EarlyStopping behavior

[vl-dud]

Is it expected that the EarlyStopping callback’s behavior depends on the display_every option, i.e. that

early_stopping = dde.callbacks.EarlyStopping(min_delta=1e-3, patience=10)
model.train(iterations=100, display_every=10, callbacks=[early_stopping])

and

early_stopping = dde.callbacks.EarlyStopping(min_delta=1e-3, patience=10)
model.train(iterations=100, display_every=100, callbacks=[early_stopping])

work differently?

[lululxvi]

Good point. Ideally it should not.

[echen5503]

@vl-dud @lululxvi

It seems that this is a bug where the frequency at which deepXDE checks for callbacks is related to display cycles, when it should not.

Can you confirm the status of this as a bug?

[echen5503]

(from docs)

[echen5503]

Looking at the code, it actually seems like this is already fixed.

on_epoch_end() is not in any way dependent on display_every as of March 20.

[vl-dud]

Can you confirm the status of this as a bug?

The issue is still relevant.

[vl-dud]

on_epoch_end() is not in any way dependent on display_every as of March 20.

EarlyStopping uses train_state under the hood.

[echen5503]

My bad, you are right about this.