NNConv Example for graph classification vs node classification

[alzaia]

Hello, in the QM9 example using NNConv layers for the molecule property task, a GRU and a set2set layer are used at the end of the pipeline. I understand these last layers are there to go from the node level to the graph aggregation (since it is a graph classification task). How would this change if I needed to use it for a node classification task? Can I simply keep the NNConv layers and then add an MLP for the node classification head at the end, or do I still need a GRU like layer in there? If yes, what other options are there instead of a GRU for this example? Thanks!

[rusty1s]

Yes, you should just drop the set2set layer. For GRU I suggest to try out both (with and without) to see which one performs best.

[alzaia]

Thanks @rusty1s for the feedback! In this example, the same learnable layers are reused multiple times during the forward pass. What would be the advantage of that, instead of having multiple different layers for each block? The reuse of the learnable layers (NNConv and GRU) here would reduce the size of the model, but also reduce the expressivity.

[rusty1s]

IMO, it is better to not share GNN layers from my own experiments, but this is what the "Neural Message Passing" paper does and so we have followed their architecture.

[alzaia]

Right. Thanks @rusty1s !