Will MLCD Loss implementation be released?

[leeesangwon]

Hi,

Amazing work on this project! I went through the repo but couldn’t spot the MLCD Loss implementation from your paper. Any plans to release it? Or maybe I just missed it—happy to be pointed in the right direction!

Thanks a lot!

[anxiangsir]

Thank you very much for your interest. We currently do not have plans to open source the implementation of this loss. We are working on new projects, but you can refer to the implementation of PartialFC to implement this loss.

[leeesangwon]

Thank you for the quick response!

I have a couple more questions:

1. For similarity-based sorting in multi-labeling, what kind of features were used? Was it the same as clustering, where (image feature + text feature) / 2 from the CLIP model was used?
2. The appendix mentions that ArcFace was used. Should I understand that the ArcFace margin was applied to the s_i term in the MLCD loss? Specifically, in the code, can I modify only the dist_cross_entropy part to match the MLCD formula while leaving the margin_softmax part unchanged?

unicom/unicom/partial_fc.py

Lines 240 to 241 in 896f95f

	logits = self.margin_softmax(logits, labels)
	loss = self.dist_cross_entropy(logits, labels)

If you could clarify these points, it would resolve all my confusion. Thank you in advance!