Can GCN and GIN run in minibatch training like GAT does?

[imethanlee]

I found an implementation of minibatch GAT (https://github.com/pyg-team/pytorch_geometric/blob/master/examples/ogbn_products_gat.py). I modify the code to GCN and GIN and it seems does not work. I was wondering whether it is possible to run GCN and GIN with minibatching as that GAT code does?

[rusty1s]

Yes, in general, all our GNN layers work in mini-batch mode. This was made easier with the new NeighborLoader class, which does no longer require bipartite message passing. Here is an example.