Callbacks for logging results to GPT2

🤗Transformers
1

I have a GPT2 model which I’m training (and works) that takes in an example input sequence and outputs a target sequence. I am able to log the model results of the validation set after the model has finished training but was hoping I could log after the end of an epoch, and was hoping to use callbacks.

What I’m hoping to log goes something like this:

class MyCallback(transformers.TrainerCallback):
    def on_epoch_end(args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
        valid_dl = trainer.get_eval_dataloader()
        
        for x, _ in valid_dl:
            x_tokenized = tokenizer(x)
            output = tokenizer.decode(model.generate(x_tokenized))
            break
       logging_fn(x, output)

I do not know how to access model, tokenizer and trainer in the above scenario. Any thoughts on how I could achieve above effect?

2

That is not supported as it would let the user modify the state of the Trainer which would then mess up with the rest of the training loop.