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

Why multiplying rstd instead of dividing by rstd? #390

Closed

gohsyi opened this issue Jul 27, 2024 · 1 comment

gohsyi commented Jul 27, 2024

OpenRLHF/openrlhf/trainer/ppo_utils/replay_buffer.py

Line 214 in a534764

setattr(item, attribute, (items[i] - mean) * rstd)

When normalizing the advantages, why multiplying the standard deviation instead of dividing by the standard deviation?

Collaborator

hijkzzz commented Jul 27, 2024

Using rstd (the reciprocal of the standard deviation's square root) instead of 1/std is primarily to avoid numerical instability.

hijkzzz closed this as completed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment