save best models based on the loss

@tiago.pereira I have noticed that sometimes somehow accuracy improves a bit but the validation loss goes up. I think using the loss is a better solution for this.

Well, I think you can use any metric.

merged

mentioned in commit c8749cdf