Add a warning when some of the modules are in eval mode before the training stage

[mszulc913]

Description & Motivation

I feel like #18951 is not visible enough. The change made it super easy to have silent correctness bugs in the codebase when using libraries like huggingface.

Pitch

Add a warning about modules that are set to eval mode before training.

Alternatives

Update the documentation and tutorials.

Additional context

No response

cc @Borda

[AndreiCComan]

Hello, thanks @mszulc913 for bringing this up. It seems like when loading a huggingface model using from_pretrained, it defaults its training state to False during the training_step. That's indeed not ideal, and could potentially have some negative consequences, especially regarding components like Dropout in the pretrained model.

[awaelchli]

I prefer not to make a warning. This could confuse users who finetune models where parts of the model are frozen (and thus have to remain in eval mode). This was the main motivation of #18951. The model summary update should give some more visibility into this #19468, it could also be made even more explicit there.

[mszulc913]

Thank you @awaelchli for the reply. Indeed, the updated model summary will help a lot. However, I'm not sure if it's enough. My main concern is that pre-trained models are ubiquitous today, and this feature poses a risk of turning users away from PL, because of a regression they don't have time or will to investigate.

The docs should be updated as well, for example here:
https://lightning.ai/docs/pytorch/stable/advanced/transfer_learning.html