How to compute per-token loss when doing language modeling?

🤗Transformers
1

I’m fine-tuning GPT-2 on a language modeling task. Given a sequence, I’d like to compute the per-token loss for each token (instead of the averaged loss over the sequence). How can I do it?

2 Likes
2

Hey, it’s been some time since you posted this. Were you able to do it? If yes, would you be able to share code for it?

3

In pytorch, you can use CrossEntropyLoss function with reduction param set to none ( CrossEntropyLoss — PyTorch 2.0 documentation ).

You can find an example in a HF course: Training a causal language model from scratch - Hugging Face NLP Course (search for keytoken_weighted_loss to find it)

1 Like
4

Thanks for your reply @chrisociepa. Yes, I was able to work it with reduction='none'