PyTorch code and models for V-JEPA self-supervised learning from video.
2.53k
stars
242
forks
source link
PCA feature map visualization of a pre-trained weights look very random, compared to without pre-trained weights loaded #66
Open
icekang opened 1 month ago
Hi,
Thank you for this amazing project.
I have been exploring the feature maps produced by the pre-trained V-JEPA, using PCA component visualization.![image](https://github.com/facebookresearch/jepa/assets/10291299/02b1ec82-dbb7-4e28-88c2-05cb9990c929)
However, the feature maps look very random, so I try doing the same thing without the pre-trained weight.![image](https://github.com/facebookresearch/jepa/assets/10291299/64e1fef7-6bfa-441a-93e6-af2cdd8aa97e)
Were the feature maps from the V-JEPA pre-training supposed to be like this, or what did I missed in loading the pretrained weight?
Here is the code I used to do the feature visualization.
The lol.csv which I downloaded from https://www.kaggle.com/datasets/ipythonx/ssv2test?resource=download