Skip to content

Fix init_value and end_value in cosine decay#884

Open
copybara-service[bot] wants to merge 1 commit intomainfrom
test_619109148
Open

Fix init_value and end_value in cosine decay#884
copybara-service[bot] wants to merge 1 commit intomainfrom
test_619109148

Conversation

@copybara-service
Copy link
Copy Markdown

Fix init_value and end_value in cosine decay

Problem:

  • the starting value was not always init_value
  • the last value was not always end_value

Solution:

Misc:
renamed alpha -> end_value in warmup_cosine_decay_schedule

Problem:
* the starting value was not always init_value
* the last value was not always end_value

Solution:
* changed the formulas (they are now compatible with the pytorch implementation https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.CosineAnnealingLR.html)
* added test to check that starting and end value coincide with `init_value` and `end_value` respectively

Misc:
   renamed alpha -> end_value in warmup_cosine_decay_schedule
PiperOrigin-RevId: 619109148
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.

1 participant