Edge attribute dimension with GINEConv

[dpinski1]

Hi, I'm trying to train a model using GINEConv that incorporates edge attributes. I'm confused by the second assert statement in the models forward method:

if isinstance(edge_index, Tensor):
            assert edge_attr is not None
            assert x[0].size(-1) == edge_attr.size(-1)

Let's say my node feature matrix X is of dimensions (10, 100) - 10 nodes with 100 features each. If there are 50 edges and I am only using one edge attribute, then my edge attribute matrix will be size 50. I don't understand why the size of my first row of the node feature matrix has to be equal to the size of my edge attribute matrix - is this claiming there should be a relationship between the number of node features and the number of edge attributes? Hoping someone can point out where I'm getting confused - thanks!

[rusty1s]

Yes, exactly. Node feature dimensions and edge feature dimensions have to match, as GINEConv utilizes summation to integrate edge features into source node features. A good way to allow that is to transform your edge features first before passing it into GINEConv, e.g.:

x = conv(x, edge_index, lin(edge_attr))

where lin denotes a Linear layer mapping to the number of node features.

[dpinski1]

Thank you!! A follow up - in this case, should the parameters for the Linear layer be learnable? I'm assuming its fine if they are fixed.

[rusty1s]

I think it's reasonable to let Linear be trainable, as the model can learn how to integrate edge features to the message passing phase this way.