Let EarlyStoppingCallback not require load_best_model_at_end

[muellerzr]

What does this PR do?

This PR tweaks EarlyStoppingCallback to still require a metric, but doesn't need load_best_model_at_end. Doing so lets users still use it, but optionally doesn't require them to load it back in at the end of training if they don't want to.

Fixes # (issue)

Request of @BenjaminBossan

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@Rocketknight1 @SunMarc

[SunMarc]

LGTM ! Thanks !

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Rocketknight1]

This one might need a real core maintainer cc @ArthurZucker @LysandreJik! The code looks good, but is there a risk of user confusion if they use EarlyStopping but actually get a checkpoint later than the best one with significantly worse metrics? I think I'd find that behaviour quite surprising.

[BenjaminBossan]

Thanks for the change, from my point of view it's much more intuitive that I can use early stopping without loading the best model at the end. The checkpoint is still there and I can decide to load it if I want to.

but is there a risk of user confusion if they use EarlyStopping but actually get a checkpoint later than the best one with significantly worse metrics? I think I'd find that behaviour quite surprising.

How about just giving a warning? So use the old check but don't error out but only warn. I'm not necessarily happy about adding more warnings, but if users have come to rely on this error to remind them to enable load_best_model_at_the_end, getting a warning could be a good compromise.

[Rocketknight1]

Hmm, I'm negative on warnings in general, but maybe a single warning at the end of training might be understandable here, given the potential performance damage if users want the best checkpoint but get the one after n epochs of metrics getting worse.

[BenjaminBossan]

@muellerzr What is your opinion on giving a warning?

[ArthurZucker]

I have no strong opinion, the less warning the better as long as something is properly documented! 🤗

[BenjaminBossan]

Let's not forget about this PR, not a lot missing for it to be merged.

[muellerzr]

I'll add in the warning now :)