Early stopping

[derekehyatt]

Stop early if the validation loss does not improve in early_stopping_iters (default: off).

[derekehyatt]

@karpathy - would love a review for this simple early stopping flag. Thanks!

[dimentox]

Observations: Correct Implementation: The implementation looks correct in terms of functionality. It allows the training to stop early when there is no improvement in validation loss, saving computational resources. Patience Logic: The patience logic is clear and straightforward. It resets patience to 0 when there’s an improvement in validation loss and increments it otherwise. Mutual Exclusivity Check: The script correctly checks that early_stopping_iters and always_save_checkpoint are not used together. Possible Improvements: Log More Information on Early Stopping: When early stopping is triggered, it might be helpful to log not just the number of patience iterations but also the final best validation loss and step number.