One Graph, and inputs are features

[hossam-zaki]

Hello! I'm learning a bit more about Pytorch geometric for a project I'm working on. My original plan was to have one graph, and the inputs would be the features for each node in the graph. But, from what I'm reading, this doesn't seem to be possible. Is it?

Another alternative would be to have the standard inputs (graphs and features in the Data object), where the adjacency matrix will not change, but the features will. Would this work too?

Any help is appreciated. Thank you so much!

[rusty1s]

Thanks for your interest. I'm not entirely sure what you are trying to do. PyG can handle both processing of a single graph (e.g., examples/gcn.py) or multiple graphs (e.g., examples/mutag_gin.py). If your adjacency matrix is fixed and only your input feature matrix vary, I think it's best to assign your edge_index to all graphs and store them in a list of data objects.

[hossam-zaki]

Thanks a lot for the comment. Yes, the adjacency matrix is fixed, and the input feature matrix will vary. So basically, you are suggesting that I process multiple graphs, each with the same adjacency matrix but different input features?

[rusty1s]

Yes, as long as you take care of sharing memory of edge_index between different data objects, there shouldn't be any memory overhead on the CPU side. On the GPU, edge_index will be replicated for each example in your batch, which is probably a negligible overhead.