How to use GINEConv with the HeteroConv wrapper?

[DogitoErgoSum]

I am using the latter to create the GINEConv convolutions but i cannot find a way to add the neural networks in this loop. How am I supposed to do that?

[rusty1s]

You can instantiate the MLP directly inside GINEConv:

from torch_geometric import MLP, GINEConv

GINEConv(MLP([-1, hidden_channels, hidden_channels]))

Keep in mind that this will only work for node types of same feature dimensionality!

[DogitoErgoSum]

If i understood correctly, can i do this?

I cant seem to use the MLP class because of the following error

[DogitoErgoSum]

Ty.
Relative to what you said there "Keep in mind that this will only work for node types of same feature dimensionality!", if node type A has dim of (1,2) and node type B has dim of (1,10) i cannot use GINEConv? I have to transform both node types to dim(1,X) and then use GINEConv?

[DogitoErgoSum]

What i mean is passing all node types to dim=(1,2), as exemplified in blue, and then i am able to use GINEConv (red)?

[DogitoErgoSum]

Guess i can do this ONLY if use GINConv instead of GINEConv. Thank you for your big patience. How can i acknowledge you or/and the pytorch geometric team in my master´s thesis?

[rusty1s]

This looks correct to me. Glad that it is working. Citing our paper is all we need :)

[rusty1s]

pip install class-resolver

Should fix this :)