Batching for GCNConv Layer

[kleingeo]

📚 Describe the documentation issue

It is unclear what to do in regards to batching when using a custom feature map and adjacency matrix. For instance, the docs help when dealing with the batching from the dataloader perspective, but not with a custom approach. I am dealing with a case when I am trying to use GCNConv inside of an existing model based on latent features (ie, not from the dataloader). It is unclear what to do with the batch dimension in this case for the adjacency matrix and/or feature map. needs to be clarified

Suggest a potential alternative/fix

Show an example when using GCNConv or other graph convolution layers outside of the pre-defined dataloading.

[Percibus]

I'm not sure about this, but I think in the case of GCNConv you can batch your data and it accepts any input shape as well as the node_dim attribute is set correctly (by default to -2). So you can use for example a tensor of [BxL, N, F] and it will operate with the nodes on N for each BxL element.

[kleingeo]

The issue doesn't seem to be the tensor itself, but the adjacency matrix. It is specified in the MessagePassing.propagate that for a torch.tensor object, the shape on edge_index needs to be [2, num_messages]. I think for this to work, the edge_index needs to work with either a list, where each edge_index in the list corresponds to a batch sample, or for edge_index to allow [3, num_messages].

[Rain920]

I have the same question. When using GCNConv (or other models based on MessagePassing), how do you handle the batch dimension? Specifically, where does the batch dimension appear in the shape of x passed to propagate method?