Thanks for your great work.
In the second stage of training teacher network, you train SS module with contrastive loss, just meantioned as Eq(5) in your paper.But I only see a normal cross_entropy loss in this module. How does this Eq reflected in your code? Does the ss_T means the t in Eq5?
I also have a question that in your code about similarity,we all have:
simi = F.cosine_similarity(s_aug_feat, s_nor_feat, dim=1)
The aug_feat means the sample after transformation, the nor_feat is the same sample expanded by 3 times to make the size consistence to calculate similarity. But I think all of the simi calcuation are positive pairs. Where are the negative pairs?
Thank you.
Thanks for your great work. In the second stage of training teacher network, you train SS module with contrastive loss, just meantioned as Eq(5) in your paper.But I only see a normal cross_entropy loss in this module. How does this Eq reflected in your code? Does the ss_T means the t in Eq5?
I also have a question that in your code about similarity,we all have:
simi = F.cosine_similarity(s_aug_feat, s_nor_feat, dim=1)
The aug_feat means the sample after transformation, the nor_feat is the same sample expanded by 3 times to make the size consistence to calculate similarity. But I think all of the simi calcuation are positive pairs. Where are the negative pairs? Thank you.