On updating edge attributes

[leonardtschora]

Hi everyone,

I am struggling to understand how exactly Edge Feature updating works in the pytorch framework.
More precisely, here is a list of questions that I hope are sufficiently clear for being answered.

1. As I understand, it is possible to update edge attributes using the 'edge_update' function of a Conv layer. Using this function works similarly than the 'forward' function. Can you confirm that, up to this date, only the GatConv layer has this function implemented? (on the latest version, 2.1.0).
2. In the GatConv paper, the edge attribute updating procedure is not put forward. Moreover, it seems to be missing from the GatConv layer documentation. Are the edge attributes replaced by the aggregation of attention coefficients?
3. Lastly, I feel that the ResGatedGraphConv layer should support edge attribute updating. Indeed, authors of the paper mention edge gates that seems to update the edge attributes. This is confirmed in this article review. Is there any particular reason why this functionality is not available?

I hope I have been clear enough. I thank you by advance for your help, any input would be very appreciated.

[rusty1s]

1. Yes, currently only GATConv makes use of it. The tests have some additional examples, see here.
2. In the current design, the edge features are used to derive at the attention coefficients between source and destination nodes.
3. I am happy to include it there if you want to make a PR :)

Let me know if this resolves your concerns.

[leonardtschora]

Thanks for the answer! Just to be sure, the GATConv layer does NOT updates edges embeddings, it only computes them, right?

[rusty1s]

That is correct, if you want to update them you would have to return them as part of forward and just them in later layers.

[leonardtschora]

Alright, thank you very much!