search

tonyhuang2022 / UPL

This repo is the official implementation of UPL (Unsupervised Prompt Learning for Vision-Language Models).
106 stars 5 forks source link

Pseudolabelling Details #13

Closed andydelworth closed 1 year ago

andydelworth commented 1 year ago

Hi,

I am wondering if you could elaborate on your top-K pseudolabelling scheme. Mainly I am wondering, if an image is in the top-K for multiple classes, does it appear in the dataset multiple times, once for each class? Please let me know if this question makes sense.

Thanks, Andy

andydelworth commented 1 year ago

Any pointers to the pseudolabelling code in your repo would be helpful as well, as I could not find it. Thanks!

tonyhuang2022 commented 1 year ago

Thanks for your question. You can find the details of top-k selection in https://github.com/tonyhuang2022/UPL/blob/master/trainers/utils.py. We only select each image once.