GNNChain doesn't work with layers that accept edge features

[ivanightingale]

Take the example in the documentation.

If we replace

l = GATv2Conv((in_channel, ein) => out_channel, add_self_loops = false)

with

l = GNNChain(GATv2Conv((in_channel, ein) => out_channel, add_self_loops = false))

then we will get a MethodError in the forward pass call y = l(g, x, e).

Versions

• GraphNeuralNetworks.jl: v1.0.0

[CarloLucibello]

For complex message passing involving passing edge features, the recommended approach is to define a struct containing the layers and then defining the forward pass, see https://juliagraphs.org/GraphNeuralNetworks.jl/docs/GraphNeuralNetworks.jl/stable/guides/models/#Explicit-modeling

This approach is more verbose than GNNChain but allows full flexibility.