Question regarding reproducibility of train, validation splits

Are the train, validations splits of the dataset guaranteed? From what I see, if I’m not missing something, randomness depends on random_split that uses PyTorch default_generator if a new PyTorch version changes this, is it possible that 2 different versions of the MNISTDataModule for example with 2 torch versions have different splits? Thanks and sorry if this is obvious!

[nateraw]

@alejandrodumas The splits you linked to will be encouraged to be done in the datamodule’s setup hook, similar to how you would do something like that in a LightningModule. This way you’d only make one call instead of two as you showed above. Great catch on that btw 😄 .

The LightningDataModule will come packaged directly with lightning in version 0.9.0 and will be removed from bolts. You can find the docs on the latest version of the built-in datamodule class in the lightning docs here.

@ananyahjha93 Let’s close this once we remove calls to random_split from dataloader hooks across this repo (if you think that’s reasonable). That way there is less confusion.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.