Why the RGCNConv does not support x_dict and edge_index_dict?

[ver228]

I have been using the heterogenous graph functionalities of PyG and they work great. Thanks for developing them!

However, now I would like to bench mark against a R-GCN model, and realise that the RGCNConv layer needs a flatten version of the graph, in contrast from the HeteroConv that can receive the HeteroData x_dict and edge_index_dict.

Is there a hard reason for this or it just that this module was developed before the HeteroData format was developed?

[rusty1s]

R-GCN is designed to only be applicable on equally sized feature matrices. However, to_hetero and R-GCN are actually equivalent given the GraphConv/SAGEConv operator, see https://github.com/pyg-team/pytorch_geometric/blob/master/test/nn/test_to_hetero_transformer.py#L384.

[ver228]

Thanks! That is a very nice little example, you should consider to add it the tutorial or at least to the examples folder :)