transformers.Trainerのカスタマイズ

To inject custom behavior you can subclass them and override the following methods:

Another way to customize the training loop behavior for the PyTorch Trainer is to use callbacks that can inspect the training loop state (...) and take decisions (like early stopping).