speedup for .class.C for MLP #121

pseyfert · 2015-12-21T17:52:41Z

I ported the changes from pseyfert/tmva-mlp to the code generation. As a test I just ran the class.C network resulting from tutorials/tmva/TMVAClassification.C (with "MLP") and evaluated it similar to tutorials/tmva/TMVAClassificationApplication.C. According to callgrind the network evaluation is ~17% cpu cycles faster.
NB: i did not port all changes from pseyfert/tmva-mlp - I did not import the SSE/AVX intrinsics and avoided what seemed too difficult (optimising the putIndices and getIndices out)

pseyfert · 2016-10-31T14:31:16Z

just added ec7918b (that switch hardly seemed to do do anything)

pseyfert · 2017-03-09T16:20:16Z

looks like the conflict is purely whitespace changes on master, mergeing by hand.

phsft-bot · 2017-03-09T16:20:19Z

Can one of the admins verify this patch?

pseyfert · 2017-05-13T16:22:50Z

this is not superseded by #572 (though merging both might result in conflicts. i can work that out in #121 once #572 is merged).

phsft-bot · 2017-05-13T16:22:52Z

Can one of the admins verify this patch?

vgvassilev · 2017-05-13T16:33:40Z

Could you use git rebase master on this branch instead. We don't like these merge commits...

pseyfert · 2017-05-13T17:13:55Z

I think that's what I meant. My idea was:

speedup for .class.C for MLP #121 is in no-conflict state
wait until [tmva] MLP codegen threadsafety #572 is merged (easier to review, more important)
speedup for .class.C for MLP #121 is in conflict state
git rebase master on pseyfert/tmva-mlp-codegen
work out the merge conflicts in my repository
push the fixed pseyfert/tmva-mlp-codegen
speedup for .class.C for MLP #121 is in no-conflict state

pseyfert · 2017-05-26T18:55:32Z

just rebased (harder than expected). Let me double check I didn't mess up.

vgvassilev · 2017-06-02T11:59:27Z

@phsft-bot build!

@pseyfert, I was wondering can we create a gtest testing that part? This is good for test coverage and documentation purposes...

phsft-bot · 2017-06-02T11:59:41Z

Starting build on gcc49/centos7, native/mac1012, gcc49/slc6, gcc62/slc6, native/ubuntu14 with CMake flags -Dvc=OFF -Dimt=ON -Dccache=ON

phsft-bot · 2017-06-02T12:11:23Z

Build failed on mac1012/native.
See console output.

Errors:

/Volumes/HDD2/ec/build/workspace/root-pullrequests-build/root/tmva/tmva/src/MethodANNBase.cxx:1120:74: error: expected ';' after expression

phsft-bot · 2017-06-02T12:15:39Z

Build failed on ubuntu14/native.
See console output.

Errors:

/mnt/vdb/lsf/workspace/root-pullrequests-build/root/tmva/tmva/src/MethodANNBase.cxx:1121:7: error: expected ‘;’ before ‘}’ token

phsft-bot · 2017-06-02T12:16:24Z

Build failed on centos7/gcc49.
See console output.

Errors:

/mnt/build/jenkins/workspace/root-pullrequests-build/root/tmva/tmva/src/MethodANNBase.cxx:1121:7: error: expected ‘;’ before ‘}’ token

phsft-bot · 2017-06-02T12:19:09Z

Build failed on slc6/gcc62.
See console output.

Errors:

/mnt/build/jenkins/workspace/root-pullrequests-build/root/tmva/tmva/src/MethodANNBase.cxx:1121:7: error: expected ‘;’ before ‘}’ token

phsft-bot · 2017-06-02T12:22:59Z

Build failed on slc6/gcc49.
See console output.

Errors:

/mnt/build/jenkins/workspace/root-pullrequests-build/root/tmva/tmva/src/MethodANNBase.cxx:1121:7: error: expected ‘;’ before ‘}’ token

vgvassilev · 2017-06-08T08:22:49Z

@pseyfert ping...

pseyfert · 2017-07-04T16:06:01Z

back on this one.

it appears my change is not super numerically stable. I ran the network from tutorials/tmva/TMVAClassification and the floating point comparison from https://stackoverflow.com/a/253874 with rapidcheck and this check

    if (!( essentiallyEqual(reference , newimplementation , 1000.f*std::numeric_limits<float>::epsilon()) )) {
      RC_ASSERT( refval == safval) ;
    }

and get a failure every ~ 500 tests (i.e. at that rate the numerical noise i generate by changing the network implementation exceeds more than 1000 times floating point precision (NB: the return value is actually a double …))

but this only if i disable the fast tanh evaluation (enabled by default). If i enable the fast tanh evaluation it appears response values change by more than that.
[background: the TMVA::Reader already uses the fast evaluation, the codegen wasn't. so this should actually improve actually the consistency between TMVA::Reader usage and the codegen - i was only testing old codegen vs new codegen]

will discuss at the next tmva meeting, with a bit more insight.

