Where did the default value for max epochs come from?

[wigging]

In mattergen/conf/data_module/mp_20.yaml the default value for max_epochs is 900 and for mattergen/conf/data_module/alex_mp_20.yaml the default value for max_epochs is 2200. Where did these max epoch values come from? Why is max epochs for MP-20 lower than for Alex MP-20?

[wigging]

@danielzuegner Do you have any comments regarding this?

[danielzuegner]

Yes, that's when the validation loss would no longer decrease or even start to go up again.

[wigging]

So if I'm training with a new dataset how do I know what value to use for max epochs? Do I just guess what the max epochs should be and keep track of the loss then adjust max epochs accordingly for subsequent training?

[danielzuegner]

Hi @wigging,

I recommend using an early stopping callback (see here) with a generous patience to avoid prematurely stopping training. I'd monitor validation loss -- the model will automatically stop training if val loss hasn't decreased in <patience> many epochs.

[wigging]

Yes, that is what I used in our previous discussion #183 except I didn't explicitly define a patience value but it looks like the default is 3. But you still have to provide some value for max epochs.

[danielzuegner]

Yes, as far as I know you do, but you can set it very high such that it is effectively no constraint on the training length. I think 3 is very low, I would set it to at least 10, perhaps more.