Open viktor-ktorvi opened 9 months ago
Currently, sparse tensors are not supported yet as part of feature fetching during NeighborLoader
. We can fix the issues with cat_dim
being a tuple, but it will still crash downstream since torch.sparse
does not support index_select
yet :(
🐛 Describe the bug
Hi,
I batch my sparse matrices block diagonally so
cat_dim
is a tuple. I want to use NeighborLoader but it expectscat_dim
to not be a tuple internally and throws an error.I should add that the block diagonal batching and using the NeighborLoader are unrelated tasks. I usually work with smaller graphs and load them with the usual DataLoader. I just so happen to have a larger graph that I want to make lots of smaller graphs from, so I'm using the NeighborLoader for that. Strange use case but seems reasonable enough.
The following snippet recreates the error:
I feel like this is an issue.
Versions
I'm on Windows so I couldn't get the above to run but my pip freeze is: