Raise PL requirement to 1.3.0

asteroid-team / asteroid

The PyTorch-based audio source separation toolkit for researchers

https://asteroid-team.github.io/

MIT License

2.26k stars 422 forks source link

Raise PL requirement to 1.3.0 #612

Closed paulfd closed 2 years ago

paulfd commented 2 years ago

Remove the deprecated output parameter from on_train_epoch_end
Add a simple custom lr_scheduler_step to the base model
Remove the PL upper version constraint

This allows also torchmetrics to follow PL, which caused problems when PL was pinned and not torchmetrics.

paulfd commented 2 years ago

Sorry for the linter, this one is simple to fix though. I don't really understand why the docs and coverage break; maybe some supporting package got upgraded through pytorch-lightning newer version or something?

mpariente commented 2 years ago

Thanks for the PR. The coverage fail is normal because you added some code without tests, but for the docs, it's probably a dependency problem, as you mention, do you mind having a look please ?

mpariente commented 2 years ago

Thanks @paulfd for the PR !

We'll be able to upgrade to lightning 1.5 and above. But first we'll need to update the recipes for the new Trainer API (strategy, accelerator and devices arguments are different from PL<1.3.0).

mpariente commented 2 years ago

I changed the base branch of the PR, because as soon as we merge this, the recipes won't work anymore (because of changes in the Trainer arguments), so I prefer to make a dev branch out of this.

Thanks again, merging this !