Skip to content

fix: use closed-form LR calculation to fix polynomial decay formula bug #67

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

fix: use closed-form LR calculation to fix polynomial decay formula bug #67

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
21 changes: 5 additions & 16 deletions training/lr_scheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,22 +27,11 @@ def get_lr(self):
UserWarning,
)

if self.last_epoch == 0 or self.last_epoch > self.total_iters:
return [group["lr"] for group in self.optimizer.param_groups]

if self.last_epoch <= self.warmup_iters:
return [
base_lr * self.last_epoch / self.warmup_iters
for base_lr in self.base_lrs
]
else:
l = self.last_epoch
w = self.warmup_iters
t = self.total_iters
decay_factor = (
(1.0 - (l - w) / (t - w)) / (1.0 - (l - 1 - w) / (t - w))
) ** self.power
return [(1/2 * group["lr"]) * (1+decay_factor) for group in self.optimizer.param_groups]
# Use closed-form calculation for correctness and numerical stability
# Previous implementation had a formula bug: (1/2 * lr) * (1 + decay_factor)
# is NOT equivalent to polynomial decay. It causes slower decay than intended.
# The closed-form directly computes: base_lr * (1 - progress)^power
return self._get_closed_form_lr()

def _get_closed_form_lr(self):
if self.last_epoch <= self.warmup_iters:
Expand Down