facebookresearch / LaViLa

Code release for "Learning Video Representations from Large Language Models"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Normalization values for CLIP models

Jazzcharles opened this issue · comments

Thanks for the great work.

I would like to confirm whether the normalization mean (108.32, 116.74, 104.09) for OpenAI clip-based models is correct.
The official normalization values in OpenAI CLIP is (0.48145, 0.45782, 0.40821). In this case, by multiplying 255, the values should be (122.77, 116.74, 104.09).

It's not a big deal though. The model is robust against small changes to normalization values.

Hi @Jazzcharles ,

You are right. 122.77 should be a more correct one. I'd keep the legacy value since re-training the models is beyond my compute resource. Also, the performance is unlikely to change much.

Best,
Yue