A few questions on the MetaLayer

[alzaia]

Hello, I was able to adapt the MetaLayer to my actual needs but still have a few practical questions to make sure that I'm making the most of it.

1. Currently I did not implement the GlobalModel and I'm sticking to the NodeModel and EdgeModel (basically doing the steps 1 to 3 of the Battaglia paper algorithm 1: update the edges, aggregate edge attributes per node, and then update the nodes). I followed the PyG example of the MetaLayer closely. One MetaLayer pass performs one update for both edges and nodes. However, I'm stacking multiple MetaLayer layers in my case to get a more expressive model (no weight sharing: different EdgeModel and NodeModel modules are used for each block). Is this the intended use of the MetaLayer?
2. On the EdgeModel: are there other options for the NN here? I want ideally something more expressive than a MLP. Or at least stacking more of them here?
3. On the NodeModel: is it doable here to use another existing node update layer (e.g. GCN, GAT) instead of the MLP examples provided in the PyG doc? If yes, what would I need to change from the doc version of the NodeModel?

Thanks!

[wsad1]

1. I think you are using it as intended. The whole idea of MetaLayer is to plug in blocks you like, depending on your use case.
2. You could stack mlps and also maybe take difference/dot product between node features to make it more expressive in practice.
3. Yes you could replace it with any model given that model has the following forward interface forward(self, x, edge_index, edge_attr, u, batch). For example, in GATConv the interface is forward(x, edge_index, edge_attr, size, return_attention_wieghts) which should work as NodeModel, given u is None. If its not None u will get passed as size which might cause issues, but I am not sure of that.

[alzaia]

Thanks a lot @wsad1!