Doubt regarding the GATConv

[arake001]

In the below code snippet. Can you explain the scenario where x_l and x_r are different? I think x will always be a tensor. Therefore, x_l and x_r will always be the same.

if isinstance(x, Tensor):
    assert x.dim() == 2, 'Static graphs not supported in GATConv.'
    x_l = x_r = self.lin_l(x).view(-1, H, C)
    alpha_l = (x_l * self.att_l).sum(dim=-1)
    alpha_r = (x_r * self.att_r).sum(dim=-1)
else:
            x_l, x_r = x[0], x[1]
            print ('not same')
            pdb.set_trace()
            assert x[0].dim() == 2, 'Static graphs not supported in `GATConv`.'
            x_l = self.lin_l(x_l).view(-1, H, C)
            alpha_l = (x_l * self.att_l).sum(dim=-1)
            if x_r is not None:
                x_r = self.lin_r(x_r).view(-1, H, C)
                alpha_r = (x_r * self.att_r).sum(dim=-1)

The if statement will always be true in my opinion, as x is always a tensor. I tried with the TUDataset example. The else statement is not executed even once. Can you check or give an example where the other use case is valid? Any help would be appreciated. @rusty1s

Thank you!

[rusty1s]

The GATConv operator supports both message passing in homogeneous graphs (where x is just a regular Tensor), as well as message passing in bipartite graphs (where x is a tuple of two tensors: the node features of source and target nodes):

conv = GATConv(in_channels=(64, 128), out_channels=256)

x_src = torch.randn(num_src_nodes, 64)
x_dst = torch.randn(num_dst_nodes, 128)

out = conv((x_src, x_dst), edge_index)

This will compute new embeddings for target nodes based on incoming features (x_src) from source nodes. Hope this clarifies your issues :)

[liujxing]

I also have a question about this: is there a specific order on the index of the nodes? For example if I have

num_src_nodes = 2
num_dst_nodes = 3
edge_index = [[0, 0, 0, 1, 1, 1], [2, 3, 4, 2, 3, 4]]
x_src = torch.randn(num_src_nodes, 64)
x_dst = torch.randn(num_dst_nodes, 99)
out = conv((x_src, x_dst), edge_index)

How does the library know how to map the nodes in edge_index to nodes in features?

[rusty1s]

This is done via the indices in edge_index, e.g., a value of 0 in edge_index maps to the feature value in x[0].

[liujxing]

Yes that I understand, my question is more like "how does it know whether x[0] should be matched to x_src[0] or x_dst[0] "?

[rusty1s]

edge_index[0] (the source nodes) will be used to index x_src, while edge_index[1] (the destination nodes) will be used to indexed x_dst.

[liujxing]

now I understand, thanks a lot!