Edge attributes for GatedGraphConv

[rubyzhou2014]

Hi,
I am building a model using GatedGraphConv operator. I want to involve edge attributes in the model training but didn't find a way. I see that it can pass edge weight into GatedGraphConv, but not edge_attr.

Is there any API provided to support this feature? Or any suggestions for coding it manually? Thanks!

[rusty1s]

You might want to extend the GatedGraphConv manually for this. Usually, this is done by concatenating edge features and source node features in message:

def message(self, x_j, edge_attr):
        return self.lin(torch.cat([x_j, edge_attr], dim=-1))

where self.lin is a Lienar layer that maps to out_channels.