How to do custom EarlyStopping?❓ [QUESTION]

[ThePauliPrinciple]

How to do custom EarlyStopping?❓ [QUESTION]

[Linux-cpp-lisp]

Hi @ThePauliPrinciple ,

Thanks for your nice question and work with our code!

Re

checking if I get close to a walltime on a compute cluster
we do have support for a fixed walltime bound: https://github.com/mir-group/nequip/blob/develop/configs/full.yaml#L210-L211. But if you want to query the job scheduler for example that will have to be custom of course.

I've just added support for custom early stopping conditions on branch: https://github.com/mir-group/nequip/tree/feature-custom-early-stop with an example at https://github.com/mir-group/nequip-example-extension/tree/earlystop. Please give this a try and let me know if it works for you, and I'll merge it down.

If this doesn't fully solve the issue (or even if it does), it might be a more complicated workflow than I'm anticipating, and maybe we should have a quick call to discuss---please feel free to send me an email at the address listed in my profile.

[ThePauliPrinciple]

This looks good to me.

Passing the trainer object to the stopper might be useful to some, although for my use case I am only interested in "external" information.

I'm not exactly certain what the comment about restarting means, in particular, when is a stopper considered "stateful"?

The original early stopper also returned values to immediately debug/print, maybe that's also nice to add.

[Linux-cpp-lisp]

Great!

A stopper is "stateful" when it maintains a state like, say, how many epochs the validation loss hasn't improved (like the patience setting) or what the minimum observed value was (see https://github.com/mir-group/nequip/blob/feature-custom-early-stop/nequip/train/early_stopping.py#L120-L121). If it only depends on the current arguments to the object, and not any state stored in your custom object, then it's not stateful. (State of the trainer, if that was passed in, will be correctly preserved across restarts.)

The original early stopper also returned values to immediately debug/print, maybe that's also nice to add.

What do you mean, exactly?

[ThePauliPrinciple]

nequip/nequip/train/early_stopping.py

Line 98 in c56f48f

return stop, stop_args, debug_args

Here debug_args is returned, which is printed to the log:

nequip/nequip/train/trainer.py

Lines 874 to 882 in c56f48f

	if self.early_stopping_conds is not None and hasattr(self, "mae_dict"):
	early_stop, early_stop_args, debug_args = self.early_stopping_conds(
	self.mae_dict
	)
	if debug_args is not None:
	self.logger.debug(debug_args)
	if early_stop:
	self.stop_arg = early_stop_args
	return True