I was going through the datamodules of lightning_bolts and realized that train/val_dataloader methods are being used to define ‘transforms’ and ‘datasets’ instead of setup hook as shown in the docs.
Is there any known issue blocking this provision or it’s just a matter of preference?
I would say this comes down to preference. If you look at something like STL10, since there are many different splits of data (labelled, unlabelled and a mix of both), keeping transforms within *_dataloader() methods is just easier to track.
The real answer here is because bolts was where the DataModule class originally lived, and it didn’t have a setup hook, if I remember correctly.
We added setup when we integrated DataModule with Lightning, and the bolts implementations just continued to work, so we figured “if it ain’t broke, don’t fix it”.
Thanks for the answer @nate