Question regarding NeighborLoader sampling

[skalyanb]

I have a question regarding the usage of the NeighborLoader. I have a non-tensor node property in my graph. So my Data object looks something like this:

Data(x=[10000, 200], edge_index=[2, 822330], y=[10000, 1], prop=[10000], train_mask=[1000], n_id=[10000])

where the prop is a list of custom objects. When I use NeighborLoader to get a batch, the prop is untouched.

> loader = NeighborLoader (g, num_neighbors=[-1], batch_size=32) 
> batch = next(iter(loader))
> batch
  Data(x=[600, 200], edge_index=[2, 5000], y=[529, 1], prop=[10000], train_mask=[600], n_id=[600], batch_size=32)

I can use the n_id to retrieve the prop for the nodes involved in the batch, but is there a way to tell the NeighborLoader to do this automatically?

[rusty1s]

Yeah, this is currently not supported (any non-tensor will be simply passed through). I'm sorry for that :(

We will have to implement logic for that here: https://github.com/pyg-team/pytorch_geometric/blob/master/torch_geometric/loader/utils.py#L59-L74, e.g., by filtering via index in case we encounter a list of length num_nodes. Let me know if you have interest in contributing this feature :)

[skalyanb]

I can contribute to this feature. Can I slack you in the PyG channel regarding this?

[rusty1s]

Sure, sounds awesome :)