gpu fix

[vitskvara]

Trying to solve #11.

[vitskvara]

This only fixes the mapping conversion, however the σ field is still not converted.

[vitskvara]

Also, this is probably going to break if one has CuArrays loaded, the model is on cpu and functor is not called in the gpu conversion process - what are the places where functor is called?

[codecov-io]

Codecov Report

Merging #12 into master will decrease coverage by 0.87%.
The diff coverage is 50%.

@@            Coverage Diff             @@
##           master      #12      +/-   ##
==========================================
- Coverage    92.3%   91.42%   -0.88%     
==========================================
  Files           8        8              
  Lines         104      105       +1     
==========================================
  Hits           96       96              
- Misses          8        9       +1

Impacted Files	Coverage Δ
src/cmeanvar_gaussian.jl	100% <100%> (ø)	⬆️
src/cmean_gaussian.jl	83.33% <33.33%> (-3.63%)	⬇️

---

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 4a674b1...749600b. Read the comment docs.

[nmheim]

functor is used by e.g. trainable and gpu. it recursively tries to adapt all Arrays to CuArrays. As far as I can see it constructs the type signature (with the cpu arrays in the mapping) before moving the mapping to the gpu... I am not sure if this is easily solvable with the type signature we have right now...

[nmheim]

I cannot test on GPU atm, but does this work?

function Flux.functor(p::CMeanGaussian{V}) where V
    fs = fieldnames(typeof(p))
    nt = (; (name=>getfield(p, name) for name in fs)...)
    nt, y -> CMeanGaussian{V}(y...)
end

[nmheim]

putting a gpu call in functor is not a good idea, because gpu itsself calls functor (through fmap and fmap1). have you tried if the snippet I posted above works on a GPU?

[vitskvara]

It helps with CMeanVarGaussian, but not with CMeanGaussian.

[vitskvara]

Now it fails with

julia> p  = CMeanGaussian{DiagVar}(mapping, var) |> gpu
ERROR: MethodError: no method matching CMeanGaussian{DiagVar,M,S} where S<:AbstractArray where M(::Dense{typeof(identity),CuArray{Float32,2,Nothing},CuArray{Float32,1,Nothing}}, ::CuArray{Float32,1,Nothing}, ::Int64, ::Dict{Symbol,Bool})
Closest candidates are:
  CMeanGaussian{DiagVar,M,S} where S<:AbstractArray where M(::M, ::Any, ::Int64) where {V, M} at /home/vit/.julia/dev/ConditionalDists/src/cmean_gaussian.jl:40
  CMeanGaussian{DiagVar,M,S} where S<:AbstractArray where M(::Any, ::Any) at /home/vit/.julia/dev/ConditionalDists/src/cmean_gaussian.jl:46
Stacktrace:
 [1] #14 at /home/vit/.julia/dev/ConditionalDists/src/cmean_gaussian.jl:68 [inlined]
 [2] fmap1(::Function, ::CMeanGaussian{DiagVar,Dense{typeof(identity),Array{Float32,2},Array{Float32,1}},Array{Float32,1}}) at /home/vit/.julia/packages/Flux/oObnA/src/functor.jl:32
 [3] #fmap#41(::IdDict{Any,Any}, ::typeof(fmap), ::typeof(cu), ::CMeanGaussian{DiagVar,Dense{typeof(identity),Array{Float32,2},Array{Float32,1}},Array{Float32,1}}) at /home/vit/.julia/packages/Flux/oObnA/src/functor.jl:37
 [4] fmap at /home/vit/.julia/packages/Flux/oObnA/src/functor.jl:36 [inlined]

[vitskvara]

I am really confused with this. When debugging, it seems that the overloaded functor is called twice before some default one is called which tries to construct the Dense layer but with the original CMeanGaussian arguments, which of course does not make any sense.

[nmheim]

gpu tried to call CMeanGaussian{V}(m,s,xlength,d) at some point, but only CMeanGaussian{V,M,S}(m,s,xlength,d) (with the types M and S) was defined.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

Attention: Patch coverage is 50.00000% with 2 lines in your changes missing coverage. Please review.

Project coverage is 91.42%. Comparing base (4a674b1) to head (749600b).
Report is 46 commits behind head on master.

Files with missing lines	Patch %	Lines
src/cmean_gaussian.jl	33.33%	2 Missing ⚠️

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #12      +/-   ##
==========================================
- Coverage   92.30%   91.42%   -0.88%     
==========================================
  Files           8        8              
  Lines         104      105       +1     
==========================================
  Hits           96       96              
- Misses          8        9       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.