Skip to content
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.

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

Add scheduler for cosine in linear envelope. #691

Merged
merged 4 commits into from
Aug 8, 2024

Conversation

dwadden
Copy link
Contributor

@dwadden dwadden commented Aug 6, 2024

This didn't prove particularly effective for annealing, but might as well have it as an option.

This didn't prove particularly effective for annealing, but might as
well have it as an option.
olmo/optim.py Outdated
t_max: Optional[int] = None

def get_lr(self, initial_lr: float, step: int, max_steps: int) -> float:
eta_min = 0.0
Copy link
Collaborator

Choose a reason for hiding this comment

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

Consider setting this to initial_lr * alpha_f, like with the cosine and linear schedulers. Warm up would probably be nice too.

Copy link
Collaborator

Choose a reason for hiding this comment

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

If you use alpha_f, you may need to update the logic for computing the envelope.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Changed, lmk if this looks good.

Copy link
Collaborator

@2015aroras 2015aroras left a comment

Choose a reason for hiding this comment

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

LGTM!

@dwadden dwadden merged commit 4332c32 into main Aug 8, 2024
12 checks passed
@dwadden dwadden deleted the dave/cosine_linear_envelope branch August 8, 2024 18:58
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants