9d7a 100644
--- a/ngraph/test/runtime/interpreter/int_executable.hpp
+++ b/ngraph/test/runtime/interpreter/int_executable.hpp
@@ -128,11 +128,6 @@ namespace ngraph
 #undef NGRAPH_OP
                 UnknownOp
             };
-
-            void run_nms5(const op::v5::NonMaxSuppression* nms5,
-                          const std::vector<std::shared_ptr<HostTensor>>& out,
-                          const std::vector<std::shared_ptr<HostTensor>>& args);
-
         } // namespace interpreter
     }     // namespace runtime
 } // namespace ngraph

From a01d6088914a26b456e1953d3aeaec15816005bf Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 27 Oct 2020 13:30:29 +0300
Subject: [PATCH 148/173] Now conversions NMS-1, NMS-3, NMS-4 -> NMS-5  and
 NMS-5 -> NMSIE3 generate NMS nodes with 5 inputs.

---
 .../src/legacy_api/include/legacy/ie_layers.h |  4 ++
 .../include/legacy/ngraph_ops/nms_ie.hpp      |  9 ++++
 .../src/convert_function_to_cnn_network.cpp   |  4 +-
 .../legacy_api/src/ie_layer_validators.cpp    |  1 +
 .../src/legacy_api/src/ngraph_ops/nms_ie.cpp  | 19 +++++++
 .../convert_nms_5_to_legacy.cpp               | 47 +++++++++++------
 .../ir_reader_v7/ie_layer_validators.cpp      |  4 +-
 .../convert_previous_nms_to_nms_5.cpp         |  3 --
 .../non_max_suppression_tests.cpp             | 50 +++++++++++++++++--
 ngraph/python/src/ngraph/opset5/ops.py        | 11 ++--
 10 files changed, 124 insertions(+), 28 deletions(-)

diff --git a/inference-engine/src/legacy_api/include/legacy/ie_layers.h b/inference-engine/src/legacy_api/include/legacy/ie_layers.h
index d02532dbcc7bd4..b298ac969fa6b3 100644
--- a/inference-engine/src/legacy_api/include/legacy/ie_layers.h
+++ b/inference-engine/src/legacy_api/include/legacy/ie_layers.h
@@ -2064,6 +2064,10 @@ class INFERENCE_ENGINE_INTERNAL_CNNLAYER_CLASS(NonMaxSuppressionLayer): public C
      * classes
      */
     bool sort_result_descending = true;
+    /**
+     * @brief Output type for first and third inputs
+     */
+    std::string output_type = "I64";
     /**
      * @brief Creates a new NonMaxSuppressionLayer instance.
      */
diff --git a/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp b/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp
index 49e689e9bd4727..f8b04c5fcca48d 100644
--- a/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp
+++ b/inference-engine/src/legacy_api/include/legacy/ngraph_ops/nms_ie.hpp
@@ -62,6 +62,15 @@ class INFERENCE_ENGINE_API_CLASS(NonMaxSuppressionIE3) : public Op {
 public:
     NGRAPH_RTTI_DECLARATION;
 
+    NonMaxSuppressionIE3(const Output<Node>& boxes,
+                         const Output<Node>& scores,
+                         const Output<Node>& max_output_boxes_per_class,
+                         const Output<Node>& iou_threshold,
+                         const Output<Node>& score_threshold,
+                         int center_point_box,
+                         bool sort_result_descending,
+                         const ngraph::element::Type& output_type = ngraph::element::i64);
+
     NonMaxSuppressionIE3(const Output<Node>& boxes,
                          const Output<Node>& scores,
                          const Output<Node>& max_output_boxes_per_class,
diff --git a/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp b/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
index 3dc5c291f477f8..b6659bd2d39559 100644
--- a/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
+++ b/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
@@ -537,9 +537,9 @@ InferenceEngine::details::CNNLayerCreator::CNNLayerCreator(const std::shared_ptr
 
     addSpecificCreator({"NonMaxSuppressionIE3"}, [](const std::shared_ptr<::ngraph::Node>& node,
         const std::map<std::string, std::string>& params) -> CNNLayerPtr {
-        LayerParams attrs = {node->get_friendly_name(), "NonMaxSuppressionIE3",
+        LayerParams attrs = {node->get_friendly_name(), "NonMaxSuppression",
             details::convertPrecision(node->get_output_element_type(0))};
-        auto res = std::make_shared<InferenceEngine::CNNLayer>(attrs);
+        auto res = std::make_shared<InferenceEngine::NonMaxSuppressionLayer>(attrs);
         res->params = params;
         return res;
     });
diff --git a/inference-engine/src/legacy_api/src/ie_layer_validators.cpp b/inference-engine/src/legacy_api/src/ie_layer_validators.cpp
index 11e96335271103..83949bf309ebc9 100644
--- a/inference-engine/src/legacy_api/src/ie_layer_validators.cpp
+++ b/inference-engine/src/legacy_api/src/ie_layer_validators.cpp
@@ -1177,6 +1177,7 @@ void NMSValidator::parseParams(CNNLayer* layer) {
 
     casted->center_point_box = layer->GetParamAsBool("center_point_box", false);
     casted->sort_result_descending = layer->GetParamAsBool("sort_result_descending", true);
+    casted->output_type = layer->GetParamAsString("output_type", "I64");
 }
 
 #define REG_LAYER_VALIDATOR_FOR_TYPE(__validator, __type) _validators[#__type] = std::make_shared<__validator>(#__type)
diff --git a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
index 0bee3d569db920..dc9ffc52eaa79a 100644
--- a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
@@ -104,6 +104,19 @@ void op::NonMaxSuppressionIE2::validate_and_infer_types() {
 
 NGRAPH_RTTI_DEFINITION(op::NonMaxSuppressionIE3, "NonMaxSuppressionIE3", 3);
 
+op::NonMaxSuppressionIE3::NonMaxSuppressionIE3(const Output<Node>& boxes,
+                                               const Output<Node>& scores,
+                                               const Output<Node>& max_output_boxes_per_class,
+                                               const Output<Node>& iou_threshold,
+                                               const Output<Node>& score_threshold,
+                                               int center_point_box,
+                                               bool sort_result_descending,
+                                               const ngraph::element::Type& output_type)
+        : Op({boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold}),
+          m_center_point_box(center_point_box), m_sort_result_descending(sort_result_descending), m_output_type(output_type) {
+    constructor_validate_and_infer_types();
+}
+
 op::NonMaxSuppressionIE3::NonMaxSuppressionIE3(const Output<Node>& boxes,
                                                const Output<Node>& scores,
                                                const Output<Node>& max_output_boxes_per_class,
@@ -139,6 +152,12 @@ static constexpr size_t max_output_boxes_per_class_port = 2;
 int64_t op::NonMaxSuppressionIE3::max_boxes_output_from_input() const {
     int64_t max_output_boxes{0};
 
+    size_t num_of_inputs = inputs().size();
+    if (num_of_inputs < 3)
+    {
+        return 0;
+    }
+
     const auto max_output_boxes_input =
         as_type_ptr<op::Constant>(input_value(max_output_boxes_per_class_port).get_node_shared_ptr());
     max_output_boxes = max_output_boxes_input->cast_vector<int64_t>().at(0);
diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index dae4adaf0510ba..f892528ce092e8 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -32,7 +32,7 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         const auto& arg2 = num_of_inputs > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
         const auto& arg3 = num_of_inputs > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg4 = num_of_inputs > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg5 = num_of_inputs > 5 ? new_args.at(5) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
+//        const auto& arg5 = num_of_inputs > 5 ? new_args.at(5) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
 
         // vector of new nGraph operations
         NodeVector new_ops;
@@ -58,8 +58,8 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         new_score_threshold = std::make_shared<opset1::Reshape>(arg4, new_shape_for_score_threshold, true);
         new_ops.push_back(new_score_threshold.get_node_shared_ptr());
 
-        new_soft_nms_sigma = std::make_shared<opset1::Reshape>(arg5, new_shape_for_soft_nms_sigma, true);
-        new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
+//        new_soft_nms_sigma = std::make_shared<opset1::Reshape>(arg5, new_shape_for_soft_nms_sigma, true);
+//        new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
 
         int center_point_box = 0;
         switch (nms_5->get_box_encoding()) {
@@ -73,17 +73,36 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
                 throw ngraph_error("NonMaxSuppression layer " + nms_5->get_friendly_name() +
                                    " has unsupported box encoding");
         }
-        const auto nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
-                new_args.at(0),
-                new_args.at(1),
-                new_max_per_class,
-                new_iou_threshold,
-                new_score_threshold,
-                new_soft_nms_sigma,
-                center_point_box,
-                nms_5->get_sort_result_descending(),
-                nms_5->get_output_type());
-        new_ops.push_back(nms_legacy);
+
+        std::shared_ptr<op::NonMaxSuppressionIE3> nms_legacy{nullptr};
+
+        if (num_of_inputs > 5) {
+            new_soft_nms_sigma = std::make_shared<opset1::Reshape>(new_args.at(5), new_shape_for_soft_nms_sigma, true);
+            new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
+            nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
+                    new_args.at(0),
+                    new_args.at(1),
+                    new_max_per_class,
+                    new_iou_threshold,
+                    new_score_threshold,
+                    new_soft_nms_sigma,
+                    center_point_box,
+                    nms_5->get_sort_result_descending(),
+                    nms_5->get_output_type());
+            new_ops.push_back(nms_legacy);
+        } else {
+            nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
+                    new_args.at(0),
+                    new_args.at(1),
+                    new_max_per_class,
+                    new_iou_threshold,
+                    new_score_threshold,
+                    center_point_box,
+                    nms_5->get_sort_result_descending(),
+                    nms_5->get_output_type());
+            new_ops.push_back(nms_legacy);
+        }
+
 
         nms_legacy->set_friendly_name(nms_5->get_friendly_name());
         ngraph::copy_runtime_info(nms_5, new_ops);
diff --git a/inference-engine/src/readers/ir_reader_v7/ie_layer_validators.cpp b/inference-engine/src/readers/ir_reader_v7/ie_layer_validators.cpp
index 7cd900a669e9ad..392b24960afedb 100644
--- a/inference-engine/src/readers/ir_reader_v7/ie_layer_validators.cpp
+++ b/inference-engine/src/readers/ir_reader_v7/ie_layer_validators.cpp
@@ -2098,9 +2098,9 @@ void NMSValidator::checkParams(const CNNLayer* layer) {
 
 void NMSValidator::checkShapes(const CNNLayer* layer, const vector<SizeVector>& inShapes) const {
     size_t numInputs = inShapes.size();
-    if (numInputs < 2 || numInputs > 5)
+    if (numInputs < 2 || numInputs > 6)
         THROW_IE_EXCEPTION << layer->name
-                           << " NonMaxSuppression can take 2 - 5 inputs, but actually it has: " << numInputs;
+                           << " NonMaxSuppression can take 2 - 6 inputs, but actually it has: " << numInputs;
 
     if (inShapes[0].size() != 3 || inShapes[0][2] != 4)
         THROW_IE_EXCEPTION << layer->name << " 'boxes' should be with shape [num_batches, spatial_dimension, 4]";
diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index e16b62daf7d22a..7dc66e44d4493a 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -142,12 +142,10 @@ namespace {
         const auto& arg2 = num_of_args > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i64, Shape{}, {0});
         const auto& arg3 = num_of_args > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg4 = num_of_args > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-        const auto& arg5 = ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
 
         // list of new nGraph operations
         std::list<std::shared_ptr<::ngraph::Node>> new_ops_list;
 
-        new_ops_list.push_front(arg5);
         if (num_of_args <= 4) {
             new_ops_list.push_front(arg4.get_node_shared_ptr());
         }
@@ -164,7 +162,6 @@ namespace {
                 arg2,
                 arg3,
                 arg4,
-                arg5,
                 attrs.box_encoding,
                 attrs.sort_result_descending,
                 attrs.output_type);
diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index 236858d1073a37..4cfcb6a978f055 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -111,6 +111,28 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </input>
         </layer>
+        <layer id="9" name="mul2" type="Multiply" version="opset1">
+            <input>
+                <port id="0" precision="I64">
+                    <dim>1</dim>
+                </port>
+                <port id="1" precision="I64">
+                    <dim>1</dim>
+                </port>
+            </input>
+            <output>
+                <port id="2" precision="I64">
+                    <dim>1</dim>
+                </port>
+            </output>
+        </layer>
+        <layer id="10" name="output2" type="Result" version="opset1">
+            <input>
+                <port id="0" precision="I64">
+                    <dim>1</dim>
+                </port>
+            </input>
+        </layer>
     </layers>
     <edges>
         <edge from-layer="0" from-port="0" to-layer="6" to-port="0"/>
@@ -121,7 +143,10 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
         <edge from-layer="5" from-port="0" to-layer="6" to-port="5"/>
         <edge from-layer="6" from-port="6" to-layer="7" to-port="0"/>
         <edge from-layer="6" from-port="6" to-layer="7" to-port="1"/>
+        <edge from-layer="6" from-port="8" to-layer="9" to-port="0"/>
+        <edge from-layer="6" from-port="8" to-layer="9" to-port="1"/>
         <edge from-layer="7" from-port="2" to-layer="8" to-port="0"/>
+        <edge from-layer="9" from-port="2" to-layer="10" to-port="0"/>
     </edges>
 </net>
 )V0G0N";
@@ -186,8 +211,8 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 <custom offset="16" precision="FP32" size="4"/>
             </blobs>
         </layer>
-        <layer id="6" name="nms" type="NonMaxSuppressionIE3" precision="I64">
-            <data box_encoding="corner" sort_result_descending="0"/>
+        <layer id="6" name="nms" type="NonMaxSuppression" precision="I64">
+            <data center_point_box="0" output_type="i64" sort_result_descending="false"/>
             <input>
                 <port id="0">
                     <dim>1</dim>
@@ -226,7 +251,8 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </output>
         </layer>
-        <layer id="7" name="mul" type="Multiply" precision="I64">
+        <layer id="7" name="mul" type="Eltwise" precision="I64">
+            <data operation="prod"/>
             <input>
                 <port id="0">
                     <dim>16000</dim>
@@ -244,6 +270,22 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </output>
         </layer>
+        <layer id="8" name="mul2" type="Eltwise" precision="I64">
+            <data operation="prod"/>
+            <input>
+                <port id="0">
+                    <dim>1</dim>
+                </port>
+                <port id="1">
+                    <dim>1</dim>
+                </port>
+            </input>
+            <output>
+                <port id="2" precision="I64">
+                    <dim>1</dim>
+                </port>
+            </output>
+        </layer>
     </layers>
     <edges>
         <edge from-layer="0" from-port="0" to-layer="6" to-port="0"/>
@@ -254,6 +296,8 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
         <edge from-layer="5" from-port="0" to-layer="6" to-port="5"/>
         <edge from-layer="6" from-port="6" to-layer="7" to-port="0"/>
         <edge from-layer="6" from-port="6" to-layer="7" to-port="1"/>
+        <edge from-layer="6" from-port="8" to-layer="8" to-port="0"/>
+        <edge from-layer="6" from-port="8" to-layer="8" to-port="1"/>
     </edges>
 </net>
 )V0G0N";
diff --git a/ngraph/python/src/ngraph/opset5/ops.py b/ngraph/python/src/ngraph/opset5/ops.py
index 2d1da821864a69..17ea4257e26378 100644
--- a/ngraph/python/src/ngraph/opset5/ops.py
+++ b/ngraph/python/src/ngraph/opset5/ops.py
@@ -153,11 +153,14 @@ def non_max_suppression(
     if score_threshold is None:
         score_threshold = make_constant_node(0, np.float32)
     if soft_nms_sigma is None:
-        soft_nms_sigma = make_constant_node(0, np.float32)
+        inputs = as_nodes(
+            boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold
+        )
+    else:
+        inputs = as_nodes(
+            boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, soft_nms_sigma
+        )
 
-    inputs = as_nodes(
-        boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, soft_nms_sigma
-    )
     attributes = {
         "box_encoding": box_encoding,
         "sort_result_descending": sort_result_descending,

From cbae8cfb45d0af81a97f0c3ff7d28a1a03e95f67 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 27 Oct 2020 13:36:11 +0300
Subject: [PATCH 149/173] Fixed ctor in MKLDNN NonMaxSuppressionImpl:
 validation of number of output edges.

