I have a dataset separated in train, test and validation splits.
After each epoch, I evaluate the loss and accuracy in the validation split.
When the loss in validation split is not better, I stop train and choose that as final model.
But, I should merge train and validation as final model? How can I choose the best model?