Make cache vector of same type as u0

[baggepinnen]

No description provided.

[ChrisRackauckas]

yup, no reason not to.

[ChrisRackauckas]

Does this fix the autodiff? I assumed it would be something simple.

[baggepinnen]

It does not due to

MethodError: no method matching procf(::ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2}, ::Int64, ::ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2})
Closest candidates are:
  procf(::Any, ::Int64, !Matched::Float64) at /home/fredrikb/.julia/packages/PoissonRandom/Ohb1g/src/PoissonRandom.jl:119
ad_rand(::RandomNumbers.Xorshifts.Xoroshiro128Plus, ::ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2}) at PoissonRandom.jl:48
pois_rand(::RandomNumbers.Xorshifts.Xoroshiro128Plus, ::ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2}) at PoissonRandom.jl:145
_broadcast_getindex_evalf at broadcast.jl:630 [inlined]
_broadcast_getindex at broadcast.jl:603 [inlined]
getindex at broadcast.jl:563 [inlined]
macro expansion at broadcast.jl:909 [inlined]
macro expansion at simdloop.jl:77 [inlined]
copyto! at broadcast.jl:908 [inlined]
copyto! at broadcast.jl:863 [inlined]
materialize! at broadcast.jl:822 [inlined]
#solve#136(::Nothing, ::Float64, ::typeof(solve), ::JumpProblem{DiscreteProblem{Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1},Tuple{Float64,Float64},true,Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1},DiscreteFunction{true,DiffEqBase.var"#161#162",Nothing,Nothing},Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}},Direct,CallbackSet{Tuple{},Tuple{}},Nothing,Tuple{},RegularJump{typeof(rate),typeof(c),Array{Float64,2},Nothing},Nothing}, ::SimpleTauLeaping) at simple_regular_solve.jl:41
(::DiffEqBase.var"#kw##solve")(::NamedTuple{(:dt,),Tuple{Float64}}, ::typeof(solve), ::JumpProblem{DiscreteProblem{Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1},Tuple{Float64,Float64},true,Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1},DiscreteFunction{true,DiffEqBase.var"#161#162",Nothing,Nothing},Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}},Direct,CallbackSet{Tuple{},Tuple{}},Nothing,Tuple{},RegularJump{typeof(rate),typeof(c),Array{Float64,2},Nothing},Nothing}, ::SimpleTauLeaping) at none:0
forward_pass(::Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1}) at jumpdiff.jl:45
vector_mode_gradient(::typeof(forward_pass), ::Array{Float64,1}, ::ForwardDiff.GradientConfig{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2,Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1}}) at apiutils.jl:37
gradient(::Function, ::Array{Float64,1}, ::ForwardDiff.GradientConfig{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2,Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1}}, ::Val{true}) at gradient.jl:17
gradient(::Function, ::Array{Float64,1}, ::ForwardDiff.GradientConfig{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2,Array{ForwardDiff.Dual{ForwardDiff.Tag{typeof(forward_pass),Float64},Float64,2},1}}) at gradient.jl:15
gradient(::Function, ::Array{Float64,1}) at gradient.jl:15
(::var"#21#22")(::Array{Float64,1}) at jumpdiff.jl:55
top-level scope at jumpdiff.jl:58

[ChrisRackauckas]

Oh yes, that would break it... I see. The difference from SDEs to this is that SDEs always use the same way to generate the random numbers (randn()) independently of any state values, and then multiply something to that, so you can treat those as constants. Here the distribution really does need the value of the state, since the rates are dependent on the states and you can't just rescale it by a factor after generating the random numbers. That would make it non-AD differentiable.

[coveralls]

Coverage decreased (-0.09%) to 78.466% when pulling 072134d on baggepinnen:patch-1 into 53b396b on JuliaDiffEq:master.

[codecov]

Codecov Report

Merging #88 into master will decrease coverage by 0.08%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master      #88      +/-   ##
==========================================
- Coverage   78.55%   78.46%   -0.09%     
==========================================
  Files          21       21              
  Lines         746      743       -3     
==========================================
- Hits          586      583       -3     
  Misses        160      160

Impacted Files	Coverage Δ
src/simple_regular_solve.jl	100% <ø> (ø)	⬆️
src/coupling.jl	73.43% <0%> (-1.19%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 53b396b...072134d. Read the comment docs.

[ChrisRackauckas]

Well this is necessary anyways. Thanks!

[ChrisRackauckas]

I'm not sure how you'd do particle sampling of a Poisson RNG? Something to think about though. Make it generate N random outputs?

[baggepinnen]

Yeah it would draw one sample from each poisson distribution represented by the different particles.

[ChrisRackauckas]

That might directly require a dispatch on the rand functions.

[baggepinnen]

Yep, me and Chad Scherrer have been doing some thinking about that here
baggepinnen/MonteCarloMeasurements.jl#22
So far we have not converged to a systematic way of doing this.