---
 .../src/mkldnn_plugin/nodes/non_max_suppression.cpp             | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
index 407eb51228ffbd..9b943f78021d9e 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
@@ -23,7 +23,7 @@ class NonMaxSuppressionImpl: public ExtLayerBase {
             if (layer->insData.size() < 2 || layer->insData.size() > 5)
                 THROW_IE_EXCEPTION << layer->name << " Incorrect number of input edges!";
 
-            if (layer->outData.size() != 1)
+            if (layer->outData.size() > 3)
                 THROW_IE_EXCEPTION << layer->name << " Incorrect number of output edges!";
 
             if (layer->insData[NMS_BOXES].lock()->getTensorDesc().getPrecision() != Precision::FP32)

From 70ffa574f9030aa16807df63c010c75af793389e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 27 Oct 2020 13:56:24 +0300
Subject: [PATCH 150/173] Added conversion of output_type for NMS-5.

---
 .../src/transformations/convert_precision.cpp          | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/inference-engine/src/transformations/src/transformations/convert_precision.cpp b/inference-engine/src/transformations/src/transformations/convert_precision.cpp
index bb7aeaca89fc15..10112e7f1cbf52 100644
--- a/inference-engine/src/transformations/src/transformations/convert_precision.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_precision.cpp
@@ -21,6 +21,7 @@ bool fuse_type_to_parameter(std::shared_ptr<ngraph::Node> & node, ngraph::elemen
 bool fuse_type_to_convert(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_nms3(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_nms4(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
+bool fuse_type_to_nms5(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_topk(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_nonzero(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_bucketize(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
@@ -81,6 +82,7 @@ bool ngraph::pass::ConvertPrecision::run_on_function(std::shared_ptr<ngraph::Fun
         {opset4::ShapeOf::type_info, fuse_type_to_shapeof},
         {opset3::NonMaxSuppression::type_info, fuse_type_to_nms3},
         {opset4::NonMaxSuppression::type_info, fuse_type_to_nms4},
+        {opset5::NonMaxSuppression::type_info, fuse_type_to_nms5},
         {opset4::TopK::type_info, fuse_type_to_topk},
         {opset4::NonZero::type_info, fuse_type_to_nonzero},
         {opset4::Bucketize::type_info, fuse_type_to_bucketize},
@@ -237,6 +239,14 @@ bool fuse_type_to_nms4(std::shared_ptr<ngraph::Node> & node, ngraph::element::Ty
     return false;
 }
 
+bool fuse_type_to_nms5(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx) {
+    if (auto nms = as_type_ptr<opset5::NonMaxSuppression>(node)) {
+        nms->set_output_type(to);
+        return true;
+    }
+    return false;
+}
+
 bool fuse_type_to_topk(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx) {
     if (auto topk = as_type_ptr<opset4::TopK>(node)) {
         if (idx == 1 && (to == element::i32 || to == element::i64)) {

From 730ec8a35eaf051bf94fa62eeee95be7dd33cdfa Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 27 Oct 2020 15:18:29 +0300
Subject: [PATCH 151/173] Fixes in the transformation NMS5 -> NMSIE3.

---
 .../transformations/convert_nms5_test.cpp      | 18 +++---------------
 1 file changed, 3 insertions(+), 15 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index c3aba0880fe847..27f787482b4996 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -102,7 +102,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -114,11 +113,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFiveInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -155,7 +151,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -167,11 +162,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticFourInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -207,7 +199,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -219,11 +210,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticThreeInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});

From 45da4c76f894e9d3f3a9cfa0021c75e65a60949c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 28 Oct 2020 11:07:52 +0300
Subject: [PATCH 152/173] Fixes in the conversion of NMS-5 to NMSIE3.

---
 .../convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp   | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index f892528ce092e8..ad37ba8f33f0cb 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -32,7 +32,6 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         const auto& arg2 = num_of_inputs > 2 ? new_args.at(2) : ngraph::opset5::Constant::create(element::i32, Shape{}, {0});
         const auto& arg3 = num_of_inputs > 3 ? new_args.at(3) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
         const auto& arg4 = num_of_inputs > 4 ? new_args.at(4) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
-//        const auto& arg5 = num_of_inputs > 5 ? new_args.at(5) : ngraph::opset5::Constant::create(element::f32, Shape{}, {.0f});
 
         // vector of new nGraph operations
         NodeVector new_ops;
@@ -58,9 +57,6 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         new_score_threshold = std::make_shared<opset1::Reshape>(arg4, new_shape_for_score_threshold, true);
         new_ops.push_back(new_score_threshold.get_node_shared_ptr());
 
-//        new_soft_nms_sigma = std::make_shared<opset1::Reshape>(arg5, new_shape_for_soft_nms_sigma, true);
-//        new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
-
         int center_point_box = 0;
         switch (nms_5->get_box_encoding()) {
             case ::ngraph::opset5::NonMaxSuppression::BoxEncodingType::CENTER:
@@ -76,7 +72,7 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
 
         std::shared_ptr<op::NonMaxSuppressionIE3> nms_legacy{nullptr};
 
-        if (num_of_inputs > 5) {
+        if (num_of_inputs > 5 && nms_5->soft_nms_sigma_from_input() != 0.0f) {
             new_soft_nms_sigma = std::make_shared<opset1::Reshape>(new_args.at(5), new_shape_for_soft_nms_sigma, true);
             new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
             nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
@@ -103,7 +99,6 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
             new_ops.push_back(nms_legacy);
         }
 
-
         nms_legacy->set_friendly_name(nms_5->get_friendly_name());
         ngraph::copy_runtime_info(nms_5, new_ops);
         ngraph::replace_node(nms_5, nms_legacy);

From b7f3de0799675187d86197b3f03c8159ede88e7f Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 28 Oct 2020 11:08:55 +0300
Subject: [PATCH 153/173] Fixes in MKLDNN NMS ctor.

---
 .../nodes/non_max_suppression.cpp             | 44 +++++++++++--------
 1 file changed, 25 insertions(+), 19 deletions(-)

diff --git a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
index 9b943f78021d9e..e899adcda4f0ff 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
@@ -20,11 +20,10 @@ class NonMaxSuppressionImpl: public ExtLayerBase {
 public:
     explicit NonMaxSuppressionImpl(const CNNLayer* layer) {
         try {
-            if (layer->insData.size() < 2 || layer->insData.size() > 5)
-                THROW_IE_EXCEPTION << layer->name << " Incorrect number of input edges!";
-
-            if (layer->outData.size() > 3)
-                THROW_IE_EXCEPTION << layer->name << " Incorrect number of output edges!";
+            if (layer->insData.size() < 2 || layer->insData.size() > 6)
+                THROW_IE_EXCEPTION << "NMS" << "has incorrect number of input edges: " << layer->insData.size();
+            if (layer->outData.size() < 1 || layer->outData.size() > 3)
+                THROW_IE_EXCEPTION << "NMS" << "has incorrect number of output edges: " << layer->outData.size();
 
             if (layer->insData[NMS_BOXES].lock()->getTensorDesc().getPrecision() != Precision::FP32)
                 THROW_IE_EXCEPTION << layer->name << " Incorrect 'boxes' input precision. Only FP32 is supported!";
@@ -73,21 +72,28 @@ class NonMaxSuppressionImpl: public ExtLayerBase {
             if (selected_indices_dims.size() != 2 || selected_indices_dims[1] != 3)
                 THROW_IE_EXCEPTION << layer->name << " 'selected_indices' should be with shape [num_selected_indices, 3]";
 
-            center_point_box = layer->GetParamAsBool("center_point_box", false);
-            sort_result_descending = layer->GetParamAsBool("sort_result_descending", true);
-
-            if (layer->insData.size() == 2) {
-                addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
-            } else if (layer->insData.size() == 3) {
-                addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN) },
-                    { DataConfigurator(ConfLayout::PLN) });
-            } else if (layer->insData.size() == 4) {
-                addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN),
-                    DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
-            } else {
-                addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN),
-                    DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
+            LayerConfig config;
+            for (size_t i = 0; i < layer->insData.size(); i++) {
+                DataConfig inConfig;
+                Precision inPrecision = Precision::FP32;
+                if (i == NMS_MAXOUTPUTBOXESPERCLASS)
+                    inPrecision = Precision::I32;
+                const SizeVector& inDims = layer->insData[i].lock()->getTensorDesc().getDims();
+                inConfig.desc = TensorDesc(inPrecision, inDims, InferenceEngine::TensorDesc::getLayoutByDims(inDims));
+                config.inConfs.push_back(inConfig);
+            }
+            for (size_t i = 0; i < layer->outData.size(); i++) {
+                DataConfig outConfig;
+                Precision outPrecision = Precision::I32;
+                if (i == 1)
+                    outPrecision = Precision::FP32;
+                const SizeVector& outDims = layer->outData[i]->getTensorDesc().getDims();
+                outConfig.desc = TensorDesc(outPrecision, outDims, InferenceEngine::TensorDesc::getLayoutByDims(outDims));
+                config.outConfs.push_back(outConfig);
             }
+
+            config.dynBatchSupport = false;
+            confs.push_back(config);
         } catch (InferenceEngine::details::InferenceEngineException &ex) {
             errorMsg = ex.what();
         }

From 230ed5f64059d23a93ce077764afaed6ba04611c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 28 Oct 2020 11:09:32 +0300
Subject: [PATCH 154/173] Small fix.

---
 .../transformations/src/transformations/convert_precision.cpp    | 1 +
 1 file changed, 1 insertion(+)

diff --git a/inference-engine/src/transformations/src/transformations/convert_precision.cpp b/inference-engine/src/transformations/src/transformations/convert_precision.cpp
index 10112e7f1cbf52..c08fe61dbd841f 100644
--- a/inference-engine/src/transformations/src/transformations/convert_precision.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_precision.cpp
@@ -7,6 +7,7 @@
 #include <memory>
 #include <vector>
 
+#include <ngraph/opsets/opset5.hpp>
 #include <ngraph/opsets/opset4.hpp>
 #include <ngraph/opsets/opset3.hpp>
 #include <ngraph/opsets/opset1.hpp>

From 29eca217a46df83904b51e7b486df3e299534b2c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 28 Oct 2020 11:11:23 +0300
Subject: [PATCH 155/173] Fixed tests for the transformation NMS5 -> NMSIE3.

---
 .../transformations/convert_nms5_test.cpp     | 54 ++++---------------
 1 file changed, 9 insertions(+), 45 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
index 27f787482b4996..6ba473952a4bf4 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_nms5_test.cpp
@@ -246,7 +246,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i32, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -258,11 +257,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEStaticTwoInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -356,7 +352,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FiveInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -368,11 +363,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FiveInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -409,7 +401,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FourInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -421,11 +412,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1FourInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -461,7 +449,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1ThreeInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -473,11 +460,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1ThreeInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -512,7 +496,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1TwoInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i32, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -524,11 +507,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic1TwoInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -619,7 +599,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FiveInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -631,11 +610,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FiveInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -671,7 +647,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FourInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -683,11 +658,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2FourInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -722,7 +694,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2ThreeInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -734,11 +705,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2ThreeInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
@@ -772,7 +740,6 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2TwoInputs) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i32, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = ngraph::opset5::Constant::create(element::f32, Shape{}, {0.0f});
 
         auto one_dim_shape = Shape{1};
         auto new_max_per_class = std::make_shared<opset5::Reshape>(max_output_boxes_per_class,
@@ -784,11 +751,8 @@ TEST(TransformationTests, ConvertNMS5ToNMSIEDynamic2TwoInputs) {
         auto new_score_threshold = std::make_shared<opset5::Reshape>(score_threshold,
                                                                      opset5::Constant::create(ngraph::element::i64, one_dim_shape,
                                                                                               one_dim_shape), true);
-        auto new_soft_nms_sigma = std::make_shared<opset5::Reshape>(soft_nms_sigma,
-                                                                    opset5::Constant::create(ngraph::element::i64, one_dim_shape,
-                                                                                             one_dim_shape), true);
         auto nms = std::make_shared<op::NonMaxSuppressionIE3>(boxes, scores, new_max_per_class, new_iou_threshold, new_score_threshold,
-                                                              new_soft_nms_sigma, 0, true);
+                                                              0, true);
         nms->set_friendly_name("nms");
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});

From 98b402f05f17ae3d1c35db8f399f5a95527d41b8 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Wed, 28 Oct 2020 11:12:40 +0300
Subject: [PATCH 156/173] Fixed tests for conversions NMS-1, NMS-3, NMS-4 ->
 NMS-5.

---
 .../convert_previous_nms_to_nms_5.cpp          | 18 ++++++------------
 1 file changed, 6 insertions(+), 12 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
index 0f96d32b690c49..a1120f41c419b4 100644
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_previous_nms_to_nms_5.cpp
@@ -49,9 +49,8 @@ TEST(TransformationTests, ConvertNMS4FiveInputsToNMS5) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {.0f});
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
-                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
@@ -82,9 +81,8 @@ TEST(TransformationTests, ConvertNMS4TwoInputsToNMS5) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
-                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
@@ -119,9 +117,8 @@ TEST(TransformationTests, ConvertNMS3FiveInputsToNMS5) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {.0f});
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
-                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
@@ -152,9 +149,8 @@ TEST(TransformationTests, ConvertNMS3TwoInputsToNMS5) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
-                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
@@ -189,9 +185,8 @@ TEST(TransformationTests, ConvertNMS1FiveInputsToNMS5) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {10});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.75});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.7});
-        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {.0f});
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
-                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }
@@ -222,9 +217,8 @@ TEST(TransformationTests, ConvertNMS1TwoInputsToNMS5) {
         auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {0});
         auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.0f});
-        auto soft_nms_sigma = opset5::Constant::create(element::f32, Shape{}, {0.0f});
         auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
-                                                               soft_nms_sigma, opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
+                                                               opset5::NonMaxSuppression::BoxEncodingType::CORNER, true);
 
         f_ref = std::make_shared<Function>(NodeVector{nms}, ParameterVector{boxes, scores});
     }

From fbba3a3006b12f56748d793a788574dcee1aa77a Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 29 Oct 2020 10:31:15 +0300
Subject: [PATCH 157/173] Small fixes in MKLDNN NMS ctor.

