Skip to content
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.

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

CoCa: Condition captioning loss on the CLIP similarity #469

Draft
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

iejMac
Copy link
Contributor

@iejMac iejMac commented Mar 17, 2023

This is an attempt at a step in the direction of models determining which samples are relevant for themselves. We can use the CLIP similarities to re-weigh the loss of the captioner. It's kind of like finer filtering.

Some notes:
prior(gs) = 1.0
evidence(gs) = clip similarity over local batch
posterior = ???

currently I just do posterior = evidence but this isn't smart, there is some clever way I just need to think about it more. It's not great during the beginning of training when your CLIP model still sucks because then evidence is just 1/bs always so you effectively just divide your learning rate by batch size. There's some clever thing to do here like the closer you are to the beginning distribution the less you pay attention to CLIP and the more the distribution becomes interesting the more you pay attention to it. Or you can just put it on a schedule like the learning rate.

Wandb: https://wandb.ai/iejmac/open-clip/reports/CLIP-conditioned-caption-loss--VmlldzozODExMzUy

@iejMac iejMac marked this pull request as draft March 17, 2023 07:57
git pus
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