Supervision Edges Attributes in LinkNeighborLoader

[YoavLotem]

Hey!

I'm working on an edge classification problem.
I'm using the LinkNeighborLoader to sample supervision edges and their corresponding labels.

Since the supervision edges (defined by edge_label_index) are not necessarily part of the message passing edges (defined by edge_index) I'm having a hard time adding their features.

I tried concatenating them to the edge_label and extract them in the training loop but I got a "ValueError: only one element tensors can be converted to Python scalars", which I guess means the code assumes the edge_label has 1-d.

Maybe it could be beneficial to add a label_edge_attr parameter to the LinkNeighborLoader.

Can think of any workaround to solve this problem?

[rusty1s]

Interesting. So it looks like you want to integrate edge-level features that are only available for supervision edges. Is that correct? What is the use-case of this? I think for now concatenating is a valid approach. Can you share an reproducible example of the error? This does not seem intended.

[YoavLotem]

@rusty1s kind reminder

[rusty1s]

Yes, will look into this ASAP, sorry for the delay!

[rusty1s]

I just pushed a fix via #5186. Thanks for being patient!

[YoavLotem]

Thank you very much! I appreciate it.

I hope your team would consider my initial suggestion - I think that many real world applications might benefit from it.

[rusty1s]

Definitely, sorry for the "hacky" workaround for now :(