---
 .../nodes/non_max_suppression.cpp             |   2 +
 .../non_max_suppression_tests.cpp             | 322 ++++++++++++++++--
 .../functional_test_utils/network_utils.cpp   |  13 +-
 3 files changed, 307 insertions(+), 30 deletions(-)

diff --git a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
index e899adcda4f0ff..e7f293075a2a7b 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
@@ -72,6 +72,8 @@ class NonMaxSuppressionImpl: public ExtLayerBase {
             if (selected_indices_dims.size() != 2 || selected_indices_dims[1] != 3)
                 THROW_IE_EXCEPTION << layer->name << " 'selected_indices' should be with shape [num_selected_indices, 3]";
 
+            center_point_box = layer->GetParamAsBool("center_point_box", false);
+            sort_result_descending = layer->GetParamAsBool("sort_result_descending", true);
             LayerConfig config;
             for (size_t i = 0; i < layer->insData.size(); i++) {
                 DataConfig inConfig;
diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index 4cfcb6a978f055..c5a434badfe1e9 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -133,6 +133,32 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </input>
         </layer>
+        <layer id="11" name="mul3" type="Multiply" version="opset1">
+            <input>
+                <port id="0" precision="I64">
+                    <dim>15130</dim>
+                    <dim>3</dim>
+                </port>
+                <port id="1" precision="I64">
+                    <dim>15130</dim>
+                    <dim>3</dim>
+                </port>
+            </input>
+            <output>
+                <port id="2" precision="I64">
+                    <dim>15130</dim>
+                    <dim>3</dim>
+                </port>
+            </output>
+        </layer>
+        <layer id="12" name="output3" type="Result" version="opset1">
+            <input>
+                <port id="0" precision="I64">
+                    <dim>15130</dim>
+                    <dim>3</dim>
+                </port>
+            </input>
+        </layer>
     </layers>
     <edges>
         <edge from-layer="0" from-port="0" to-layer="6" to-port="0"/>
@@ -143,10 +169,13 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
         <edge from-layer="5" from-port="0" to-layer="6" to-port="5"/>
         <edge from-layer="6" from-port="6" to-layer="7" to-port="0"/>
         <edge from-layer="6" from-port="6" to-layer="7" to-port="1"/>
+        <edge from-layer="6" from-port="7" to-layer="11" to-port="0"/>
+        <edge from-layer="6" from-port="7" to-layer="11" to-port="1"/>
         <edge from-layer="6" from-port="8" to-layer="9" to-port="0"/>
         <edge from-layer="6" from-port="8" to-layer="9" to-port="1"/>
         <edge from-layer="7" from-port="2" to-layer="8" to-port="0"/>
         <edge from-layer="9" from-port="2" to-layer="10" to-port="0"/>
+        <edge from-layer="11" from-port="2" to-layer="12" to-port="0"/>
     </edges>
 </net>
 )V0G0N";
@@ -201,17 +230,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 <custom offset="12" precision="FP32" size="4"/>
             </blobs>
         </layer>
-        <layer id="5" name="soft_nms_sigma" precision="FP32" type="Const">
-            <output>
-                <port id="0" precision="FP32">
-                    <dim>1</dim>
-                </port>
-            </output>
-            <blobs>
-                <custom offset="16" precision="FP32" size="4"/>
-            </blobs>
-        </layer>
-        <layer id="6" name="nms" type="NonMaxSuppression" precision="I64">
+        <layer id="5" name="nms" type="NonMaxSuppression" precision="I64">
             <data center_point_box="0" output_type="i64" sort_result_descending="false"/>
             <input>
                 <port id="0">
@@ -233,25 +252,22 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 <port id="4">
                     <dim>1</dim>
                 </port>
-                <port id="5">
-                    <dim>1</dim>
-                </port>
             </input>
             <output>
-                <port id="6" precision="I64">
+                <port id="5" precision="I64">
                     <dim>16000</dim>
                     <dim>3</dim>
                 </port>
-                <port id="7" precision="FP32">
+                <port id="6" precision="FP32">
                     <dim>16000</dim>
                     <dim>3</dim>
                 </port>
-                <port id="8" precision="I64">
+                <port id="7" precision="I64">
                     <dim>1</dim>
                 </port>
             </output>
         </layer>
-        <layer id="7" name="mul" type="Eltwise" precision="I64">
+        <layer id="6" name="mul" type="Eltwise" precision="I64">
             <data operation="prod"/>
             <input>
                 <port id="0">
@@ -270,7 +286,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </output>
         </layer>
-        <layer id="8" name="mul2" type="Eltwise" precision="I64">
+        <layer id="7" name="mul2" type="Eltwise" precision="I64">
             <data operation="prod"/>
             <input>
                 <port id="0">
@@ -286,18 +302,38 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
                 </port>
             </output>
         </layer>
+        <layer id="8" name="mul3" type="Eltwise" precision="FP32">
+            <data operation="prod"/>
+            <input>
+                <port id="0">
+                    <dim>16000</dim>
+                    <dim>3</dim>
+                </port>
+                <port id="1">
+                    <dim>16000</dim>
+                    <dim>3</dim>
+                </port>
+            </input>
+            <output>
+                <port id="2" precision="FP32">
+                    <dim>16000</dim>
+                    <dim>3</dim>
+                </port>
+            </output>
+        </layer>
     </layers>
     <edges>
-        <edge from-layer="0" from-port="0" to-layer="6" to-port="0"/>
-        <edge from-layer="1" from-port="0" to-layer="6" to-port="1"/>
-        <edge from-layer="2" from-port="0" to-layer="6" to-port="2"/>
-        <edge from-layer="3" from-port="0" to-layer="6" to-port="3"/>
-        <edge from-layer="4" from-port="0" to-layer="6" to-port="4"/>
-        <edge from-layer="5" from-port="0" to-layer="6" to-port="5"/>
-        <edge from-layer="6" from-port="6" to-layer="7" to-port="0"/>
-        <edge from-layer="6" from-port="6" to-layer="7" to-port="1"/>
-        <edge from-layer="6" from-port="8" to-layer="8" to-port="0"/>
-        <edge from-layer="6" from-port="8" to-layer="8" to-port="1"/>
+        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
+        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
+        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
+        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
+        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
+        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
+        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
+        <edge from-layer="5" from-port="6" to-layer="8" to-port="0"/>
+        <edge from-layer="5" from-port="6" to-layer="8" to-port="1"/>
+        <edge from-layer="5" from-port="7" to-layer="7" to-port="0"/>
+        <edge from-layer="5" from-port="7" to-layer="7" to-port="1"/>
     </edges>
 </net>
 )V0G0N";
@@ -312,3 +348,231 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
         fp32w[4] = 0.0;
     });
 }
+
+//TEST_F(NGraphReaderTests, ReadNonMaxSuppression4) {
+//    std::string model = R"V0G0N(
+//<net name="Network" version="10">
+//    <layers>
+//        <layer id="0" name="in1" type="Parameter" version="opset1">
+//            <data element_type="f32" shape="1,15130,4"/>
+//            <output>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                    <dim>15130</dim>
+//                    <dim>4</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="1" name="in2" type="Parameter" version="opset1">
+//            <data element_type="f32" shape="1,80,15130"/>
+//            <output>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                    <dim>80</dim>
+//                    <dim>15130</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const" version="opset1">
+//            <data offset="0" size="8"/>
+//            <output>
+//                <port id="0" precision="I64"/>
+//            </output>
+//        </layer>
+//        <layer id="3" name="iou_threshold" precision="FP32" type="Const" version="opset1">
+//            <data offset="8" size="4"/>
+//            <output>
+//                <port id="0" precision="FP32"/>
+//            </output>
+//        </layer>
+//        <layer id="4" name="score_threshold" precision="FP32" type="Const" version="opset1">
+//            <data offset="12" size="4"/>
+//            <output>
+//                <port id="0" precision="FP32"/>
+//            </output>
+//        </layer>
+//        <layer id="5" name="nms" type="NonMaxSuppression" version="opset4">
+//            <data box_encoding="corner" output_type="i32" sort_result_descending="0"/>
+//            <input>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                    <dim>15130</dim>
+//                    <dim>4</dim>
+//                </port>
+//                <port id="1" precision="FP32">
+//                    <dim>1</dim>
+//                    <dim>80</dim>
+//                    <dim>15130</dim>
+//                </port>
+//                <port id="2" precision="I64"/>
+//                <port id="3" precision="FP32"/>
+//                <port id="4" precision="FP32"/>
+//            </input>
+//            <output>
+//                <port id="5" precision="I32">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="6" name="mul" type="Multiply" version="opset1">
+//            <input>
+//                <port id="0" precision="I32">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//                <port id="1" precision="I32">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </input>
+//            <output>
+//                <port id="2" precision="I32">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="7" name="output" type="Result" version="opset1">
+//            <input>
+//                <port id="0" precision="I64">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </input>
+//        </layer>
+//    </layers>
+//    <edges>
+//        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
+//        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
+//        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
+//        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
+//        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
+//        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
+//        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
+//        <edge from-layer="6" from-port="2" to-layer="7" to-port="0"/>
+//    </edges>
+//</net>
+//)V0G0N";
+//    std::string modelV5 = R"V0G0N(
+//<net name="Network" version="7">
+//    <layers>
+//        <layer id="0" name="in1" precision="FP32" type="Input" >
+//            <output>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                    <dim>15130</dim>
+//                    <dim>4</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="1" name="in2" precision="FP32" type="Input" >
+//            <output>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                    <dim>80</dim>
+//                    <dim>15130</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const">
+//            <output>
+//                <port id="0" precision="I64">
+//                    <dim>1</dim>
+//                </port>
+//            </output>
+//            <blobs>
+//                <custom offset="0" precision="I64" size="8"/>
+//            </blobs>
+//        </layer>
+//        <layer id="3" name="iou_threshold" precision="FP32" type="Const">
+//            <output>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                </port>
+//            </output>
+//            <blobs>
+//                <custom offset="8" precision="FP32" size="4"/>
+//            </blobs>
+//        </layer>
+//        <layer id="4" name="score_threshold" precision="FP32" type="Const">
+//            <output>
+//                <port id="0" precision="FP32">
+//                    <dim>1</dim>
+//                </port>
+//            </output>
+//            <blobs>
+//                <custom offset="12" precision="FP32" size="4"/>
+//            </blobs>
+//        </layer>
+//        <layer id="5" name="nms" type="NonMaxSuppression" precision="I32">
+//            <data center_point_box="0" output_type="i32" sort_result_descending="false"/>
+//            <input>
+//                <port id="0">
+//                    <dim>1</dim>
+//                    <dim>15130</dim>
+//                    <dim>4</dim>
+//                </port>
+//                <port id="1">
+//                    <dim>1</dim>
+//                    <dim>80</dim>
+//                    <dim>15130</dim>
+//                </port>
+//                <port id="2">
+//                    <dim>1</dim>
+//                </port>
+//                <port id="3">
+//                    <dim>1</dim>
+//                </port>
+//                <port id="4">
+//                    <dim>1</dim>
+//                </port>
+//            </input>
+//            <output>
+//                <port id="5" precision="I32">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </output>
+//        </layer>
+//        <layer id="6" name="mul" type="Eltwise" precision="I32">
+//            <data operation="prod"/>
+//            <input>
+//                <port id="0">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//                <port id="1">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </input>
+//            <output>
+//                <port id="2" precision="I32">
+//                    <dim>16000</dim>
+//                    <dim>3</dim>
+//                </port>
+//            </output>
+//        </layer>
+//    </layers>
+//    <edges>
+//        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
+//        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
+//        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
+//        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
+//        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
+//        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
+//        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
+//    </edges>
+//</net>
+//)V0G0N";
+//
+//    compareIRs(model, modelV5, 16, [](Blob::Ptr& weights) {
+//        auto * i64w = weights->buffer().as<int64_t*>();
+//        i64w[0] = 200;
+//
+//        auto * fp32w = weights->buffer().as<float*>();
+//        fp32w[2] = 0.5;
+//        fp32w[3] = 0.05;
+//    });
+//}
diff --git a/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp b/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
index 83709dac917a5a..2fc65edcdb3519 100644
--- a/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
+++ b/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
@@ -202,8 +202,19 @@ namespace FuncTestUtils {
                 break;
             success = success && old_info.find(it.first) != old_info.end();
         }
-        if (!success)
+        if (!success) {
+            std::cout << "new output: ";
+            for (const auto &p : new_info) {
+                std::cout << p.first << " ";
+            }
+            std::cout << "\n";
+            std::cout << "old output: ";
+            for (const auto &p : old_info) {
+                std::cout << p.first << " ";
+            }
+            std::cout << "\n";
             THROW_IE_EXCEPTION << err_msg;
+        }
     }
 
     void

From 259ba7fdc163d9cae738e3920ca47567fca71a91 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 29 Oct 2020 17:38:40 +0300
Subject: [PATCH 158/173] Rewritten create_layer() functions for NMS-5 and
 NMSIE3 as addSpecificCreator() functions.

---
 .../src/convert_function_to_cnn_network.cpp   | 69 +++++++++++++++++-
 .../src/ie_cnn_layer_builder_ngraph.cpp       | 70 -------------------
 .../non_max_suppression_tests.cpp             |  2 +-
 3 files changed, 69 insertions(+), 72 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp b/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
index b6659bd2d39559..87e09cd531fd04 100644
--- a/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
+++ b/inference-engine/src/legacy_api/src/convert_function_to_cnn_network.cpp
@@ -539,8 +539,76 @@ InferenceEngine::details::CNNLayerCreator::CNNLayerCreator(const std::shared_ptr
         const std::map<std::string, std::string>& params) -> CNNLayerPtr {
         LayerParams attrs = {node->get_friendly_name(), "NonMaxSuppression",
             details::convertPrecision(node->get_output_element_type(0))};
+
+        auto castedLayer = ::ngraph::as_type_ptr<::ngraph::op::NonMaxSuppressionIE3>(node);
+        IE_ASSERT(castedLayer) << " Operation " << node->description() << " with name "
+            << node->get_friendly_name() << " cannot be casted to ngraph::op::NonMaxSuppressionIE3";
+
+        auto res = std::make_shared<InferenceEngine::NonMaxSuppressionLayer>(attrs);
+        res->params = params;
+
+        res->params["center_point_box"] = castedLayer->m_center_point_box ? "true" : "false";
+        res->params["sort_result_descending"] = castedLayer->m_sort_result_descending ? "true" : "false";
+
+        auto output_type = details::convertPrecision(castedLayer->m_output_type);
+        std::string output_type_str;
+        switch (output_type) {
+        case Precision::I32:
+            output_type_str = "I32";
+            break;
+        case Precision::I64:
+            output_type_str = "I64";
+            break;
+        default:
+            THROW_IE_EXCEPTION << "Unsupported output type";
+        }
+        res->params["output_type"] = output_type_str;
+
+        return res;
+    });
+
+    addSpecificCreator({"NonMaxSuppression"}, [](const std::shared_ptr<::ngraph::Node>& node,
+        const std::map<std::string, std::string>& params) -> CNNLayerPtr {
+        LayerParams attrs = {node->get_friendly_name(), "NonMaxSuppression",
+            details::convertPrecision(node->get_output_element_type(0))};
+
+        auto castedLayer = ::ngraph::as_type_ptr<::ngraph::op::v5::NonMaxSuppression>(node);
+        IE_ASSERT(castedLayer) << " Operation " << node->description() << " with name "
+            << node->get_friendly_name() << " cannot be casted to ngraph::op::v5::NonMaxSuppression";
+
         auto res = std::make_shared<InferenceEngine::NonMaxSuppressionLayer>(attrs);
         res->params = params;
+
+        auto box_encoding = castedLayer->get_box_encoding();
+        switch (box_encoding) {
+            case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CORNER:
+                res->params["center_point_box"] = "false";
+                break;
+            case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CENTER:
+                res->params["center_point_box"] = "true";
+                break;
+            default:
+                THROW_IE_EXCEPTION << "Unsupported box encoding for NonMaxSuppression op";
+                break;
+        }
+
+        auto output_type = details::convertPrecision(castedLayer->get_output_type());
+        std::string output_type_str;
+        switch (output_type) {
+        case Precision::I32:
+            output_type_str = "I32";
+            break;
+        case Precision::I64:
+            output_type_str = "I64";
+            break;
+        default:
+            THROW_IE_EXCEPTION << "Unsupported output type";
+        }
+        res->params["output_type"] = output_type_str;
+
+        bool sort_result_descending = castedLayer->get_sort_result_descending();
+        res->params["sort_result_descending"] = sort_result_descending ? "true" : "false";
+
         return res;
     });
 
