I am training a neural network with Keras using EarlyStopping based on val_acc and patience=0. EarlyStopping stops the training as soon as val_acc decreases.
However the final model that I obtain is not the best model, namely the one with the highest val_acc. But I rather have the model corresponding to the epoch after, namely the one corresponding to a val_acc just a bit lower than the best one and that caused the early stopping!
How do I get the best one?
I tried to use the save the best model using the call back:
ModelCheckpoint(filepath='best_model.h5', monitor='val_loss', save_best_only=True)]
But I get the same results.
If you would like to save the highest accuracy then you should set the checkpoint monitor='val_acc' it will automatically save on highest. Lowest loss might not necessarily correspond to highest accuracy. You can also set verbose=1 to see which model is being saved and why.
In Keras 2.2.3, a new argument called restore_best_weights have been introduced for EarlyStopping callback that if set to True (defaults to False), it would restore the weights from the epoch with the best monitored quantity:
restore_best_weights: whether to restore model weights from the epoch with the best value of the monitored quantity. If False, the
model weights obtained at the last step of training are used.
