GCNConv and SimpleWeightedGraphs

[CarstenHUB]

Having trouble using a SimpleWeightedGraph as input for the GCNConv-layer.
My graph is an undirected simple Int64 graph with Float64 weights and calling adjacency_matrix() on this works fine.
But during training I get the error

InexactError: Int64(-1.0384249352874253)
Stacktrace:
[1] Int64 at .\float.jl:710 [inlined]
[2] _map_zeropres!(::Type{Int64}, ::SparseMatrixCSC{Int64,Int64}, ::SparseMatrixCSC{Float64,Int64}) at D:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.5\SparseArrays\src\higherorderfns.jl:246
[3] _noshapecheck_map(::Type{Int64}, ::SparseMatrixCSC{Float64,Int64}) at D:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.5\SparseArrays\src\higherorderfns.jl:165
[4] copy at D:\buildbot\worker\package_win64\build\usr\share\julia\stdlib\v1.5\SparseArrays\src\higherorderfns.jl:169 [inlined]
[5] materialize(::Base.Broadcast.Broadcasted{SparseArrays.HigherOrderFns.SparseMatStyle,Nothing,Type{Int64},Tuple{SparseMatrixCSC{Float64,Int64}}}) at .\broadcast.jl:837
[6] adjacency_matrix(::SimpleWeightedGraph{Int64,Float64}, ::DataType; dir::Symbol) at C:\Users\rcars.julia\packages\SimpleWeightedGraphs\IDzOp\src\overrides.jl:32
[7] adjacency_matrix at C:\Users\rcars.julia\packages\SimpleWeightedGraphs\IDzOp\src\overrides.jl:31 [inlined]
[8] adjacency_matrix at C:\Users\rcars.julia\packages\GraphSignals\dB9OV\src\linalg.jl:7 [inlined] (repeats 2 times)
[9] _pullback at C:\Users\rcars.julia\packages\Zygote\chgvX\src\lib\grad.jl:8 [inlined]

and I don't see why it tries to change the values to Int64.
Does anyone have an idea?

[yuehhua]

Sorry to be late. Could you update GeometricFlux and try that again?

[CarstenHUB]

Unfortunately, I get the same error even after updating.

[yuehhua]

Should be resolved.

[yuehhua]

If you're still interested, you could try new version. If you have any questions just comment down.

[eahenle]

This error (or something like it) seems to be present in the most recent version (0.13.10).

Given graphs::Vector{Metagraph}:

fgs = [
	FeaturedGraph(
		SimpleGraph(g); 
		nf=[get_prop(g, v, :label) for v in vertices(g)]',
		ef=[get_prop(g, e, :label) for e in edges(g)]'
	)
	for g in graphs
]

GCNConv(1 => 10, relu)(fgs[1]) # ❌ InexactError

Casting the node features to an actual matrix solves the problem:

fgs = [
	FeaturedGraph(
		SimpleGraph(g); 
		nf=Float32.([get_prop(g, v, :label) for v in vertices(g)]'),
		ef=[get_prop(g, e, :label) for e in edges(g)]'
	)
	for g in graphs
]

GCNConv(1 => 10, relu)(fgs[1]) # ✔️