@@ -823,7 +891,6 @@ void convertFunctionToICNNNetwork(const std::shared_ptr<const ::ngraph::Function
                 std::make_shared<Builder::NodeConverter<::ngraph::op::v1::LogicalNot>>(),
                 std::make_shared<Builder::NodeConverter<::ngraph::op::ShuffleChannels>>(),
                 std::make_shared<Builder::NodeConverter<::ngraph::op::v4::Interpolate>>(),
-                std::make_shared<Builder::NodeConverter<::ngraph::op::v5::NonMaxSuppression>>(),
                 std::make_shared<Builder::NodeConverter<::ExecGraphInfoSerialization::ExecutionNode>>(),
         };
         CNNLayerPtr result;
diff --git a/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp b/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
index 7ce5f648926f72..e846c646a773b3 100644
--- a/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
+++ b/inference-engine/src/legacy_api/src/ie_cnn_layer_builder_ngraph.cpp
@@ -1726,76 +1726,6 @@ CNNLayer::Ptr NodeConverter<ngraph::op::v0::Interpolate>::createLayer(const std:
     THROW_IE_EXCEPTION << "Interpolate operation should be converted to Interp";
 }
 
-template <>
-CNNLayer::Ptr NodeConverter<ngraph::op::NonMaxSuppressionIE3>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
-    LayerParams params = {layer->get_friendly_name(), "NonMaxSuppressionIE3",
-                          details::convertPrecision(layer->get_output_element_type(0))};
-    auto castedLayer = ngraph::as_type_ptr<ngraph::op::NonMaxSuppressionIE3>(layer);
-    if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
-
-    auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
-    res->params["center_point_box"] = asString(castedLayer->m_center_point_box);
-    res->params["sort_result_descending"] = castedLayer->m_sort_result_descending ? "1" : "0";
-
-    auto output_type = details::convertPrecision(castedLayer->m_output_type);
-    std::string output_type_str;
-    switch (output_type) {
-    case Precision::I32:
-        output_type_str = "I32";
-        break;
-    case Precision::I64:
-        output_type_str = "I64";
-        break;
-    default:
-        THROW_IE_EXCEPTION << "Unsupported output type";
-    }
-    res->params["output_type"] = output_type_str;
-
-    return res;
-}
-
-template <>
-CNNLayer::Ptr NodeConverter<ngraph::op::v5::NonMaxSuppression>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
-    LayerParams params = {layer->get_friendly_name(), "NonMaxSuppressionIE3",
-                          details::convertPrecision(layer->get_output_element_type(0))};
-    auto castedLayer = ngraph::as_type_ptr<ngraph::op::v5::NonMaxSuppression>(layer);
-    if (castedLayer == nullptr) THROW_IE_EXCEPTION << "Cannot get " << params.type << " layer " << params.name;
-
-    auto box_encoding = castedLayer->get_box_encoding();
-
-    auto res = std::make_shared<InferenceEngine::CNNLayer>(params);
-
-    switch (box_encoding) {
-        case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CORNER:
-            res->params["center_point_box"] = "0";
-            break;
-        case ngraph::op::v5::NonMaxSuppression::BoxEncodingType::CENTER:
-            res->params["center_point_box"] = "1";
-            break;
-        default:
-            THROW_IE_EXCEPTION << "Unsupported box encoding for NonMaxSuppression op";
-            break;
-    }
-
-    auto output_type = details::convertPrecision(castedLayer->get_output_type());
-    std::string output_type_str;
-    switch (output_type) {
-    case Precision::I32:
-        output_type_str = "I32";
-        break;
-    case Precision::I64:
-        output_type_str = "I64";
-        break;
-    default:
-        THROW_IE_EXCEPTION << "Unsupported output type";
-    }
-    res->params["output_type"] = output_type_str;
-
-    bool sort_result_descending = castedLayer->get_sort_result_descending();
-    res->params["sort_result_descending"] = sort_result_descending ? "1" : "0";
-    return res;
-}
-
 template <>
 CNNLayer::Ptr NodeConverter<ngraph::op::v4::Interpolate>::createLayer(const std::shared_ptr<ngraph::Node>& layer) const {
     LayerParams params = {layer->get_friendly_name(), "Interpolate",
diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index c5a434badfe1e9..deec162a6cfdc4 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -231,7 +231,7 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
             </blobs>
         </layer>
         <layer id="5" name="nms" type="NonMaxSuppression" precision="I64">
-            <data center_point_box="0" output_type="i64" sort_result_descending="false"/>
+            <data center_point_box="false" output_type="I64" sort_result_descending="false"/>
             <input>
                 <port id="0">
                     <dim>1</dim>

From 13df494ad45065677716a15e40b26c4c197a5991 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 30 Oct 2020 16:38:47 +0300
Subject: [PATCH 159/173] Disabled tests for IE IR reader for NMS-4.

---
 .../non_max_suppression_tests.cpp             | 525 ++++++++++--------
 1 file changed, 298 insertions(+), 227 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index deec162a6cfdc4..eaa96c33397b2d 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -4,6 +4,15 @@
 
 #include <string>
 #include "ngraph_reader_tests.hpp"
+#include <ngraph/graph_util.hpp>
+#include <ngraph/pass/manager.hpp>
+
+#include "generic_ie.hpp"
+#include <transformations/init_node_info.hpp>
+#include <transformations/common_optimizations/common_optimizations.hpp>
+#include <legacy/transformations/convert_opset1_to_legacy/convert_prior_to_ie_prior.hpp>
+
+#include "legacy/convert_function_to_cnn_network.hpp"
 
 TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     std::string model = R"V0G0N(
@@ -349,230 +358,292 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     });
 }
 
-//TEST_F(NGraphReaderTests, ReadNonMaxSuppression4) {
-//    std::string model = R"V0G0N(
-//<net name="Network" version="10">
-//    <layers>
-//        <layer id="0" name="in1" type="Parameter" version="opset1">
-//            <data element_type="f32" shape="1,15130,4"/>
-//            <output>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                    <dim>15130</dim>
-//                    <dim>4</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="1" name="in2" type="Parameter" version="opset1">
-//            <data element_type="f32" shape="1,80,15130"/>
-//            <output>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                    <dim>80</dim>
-//                    <dim>15130</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const" version="opset1">
-//            <data offset="0" size="8"/>
-//            <output>
-//                <port id="0" precision="I64"/>
-//            </output>
-//        </layer>
-//        <layer id="3" name="iou_threshold" precision="FP32" type="Const" version="opset1">
-//            <data offset="8" size="4"/>
-//            <output>
-//                <port id="0" precision="FP32"/>
-//            </output>
-//        </layer>
-//        <layer id="4" name="score_threshold" precision="FP32" type="Const" version="opset1">
-//            <data offset="12" size="4"/>
-//            <output>
-//                <port id="0" precision="FP32"/>
-//            </output>
-//        </layer>
-//        <layer id="5" name="nms" type="NonMaxSuppression" version="opset4">
-//            <data box_encoding="corner" output_type="i32" sort_result_descending="0"/>
-//            <input>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                    <dim>15130</dim>
-//                    <dim>4</dim>
-//                </port>
-//                <port id="1" precision="FP32">
-//                    <dim>1</dim>
-//                    <dim>80</dim>
-//                    <dim>15130</dim>
-//                </port>
-//                <port id="2" precision="I64"/>
-//                <port id="3" precision="FP32"/>
-//                <port id="4" precision="FP32"/>
-//            </input>
-//            <output>
-//                <port id="5" precision="I32">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="6" name="mul" type="Multiply" version="opset1">
-//            <input>
-//                <port id="0" precision="I32">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//                <port id="1" precision="I32">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </input>
-//            <output>
-//                <port id="2" precision="I32">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="7" name="output" type="Result" version="opset1">
-//            <input>
-//                <port id="0" precision="I64">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </input>
-//        </layer>
-//    </layers>
-//    <edges>
-//        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
-//        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
-//        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
-//        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
-//        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
-//        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
-//        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
-//        <edge from-layer="6" from-port="2" to-layer="7" to-port="0"/>
-//    </edges>
-//</net>
-//)V0G0N";
-//    std::string modelV5 = R"V0G0N(
-//<net name="Network" version="7">
-//    <layers>
-//        <layer id="0" name="in1" precision="FP32" type="Input" >
-//            <output>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                    <dim>15130</dim>
-//                    <dim>4</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="1" name="in2" precision="FP32" type="Input" >
-//            <output>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                    <dim>80</dim>
-//                    <dim>15130</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const">
-//            <output>
-//                <port id="0" precision="I64">
-//                    <dim>1</dim>
-//                </port>
-//            </output>
-//            <blobs>
-//                <custom offset="0" precision="I64" size="8"/>
-//            </blobs>
-//        </layer>
-//        <layer id="3" name="iou_threshold" precision="FP32" type="Const">
-//            <output>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                </port>
-//            </output>
-//            <blobs>
-//                <custom offset="8" precision="FP32" size="4"/>
-//            </blobs>
-//        </layer>
-//        <layer id="4" name="score_threshold" precision="FP32" type="Const">
-//            <output>
-//                <port id="0" precision="FP32">
-//                    <dim>1</dim>
-//                </port>
-//            </output>
-//            <blobs>
-//                <custom offset="12" precision="FP32" size="4"/>
-//            </blobs>
-//        </layer>
-//        <layer id="5" name="nms" type="NonMaxSuppression" precision="I32">
-//            <data center_point_box="0" output_type="i32" sort_result_descending="false"/>
-//            <input>
-//                <port id="0">
-//                    <dim>1</dim>
-//                    <dim>15130</dim>
-//                    <dim>4</dim>
-//                </port>
-//                <port id="1">
-//                    <dim>1</dim>
-//                    <dim>80</dim>
-//                    <dim>15130</dim>
-//                </port>
-//                <port id="2">
-//                    <dim>1</dim>
-//                </port>
-//                <port id="3">
-//                    <dim>1</dim>
-//                </port>
-//                <port id="4">
-//                    <dim>1</dim>
-//                </port>
-//            </input>
-//            <output>
-//                <port id="5" precision="I32">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </output>
-//        </layer>
-//        <layer id="6" name="mul" type="Eltwise" precision="I32">
-//            <data operation="prod"/>
-//            <input>
-//                <port id="0">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//                <port id="1">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </input>
-//            <output>
-//                <port id="2" precision="I32">
-//                    <dim>16000</dim>
-//                    <dim>3</dim>
-//                </port>
-//            </output>
-//        </layer>
-//    </layers>
-//    <edges>
-//        <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
-//        <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
-//        <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
-//        <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
-//        <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
-//        <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
-//        <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
-//    </edges>
-//</net>
-//)V0G0N";
-//
-//    compareIRs(model, modelV5, 16, [](Blob::Ptr& weights) {
-//        auto * i64w = weights->buffer().as<int64_t*>();
-//        i64w[0] = 200;
-//
-//        auto * fp32w = weights->buffer().as<float*>();
-//        fp32w[2] = 0.5;
-//        fp32w[3] = 0.05;
-//    });
-//}
+static InferenceEngine::details::CNNNetworkImplPtr read_network_v10(const std::string& model, Blob::Ptr weights) {
+    Core ie;
+    auto result = std::make_shared<InferenceEngine::details::CNNNetworkImpl>();
+
+    auto ngraphImpl = ie.ReadNetwork(model, weights);
+    auto graph = ngraph::clone_function(*ngraphImpl.getFunction());
+
+    ::ngraph::op::GenericIE::DisableReshape noReshape(graph);
+
+    ::ngraph::pass::Manager manager;
+    manager.register_pass<::ngraph::pass::InitNodeInfo>();
+    // WA: ConvertPriorBox must be executed before the 1st ConstantFolding pass
+    manager.register_pass<::ngraph::pass::ConvertPriorBox>();
+    manager.register_pass<::ngraph::pass::CommonOptimizations>();
+    manager.run_passes(graph);
+
+    IE_SUPPRESS_DEPRECATED_START
+    InferenceEngine::details::convertFunctionToICNNNetwork(graph, ngraphImpl, result.get(), false);
+    IE_SUPPRESS_DEPRECATED_END
+
+    return result;
+}
+
+static void compareIRsV10(const std::string& model, const std::string& ref, size_t weightsSize = 0, const std::function<void(Blob::Ptr&)>& fillBlob = {}) {
+    Core ie;
+    Blob::Ptr weights;
+
+    if (weightsSize) {
+        weights = make_shared_blob<uint8_t>(TensorDesc(Precision::U8, {weightsSize}, Layout::C));
+        weights->allocate();
+        CommonTestUtils::fill_data(weights->buffer().as<float *>(), weights->size() / sizeof(float));
+        if (fillBlob)
+            fillBlob(weights);
+    }
+
+    auto network = read_network_v10(model, weights);
+    auto cnnNetwork = ie.ReadNetwork(ref, weights);
+
+    IE_SUPPRESS_DEPRECATED_START
+    FuncTestUtils::compareCNNNetworks(InferenceEngine::CNNNetwork(network), cnnNetwork, false);
+
+    for (auto it = details::CNNNetworkIterator(network.get()); it != details::CNNNetworkIterator(); it++) {
+        InferenceEngine::CNNLayerPtr layer = *it;
+        ASSERT_NE(nullptr, layer->getNode());
+    }
+
+    ASSERT_EQ(nullptr, cnnNetwork.getFunction());
+    for (auto it = details::CNNNetworkIterator(cnnNetwork); it != details::CNNNetworkIterator(); it++) {
+        InferenceEngine::CNNLayerPtr layer = *it;
+        ASSERT_EQ(nullptr, layer->getNode());
+    }
+    IE_SUPPRESS_DEPRECATED_END
+}
+
+// When IRv10 model contains NMS-1, NMS-3, or NMS-4, we read the network, and then CNNNetworkImpl constructor
+// performs all nGraph transformations, and converts a resulting graph to legacy IR. In particular, NMS-1, NMS-3, NMS-4
+// are converted into NMS-5, and then NMS-5 will be converted into the inner operation NMSIE3. According to
+// specifications, each of operations NMS-1, NMS-3, NMS-4 has one output port, but NMS-5 has three output ports,
+// and one or two output ports can be disconnected. Hence, we will get legacy IR with CNNLayer "NonMaxSuppression"
+// with three output ports and with one element in OutputDataMap. Hence, we can create neither reference legacy IR with
+// one output, nor legacy IR with three outputs, because the function that compares instances of CNNLayer, uses
+// a comparison of the number of all output ports, not only connected ports. Moreover, we cannot creates IRs V10
+// with NMS-5, because output shapes of NMS-5 are dynamic.
+TEST_F(NGraphReaderTests, DISABLED_ReadNonMaxSuppression4) {
+   std::string model = R"V0G0N(
+<net name="Network" version="10">
+   <layers>
+       <layer id="0" name="in1" type="Parameter" version="opset1">
+           <data element_type="f32" shape="1,15130,4"/>
+           <output>
+               <port id="0" precision="FP32">
+                   <dim>1</dim>
+                   <dim>15130</dim>
+                   <dim>4</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="1" name="in2" type="Parameter" version="opset1">
+           <data element_type="f32" shape="1,80,15130"/>
+           <output>
+               <port id="0" precision="FP32">
+                   <dim>1</dim>
+                   <dim>80</dim>
+                   <dim>15130</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const" version="opset1">
+           <data offset="0" size="8"/>
+           <output>
+               <port id="0" precision="I64"/>
+           </output>
+       </layer>
+       <layer id="3" name="iou_threshold" precision="FP32" type="Const" version="opset1">
+           <data offset="8" size="4"/>
+           <output>
+               <port id="0" precision="FP32"/>
+           </output>
+       </layer>
+       <layer id="4" name="score_threshold" precision="FP32" type="Const" version="opset1">
+           <data offset="12" size="4"/>
+           <output>
+               <port id="0" precision="FP32"/>
+           </output>
+       </layer>
+       <layer id="5" name="nms" type="NonMaxSuppression" version="opset4">
+           <data box_encoding="corner" output_type="i32" sort_result_descending="0"/>
+           <input>
+               <port id="0" precision="FP32">
+                   <dim>1</dim>
+                   <dim>15130</dim>
+                   <dim>4</dim>
+               </port>
+               <port id="1" precision="FP32">
+                   <dim>1</dim>
+                   <dim>80</dim>
+                   <dim>15130</dim>
+               </port>
+               <port id="2" precision="I64"/>
+               <port id="3" precision="FP32"/>
+               <port id="4" precision="FP32"/>
+           </input>
+           <output>
+               <port id="5" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="6" name="mul" type="Multiply" version="opset1">
+           <input>
+               <port id="0" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+               <port id="1" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </input>
+           <output>
+               <port id="2" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="7" name="output" type="Result" version="opset1">
+           <input>
+               <port id="0" precision="I64">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </input>
+       </layer>
+   </layers>
+   <edges>
+       <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
+       <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
+       <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
+       <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
+       <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
+       <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
+       <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
+       <edge from-layer="6" from-port="2" to-layer="7" to-port="0"/>
+   </edges>
+</net>
+)V0G0N";
+   std::string refModel = R"V0G0N(
+<net name="Network" version="10">
+   <layers>
+       <layer id="0" name="in1" type="Parameter" version="opset1">
+           <data element_type="f32" shape="1,15130,4"/>
+           <output>
+               <port id="0" precision="FP32">
+                   <dim>1</dim>
+                   <dim>15130</dim>
+                   <dim>4</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="1" name="in2" type="Parameter" version="opset1">
+           <data element_type="f32" shape="1,80,15130"/>
+           <output>
+               <port id="0" precision="FP32">
+                   <dim>1</dim>
+                   <dim>80</dim>
+                   <dim>15130</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const" version="opset1">
+           <data offset="0" size="8"/>
+           <output>
+               <port id="0" precision="I64"/>
+           </output>
+       </layer>
+       <layer id="3" name="iou_threshold" precision="FP32" type="Const" version="opset1">
+           <data offset="8" size="4"/>
+           <output>
+               <port id="0" precision="FP32"/>
+           </output>
+       </layer>
+       <layer id="4" name="score_threshold" precision="FP32" type="Const" version="opset1">
+           <data offset="12" size="4"/>
+           <output>
+               <port id="0" precision="FP32"/>
+           </output>
+       </layer>
+       <layer id="5" name="nms" type="NonMaxSuppression" version="opset5">
+           <data box_encoding="corner" output_type="i32" sort_result_descending="0"/>
+           <input>
+               <port id="0" precision="FP32">
+                   <dim>1</dim>
+                   <dim>15130</dim>
+                   <dim>4</dim>
+               </port>
+               <port id="1" precision="FP32">
+                   <dim>1</dim>
+                   <dim>80</dim>
+                   <dim>15130</dim>
+               </port>
+               <port id="2" precision="I64"/>
+               <port id="3" precision="FP32"/>
+               <port id="4" precision="FP32"/>
+           </input>
+           <output>
+               <port id="5" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+               <port id="6" precision="FP32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+               <port id="7" precision="I32">
+                   <dim>1</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="6" name="mul" type="Multiply" version="opset1">
+           <input>
+               <port id="0" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+               <port id="1" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </input>
+           <output>
+               <port id="2" precision="I32">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </output>
+       </layer>
+       <layer id="7" name="output" type="Result" version="opset1">
+           <input>
+               <port id="0" precision="I64">
+                   <dim>16000</dim>
+                   <dim>3</dim>
+               </port>
+           </input>
+       </layer>
+   </layers>
+   <edges>
+       <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
+       <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
+       <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
+       <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
+       <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
+       <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
+       <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
+       <edge from-layer="6" from-port="2" to-layer="7" to-port="0"/>
+   </edges>
+</net>
+)V0G0N";
+
+    compareIRsV10(model, refModel, 16, [](Blob::Ptr& weights) {
+        auto * i64w = weights->buffer().as<int64_t*>();
+        i64w[0] = 200;
+
+        auto * fp32w = weights->buffer().as<float*>();
+        fp32w[2] = 0.5;
+        fp32w[3] = 0.05;
+    });
+}

