Does PyG offer an official dense-equivalent to its DataLoader class?

[amorehead]

Hello.

I was curious, does PyG offer an official dense-equivalent to its DataLoader class? In particular, does PyG offer a collate function that can take a list of Data or HeteroData objects and create a Batch object from them? This Batch object, instead of sparsely concatenating all graph features, would concatenate them along a new batch_size first dimension for the corresponding feature tensor. Moreover, each (sub)graph may have a different number of nodes, so DenseDataLoader wouldn't seem to work for this use case.

[rusty1s]

We provide support for batching along a new dimension via https://pytorch-geometric.readthedocs.io/en/latest/advanced/batching.html#batching-along-new-dimensions. Does this work for your use-case? However, note that we don't pad features by default.

[amorehead]

Hello. Thank you for pointing this out. I did try this approach, but it doesn't seem to work when you need to pad features (when they have different numbers of elements).

[amorehead]

This new PR seems to fix the issue: #8518.