Merge pull request #343 from FluxML/gatedgraphconv

Fix GatedGraphConv

Author: yuehhua

src/layers/graph_conv.jl

@@ -556,17 +556,23 @@ function (l::GatedGraphConv)(el::NamedTuple, H::AbstractArray{T}) where {T<:Real
         H = vcat(H, Hpad)
     end
     for i = 1:l.num_layers
-        M = _matmul(view(l.weight, :, :, i), H)
+        M = _matmul(l.weight[:, :, i], H)
         _, M = propagate(l, el, nothing, M, nothing, l.aggr, nothing, nothing)
-        H, _ = l.gru(H, M)
+        H = apply_gru(l.gru, H, M)
     end
     return H
 end
 
+function apply_gru(gru, H::AbstractArray, M::AbstractArray)
+    H′ = apply_gru(gru, reshape(H, size(H, 1), :), reshape(M, size(M, 1), :))
+    return reshape(H′, size(H′, 1), size(H)[2:end]...)
+end
+
+apply_gru(gru, H::AbstractMatrix, M::AbstractMatrix) = gru(H, M)[1]
+
 function Base.show(io::IO, l::GatedGraphConv)
     print(io, "GatedGraphConv(($(l.out_ch) => $(l.out_ch))^$(l.num_layers)")
-    print(io, ", aggr=", l.aggr)
-    print(io, ")")
+    print(io, ", aggr=", l.aggr, ")")
 end
 
 

test/cuda/graph_conv.jl

@@ -3,7 +3,7 @@
     in_channel = 3
     out_channel = 5
     batch_size = 10
-    
+
     N = 4
     adj = T[0 1 0 1;
            1 0 1 0;
@@ -48,7 +48,7 @@
             @test size(cc.weight) == (out_channel, in_channel, k)
             @test size(cc.bias) == (out_channel,)
             @test cc.k == k
-            
+
             fg = FeaturedGraph(adj, nf=X) |> gpu
             fg_ = cc(fg)
             @test size(node_feature(fg_)) == (out_channel, N)
@@ -106,12 +106,12 @@
             @test size(gat.weight) == (out_channel * heads, in_channel)
             @test size(gat.bias) == (out_channel * heads,)
             @test size(gat.a) == (2*out_channel, heads)
-            
+
             X = rand(T, in_channel, N)
             fg = FeaturedGraph(adj, nf=X) |> gpu
             fg_ = gat(fg)
             @test size(node_feature(fg_)) == (out_channel * heads, N)
-    
+
             g = gradient(() -> sum(node_feature(gat(fg))), Flux.params(gat))
             @test length(g.grads) == 5
         end
@@ -121,7 +121,7 @@
             gat = WithGraph(fg, GATConv(in_channel=>out_channel, heads=2)) |> gpu
             Y = gat(X |> gpu)
             @test size(Y) == (out_channel * heads, N, batch_size)
-    
+
             g = gradient(() -> sum(gat(X |> gpu)), Flux.params(gat))
             @test length(g.grads) == 4
         end
@@ -145,11 +145,11 @@
         @testset "layer with static graph" begin
             X = rand(T, in_channel, N, batch_size)
             ggc = WithGraph(fg, GatedGraphConv(out_channel, num_layers)) |> gpu
-            @test_broken Y = ggc(X |> gpu)
-            @test_broken size(Y) == (out_channel, N, batch_size)
+            Y = ggc(X |> gpu)
+            @test size(Y) == (out_channel, N, batch_size)
 
-            @test_broken g = gradient(() -> sum(ggc(X |> gpu)), Flux.params(ggc))
-            @test_broken length(g.grads) == 6
+            g = gradient(() -> sum(ggc(X |> gpu)), Flux.params(ggc))
+            @test length(g.grads) == 7
         end
     end
 

test/layers/graph_conv.jl

@@ -265,11 +265,11 @@
         @testset "layer with static graph" begin
             X = rand(T, in_channel, N, batch_size)
             ggc = WithGraph(fg, GatedGraphConv(out_channel, num_layers))
-            @test_broken Y = ggc(X)
-            @test_broken size(Y) == (out_channel, N, batch_size)
+            Y = ggc(X)
+            @test size(Y) == (out_channel, N, batch_size)
 
-            @test_broken g = gradient(() -> sum(ggc(X)), Flux.params(ggc))
-            @test_broken length(g.grads) == 6
+            g = gradient(() -> sum(ggc(X)), Flux.params(ggc))
+            @test length(g.grads) == 6
         end
     end