on_epoch_finished are not notified when fitting is complete about nolearn HOT 5 CLOSED

I really need to get around documenting the interface. (Been waiting for Lasagne to add names to layers so that I could simplify a few things before.) But what you're looking for is already there and it's called on_training_finished.

from nolearn.

So in order to write early stopping to set best on training finished you would want to set the same early stopping object in both lists?

Eg I would write:

early = EarlyStopping(...)

NeuralNet(
....
on_training_finished=[early],
on_epoch_finished=[early],
)

from nolearn.

You'd have to add a method def load_best_weights(self, nn, train_history) to EarlyStopping that does only the nn.load_weights_from(self.best_weights) call. Then use that method in the on_epoch_finished handler: on_epoch_finished=[early.load_best_weights].

from nolearn.

More concretely, this would be the extra method on EarlyStopping:

    def load_best_weights(self, nn, train_history):
        nn.load_weights_from(self.best_weights)

I think it'll be useful to move EarlyStopping into nolearn itself (#18), in which case we might provide an implementation that has the above method, and documentation on how to use it.

Please reopen if you think this issue is still valid on its own.

from nolearn.

Here is the origin of the early stopping. Concise!

from nolearn.