New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

#

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Jump to bottom

Shouldn't loss for each example be weighed by the sqrt of the length? #517

Open

tsjain opened this issue Jan 18, 2025 · 0 comments

tsjain commented Jan 18, 2025 •

edited

Loading

From the code it appears that the loss for an entire batch is scaled by either the crop length or the average of the sequence length for the batch.

Shouldn't each individual sample's loss be scaled by the sqrt of its length?

https://github.com/aqlaboratory/openfold/blob/main/openfold/utils/loss.py#L1776

# for free to join this conversation on GitHub. Already have an account? # to comment