Method error when training GCNConv with FeaturedGraph input

[tdfalc1]

I made the following modifications to the GCNConv layer example to incorporate FeaturedGraph input:

train_X_fg = GeometricFlux.FeaturedGraph(adj_mat, train_X)

model = Chain(GCNConv(num_features => hidden, relu),
              Dropout(0.5),
              GCNConv(hidden => target_catg),
              x -> softmax(x.nf[])) |> gpu

train_data = [(train_X_fg, train_y)]
@epochs epochs Flux.train!(loss, ps, train_data, opt, cb=throttle(evalcb, 10))

I get the following error:

ERROR: MethodError: no method matching (::GeometricFlux.var"#adjacency_matrix#76")(::FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}})
Closest candidates are:
  adjacency_matrix(::AbstractArray{T,2} where T) at /Users/thomasfalconer/.julia/packages/GeometricFlux/k4atN/src/operations/linalg.jl:6
  adjacency_matrix(::AbstractArray{T,2} where T, ::DataType) at /Users/thomasfalconer/.julia/packages/GeometricFlux/k4atN/src/operations/linalg.jl:6
Stacktrace:
 [1] _pullback at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/lib/grad.jl:8 [inlined]
 [2] GCNConv at /Users/thomasfalconer/.julia/packages/GeometricFlux/k4atN/src/layers/conv.jl:55 [inlined]
 [3] _pullback(::Zygote.Context, ::GCNConv{Float32,typeof(relu),FeaturedGraph{Array{Float64,2},Array{Float64,2},Array{Float64,2},Array{Float64,1}}}, ::FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}}) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface2.jl:0
 [4] applychain at /Users/thomasfalconer/.julia/packages/Flux/IjMZL/src/layers/basic.jl:36 [inlined]
 [5] _pullback(::Zygote.Context, ::typeof(Flux.applychain), ::Tuple{GCNConv{Float32,typeof(relu),FeaturedGraph{Array{Float64,2},Array{Float64,2},Array{Float64,2},Array{Float64,1}}},Dropout{Float64,Colon},GCNConv{Float32,typeof(identity),FeaturedGraph{Array{Float64,2},Array{Float64,2},Array{Float64,2},Array{Float64,1}}},var"#11#12"}, ::FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}}) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface2.jl:0
 [6] Chain at /Users/thomasfalconer/.julia/packages/Flux/IjMZL/src/layers/basic.jl:38 [inlined]
 [7] _pullback(::Zygote.Context, ::Chain{Tuple{GCNConv{Float32,typeof(relu),FeaturedGraph{Array{Float64,2},Array{Float64,2},Array{Float64,2},Array{Float64,1}}},Dropout{Float64,Colon},GCNConv{Float32,typeof(identity),FeaturedGraph{Array{Float64,2},Array{Float64,2},Array{Float64,2},Array{Float64,1}}},var"#11#12"}}, ::FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}}) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface2.jl:0
 [8] loss at ./REPL[42]:2 [inlined]
 [9] _pullback(::Zygote.Context, ::typeof(loss), ::FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}}, ::SparseMatrixCSC{Float32,Int64}) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface2.jl:0
 [10] adjoint at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/lib/lib.jl:175 [inlined]
 [11] _pullback at /Users/thomasfalconer/.julia/packages/ZygoteRules/6nssF/src/adjoint.jl:47 [inlined]
 [12] #15 at /Users/thomasfalconer/.julia/packages/Flux/IjMZL/src/optimise/train.jl:83 [inlined]
 [13] _pullback(::Zygote.Context, ::Flux.Optimise.var"#15#21"{typeof(loss),Tuple{FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}},SparseMatrixCSC{Float32,Int64}}}) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface2.jl:0
 [14] pullback(::Function, ::Zygote.Params) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface.jl:172
 [15] gradient(::Function, ::Zygote.Params) at /Users/thomasfalconer/.julia/packages/Zygote/iFibI/src/compiler/interface.jl:53
 [16] macro expansion at /Users/thomasfalconer/.julia/packages/Flux/IjMZL/src/optimise/train.jl:82 [inlined]
 [17] macro expansion at /Users/thomasfalconer/.julia/packages/Juno/tLMZd/src/progress.jl:134 [inlined]
 [18] train!(::Function, ::Zygote.Params, ::Array{Tuple{FeaturedGraph{Array{Float32,2},SparseMatrixCSC{Float32,Int64},Array{Float64,2},Array{Float64,1}},SparseMatrixCSC{Float32,Int64}},1}, ::ADAM; cb::Flux.var"#throttled#26"{Flux.var"#throttled#22#27"{Bool,Bool,typeof(evalcb),Int64}}) at /Users/thomasfalconer/.julia/packages/Flux/IjMZL/src/optimise/train.jl:80
 [19] top-level scope at /Users/thomasfalconer/.julia/packages/Flux/IjMZL/src/optimise/train.jl:115
 [20] top-level scope at /Users/thomasfalconer/.julia/packages/Juno/tLMZd/src/progress.jl:134

Having trouble investigating the cause of this error @yuehhua

[yuehhua]

We need some utility functions for this.

[tdfalc1]

@yuehhua Is this a problem with the differentiability of the layer? Im currently using the FeaturedGraph type in order to have two channels as input (node features and edge features), is this currently the only way to achieve this?

[yuehhua]

It is not related with differentiability.

Currently, FeaturedGraph approach and regular array approach are not compatible with each other. So, we should have some way to convert FeaturedGraph to array. BTW, I don't expect the reverse conversion of converting array to FeaturedGraph is reasonable. Once the FeaturedGraph flow is converted to array flow, it will go until the end of model.

[tdfalc1]

Makes sense, thanks! Would you be able to recommend any interim solutions for this?

[yuehhua]

I don't think there are any interim solutions existing but I will implement these functionalities recently.

[rezabayani]

@yuehhua @tdfalc1 Does this mean training FeaturedGraph is not possible with this package? I seem to encounter the same error.

[rezabayani]

@yuehhua @tejasvaidhyadev

I don't think there are any interim solutions existing but I will implement these functionalities recently.

Does this mean training FeaturedGraph is not possible with this package? Or are there "manual" fixes? I seem to encounter the same error.

[yuehhua]

OK, I will get manual update recently. It is possible to train data within FeaturedGraph with Geometriclux model.

[yuehhua]

@rezabayani and other users,
I have updated the documentation about the cooperation with Flux layers.

[yuehhua]

I think the documentation explain the mechanism well. If there are any insufficient, just let me know or file an issue. I just close this issue.