GCNConv layer fails when the GNNGraph comes from an adjacency matrix

[gdalle]

MWE:

using GraphNeuralNetworks, Flux, Zygote
using LinearAlgebra, Random, SparseArrays, Statistics

rng = Random.seed!(63)

x = randn(rng, Float32, 5, 10)
y = randn(10)

A = sparse(Symmetric(sprand(rng, Bool, 10, 10, 0.2)))

g1 = rand_graph(10, 20; ndata=(; x, y))  # random graph
g2 = GNNGraph(A; ndata=(; x, y))  # from sparse adjacency matrix

model1 = GNNChain(Dense(5 => 1))  # without GCNConv
model2 = GNNChain(GCNConv(5 => 5), Dense(5 => 1))  # with GCNConv

myloss(model, g::GNNGraph) = mean(abs2, vec(model(g, g.x)) .- g.y)

model1(g1, g1.x), model2(g1, g1.x), model2(g2, g2.x), model2(g2, g2.x);  # all works

gradient(model -> myloss(model, g1), model1)  # works
gradient(model -> myloss(model, g1), model2)  # works
gradient(model -> myloss(model, g2), model1)  # works
gradient(model -> myloss(model, g2), model2)  # fails

Error:

julia> gradient(model -> myloss(model, g2), model2)  # fails
ERROR: MethodError: no method matching zero(::Nothing)

Closest candidates are:
  zero(::Type{Union{}}, Any...)
   @ Base number.jl:310
  zero(::Type{Dates.Time})
   @ Dates ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/Dates/src/types.jl:440
  zero(::Type{Pkg.Resolve.FieldValue})
   @ Pkg ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/Pkg/src/Resolve/fieldvalues.jl:38
  ...

Stacktrace:
  [1] iszero(x::Nothing)
    @ Base ./number.jl:42
  [2] _iszero(x::Nothing)
    @ SparseArrays ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/SparseArrays/src/SparseArrays.jl:41
  [3] _noshapecheck_map(::typeof(Zygote.wrap_chainrules_output), ::SparseMatrixCSC{ChainRulesCore.NoTangent, Int64})
    @ SparseArrays.HigherOrderFns ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/SparseArrays/src/higherorderfns.jl:181
  [4] map
    @ ~/.julia/juliaup/julia-1.10.4+0.x64.linux.gnu/share/julia/stdlib/v1.10/SparseArrays/src/higherorderfns.jl:1187 [inlined]
  [5] wrap_chainrules_output
    @ ~/.julia/packages/Zygote/nsBv0/src/compiler/chainrules.jl:127 [inlined]
  [6] wrap_chainrules_output
    @ ~/.julia/packages/Zygote/nsBv0/src/compiler/chainrules.jl:110 [inlined]
  [7] map
    @ ./tuple.jl:293 [inlined]
  [8] wrap_chainrules_output
    @ ~/.julia/packages/Zygote/nsBv0/src/compiler/chainrules.jl:111 [inlined]
  [9] ZBack
    @ ~/.julia/packages/Zygote/nsBv0/src/compiler/chainrules.jl:211 [inlined]
 [10] propagate
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/msgpass.jl:251 [inlined]
 [11] #propagate#128
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/msgpass.jl:77 [inlined]
 [12] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [13] propagate
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/msgpass.jl:76 [inlined]
 [14] (::Zygote.Pullback{Tuple{…}, Any})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [15] #_#24
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/layers/conv.jl:160 [inlined]
 [16] (::Zygote.Pullback{Tuple{…}, Any})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [17] GCNConv (repeats 2 times)
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/layers/conv.jl:103 [inlined]
 [18] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [19] _applylayer
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/layers/basic.jl:158 [inlined]
 [20] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [21] _applychain
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/layers/basic.jl:144 [inlined]
 [22] (::Zygote.Pullback{Tuple{…}, Any})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [23] GNNChain
    @ ~/.julia/packages/GraphNeuralNetworks/AxIf2/src/layers/basic.jl:130 [inlined]
 [24] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Matrix{Float32})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [25] myloss
    @ ~/Work/GitHub/Julia/Oversmoothing.jl/experiments/gnn.jl:43 [inlined]
 [26] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Float64)
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [27] #25
    @ ~/Work/GitHub/Julia/Oversmoothing.jl/experiments/gnn.jl:50 [inlined]
 [28] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Float64)
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
 [29] (::Zygote.var"#75#76"{Zygote.Pullback{Tuple{…}, Tuple{…}}})(Δ::Float64)
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface.jl:91
 [30] gradient(f::Function, args::GNNChain{Tuple{GCNConv{…}, Dense{…}}})
    @ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface.jl:148
 [31] top-level scope
    @ ~/Work/GitHub/Julia/Oversmoothing.jl/experiments/gnn.jl:50
Some type information was truncated. Use `show(err)` to see complete types.

[gdalle]

My bad, I had not specified graph_type=:sparse