Fix Pkg.test failure for registered packages (#700)

GNNLux/test/Project.toml

@@ -22,9 +22,7 @@ TestItems = "1c621080-faea-4a02-84b6-bbd5e436b8fe"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [sources]
-GNNGraphs = {path = "../../GNNGraphs"}
 GNNLux = {path = ".."}
-GNNlib = {path = "../../GNNlib"}
 
 [compat]
 NNlib = "0.9.38"

GNNlib/src/layers/temporalconv.jl

@@ -8,15 +8,13 @@ and the Lux (`GNNLux`) frontends build such a carrier and call into these
 functions, so the recurrence math lives in a single place.
 =#
 
-"""
-    tgcn(l, cz, cr, ch, h)
-
-GRU gating of the T-GCN cell. `cz`, `cr`, `ch` are the spatial-convolution
-outputs of the three gates (computed by the frontend), `h` is the current
-hidden state, and `l` carries the three gate `Dense` sub-modules
-(`l.dense_z`, `l.dense_r`, `l.dense_h`, each exposing `weight`, `bias`, `σ`).
-Returns the updated hidden state.
-"""
+# tgcn(l, cz, cr, ch, h)
+#
+# GRU gating of the T-GCN cell. `cz`, `cr`, `ch` are the spatial-convolution
+# outputs of the three gates (computed by the frontend), `h` is the current
+# hidden state, and `l` carries the three gate `Dense` sub-modules
+# (`l.dense_z`, `l.dense_r`, `l.dense_h`, each exposing `weight`, `bias`, `σ`).
+# Returns the updated hidden state.
 function tgcn(l, cz, cr, ch, h)
     z = l.dense_z.σ.(l.dense_z.weight * vcat(cz, h) .+ l.dense_z.bias)
     r = l.dense_r.σ.(l.dense_r.weight * vcat(cr, h) .+ l.dense_r.bias)
@@ -25,13 +23,11 @@ function tgcn(l, cz, cr, ch, h)
     return h
 end
 
-"""
-    gconv_gru(l, g, x, h)
-
-Forward pass of the GConvGRU cell. `l` carries the six `ChebConv` sub-modules
-(`conv_x_r`, `conv_h_r`, `conv_x_z`, `conv_h_z`, `conv_x_h`, `conv_h_h`).
-Returns the updated hidden state.
-"""
+# gconv_gru(l, g, x, h)
+#
+# Forward pass of the GConvGRU cell. `l` carries the six `ChebConv` sub-modules
+# (`conv_x_r`, `conv_h_r`, `conv_x_z`, `conv_h_z`, `conv_x_h`, `conv_h_h`).
+# Returns the updated hidden state.
 function gconv_gru(l, g::GNNGraph, x, h)
     r = NNlib.sigmoid_fast.(cheb_conv(l.conv_x_r, g, x) .+ cheb_conv(l.conv_h_r, g, h))
     z = NNlib.sigmoid_fast.(cheb_conv(l.conv_x_z, g, x) .+ cheb_conv(l.conv_h_z, g, h))
@@ -40,13 +36,11 @@ function gconv_gru(l, g::GNNGraph, x, h)
     return h
 end
 
-"""
-    gconv_lstm(l, g, x, h, c)
-
-Forward pass of the GConvLSTM cell. `l` carries the eight `ChebConv` sub-modules
-and the four peephole scalings/biases (`w_i`, `b_i`, `w_f`, `b_f`, `w_c`, `b_c`,
-`w_o`, `b_o`). Returns the updated `(h, c)` state.
-"""
+# gconv_lstm(l, g, x, h, c)
+#
+# Forward pass of the GConvLSTM cell. `l` carries the eight `ChebConv` sub-modules
+# and the four peephole scalings/biases (`w_i`, `b_i`, `w_f`, `b_f`, `w_c`, `b_c`,
+# `w_o`, `b_o`). Returns the updated `(h, c)` state.
 function gconv_lstm(l, g::GNNGraph, x, h, c)
     # input gate
     i = cheb_conv(l.conv_x_i, g, x) .+ cheb_conv(l.conv_h_i, g, h) .+ l.w_i .* c .+ l.b_i
@@ -63,12 +57,10 @@ function gconv_lstm(l, g::GNNGraph, x, h, c)
     return h, c
 end
 
-"""
-    dcgru(l, g, x, h)
-
-Forward pass of the DCGRU cell. `l` carries the three `DConv` sub-modules
-(`dconv_u`, `dconv_r`, `dconv_c`). Returns the updated hidden state.
-"""
+# dcgru(l, g, x, h)
+#
+# Forward pass of the DCGRU cell. `l` carries the three `DConv` sub-modules
+# (`dconv_u`, `dconv_r`, `dconv_c`). Returns the updated hidden state.
 function dcgru(l, g::GNNGraph, x, h)
     h̃ = vcat(x, h)
     z = NNlib.sigmoid_fast.(d_conv(l.dconv_u, g, h̃))

GNNlib/test/Project.toml

@@ -19,7 +19,6 @@ TestItemRunner = "f8b46487-2199-4994-9208-9a1283c18c0a"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [sources]
-GNNGraphs = {path = "../../GNNGraphs"}
 GNNlib = {path = ".."}
 
