🚀 Add early stopping to the trainer

[BramVanroy]

🚀 Feature request

The trainer (pt, tf) is an easy access point for users who rather not spend too much time building their own trainer class but prefer an out-of-the-box solution. Even though transformers was never meant to be a fully fletched training library, it might please users to add an additional feature: early stopping.

Motivation

Early stopping ensures that the trainer does not needlessly keep training when the loss does not improve. This saves time, money, and let's not forget the trees. 😉 Performance-wise this should not lead to different results.

Your contribution

At the moment I cannot work on this, but here are my thoughts:

• a training argument should be added (pt, tf). This would only work when evaluate_during_training is enabled.
• for PyTorch: at every evaluation step, an early stopper (can be a separate class even) checks if the loss has improved in the last n steps. Potentially with a minimal threshold that the loss should have improved. If not, the trainer should stop
• for Tensorflow: I don't have experience with TF myself, but I assume one could use tf.keras.callbacks.EarlyStopping.

[stale]

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.

[BramVanroy]

Looking at the interest this topic has, I am bumping it to re-open it.

[san7988]

Hi,

So when #4186 is closed, this will close as well? Or is there any more changes expected. on this issue, apart from what #4186 adds?

Thanks

[KMFODA]

If I've understood things correctly, I think #4186 only addresses the Pytorch implementation of the trainer. @BramVanroy if that's the case I'm happy to work on implementing this feature in Tensorflow (trainer_tf.py).

[BramVanroy]

@san7988 @KMFODA This issue should not directly be closed when that PR is merged because as @KMFODA mentions, it only seems to address PyTorch. A PR for Tensorflow is also welcome!

[KMFODA]

Thanks for clarifying @BramVanroy. Apologies I was out for the past month due to a personal issue. I'll submit a PR for Tensorflow early stopping now.

[BramVanroy]

An early stopping callback has now been introduced in the PyTorch trainer by @cbrochtrup! 👏

AFAIK the implementation the TF Trainer is still under way (#7533) so I'll keep this topic open for now.

[djrochford]

I gather from the conversation on #7533 that this issue should now be closed; is that correct, @BramVanroy ?