From d82f3e1c9e2d6ea36b7ec0d025f465f8a5b8b869 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 3 Nov 2020 15:12:27 +0300
Subject: [PATCH 160/173] Deleted debug code.

---
 .../functional_test_utils/network_utils.cpp   | 30 +------------------
 1 file changed, 1 insertion(+), 29 deletions(-)

diff --git a/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp b/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
index 2fc65edcdb3519..8f4a507eec94ed 100644
--- a/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
+++ b/inference-engine/tests/ie_test_utils/functional_test_utils/network_utils.cpp
@@ -119,23 +119,6 @@ namespace FuncTestUtils {
             if (refLayer->outData[i]->getTensorDesc() != layer->outData[i]->getTensorDesc()) {
                 err_log.push_back("Layer " + layer->name + " and ref layer " + refLayer->name +
                                   " have different tensor desc for out Data");
-                std::cout << "For output " << i << "\n";
-                auto refDescriptor = refLayer->outData[i]->getTensorDesc();
-                auto layerDescriptor = layer->outData[i]->getTensorDesc();
-                std::cout << "ref precision:   " << refDescriptor.getPrecision() << "\n";
-                std::cout << "layer precision: " << layerDescriptor.getPrecision() << "\n";
-                auto refDims = refDescriptor.getDims();
-                auto layerDims = layerDescriptor.getDims();
-                std::cout <<"ref dims:    ";
-                for (size_t d : refDims) {
-                    std::cout << d << " ";
-                }
-                std::cout << "\n";
-                std::cout <<"layer dims:  ";
-                for (size_t d : layerDims) {
-                    std::cout << d << " ";
-                }
-                std::cout << "\n";
             }
             success = success && refLayer->outData[i]->getTensorDesc() == layer->outData[i]->getTensorDesc();
         }
@@ -202,19 +185,8 @@ namespace FuncTestUtils {
                 break;
             success = success && old_info.find(it.first) != old_info.end();
         }
-        if (!success) {
-            std::cout << "new output: ";
-            for (const auto &p : new_info) {
-                std::cout << p.first << " ";
-            }
-            std::cout << "\n";
-            std::cout << "old output: ";
-            for (const auto &p : old_info) {
-                std::cout << p.first << " ";
-            }
-            std::cout << "\n";
+        if (!success)
             THROW_IE_EXCEPTION << err_msg;
-        }
     }
 
     void

From 007b895c5cec8f75f4568c79591c9f1a62a0cd96 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 3 Nov 2020 15:31:37 +0300
Subject: [PATCH 161/173] Added comment about disabling tests
 IE_CPU.nonmaxsuppression.

---
 ngraph/test/runtime/ie/unit_test.manifest | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/ngraph/test/runtime/ie/unit_test.manifest b/ngraph/test/runtime/ie/unit_test.manifest
index 9055a893f2543a..3d66b925a56e9a 100644
--- a/ngraph/test/runtime/ie/unit_test.manifest
+++ b/ngraph/test/runtime/ie/unit_test.manifest
@@ -1136,7 +1136,16 @@ IE_CPU.log_softmax_2d_axis_neg2
 IE_CPU.log_softmax_3d_axis_0
 IE_CPU.log_softmax_3d_axis_neg3
 
-# Unsupported NMS-5
+# Next nine tests fails in CPU for the following reason. The nGraph function
+# for NMS-5 are passed to the method compile() of the backend, but this
+# method doesn't apply any nGraph transformations to the passed function,
+# and the plugin backend gets CNNNetwork with NMS-5, NMS-5 has dynamic shapes
+# for two of three outputs, and results of these two outputs are interpreted
+# as scalars. If we apply all needed nGraph transformations to the nGraph
+# function with NMS-5 to get the nGraph function with NMSIE3 (internal
+# operation, similar with NMS-5, but with all static output shapes), before
+# the method compile() call, then tests for INTERPRETER backend for NMS-5 will
+# fail, because NMSIE3 has not the reference implementation, but NMS-5 has one.
 IE_CPU.nonmaxsuppression_center_point_box_format
 IE_CPU.nonmaxsuppression_flipped_coordinates
 IE_CPU.nonmaxsuppression_identical_boxes

From 2d7a6342e6c1eaafd3fd7e75b847bc5123bbf453 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Tue, 3 Nov 2020 18:01:03 +0300
Subject: [PATCH 162/173] Written IE IR Reader test for NMS-4.

---
 .../non_max_suppression_tests.cpp             | 199 ++++--------------
 1 file changed, 39 insertions(+), 160 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index eaa96c33397b2d..55e6b00091862e 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -4,16 +4,25 @@
 
 #include <string>
 #include "ngraph_reader_tests.hpp"
+
+#include <ngraph/function.hpp>
 #include <ngraph/graph_util.hpp>
+#include <ngraph/opsets/opset1.hpp>
+#include <ngraph/opsets/opset3.hpp>
+#include <ngraph/opsets/opset4.hpp>
+#include <ngraph/opsets/opset5.hpp>
+#include <transformations/common_optimizations/common_optimizations.hpp>
+#include <transformations/init_node_info.hpp>
+#include <transformations/utils/utils.hpp>
 #include <ngraph/pass/manager.hpp>
 
+#include "common_test_utils/ngraph_test_utils.hpp"
 #include "generic_ie.hpp"
-#include <transformations/init_node_info.hpp>
-#include <transformations/common_optimizations/common_optimizations.hpp>
-#include <legacy/transformations/convert_opset1_to_legacy/convert_prior_to_ie_prior.hpp>
 
 #include "legacy/convert_function_to_cnn_network.hpp"
 
+using namespace ngraph;
+
 TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     std::string model = R"V0G0N(
 <net name="Network" version="10">
@@ -358,31 +367,16 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     });
 }
 
-static InferenceEngine::details::CNNNetworkImplPtr read_network_v10(const std::string& model, Blob::Ptr weights) {
+static std::shared_ptr<ngraph::Function> function_from_model(const std::string& model, Blob::Ptr weights) {
     Core ie;
-    auto result = std::make_shared<InferenceEngine::details::CNNNetworkImpl>();
 
     auto ngraphImpl = ie.ReadNetwork(model, weights);
     auto graph = ngraph::clone_function(*ngraphImpl.getFunction());
 
-    ::ngraph::op::GenericIE::DisableReshape noReshape(graph);
-
-    ::ngraph::pass::Manager manager;
-    manager.register_pass<::ngraph::pass::InitNodeInfo>();
-    // WA: ConvertPriorBox must be executed before the 1st ConstantFolding pass
-    manager.register_pass<::ngraph::pass::ConvertPriorBox>();
-    manager.register_pass<::ngraph::pass::CommonOptimizations>();
-    manager.run_passes(graph);
-
-    IE_SUPPRESS_DEPRECATED_START
-    InferenceEngine::details::convertFunctionToICNNNetwork(graph, ngraphImpl, result.get(), false);
-    IE_SUPPRESS_DEPRECATED_END
-
-    return result;
+    return graph;
 }
 
-static void compareIRsV10(const std::string& model, const std::string& ref, size_t weightsSize = 0, const std::function<void(Blob::Ptr&)>& fillBlob = {}) {
-    Core ie;
+static Blob::Ptr construct_weights(size_t weightsSize = 0, const std::function<void(Blob::Ptr&)>& fillBlob = {}) {
     Blob::Ptr weights;
 
     if (weightsSize) {
@@ -393,35 +387,10 @@ static void compareIRsV10(const std::string& model, const std::string& ref, size
             fillBlob(weights);
     }
 
-    auto network = read_network_v10(model, weights);
-    auto cnnNetwork = ie.ReadNetwork(ref, weights);
-
-    IE_SUPPRESS_DEPRECATED_START
-    FuncTestUtils::compareCNNNetworks(InferenceEngine::CNNNetwork(network), cnnNetwork, false);
-
-    for (auto it = details::CNNNetworkIterator(network.get()); it != details::CNNNetworkIterator(); it++) {
-        InferenceEngine::CNNLayerPtr layer = *it;
-        ASSERT_NE(nullptr, layer->getNode());
-    }
-
-    ASSERT_EQ(nullptr, cnnNetwork.getFunction());
-    for (auto it = details::CNNNetworkIterator(cnnNetwork); it != details::CNNNetworkIterator(); it++) {
-        InferenceEngine::CNNLayerPtr layer = *it;
-        ASSERT_EQ(nullptr, layer->getNode());
-    }
-    IE_SUPPRESS_DEPRECATED_END
+    return weights;
 }
 
-// When IRv10 model contains NMS-1, NMS-3, or NMS-4, we read the network, and then CNNNetworkImpl constructor
-// performs all nGraph transformations, and converts a resulting graph to legacy IR. In particular, NMS-1, NMS-3, NMS-4
-// are converted into NMS-5, and then NMS-5 will be converted into the inner operation NMSIE3. According to
-// specifications, each of operations NMS-1, NMS-3, NMS-4 has one output port, but NMS-5 has three output ports,
-// and one or two output ports can be disconnected. Hence, we will get legacy IR with CNNLayer "NonMaxSuppression"
-// with three output ports and with one element in OutputDataMap. Hence, we can create neither reference legacy IR with
-// one output, nor legacy IR with three outputs, because the function that compares instances of CNNLayer, uses
-// a comparison of the number of all output ports, not only connected ports. Moreover, we cannot creates IRs V10
-// with NMS-5, because output shapes of NMS-5 are dynamic.
-TEST_F(NGraphReaderTests, DISABLED_ReadNonMaxSuppression4) {
+TEST_F(NGraphReaderTests, ReadNonMaxSuppression4) {
    std::string model = R"V0G0N(
 <net name="Network" version="10">
    <layers>
@@ -506,122 +475,11 @@ TEST_F(NGraphReaderTests, DISABLED_ReadNonMaxSuppression4) {
            </output>
        </layer>
        <layer id="7" name="output" type="Result" version="opset1">
-           <input>
-               <port id="0" precision="I64">
-                   <dim>16000</dim>
-                   <dim>3</dim>
-               </port>
-           </input>
-       </layer>
-   </layers>
-   <edges>
-       <edge from-layer="0" from-port="0" to-layer="5" to-port="0"/>
-       <edge from-layer="1" from-port="0" to-layer="5" to-port="1"/>
-       <edge from-layer="2" from-port="0" to-layer="5" to-port="2"/>
-       <edge from-layer="3" from-port="0" to-layer="5" to-port="3"/>
-       <edge from-layer="4" from-port="0" to-layer="5" to-port="4"/>
-       <edge from-layer="5" from-port="5" to-layer="6" to-port="0"/>
-       <edge from-layer="5" from-port="5" to-layer="6" to-port="1"/>
-       <edge from-layer="6" from-port="2" to-layer="7" to-port="0"/>
-   </edges>
-</net>
-)V0G0N";
-   std::string refModel = R"V0G0N(
-<net name="Network" version="10">
-   <layers>
-       <layer id="0" name="in1" type="Parameter" version="opset1">
-           <data element_type="f32" shape="1,15130,4"/>
-           <output>
-               <port id="0" precision="FP32">
-                   <dim>1</dim>
-                   <dim>15130</dim>
-                   <dim>4</dim>
-               </port>
-           </output>
-       </layer>
-       <layer id="1" name="in2" type="Parameter" version="opset1">
-           <data element_type="f32" shape="1,80,15130"/>
-           <output>
-               <port id="0" precision="FP32">
-                   <dim>1</dim>
-                   <dim>80</dim>
-                   <dim>15130</dim>
-               </port>
-           </output>
-       </layer>
-       <layer id="2" name="max_output_boxes_per_class" precision="I64" type="Const" version="opset1">
-           <data offset="0" size="8"/>
-           <output>
-               <port id="0" precision="I64"/>
-           </output>
-       </layer>
-       <layer id="3" name="iou_threshold" precision="FP32" type="Const" version="opset1">
-           <data offset="8" size="4"/>
-           <output>
-               <port id="0" precision="FP32"/>
-           </output>
-       </layer>
-       <layer id="4" name="score_threshold" precision="FP32" type="Const" version="opset1">
-           <data offset="12" size="4"/>
-           <output>
-               <port id="0" precision="FP32"/>
-           </output>
-       </layer>
-       <layer id="5" name="nms" type="NonMaxSuppression" version="opset5">
-           <data box_encoding="corner" output_type="i32" sort_result_descending="0"/>
-           <input>
-               <port id="0" precision="FP32">
-                   <dim>1</dim>
-                   <dim>15130</dim>
-                   <dim>4</dim>
-               </port>
-               <port id="1" precision="FP32">
-                   <dim>1</dim>
-                   <dim>80</dim>
-                   <dim>15130</dim>
-               </port>
-               <port id="2" precision="I64"/>
-               <port id="3" precision="FP32"/>
-               <port id="4" precision="FP32"/>
-           </input>
-           <output>
-               <port id="5" precision="I32">
-                   <dim>16000</dim>
-                   <dim>3</dim>
-               </port>
-               <port id="6" precision="FP32">
-                   <dim>16000</dim>
-                   <dim>3</dim>
-               </port>
-               <port id="7" precision="I32">
-                   <dim>1</dim>
-               </port>
-           </output>
-       </layer>
-       <layer id="6" name="mul" type="Multiply" version="opset1">
            <input>
                <port id="0" precision="I32">
                    <dim>16000</dim>
                    <dim>3</dim>
                </port>
-               <port id="1" precision="I32">
-                   <dim>16000</dim>
-                   <dim>3</dim>
-               </port>
-           </input>
-           <output>
-               <port id="2" precision="I32">
-                   <dim>16000</dim>
-                   <dim>3</dim>
-               </port>
-           </output>
-       </layer>
-       <layer id="7" name="output" type="Result" version="opset1">
-           <input>
-               <port id="0" precision="I64">
-                   <dim>16000</dim>
-                   <dim>3</dim>
-               </port>
            </input>
        </layer>
    </layers>
@@ -638,7 +496,7 @@ TEST_F(NGraphReaderTests, DISABLED_ReadNonMaxSuppression4) {
 </net>
 )V0G0N";
 
