About the usage of pseudolabels to enhance CLIP

[CriMenghini]

This repository is very useful to learn about the works bootstrapping off CLIP, thank you for curating it!

We have just published on arXiv a work that investigates how to best use pseudolabels generated by CLIP to enhance CLIP itself. We believe this work to have good applicability for practitioners that want to adapt CLIP to novel tasks efficiently and with limited, or no, labeled data.

You can find the paper here and the code here

I'm happy to submit a pull request if needed :)