for the build failure, sorry that was a pretty stupid merging mistake i should've seen (… if i had compiled it locally before pushing … instead of postponing that to running the numeric test).

vgvassilev · 2017-07-13T21:37:27Z

ping.

pseyfert · 2017-08-16T02:56:31Z

found the issue. ran numeric test here https://github.com/pseyfert/tmva-codegen-rapidcheck

vgvassilev · 2017-08-16T08:06:05Z

Great! Could you add a gtest incorporating the tmva-codegen-rapidcheck? You can see how it's done here.

vgvassilev · 2017-08-16T08:09:50Z

@phsft-bot build!

phsft-bot · 2017-08-16T08:10:18Z

Starting build on centos7/gcc49, mac1012/native, slc6/gcc49, slc6/gcc62, ubuntu14/native with flags -Dvc=OFF -Dimt=ON -Dccache=ON
How to customize builds

phsft-bot · 2017-08-16T08:58:57Z

Build failed on slc6/gcc49.
See console output.

Errors:

ERROR: Timeout after 10 minutes
ERROR: Error fetching remote repo 'origin'
stderr: error: RPC failed; result=18, HTTP code = 200
ERROR: Error fetching remote repo 'origin'

ashlaban · 2017-09-08T14:45:12Z

@phsft-bot build!

phsft-bot · 2018-04-30T09:47:44Z

Build failed on slc6/gcc62.
See console output.

Errors:

ERROR: Timeout after 30 minutes
ERROR: Error fetching remote repo 'origin'
stderr: error: RPC failed; result=18, HTTP code = 200
ERROR: Error fetching remote repo 'origin'

phsft-bot · 2018-04-30T11:12:08Z

Build failed on slc6/gcc62.
See console output.

Errors:

ERROR: Timeout after 30 minutes
ERROR: Error fetching remote repo 'origin'
stderr: error: RPC failed; result=18, HTTP code = 200
ERROR: Error fetching remote repo 'origin'

etejedor · 2018-07-26T10:59:23Z

@ashlaban @lmoneta what is the status of this PR? Should we try again to build and merge?

etejedor · 2018-07-26T10:59:32Z

@phsft-bot build

ashlaban · 2018-07-26T11:02:15Z

LGTM. I think Lorenzo will be back from vacation on the 28th.

phsft-bot · 2018-07-26T11:02:54Z

Starting build on slc6/gcc48, slc6-i686/gcc49, centos7/clang39, centos7/gcc62, centos7/gcc7, fedora28/native, ubuntu16/native, mac1013/native, windows10/vc15 with flags -Dimt=ON -Dccache=ON
How to customize builds

etejedor · 2018-07-26T11:21:37Z

Great @ashlaban , can you ping @lmoneta when he is back about this? This is a senior PR, opened in 2015! :)

ashlaban · 2018-07-31T14:16:18Z

@phsft-bot build!

phsft-bot · 2018-07-31T14:16:28Z

Starting build on slc6/gcc48, slc6-i686/gcc49, centos7/clang39, centos7/gcc62, centos7/gcc7, fedora28/native, ubuntu16/native, mac1013/native, windows10/vc15 with flags -Dimt=ON -Dccache=ON
How to customize builds

phsft-bot · 2018-07-31T14:18:39Z

Build failed on ubuntu16/native.
See console output.

Errors:

Error: could not find CMAKE_PROJECT_NAME in Cache

phsft-bot · 2018-07-31T14:56:00Z

Build failed on fedora28/native.
See console output.

Errors:

Error: could not find CMAKE_PROJECT_NAME in Cache

phsft-bot · 2018-07-31T15:08:26Z

Build failed on windows10/vc15.
See console output.

phsft-bot · 2018-07-31T15:09:39Z

Build failed on mac1013/native.
See console output.

Errors:

Error: could not find CMAKE_PROJECT_NAME in Cache

phsft-bot · 2018-07-31T15:28:18Z

Build failed on slc6-i686/gcc49.
See console output.

Errors:

Error: could not find CMAKE_PROJECT_NAME in Cache

Axel-Naumann · 2018-08-01T09:07:23Z

@phsft-bot build!

phsft-bot · 2018-08-01T09:07:48Z

Starting build on slc6/gcc48, slc6-i686/gcc49, centos7/clang39, centos7/gcc62, centos7/gcc7, fedora28/native, ubuntu16/native, mac1013/native, windows10/vc15 with flags -Dimt=ON -Dccache=ON
How to customize builds

phsft-bot · 2018-08-01T09:11:09Z

Build failed on ubuntu16/native.
See console output.