-    compareIRsV10(model, refModel, 16, [](Blob::Ptr& weights) {
+    auto weights = construct_weights(16, [](Blob::Ptr& weights) {
         auto * i64w = weights->buffer().as<int64_t*>();
         i64w[0] = 200;
 
@@ -646,4 +504,25 @@ TEST_F(NGraphReaderTests, DISABLED_ReadNonMaxSuppression4) {
         fp32w[2] = 0.5;
         fp32w[3] = 0.05;
     });
+
+    auto graph = function_from_model(model, weights);
+
+    ::ngraph::pass::Manager manager;
+    manager.register_pass<::ngraph::pass::InitNodeInfo>();
+    manager.register_pass<::ngraph::pass::CommonOptimizations>();
+    manager.run_passes(graph);
+
+    auto boxes = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 15130, 4});
+    auto scores = std::make_shared<opset5::Parameter>(element::f32, Shape{1, 80, 15130});
+    auto max_output_boxes_per_class = opset5::Constant::create(element::i64, Shape{}, {200});
+    auto iou_threshold = opset5::Constant::create(element::f32, Shape{}, {0.5});
+    auto score_threshold = opset5::Constant::create(element::f32, Shape{}, {0.05});
+    auto nms = std::make_shared<opset5::NonMaxSuppression>(boxes, scores, max_output_boxes_per_class,  iou_threshold, score_threshold,
+                                                           opset5::NonMaxSuppression::BoxEncodingType::CORNER, true, ngraph::element::i32);
+
+    auto mul = std::make_shared<opset5::Multiply>(nms, nms);
+    auto graph_ref = std::make_shared<Function>(NodeVector{mul}, ParameterVector{boxes, scores});
+
+    auto res = compare_functions(graph, graph_ref);
+    ASSERT_TRUE(res.first) << res.second;
 }

From 0e1e135bb5bf26d9d3a92fbf406f941ae054f5d6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 09:52:33 +0300
Subject: [PATCH 163/173] Deleted function float_from_constant_node.

---
 ngraph/core/src/op/non_max_suppression.cpp | 27 +++-------------------
 1 file changed, 3 insertions(+), 24 deletions(-)

diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 8adc25f0e16770..8afabca37cce04 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -709,24 +709,6 @@ namespace
 
         return is_scalar(shape) || (is_vector(shape) && (shape[0] == 1));
     }
-
-    float float_from_constant_node(const std::shared_ptr<op::Constant>& constant,
-                                   const element::Type& et)
-    {
-        float result = 0.0f;
-        switch (et)
-        {
-        case element::Type_t::bf16:
-            result = static_cast<float>(constant->cast_vector<bfloat16>().at(0));
-            break;
-        case element::Type_t::f16:
-            result = static_cast<float>(constant->cast_vector<float16>().at(0));
-            break;
-        case element::Type_t::f32: result = constant->cast_vector<float>().at(0); break;
-        default: throw std::runtime_error("Unsupported data type in op NonMaxSuppression-5"); break;
-        }
-        return result;
-    }
 }
 
 void op::v5::NonMaxSuppression::validate()
@@ -868,8 +850,7 @@ float op::v5::NonMaxSuppression::iou_threshold_from_input() const
 
     const auto iou_threshold_input =
         as_type_ptr<op::Constant>(input_value(iou_threshold_port).get_node_shared_ptr());
-    iou_threshold =
-        float_from_constant_node(iou_threshold_input, get_input_element_type(iou_threshold_port));
+    iou_threshold = iou_threshold_input->cast_vector<float>().at(0);
 
     return iou_threshold;
 }
@@ -886,8 +867,7 @@ float op::v5::NonMaxSuppression::score_threshold_from_input() const
 
     const auto score_threshold_input =
         as_type_ptr<op::Constant>(input_value(score_threshold_port).get_node_shared_ptr());
-    score_threshold = float_from_constant_node(score_threshold_input,
-                                               get_input_element_type(score_threshold_port));
+    score_threshold = score_threshold_input->cast_vector<float>().at(0);
 
     return score_threshold;
 }
@@ -904,8 +884,7 @@ float op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const
 
     const auto soft_nms_sigma_input =
         as_type_ptr<op::Constant>(input_value(soft_nms_sigma_port).get_node_shared_ptr());
-    soft_nms_sigma =
-        float_from_constant_node(soft_nms_sigma_input, get_input_element_type(soft_nms_sigma_port));
+    soft_nms_sigma = soft_nms_sigma_input->cast_vector<float>().at(0);
 
     return soft_nms_sigma;
 }

From c10979b6179e8f12825a5e32fb33b2f39e50485e Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 13:40:41 +0300
Subject: [PATCH 164/173] Small fixes.

---
 .../src/legacy_api/src/ngraph_ops/nms_ie.cpp       |  3 +--
 ngraph/core/src/op/non_max_suppression.cpp         | 14 ++++----------
 2 files changed, 5 insertions(+), 12 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
index dc9ffc52eaa79a..be26e1c291e151 100644
--- a/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
+++ b/inference-engine/src/legacy_api/src/ngraph_ops/nms_ie.cpp
@@ -153,8 +153,7 @@ int64_t op::NonMaxSuppressionIE3::max_boxes_output_from_input() const {
     int64_t max_output_boxes{0};
 
     size_t num_of_inputs = inputs().size();
-    if (num_of_inputs < 3)
-    {
+    if (num_of_inputs < 3) {
         return 0;
     }
 
diff --git a/ngraph/core/src/op/non_max_suppression.cpp b/ngraph/core/src/op/non_max_suppression.cpp
index 8afabca37cce04..f5b94ac263c70f 100644
--- a/ngraph/core/src/op/non_max_suppression.cpp
+++ b/ngraph/core/src/op/non_max_suppression.cpp
@@ -682,8 +682,6 @@ std::shared_ptr<Node>
     }
 }
 
-using V5BoxEncoding = op::v5::NonMaxSuppression::BoxEncodingType;
-
 namespace
 {
     constexpr size_t boxes_port = 0;
@@ -825,8 +823,7 @@ int64_t op::v5::NonMaxSuppression::max_boxes_output_from_input() const
 {
     int64_t max_output_boxes{0};
 
-    size_t num_of_inputs = inputs().size();
-    if (num_of_inputs < 3)
+    if (inputs().size() < 3)
     {
         return 0;
     }
@@ -842,8 +839,7 @@ float op::v5::NonMaxSuppression::iou_threshold_from_input() const
 {
     float iou_threshold = 0.0f;
 
-    size_t num_of_inputs = inputs().size();
-    if (num_of_inputs < 4)
+    if (inputs().size() < 4)
     {
         return iou_threshold;
     }
@@ -859,8 +855,7 @@ float op::v5::NonMaxSuppression::score_threshold_from_input() const
 {
     float score_threshold = 0.0f;
 
-    size_t num_of_inputs = inputs().size();
-    if (num_of_inputs < 5)
+    if (inputs().size() < 5)
     {
         return score_threshold;
     }
@@ -876,8 +871,7 @@ float op::v5::NonMaxSuppression::soft_nms_sigma_from_input() const
 {
     float soft_nms_sigma = 0.0f;
 
-    size_t num_of_inputs = inputs().size();
-    if (num_of_inputs < 6)
+    if (inputs().size() < 6)
     {
         return soft_nms_sigma;
     }

From 1e10629ecd09f5fe01afe3c39fa29764a6b442e6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 15:42:43 +0300
Subject: [PATCH 165/173] Deleted functions function_from_model and
 construct_weights.

---
 .../non_max_suppression_tests.cpp             | 48 +++++++------------
 1 file changed, 18 insertions(+), 30 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index 55e6b00091862e..f37713d5c8feed 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -367,29 +367,6 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression5) {
     });
 }
 
-static std::shared_ptr<ngraph::Function> function_from_model(const std::string& model, Blob::Ptr weights) {
-    Core ie;
-
-    auto ngraphImpl = ie.ReadNetwork(model, weights);
-    auto graph = ngraph::clone_function(*ngraphImpl.getFunction());
-
-    return graph;
-}
-
-static Blob::Ptr construct_weights(size_t weightsSize = 0, const std::function<void(Blob::Ptr&)>& fillBlob = {}) {
-    Blob::Ptr weights;
-
-    if (weightsSize) {
-        weights = make_shared_blob<uint8_t>(TensorDesc(Precision::U8, {weightsSize}, Layout::C));
-        weights->allocate();
-        CommonTestUtils::fill_data(weights->buffer().as<float *>(), weights->size() / sizeof(float));
-        if (fillBlob)
-            fillBlob(weights);
-    }
-
-    return weights;
-}
-
 TEST_F(NGraphReaderTests, ReadNonMaxSuppression4) {
    std::string model = R"V0G0N(
 <net name="Network" version="10">
@@ -496,14 +473,25 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression4) {
 </net>
 )V0G0N";
 
-    auto weights = construct_weights(16, [](Blob::Ptr& weights) {
-        auto * i64w = weights->buffer().as<int64_t*>();
-        i64w[0] = 200;
+    constexpr size_t weightsSize = 16;
 
-        auto * fp32w = weights->buffer().as<float*>();
-        fp32w[2] = 0.5;
-        fp32w[3] = 0.05;
-    });
+    Blob::Ptr weights;
+
+    weights = make_shared_blob<uint8_t>(TensorDesc(Precision::U8, {weightsSize}, Layout::C));
+    weights->allocate();
+    CommonTestUtils::fill_data(weights->buffer().as<float *>(), weights->size() / sizeof(float));
+
+    auto * i64w = weights->buffer().as<int64_t*>();
+    i64w[0] = 200;
+
+    auto * fp32w = weights->buffer().as<float*>();
+    fp32w[2] = 0.5;
+    fp32w[3] = 0.05;
+
+    Core ie;
+
+    auto ngraphImpl = ie.ReadNetwork(model, weights);
+    auto graph = ngraph::clone_function(*ngraphImpl.getFunction());
 
     auto graph = function_from_model(model, weights);
 

From a11cd97ce769cc97c2689be5c0750130390fdc4d Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 15:44:02 +0300
Subject: [PATCH 166/173] Small fix.

---
 .../ngraph_reader/non_max_suppression_tests.cpp                 | 2 --
 1 file changed, 2 deletions(-)

diff --git a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
index f37713d5c8feed..27ee39178690ed 100644
--- a/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
+++ b/inference-engine/tests/functional/inference_engine/ngraph_reader/non_max_suppression_tests.cpp
@@ -493,8 +493,6 @@ TEST_F(NGraphReaderTests, ReadNonMaxSuppression4) {
     auto ngraphImpl = ie.ReadNetwork(model, weights);
     auto graph = ngraph::clone_function(*ngraphImpl.getFunction());
 
-    auto graph = function_from_model(model, weights);
-
     ::ngraph::pass::Manager manager;
     manager.register_pass<::ngraph::pass::InitNodeInfo>();
     manager.register_pass<::ngraph::pass::CommonOptimizations>();

From c86a7cdb3da7c42fe9a179bb3460f74b0dfa21e1 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 17:08:38 +0300
Subject: [PATCH 167/173] Replaced push_back with emplace_back in the
 conversion of NMS-5 to legacy.

---
 .../convert_nms_5_to_legacy.cpp                      | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
index ad37ba8f33f0cb..8ef630e16879d0 100644
--- a/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
+++ b/inference-engine/src/legacy_api/src/transformations/convert_opset1_to_legacy/convert_nms_5_to_legacy.cpp
@@ -49,13 +49,13 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
         Output<Node> new_shape_for_soft_nms_sigma = opset1::Constant::create(ngraph::element::i64, Shape{1}, {1});
 
         new_max_per_class = std::make_shared<opset1::Reshape>(arg2, new_shape_for_max_per_class, true);
-        new_ops.push_back(new_max_per_class.get_node_shared_ptr());
+        new_ops.emplace_back(new_max_per_class.get_node_shared_ptr());
 
         new_iou_threshold = std::make_shared<opset1::Reshape>(arg3, new_shape_for_iou_threshold, true);
-        new_ops.push_back(new_iou_threshold.get_node_shared_ptr());
+        new_ops.emplace_back(new_iou_threshold.get_node_shared_ptr());
 
         new_score_threshold = std::make_shared<opset1::Reshape>(arg4, new_shape_for_score_threshold, true);
-        new_ops.push_back(new_score_threshold.get_node_shared_ptr());
+        new_ops.emplace_back(new_score_threshold.get_node_shared_ptr());
 
         int center_point_box = 0;
         switch (nms_5->get_box_encoding()) {
@@ -74,7 +74,7 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
 
         if (num_of_inputs > 5 && nms_5->soft_nms_sigma_from_input() != 0.0f) {
             new_soft_nms_sigma = std::make_shared<opset1::Reshape>(new_args.at(5), new_shape_for_soft_nms_sigma, true);
-            new_ops.push_back(new_soft_nms_sigma.get_node_shared_ptr());
+            new_ops.emplace_back(new_soft_nms_sigma.get_node_shared_ptr());
             nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
                     new_args.at(0),
                     new_args.at(1),
@@ -85,7 +85,7 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
                     center_point_box,
                     nms_5->get_sort_result_descending(),
                     nms_5->get_output_type());
-            new_ops.push_back(nms_legacy);
+            new_ops.emplace_back(nms_legacy);
         } else {
             nms_legacy = std::make_shared<op::NonMaxSuppressionIE3>(
                     new_args.at(0),
@@ -96,7 +96,7 @@ ngraph::pass::ConvertNMS5ToLegacyMatcher::ConvertNMS5ToLegacyMatcher() {
                     center_point_box,
                     nms_5->get_sort_result_descending(),
                     nms_5->get_output_type());
-            new_ops.push_back(nms_legacy);
+            new_ops.emplace_back(nms_legacy);
         }
 
         nms_legacy->set_friendly_name(nms_5->get_friendly_name());

From 8a1f6a85a804d7ac7242c3441ffe9c0e3138ab0c Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 17:09:31 +0300
Subject: [PATCH 168/173] Small changes.

