Training using Supcon loss is slower than cross_entropy loss

[yinbing668]

I find that in the same case(same dataloader, same network architecture), training using Supcon loss, the accuracy improves slower than the cross_entropy loss. Do you find the same case?
Besides, i cannot run the code, it report"segment error".

[tian1327]

check this ICML'21 paper: https://arxiv.org/abs/2102.08817