search

LTH14 / targeted-supcon

A PyTorch implementation of the paper Targeted Supervised Contrastive Learning for Long-tailed Recognition
MIT License
93 stars 13 forks source link

Question about the target centers. #9

Open wjun0830 opened 1 year ago

wjun0830 commented 1 year ago

Thanks for the interesting paper.

I have 2 questions.

  1. The first one is about the assignment process. In the paper, it says that the class centers are updated in a way of the weighted moving average.

However, if we softly assign the target center e.g., 0.9 c1 + 0.1 c2, isn't the property of uniformity violated? Or do you assign "c1" if the weight for c1 is 0.9?

  1. How does Eq. 1 satisfy the uniformity? If the feature dimension is 2048 and there are 1000 classes, making 1000 centers orthogonal does not seem to make them uniformly distributed.

Thank you.

whper07 commented 4 months ago

In my opinion, the relation between the relative size of the dimension and the number of categories does not seem to matter. I'm not sure