1

This looks like overfitting. You can try stop training earlier by using a validation dataset to prevent this, or you can try other regularization effects such as weight-decay, dropout etc.


Only top voted, non community-wiki answers of a minimum length are eligible