 [compat]

GraphNeuralNetworks/test/Project.toml

@@ -18,6 +18,4 @@ TestItemRunner = "f8b46487-2199-4994-9208-9a1283c18c0a"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [sources]
-GNNGraphs = {path = "../../GNNGraphs"}
-GNNlib = {path = "../../GNNlib"}
 GraphNeuralNetworks = {path = ".."}

GraphNeuralNetworks/test/layers/temporalconv.jl

@@ -25,17 +25,24 @@ end
 
 @testitem "TGCNCell" setup=[TemporalConvTestModule, TestModule] begin
     using .TemporalConvTestModule, .TestModule
-
+
+    # Mooncake is skipped here: it returns wrong gradients through the cell's
+    # sigmoid gates (`Dense(_, sigmoid)`) on a fraction of inputs, so the check is
+    # flaky (Zygote and finite differences agree). Upstream Mooncake bug with
+    # NNlib.sigmoid: https://github.com/chalk-lab/Mooncake.jl/issues/1257.
+    # Reference against Zygote only, matching the other temporal cells.
+    ad_backends = [Flux.AutoZygote()]
+
     # Test with default activation function
     cell = GraphNeuralNetworks.TGCNCell(in_channel => out_channel)
     y, h = cell(g, g.x)
     @test y === h
     @test size(h) == (out_channel, g.num_nodes)
     # with no initial state
-    test_gradients(cell, g, g.x, loss=cell_loss, rtol=RTOL_HIGH)
+    test_gradients(cell, g, g.x, loss=cell_loss, rtol=RTOL_HIGH; ad_backends)
     # with initial state
-    test_gradients(cell, g, g.x, h, loss=cell_loss, rtol=RTOL_HIGH)
-    
+    test_gradients(cell, g, g.x, h, loss=cell_loss, rtol=RTOL_HIGH; ad_backends)
+
     # Test with custom activation function
     custom_activation = tanh
     cell_custom = GraphNeuralNetworks.TGCNCell(in_channel => out_channel, act = custom_activation)
@@ -45,14 +52,19 @@ end
     # Test that outputs differ when using different activation functions
     @test !isapprox(y, y_custom, rtol=RTOL_HIGH)
     # with no initial state
-    test_gradients(cell_custom, g, g.x, loss=cell_loss, rtol=RTOL_HIGH)
+    test_gradients(cell_custom, g, g.x, loss=cell_loss, rtol=RTOL_HIGH; ad_backends)
     # with initial state
-    test_gradients(cell_custom, g, g.x, h_custom, loss=cell_loss, rtol=RTOL_HIGH)
+    test_gradients(cell_custom, g, g.x, h_custom, loss=cell_loss, rtol=RTOL_HIGH; ad_backends)
 end
 
 @testitem "TGCN" setup=[TemporalConvTestModule, TestModule] begin
     using .TemporalConvTestModule, .TestModule
-
+
+    # Mooncake is skipped here (see the TGCNCell test item and
+    # https://github.com/chalk-lab/Mooncake.jl/issues/1257): it returns wrong
+    # gradients through the cell's sigmoid gates on some inputs. Zygote only.
+    ad_backends = [Flux.AutoZygote()]
+
     # Test with default activation function
     layer = TGCN(in_channel => out_channel)
     x = rand(Float32, in_channel, timesteps, g.num_nodes)
@@ -61,9 +73,9 @@ end
     @test layer isa GNNRecurrence
     @test size(y) == (out_channel, timesteps, g.num_nodes)
     # with no initial state
-    test_gradients(layer, g, x, rtol = RTOL_HIGH)
+    test_gradients(layer, g, x, rtol = RTOL_HIGH; ad_backends)
     # with initial state
-    test_gradients(layer, g, x, state0, rtol = RTOL_HIGH)
+    test_gradients(layer, g, x, state0, rtol = RTOL_HIGH; ad_backends)
 
     # Test with custom activation function
     custom_activation = tanh
@@ -74,15 +86,15 @@ end
     # Test that outputs differ when using different activation functions
     @test !isapprox(y, y_custom, rtol = RTOL_HIGH)
     # with no initial state
-    test_gradients(layer_custom, g, x, rtol = RTOL_HIGH)
+    test_gradients(layer_custom, g, x, rtol = RTOL_HIGH; ad_backends)
     # with initial state
-    test_gradients(layer_custom, g, x, state0, rtol = RTOL_HIGH)
+    test_gradients(layer_custom, g, x, state0, rtol = RTOL_HIGH; ad_backends)
 
     # interplay with GNNChain
     model = GNNChain(TGCN(in_channel => out_channel), Dense(out_channel, 1))
     y = model(g, x)
     @test size(y) == (1, timesteps, g.num_nodes)
-    test_gradients(model, g, x, rtol = RTOL_HIGH, atol = ATOL_LOW)
+    test_gradients(model, g, x, rtol = RTOL_HIGH, atol = ATOL_LOW; ad_backends)
 end
 
 @testitem "GConvLSTMCell" setup=[TemporalConvTestModule, TestModule] begin