Skip to content

Fix ppo ratio inaccuracy #108

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

Jump to bottom
Merged
merged 3 commits into from
Nov 22, 2022
Merged

Fix ppo ratio inaccuracy #108

merged 3 commits into from
Nov 22, 2022

Conversation

maxreciprocate
Copy link
Collaborator

This pr fixes #107

  • lets models in orchestrator use attention_mask for computing old_logprobs

but also

  • removes entity_name from ppo_config (which I suspect wasn't intentional)
  • enables more thorough logging, synced with orchestrator's logging with iter_count
  • adds git commit as tag to wandb runs

https://wandb.ai/sorry/public/reports/Ratio-fix--VmlldzozMDE4NzI2

@Dahoas
Copy link
Collaborator

Dahoas commented Nov 22, 2022

Perhaps add the ratio == 1 check in a unittest?

@jon-tow jon-tow self-requested a review November 22, 2022 20:23
jon-tow
jon-tow approved these changes Nov 22, 2022
View reviewed changes
Copy link
Collaborator

@jon-tow jon-tow left a comment •
edited
Loading

Choose a reason for hiding this comment

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

Looks great @reciprocated 👏

@jon-tow jon-tow merged commit 8edcd9d into main Nov 22, 2022
@maxreciprocate maxreciprocate deleted the 107-fix-ppo-ratio branch November 22, 2022 21:31
@maxreciprocate maxreciprocate mentioned this pull request Nov 28, 2022
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@jon-tow jon-tow jon-tow approved these changes

@Dahoas Dahoas Awaiting requested review from Dahoas

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

Ratio != 1 at start of PPO training (during loss function calculation)
3 participants
@maxreciprocate @Dahoas @jon-tow