No self.best_weights in the function train_loop() ? about nolearn HOT 3 CLOSED

It is correct that the train loop does not keep track of the best weights seen so far. That being said, "best" is not uniquely defined. Do you mean best classification score or best objective score? Or some other notion of best? That decision is left to the user.

You can add this functionality using something like I proposed in: #18 (comment) (heavily inspired by Daniel's implementation in his blog post) and then:

    early_stopping = EarlyStopping(patience=100)

    net = NeuralNet(...
        on_epoch_finished=[
            early_stopping,
        ],
        on_training_finished=[early_stopping.load_best_weights]
     )

from nolearn.

Thanks for your comment. I believe we can close this one now.

from nolearn.

@cancan101 , thank you for your timely reply and you are right. This is defined by developer, which keeps the main function clean and makes it more flexible and extendible. Inspired by your early_stopping a lot. Also thanks @dnouri for your amazing blog post.

from nolearn.