implement early stopping?

[florian-huber]

I ran a model training with early stopping and than got unexpected behavior of model.save() since that will now save the final model after training, hence ignoring any save_best_only=True setting.