---
 .../src/mkldnn_plugin/nodes/non_max_suppression.cpp           | 4 ++--
 .../op_conversions/convert_previous_nms_to_nms_5.cpp          | 4 +---
 2 files changed, 3 insertions(+), 5 deletions(-)

diff --git a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
index e7f293075a2a7b..920196ad067908 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
@@ -21,9 +21,9 @@ class NonMaxSuppressionImpl: public ExtLayerBase {
     explicit NonMaxSuppressionImpl(const CNNLayer* layer) {
         try {
             if (layer->insData.size() < 2 || layer->insData.size() > 6)
-                THROW_IE_EXCEPTION << "NMS" << "has incorrect number of input edges: " << layer->insData.size();
+                THROW_IE_EXCEPTION << layer->name << "has incorrect number of input edges: " << layer->insData.size();
             if (layer->outData.size() < 1 || layer->outData.size() > 3)
-                THROW_IE_EXCEPTION << "NMS" << "has incorrect number of output edges: " << layer->outData.size();
+                THROW_IE_EXCEPTION << layer->name << "has incorrect number of output edges: " << layer->outData.size();
 
             if (layer->insData[NMS_BOXES].lock()->getTensorDesc().getPrecision() != Precision::FP32)
                 THROW_IE_EXCEPTION << layer->name << " Incorrect 'boxes' input precision. Only FP32 is supported!";
diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 7dc66e44d4493a..890ce49c55a20b 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -15,11 +15,9 @@
 #include <ngraph/rt_info.hpp>
 #include <ngraph/pattern/op/wrap_type.hpp>
 
-using NMS5BoxEncoding = ngraph::opset5::NonMaxSuppression::BoxEncodingType;
-
 struct NMSAttributes {
     ngraph::element::Type output_type;
-    NMS5BoxEncoding box_encoding;
+    ngraph::opset5::NonMaxSuppression::BoxEncodingType box_encoding;
     bool sort_result_descending;
     bool is_supported_nms;
 };

From 4e4fdee1fddadce48cb3ceb866b2ba57d954affd Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Thu, 5 Nov 2020 17:18:18 +0300
Subject: [PATCH 169/173] Some fixes.

---
 .../src/mkldnn_plugin/nodes/non_max_suppression.cpp           | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
index 920196ad067908..c3d1bd949fcb13 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/non_max_suppression.cpp
@@ -80,7 +80,9 @@ class NonMaxSuppressionImpl: public ExtLayerBase {
                 Precision inPrecision = Precision::FP32;
                 if (i == NMS_MAXOUTPUTBOXESPERCLASS)
                     inPrecision = Precision::I32;
-                const SizeVector& inDims = layer->insData[i].lock()->getTensorDesc().getDims();
+                auto inData = layer->insData[i].lock();
+                if (!inData) THROW_IE_EXCEPTION << "Layer " << layer->name << " has empty input with index: " << i;
+                const SizeVector& inDims = inData ->getTensorDesc().getDims();
                 inConfig.desc = TensorDesc(inPrecision, inDims, InferenceEngine::TensorDesc::getLayoutByDims(inDims));
                 config.inConfs.push_back(inConfig);
             }

From dddf4dc3a3527a059f14eff0a776bb80bd047eff Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 6 Nov 2020 11:57:24 +0300
Subject: [PATCH 170/173] Refactored reference implementation of NMS-5.

---
 .../runtime/reference/non_max_suppression.hpp |  46 +-
 .../runtime/reference/non_max_suppression.cpp | 476 +++++++++---------
 .../runtime/interpreter/int_executable.hpp    |  31 +-
 3 files changed, 299 insertions(+), 254 deletions(-)

diff --git a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
index a0e370c8b7d950..db343abd46dce3 100644
--- a/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
+++ b/ngraph/core/reference/include/ngraph/runtime/reference/non_max_suppression.hpp
@@ -24,7 +24,8 @@
 #include <functional>
 #include <map>
 #include <ngraph/runtime/host_tensor.hpp>
-#include "ngraph/coordinate_transform.hpp"
+#include <vector>
+#include "ngraph/node.hpp"
 #include "ngraph/op/util/op_types.hpp"
 #include "ngraph/ops.hpp"
 #include "ngraph/shape_util.hpp"
@@ -35,9 +36,46 @@ namespace ngraph
     {
         namespace reference
         {
-            void non_max_suppression(const op::v5::NonMaxSuppression* nms5,
-                                     const std::vector<std::shared_ptr<HostTensor>>& out,
-                                     const std::vector<std::shared_ptr<HostTensor>>& args);
+            struct InfoForNMS5
+            {
+                int64_t max_output_boxes_per_class;
+                float iou_threshold;
+                float score_threshold;
+                float soft_nms_sigma;
+                Shape out_shape;
+                Shape boxes_shape;
+                Shape scores_shape;
+                std::vector<float> boxes_data;
+                std::vector<float> scores_data;
+                size_t out_shape_size;
+                bool sort_result_descending;
+                ngraph::element::Type output_type;
+            };
+
+            InfoForNMS5 get_info_for_nms5_evaluation(const op::v5::NonMaxSuppression* nms5,
+                                                     const HostTensorVector& inputs);
+
+            void non_max_suppression(const float* boxes_data,
+                                     const Shape& boxes_data_shape,
+                                     const float* scores_data,
+                                     const Shape& scores_data_shape,
+                                     int64_t max_output_boxes_per_class,
+                                     float iou_threshold,
+                                     float score_threshold,
+                                     float soft_nms_sigma,
+                                     int64_t* selected_indices,
+                                     const Shape& selected_indices_shape,
+                                     float* selected_scores,
+                                     const Shape& selected_scores_shape,
+                                     int64_t* valid_outputs,
+                                     const bool sort_result_descending);
+
+            void nms5_postprocessing(const HostTensorVector& outputs,
+                                     const ngraph::element::Type output_type,
+                                     const std::vector<int64_t>& selected_indices,
+                                     const std::vector<float>& selected_scores,
+                                     int64_t valid_outputs,
+                                     const ngraph::element::Type selected_scores_type);
         }
     }
 }
diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index a6a113c7ca04f4..233fda5c81d868 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -280,298 +280,215 @@ namespace
         auto result = get_floats(scores, scores_shape);
         return result;
     }
+}
 
-    void nms_postprocessing(const std::vector<std::shared_ptr<HostTensor>>& outputs,
-                            const ngraph::element::Type output_type,
-                            const std::vector<int64_t>& selected_indices,
-                            const std::vector<float>& selected_scores,
-                            int64_t valid_outputs,
-                            const ngraph::element::Type selected_scores_type)
+namespace ngraph
+{
+    namespace runtime
     {
-        size_t num_of_outputs = outputs.size();
-        size_t selected_size = valid_outputs * 3;
-
-        if (output_type == ngraph::element::i64)
-        {
-            int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
-            memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
-        }
-        else
+        namespace reference
         {
-            int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
-            for (size_t i = 0; i < selected_size; ++i)
+            InfoForNMS5 get_info_for_nms5_evaluation(const op::v5::NonMaxSuppression* nms5,
+                                                     const HostTensorVector& inputs)
             {
-                indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
-            }
-        }
+                InfoForNMS5 result;
 
-        if (num_of_outputs < 2)
-        {
-            return;
-        }
+                result.max_output_boxes_per_class = nms5->max_boxes_output_from_input();
+                result.iou_threshold = nms5->iou_threshold_from_input();
+                result.score_threshold = nms5->score_threshold_from_input();
+                result.soft_nms_sigma = nms5->soft_nms_sigma_from_input();
 
-        size_t selected_scores_size = selected_scores.size();
+                auto selected_indices_shape =
+                    infer_selected_indices_shape(inputs, result.max_output_boxes_per_class);
+                result.out_shape = selected_indices_shape.to_shape();
 
-        switch (selected_scores_type)
-        {
-        case element::Type_t::bf16:
-        {
-            bfloat16* scores_ptr = outputs[1]->get_data_ptr<bfloat16>();
-            for (size_t i = 0; i < selected_scores_size; ++i)
-            {
-                scores_ptr[i] = bfloat16(selected_scores[i]);
-            }
-        }
-        break;
-        case element::Type_t::f16:
-        {
-            float16* scores_ptr = outputs[1]->get_data_ptr<float16>();
-            for (size_t i = 0; i < selected_scores_size; ++i)
-            {
-                scores_ptr[i] = float16(selected_scores[i]);
-            }
-        }
-        break;
-        case element::Type_t::f32:
-        {
-            float* scores_ptr = outputs[1]->get_data_ptr<float>();
-            memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
-        }
-        break;
-        default:;
-        }
+                result.boxes_shape = inputs[boxes_port]->get_shape();
+                result.scores_shape = inputs[scores_port]->get_shape();
 
-        if (num_of_outputs < 3)
-        {
-            return;
-        }
+                result.boxes_data = prepare_boxes_data(inputs[boxes_port],
+                                                       result.boxes_shape,
+                                                       nms5->get_box_encoding());
+                result.scores_data = prepare_scores_data(inputs[scores_port], result.scores_shape);
 
-        if (output_type == ngraph::element::i64)
-        {
-            int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
-            *valid_outputs_ptr = valid_outputs;
-        }
-        else
-        {
-            int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
-            *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
-        }
-    }
+                result.out_shape_size = shape_size(result.out_shape);
 
-    void eval(const float* boxes_data,
-              const Shape& boxes_data_shape,
-              const float* scores_data,
-              const Shape& scores_data_shape,
-              int64_t max_output_boxes_per_class,
-              float iou_threshold,
-              float score_threshold,
-              float soft_nms_sigma,
-              int64_t* selected_indices,
-              const Shape& selected_indices_shape,
-              float* selected_scores,
-              const Shape& selected_scores_shape,
-              int64_t* valid_outputs,
-              const bool sort_result_descending)
-    {
-        float scale = 0.0f;
-        if (soft_nms_sigma > 0.0f)
-        {
-            scale = -0.5f / soft_nms_sigma;
-        }
+                result.sort_result_descending = nms5->get_sort_result_descending();
 
-        auto func = [iou_threshold, scale](float iou) {
-            const float weight = std::exp(scale * iou * iou);
-            return iou <= iou_threshold ? weight : 0.0f;
-        };
+                result.output_type = nms5->get_output_type();
 
-        // boxes shape: {num_batches, num_boxes, 4}
-        // scores shape: {num_batches, num_classes, num_boxes}
-        int64_t num_batches = static_cast<int64_t>(scores_data_shape[0]);
-        int64_t num_classes = static_cast<int64_t>(scores_data_shape[1]);
-        int64_t num_boxes = static_cast<int64_t>(boxes_data_shape[1]);
+                return result;
+            }
 
-        SelectedIndex* selected_indices_ptr = reinterpret_cast<SelectedIndex*>(selected_indices);
-        SelectedScore* selected_scores_ptr = reinterpret_cast<SelectedScore*>(selected_scores);
+            void non_max_suppression(const float* boxes_data,
+                                     const Shape& boxes_data_shape,
+                                     const float* scores_data,
+                                     const Shape& scores_data_shape,
+                                     int64_t max_output_boxes_per_class,
+                                     float iou_threshold,
+                                     float score_threshold,
+                                     float soft_nms_sigma,
+                                     int64_t* selected_indices,
+                                     const Shape& selected_indices_shape,
+                                     float* selected_scores,
+                                     const Shape& selected_scores_shape,
+                                     int64_t* valid_outputs,
+                                     const bool sort_result_descending)
+            {
+                float scale = 0.0f;
+                if (soft_nms_sigma > 0.0f)
+                {
+                    scale = -0.5f / soft_nms_sigma;
+                }
 
-        size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
+                auto func = [iou_threshold, scale](float iou) {
+                    const float weight = std::exp(scale * iou * iou);
+                    return iou <= iou_threshold ? weight : 0.0f;
+                };
 
-        int64_t num_of_valid_boxes = 0;
+                // boxes shape: {num_batches, num_boxes, 4}
+                // scores shape: {num_batches, num_classes, num_boxes}
+                int64_t num_batches = static_cast<int64_t>(scores_data_shape[0]);
+                int64_t num_classes = static_cast<int64_t>(scores_data_shape[1]);
+                int64_t num_boxes = static_cast<int64_t>(boxes_data_shape[1]);
 
-        std::vector<BoxInfo> filteredBoxes;
+                SelectedIndex* selected_indices_ptr =
+                    reinterpret_cast<SelectedIndex*>(selected_indices);
+                SelectedScore* selected_scores_ptr =
+                    reinterpret_cast<SelectedScore*>(selected_scores);
 
-        for (int64_t batch = 0; batch < num_batches; batch++)
-        {
-            const float* boxesPtr = boxes_data + batch * num_boxes * 4;
-            Rectangle* r = reinterpret_cast<Rectangle*>(const_cast<float*>(boxesPtr));
+                size_t boxes_per_class = static_cast<size_t>(max_output_boxes_per_class);
 
-            for (int64_t class_idx = 0; class_idx < num_classes; class_idx++)
-            {
-                const float* scoresPtr =
-                    scores_data + batch * (num_classes * num_boxes) + class_idx * num_boxes;
+                int64_t num_of_valid_boxes = 0;
 
-                std::vector<BoxInfo> candidate_boxes;
-                candidate_boxes.reserve(num_boxes);
+                std::vector<BoxInfo> filteredBoxes;
 
-                for (size_t box_idx = 0; box_idx < num_boxes; box_idx++)
+                for (int64_t batch = 0; batch < num_batches; batch++)
                 {
-                    if (scoresPtr[box_idx] > score_threshold)
-                    {
-                        candidate_boxes.emplace_back(
-                            r[box_idx], box_idx, scoresPtr[box_idx], 0, batch, class_idx);
-                    }
-                }
-
-                std::priority_queue<BoxInfo> sorted_boxes(std::less<BoxInfo>(),
-                                                          std::move(candidate_boxes));
-
-                std::vector<BoxInfo> selected;
-                // Get the next box with top score, filter by iou_threshold
+                    const float* boxesPtr = boxes_data + batch * num_boxes * 4;
+                    Rectangle* r = reinterpret_cast<Rectangle*>(const_cast<float*>(boxesPtr));
 
-                BoxInfo next_candidate;
-                float original_score;
-
-                while (!sorted_boxes.empty() && selected.size() < boxes_per_class)
-                {
-                    next_candidate = sorted_boxes.top();
-                    original_score = next_candidate.score;
-                    sorted_boxes.pop();
-
-                    bool should_hard_suppress = false;
-                    for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
-                         j >= next_candidate.suppress_begin_index;
-                         --j)
+                    for (int64_t class_idx = 0; class_idx < num_classes; class_idx++)
                     {
-                        float iou = intersectionOverUnion(next_candidate.box, selected[j].box);
-                        next_candidate.score *= func(iou);
+                        const float* scoresPtr =
+                            scores_data + batch * (num_classes * num_boxes) + class_idx * num_boxes;
 
-                        if (iou >= iou_threshold)
-                        {
-                            should_hard_suppress = true;
-                            break;
-                        }
+                        std::vector<BoxInfo> candidate_boxes;
+                        candidate_boxes.reserve(num_boxes);
 
-                        if (next_candidate.score <= score_threshold)
+                        for (size_t box_idx = 0; box_idx < num_boxes; box_idx++)
                         {
-                            break;
+                            if (scoresPtr[box_idx] > score_threshold)
+                            {
+                                candidate_boxes.emplace_back(
+                                    r[box_idx], box_idx, scoresPtr[box_idx], 0, batch, class_idx);
+                            }
                         }
-                    }
 
-                    next_candidate.suppress_begin_index = selected.size();
+                        std::priority_queue<BoxInfo> sorted_boxes(std::less<BoxInfo>(),
+                                                                std::move(candidate_boxes));
 
