CODEWITHSUNDEEP
pythontensorflowmachine-learningkerasdeep-learning
Javier Naranjo Alcázar
Javier Naranjo Alcázar

Reputation: 105

start callback val acc after specific epoch

I have a doubt when implementing a early stopping using validation accuracy.

Suppose I want to stop a training when the validation accuracy does not improve after 50 epochs. However I can get randomly a good validation accuracy in the second epcoch. For instance:

accuracy training: 0.76 accuracy validation: 0.80

Maybe I do not get a better validation accuracy while the accuracy training is improving. So, when I reach 99% accuracy on training it may happen that the training stops because the 80% of validation accuracy in the second epoch has been too high (epoch = 52).

I would like to insert an offset. For example, early stopping using validation accuracy but starting in epoch = 100.

Is that possible in Keras?

Thanks in advance

Upvotes: 1

Views: 103

Answers (1)

SMX
SMX

Reputation: 1442

Keras callbacks are very flexible, so you could simply modify the built-in EarlyStopping callback:

from keras.callbacks import EarlyStopping

class DelayedEarlyStopping(EarlyStopping):
    def on_epoch_end(self, epoch, logs=None):
        if epoch >= 100:
            super().on_epoch_end(epoch, logs=logs)

Upvotes: 1

Related Questions