Edge attributes in multi-layered GNNs

[antoni4040]

Hello. I'm doing this for the layers:

self.convs = torch.nn.ModuleList([
            TransformerConv(output_features, output_features, 4, False, edge_dim=EDGE_ATTRS_NUMBER)
            for _ in range(num_layers)
        ])

And I do this in the forward pass:

 for i in self.convs:
            x = i(x.relu(), edge_index, edge_attrs)
            xs.append(x.clone())

Now, my problem is that the node data change every time, but the edge_attrs remains the same. Is this the intended way to do? Or maybe should I pass the edge_attrs through a linear layer manually? Thanks

[rusty1s]

I think this is the intended way. Each layer will still learn to transform edge features on its own by passing it through a Linear layer.