Skip to content

Shift labels during finetuning and provide correct padding mask #101

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
LianaMikael wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Shift labels during finetuning and provide correct padding mask #101

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 16 additions & 1 deletion experiments/run_finetuning.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,7 @@ def get_optimizer_and_scheduler(model, train_dataset, config):
class CustomTrainer(Trainer):
def __init__(self, *args, train_loader=None, test_loader=None, **kwargs):
super().__init__(*args, **kwargs)
self.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=self.model.config.pad_token_id)
self.loss_fn = torch.nn.CrossEntropyLoss()
self.train_loader = train_loader
self.test_loader = test_loader

Expand All @@ -62,6 +62,21 @@ def get_train_dataloader(self) -> DataLoader:

def get_eval_dataloader(self, _) -> DataLoader:
return self.test_loader

def compute_loss(self, model, inputs, return_outputs=False):
Copy link
Collaborator

@nailimixaM nailimixaM Feb 16, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

As this is most of the calculation for ppl can we make this one and our gpu_utils implementation consistent, and add a test? Better still would be to call a common subroutine from here and from the the ppl calc implementation. NB Pashmin'as ppl refactoring PR should go in first.

Copy link
Contributor Author

@LianaMikael LianaMikael Feb 19, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, the perplexity calculation should have the same changes, good idea to combine them

labels = inputs.pop('labels')
attention_mask = inputs["attention_mask"]
outputs = model(**inputs)
labels = labels[..., 1:].contiguous()
Copy link
Collaborator

@nailimixaM nailimixaM Feb 16, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: we should double check whether contiguous is required here, as we don't use it in our ppl implementation.

logits = outputs.logits[..., :-1, :].contiguous()
attention_mask = attention_mask[..., :-1].contiguous()

# ignore padding tokens when computing the loss
logits = logits * attention_mask.unsqueeze(-1)

loss = self.loss_fn(logits.view(-1, logits.shape[-1]), labels.view(-1))

return (loss, outputs) if return_outputs else loss


def argparser():
Expand Down