Open jancylee opened 3 years ago
Replacing contrastive learning from image level to pixel level,it's going to be a lot less efficient, especially when calculating comparative loss. So how do you solve the thorny efficiency issues?
Actually, it is very efficient with simple parallel matrix operations. Please see our paper for a detailed running time comparison.
Replacing contrastive learning from image level to pixel level,it's going to be a lot less efficient, especially when calculating comparative loss. So how do you solve the thorny efficiency issues?