Axel-Naumann · 2018-08-01T09:19:36Z

@phsft-bot build just on ubuntu16/native

phsft-bot · 2018-08-01T09:19:43Z

Starting build on ubuntu16/native with flags -Dimt=ON -Dccache=ON
How to customize builds

phsft-bot · 2018-08-01T09:20:58Z

Build failed on ubuntu16/native.
See console output.

Axel-Naumann · 2018-08-01T09:21:50Z

@phsft-bot build just on ubuntu16/native

phsft-bot · 2018-08-01T09:22:10Z

Starting build on ubuntu16/native with flags -Dimt=ON -Dccache=ON
How to customize builds

phsft-bot · 2018-08-01T09:23:29Z

Build failed on ubuntu16/native.
See console output.

Axel-Naumann · 2018-08-01T09:30:37Z

@phsft-bot build just on ubuntu16/native

phsft-bot · 2018-08-01T09:30:52Z

Starting build on ubuntu16/native with flags -Dimt=ON -Dccache=ON
How to customize builds

phsft-bot · 2018-08-01T09:31:59Z

Build failed on ubuntu16/native.
See console output.

Axel-Naumann · 2018-08-01T11:23:41Z

@phsft-bot build just on ubuntu16/native

phsft-bot · 2018-08-01T11:23:48Z

Starting build on ubuntu16/native with flags -Dimt=ON -Dccache=ON
How to customize builds

Axel-Naumann assigned lmoneta Jan 7, 2016

pseyfert force-pushed the tmva-mlp-codegen branch from 863ed7c to b5cf0ea Compare October 31, 2016 14:21

peremato assigned lmoneta and unassigned lmoneta Mar 1, 2017

pseyfert force-pushed the tmva-mlp-codegen branch from 9ccce9e to ef160e4 Compare May 26, 2017 18:54

pseyfert force-pushed the tmva-mlp-codegen branch from ef160e4 to f2f2ceb Compare August 16, 2017 01:05

lmoneta merged commit b7a3854 into root-project:master Aug 2, 2018

ethereal-space-cadet16 mentioned this pull request May 31, 2022

Accessing pyROOT #10676

Closed

speedup for .class.C for MLP #121

speedup for .class.C for MLP #121

Conversation

pseyfert commented Dec 21, 2015

pseyfert commented Oct 31, 2016

pseyfert commented Mar 9, 2017

phsft-bot commented Mar 9, 2017

pseyfert commented May 13, 2017

phsft-bot commented May 13, 2017

vgvassilev commented May 13, 2017

pseyfert commented May 13, 2017

pseyfert commented May 26, 2017

vgvassilev commented Jun 2, 2017

phsft-bot commented Jun 2, 2017

phsft-bot commented Jun 2, 2017

Errors:

phsft-bot commented Jun 2, 2017

Errors:

phsft-bot commented Jun 2, 2017

Errors:

phsft-bot commented Jun 2, 2017

Errors:

phsft-bot commented Jun 2, 2017

Errors:

vgvassilev commented Jun 8, 2017

pseyfert commented Jul 4, 2017

vgvassilev commented Jul 13, 2017

pseyfert commented Aug 16, 2017

vgvassilev commented Aug 16, 2017

vgvassilev commented Aug 16, 2017

phsft-bot commented Aug 16, 2017

phsft-bot commented Aug 16, 2017

Errors:

ashlaban commented Sep 8, 2017

phsft-bot commented Apr 30, 2018

Errors:

phsft-bot commented Apr 30, 2018

Errors:

etejedor commented Jul 26, 2018

etejedor commented Jul 26, 2018

ashlaban commented Jul 26, 2018

phsft-bot commented Jul 26, 2018

etejedor commented Jul 26, 2018

ashlaban commented Jul 31, 2018

phsft-bot commented Jul 31, 2018

phsft-bot commented Jul 31, 2018

Errors:

phsft-bot commented Jul 31, 2018

Errors:

phsft-bot commented Jul 31, 2018

phsft-bot commented Jul 31, 2018

Errors:

phsft-bot commented Jul 31, 2018

Errors:

Axel-Naumann commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

Axel-Naumann commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

Axel-Naumann commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

Axel-Naumann commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

phsft-bot commented Aug 1, 2018

Axel-Naumann commented Aug 1, 2018

phsft-bot commented Aug 1, 2018