-                    if (!should_hard_suppress)
-                    {
-                        if (next_candidate.score == original_score)
+                        std::vector<BoxInfo> selected;
+                        // Get the next box with top score, filter by iou_threshold
+
+                        BoxInfo next_candidate;
+                        float original_score;
+
+                        while (!sorted_boxes.empty() && selected.size() < boxes_per_class)
                         {
-                            selected.push_back(next_candidate);
-                            continue;
+                            next_candidate = sorted_boxes.top();
+                            original_score = next_candidate.score;
+                            sorted_boxes.pop();
+
+                            bool should_hard_suppress = false;
+                            for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
+                                j >= next_candidate.suppress_begin_index;
+                                --j)
+                            {
+                                float iou = intersectionOverUnion(next_candidate.box, selected[j].box);
+                                next_candidate.score *= func(iou);
+
+                                if (iou >= iou_threshold)
+                                {
+                                    should_hard_suppress = true;
+                                    break;
+                                }
+
+                                if (next_candidate.score <= score_threshold)
+                                {
+                                    break;
+                                }
+                            }
+
+                            next_candidate.suppress_begin_index = selected.size();
+
+                            if (!should_hard_suppress)
+                            {
+                                if (next_candidate.score == original_score)
+                                {
+                                    selected.push_back(next_candidate);
+                                    continue;
+                                }
+                                if (next_candidate.score > score_threshold)
+                                {
+                                    sorted_boxes.push(next_candidate);
+                                }
+                            }
                         }
-                        if (next_candidate.score > score_threshold)
+
+                        for (const auto& box_info : selected)
                         {
-                            sorted_boxes.push(next_candidate);
+                            filteredBoxes.push_back(box_info);
                         }
                     }
                 }
 
-                for (const auto& box_info : selected)
+                if (sort_result_descending)
                 {
-                    filteredBoxes.push_back(box_info);
+                    std::sort(filteredBoxes.begin(),
+                            filteredBoxes.end(),
+                            [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
                 }
-            }
-        }
 
-        if (sort_result_descending)
-        {
-            std::sort(filteredBoxes.begin(),
-                      filteredBoxes.end(),
-                      [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
-        }
+                size_t max_num_of_selected_indices = selected_indices_shape[0];
+                size_t output_size = std::min(filteredBoxes.size(), max_num_of_selected_indices);
 
-        size_t max_num_of_selected_indices = selected_indices_shape[0];
-        size_t output_size = std::min(filteredBoxes.size(), max_num_of_selected_indices);
+                *valid_outputs = output_size;
 
-        *valid_outputs = output_size;
-
-        size_t idx;
-        for (idx = 0; idx < output_size; idx++)
-        {
-            const auto& box_info = filteredBoxes[idx];
-            SelectedIndex selected_index{
-                box_info.batch_index, box_info.class_index, box_info.index};
-            SelectedScore selected_score{static_cast<float>(box_info.batch_index),
-                                         static_cast<float>(box_info.class_index),
-                                         box_info.score};
-
-            selected_indices_ptr[idx] = selected_index;
-            selected_scores_ptr[idx] = selected_score;
-        }
+                size_t idx;
+                for (idx = 0; idx < output_size; idx++)
+                {
+                    const auto& box_info = filteredBoxes[idx];
+                    SelectedIndex selected_index{
+                        box_info.batch_index, box_info.class_index, box_info.index};
+                    SelectedScore selected_score{static_cast<float>(box_info.batch_index),
+                                                static_cast<float>(box_info.class_index),
+                                                box_info.score};
+
+                    selected_indices_ptr[idx] = selected_index;
+                    selected_scores_ptr[idx] = selected_score;
+                }
 
-        SelectedIndex selected_index_filler{0, 0, 0};
-        SelectedScore selected_score_filler{0.0f, 0.0f, 0.0f};
-        for (; idx < max_num_of_selected_indices; idx++)
-        {
-            selected_indices_ptr[idx] = selected_index_filler;
-            selected_scores_ptr[idx] = selected_score_filler;
-        }
-    }
-}
+                SelectedIndex selected_index_filler{0, 0, 0};
+                SelectedScore selected_score_filler{0.0f, 0.0f, 0.0f};
+                for (; idx < max_num_of_selected_indices; idx++)
+                {
+                    selected_indices_ptr[idx] = selected_index_filler;
+                    selected_scores_ptr[idx] = selected_score_filler;
+                }
+            }
 
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace reference
-        {
-            void non_max_suppression(const op::v5::NonMaxSuppression* nms5,
-                                     const std::vector<std::shared_ptr<HostTensor>>& outputs,
-                                     const std::vector<std::shared_ptr<HostTensor>>& inputs)
+            void nms5_postprocessing(const HostTensorVector& outputs,
+                                     const ngraph::element::Type output_type,
+                                     const std::vector<int64_t>& selected_indices,
+                                     const std::vector<float>& selected_scores,
+                                     int64_t valid_outputs,
+                                     const ngraph::element::Type selected_scores_type)
             {
-                int64_t max_output_boxes_per_class = nms5->max_boxes_output_from_input();
-                float iou_threshold = nms5->iou_threshold_from_input();
-                float score_threshold = nms5->score_threshold_from_input();
-                float soft_nms_sigma = nms5->soft_nms_sigma_from_input();
-
-                auto selected_indices_shape =
-                    infer_selected_indices_shape(inputs, max_output_boxes_per_class);
-                Shape out_shape = selected_indices_shape.to_shape();
-
-                Shape boxes_shape = inputs[boxes_port]->get_shape();
-                Shape scores_shape = inputs[scores_port]->get_shape();
-
-                auto boxes_data =
-                    prepare_boxes_data(inputs[boxes_port], boxes_shape, nms5->get_box_encoding());
-                auto scores_data = prepare_scores_data(inputs[scores_port], scores_shape);
-
-                size_t out_shape_size = shape_size(out_shape);
-
-                std::vector<int64_t> selected_indices(out_shape_size);
-                std::vector<float> selected_scores(out_shape_size);
-                int64_t valid_outputs = 0;
-
-                eval(boxes_data.data(),
-                     boxes_shape,
-                     scores_data.data(),
-                     scores_shape,
-                     max_output_boxes_per_class,
-                     iou_threshold,
-                     score_threshold,
-                     soft_nms_sigma,
-                     selected_indices.data(),
-                     out_shape,
-                     selected_scores.data(),
-                     out_shape,
-                     &valid_outputs,
-                     nms5->get_sort_result_descending());
-
-                auto output_type = nms5->get_output_type();
-
                 outputs[0]->set_element_type(output_type);
                 outputs[0]->set_shape(Shape{static_cast<size_t>(valid_outputs), 3});
 
                 size_t num_of_outputs = outputs.size();
 
-                auto selected_scores_type =
-                    (inputs.size() < 4) ? element::f32 : inputs[3]->get_element_type();
-
                 if (num_of_outputs >= 2)
                 {
                     outputs[1]->set_element_type(selected_scores_type);
@@ -584,12 +501,73 @@ namespace ngraph
                     outputs[2]->set_shape(Shape{1});
                 }
 
-                nms_postprocessing(outputs,
-                                   output_type,
-                                   selected_indices,
-                                   selected_scores,
-                                   valid_outputs,
-                                   selected_scores_type);
+                size_t selected_size = valid_outputs * 3;
+
+                if (output_type == ngraph::element::i64)
+                {
+                    int64_t* indices_ptr = outputs[0]->get_data_ptr<int64_t>();
+                    memcpy(indices_ptr, selected_indices.data(), selected_size * sizeof(int64_t));
+                }
+                else
+                {
+                    int32_t* indices_ptr = outputs[0]->get_data_ptr<int32_t>();
+                    for (size_t i = 0; i < selected_size; ++i)
+                    {
+                        indices_ptr[i] = static_cast<int32_t>(selected_indices[i]);
+                    }
+                }
+
+                if (num_of_outputs < 2)
+                {
+                    return;
+                }
+
+                size_t selected_scores_size = selected_scores.size();
+
+                switch (selected_scores_type)
+                {
+                case element::Type_t::bf16:
+                {
+                    bfloat16* scores_ptr = outputs[1]->get_data_ptr<bfloat16>();
+                    for (size_t i = 0; i < selected_scores_size; ++i)
+                    {
+                        scores_ptr[i] = bfloat16(selected_scores[i]);
+                    }
+                }
+                break;
+                case element::Type_t::f16:
+                {
+                    float16* scores_ptr = outputs[1]->get_data_ptr<float16>();
+                    for (size_t i = 0; i < selected_scores_size; ++i)
+                    {
+                        scores_ptr[i] = float16(selected_scores[i]);
+                    }
+                }
+                break;
+                case element::Type_t::f32:
+                {
+                    float* scores_ptr = outputs[1]->get_data_ptr<float>();
+                    memcpy(scores_ptr, selected_scores.data(), selected_size * sizeof(float));
+                }
+                break;
+                default:;
+                }
+
+                if (num_of_outputs < 3)
+                {
+                    return;
+                }
+
+                if (output_type == ngraph::element::i64)
+                {
+                    int64_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int64_t>();
+                    *valid_outputs_ptr = valid_outputs;
+                }
+                else
+                {
+                    int32_t* valid_outputs_ptr = outputs[2]->get_data_ptr<int32_t>();
+                    *valid_outputs_ptr = static_cast<int32_t>(valid_outputs);
+                }
             }
         }
     }
diff --git a/ngraph/test/runtime/interpreter/int_executable.hpp b/ngraph/test/runtime/interpreter/int_executable.hpp
index 075bb78f5606ec..1e6ce57a313d5f 100644
--- a/ngraph/test/runtime/interpreter/int_executable.hpp
+++ b/ngraph/test/runtime/interpreter/int_executable.hpp
@@ -1342,7 +1342,36 @@ class INTERPRETER_BACKEND_API ngraph::runtime::interpreter::INTExecutable : publ
             const op::v5::NonMaxSuppression* nms =
                 static_cast<const op::v5::NonMaxSuppression*>(&node);
 
-            reference::non_max_suppression(nms, out, args);
+            auto info = reference::get_info_for_nms5_evaluation(nms, args);
+
+            std::vector<int64_t> selected_indices(info.out_shape_size);
+            std::vector<float> selected_scores(info.out_shape_size);
+            int64_t valid_outputs = 0;
+
+            reference::non_max_suppression(info.boxes_data.data(),
+                                           info.boxes_shape,
+                                           info.scores_data.data(),
+                                           info.scores_shape,
+                                           info.max_output_boxes_per_class,
+                                           info.iou_threshold,
+                                           info.score_threshold,
+                                           info.soft_nms_sigma,
+                                           selected_indices.data(),
+                                           info.out_shape,
+                                           selected_scores.data(),
+                                           info.out_shape,
+                                           &valid_outputs,
+                                           info.sort_result_descending);
+
+            auto selected_scores_type =
+                (args.size() < 4) ? element::f32 : args[3]->get_element_type();
+
+            reference::nms5_postprocessing(out,
+                                           info.output_type,
+                                           selected_indices,
+                                           selected_scores,
+                                           valid_outputs,
+                                           selected_scores_type);
             break;
         }
 

From 11863f1c8b043e6ee24594d56863758936b69cc6 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 6 Nov 2020 12:31:51 +0300
Subject: [PATCH 171/173] Moved structure NMSAttributes to unnamed namespace.

---
 .../op_conversions/convert_previous_nms_to_nms_5.cpp        | 6 +++---
 .../reference/src/runtime/reference/non_max_suppression.cpp | 4 ++--
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
index 890ce49c55a20b..3328f02058d2a7 100644
--- a/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
+++ b/inference-engine/src/transformations/src/transformations/op_conversions/convert_previous_nms_to_nms_5.cpp
@@ -15,6 +15,9 @@
 #include <ngraph/rt_info.hpp>
 #include <ngraph/pattern/op/wrap_type.hpp>
 
+using namespace ngraph;
+
+namespace {
 struct NMSAttributes {
     ngraph::element::Type output_type;
     ngraph::opset5::NonMaxSuppression::BoxEncodingType box_encoding;
@@ -22,9 +25,6 @@ struct NMSAttributes {
     bool is_supported_nms;
 };
 
-using namespace ngraph;
-
-namespace {
     NMSAttributes get_nms4_attrs(const std::shared_ptr<ngraph::opset4::NonMaxSuppression>& nms4) {
         NMSAttributes attrs;
 
diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 233fda5c81d868..39af8cef8ec048 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -461,8 +461,8 @@ namespace ngraph
                     SelectedIndex selected_index{
                         box_info.batch_index, box_info.class_index, box_info.index};
                     SelectedScore selected_score{static_cast<float>(box_info.batch_index),
-                                                static_cast<float>(box_info.class_index),
-                                                box_info.score};
+                                                 static_cast<float>(box_info.class_index),
+                                                 box_info.score};
 
                     selected_indices_ptr[idx] = selected_index;
                     selected_scores_ptr[idx] = selected_score;

From db454e9b6ccf08badfcba556bf08fdfb5fe2029a Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 6 Nov 2020 12:55:50 +0300
Subject: [PATCH 172/173] Code style fixes.

---
 .../runtime/reference/non_max_suppression.cpp  | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 39af8cef8ec048..81dbf707253319 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -305,9 +305,8 @@ namespace ngraph
                 result.boxes_shape = inputs[boxes_port]->get_shape();
                 result.scores_shape = inputs[scores_port]->get_shape();
 
-                result.boxes_data = prepare_boxes_data(inputs[boxes_port],
-                                                       result.boxes_shape,
-                                                       nms5->get_box_encoding());
+                result.boxes_data =  prepare_boxes_data(
+                    inputs[boxes_port], result.boxes_shape, nms5->get_box_encoding());
                 result.scores_data = prepare_scores_data(inputs[scores_port], result.scores_shape);
 
                 result.out_shape_size = shape_size(result.out_shape);
@@ -385,7 +384,7 @@ namespace ngraph
                         }
 
                         std::priority_queue<BoxInfo> sorted_boxes(std::less<BoxInfo>(),
-                                                                std::move(candidate_boxes));
+                                                                  std::move(candidate_boxes));
 
                         std::vector<BoxInfo> selected;
                         // Get the next box with top score, filter by iou_threshold
@@ -401,10 +400,11 @@ namespace ngraph
 
                             bool should_hard_suppress = false;
                             for (int64_t j = static_cast<int64_t>(selected.size()) - 1;
-                                j >= next_candidate.suppress_begin_index;
-                                --j)
+                                 j >= next_candidate.suppress_begin_index;
+                                 --j)
                             {
-                                float iou = intersectionOverUnion(next_candidate.box, selected[j].box);
+                                float iou =
+                                    intersectionOverUnion(next_candidate.box, selected[j].box);
                                 next_candidate.score *= func(iou);
 
                                 if (iou >= iou_threshold)
@@ -445,8 +445,8 @@ namespace ngraph
                 if (sort_result_descending)
                 {
                     std::sort(filteredBoxes.begin(),
-                            filteredBoxes.end(),
-                            [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
+                              filteredBoxes.end(),
+                              [](const BoxInfo& l, const BoxInfo& r) { return l.score > r.score; });
                 }
 
                 size_t max_num_of_selected_indices = selected_indices_shape[0];

From 279c2d9061a4fc1e7ea6274dc8f2438c3db44eb1 Mon Sep 17 00:00:00 2001
From: VGavrilov <vladimir.gavrilov@intel.com>
Date: Fri, 6 Nov 2020 13:03:34 +0300
Subject: [PATCH 173/173] Small fix.

---
 .../reference/src/runtime/reference/non_max_suppression.cpp     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
index 81dbf707253319..44fafd726afa43 100644
--- a/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
+++ b/ngraph/core/reference/src/runtime/reference/non_max_suppression.cpp
@@ -305,7 +305,7 @@ namespace ngraph
                 result.boxes_shape = inputs[boxes_port]->get_shape();
                 result.scores_shape = inputs[scores_port]->get_shape();
 
-                result.boxes_data =  prepare_boxes_data(
+                result.boxes_data = prepare_boxes_data(
                     inputs[boxes_port], result.boxes_shape, nms5->get_box_encoding());
                 result.scores_data = prepare_scores_data(inputs[scores_port], result.scores_shape);