GINE does not work beyond the first convolution

[mdanb]

I am trying to use GINE conv for a weighted graph where the node features are 1 dimensional, and the edge_attr (i.e edge weights) of my Data object are also 1D. My hidden dimensions are 16D. The following shows my code for my forward method, which uses GINE convs.

def forward(self, data):
     x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
     for i in range(self.num_layers):
         x = self.convs[i](x, edge_index, edge_attr) # convs is just a list of GINE convs, where the first GINE conv is 1x16 and the rest are 16 x 16
         x = F.relu(x)
         x = F.dropout(x, p=self.dropout, training=self.training)
         if not i == self.num_layers - 1:
             x = self.lns[i](x) # lns is just a list of LayerNorms with hidden dim = 16

     x = global_add_pool(x)

     x = self.post_mp(x) # this is just an MLP

     return F.log_softmax(x, dim=1)

The first iteration through the loop works, but beyond the first I get the error:

  File "/home/mdanb/.local/lib/python3.8/site-packages/torch_geometric/nn/conv/gin_conv.py", line 109, in forward
    assert x.size(-1) == edge_attr.size(-1)
AssertionError

x.size(-1) is 16 and edge_attr.size(-1) is 1
The error makes sense I guess since the update/aggregation for GINE involves adding up the node embeddings and edge embeddings (as opposed to GCN where we consider weights as scalars by which we scale node embeddings). So how exactly should I incorporate processing the edge weights in the code above?

[rusty1s]

In general, you need to take care about this yourself. I suggest to use a Linear layer to transform edge features to the appropriate size, e.g., Linear(num_edge_features, num_hidden_channels):

x = self.convs[i](x, edge_index, self.edge_lins[i](edge_attr))

[mdanb]

I see, thank you! I'm assuming you meant doing this:

x = self.convs[i](x, edge_index, edge_attr)
edge_attr = self.edge_lins[i](edge_attr)

(The other one didn't work because the edge attributes were being transformed before the convolution so the node attributes were still in the "old" dimension)