Hello, in the process of reproduction, I found that there was a problem when the cifar-10 experiment was reproduced. I reproduced 66.02%, Symmetric=30%. What steps did I miss? My other parameters are the same as your code. Thank you for your answer.

1. The L_e in the code i think is the Loss_reg = 𝐿𝑜𝑠𝑠𝑟𝑒𝑔 = 𝐽𝑆 (𝑞1, 𝑞2) =1/2(𝐾𝐿(𝑞1 ∥𝑈 ) + 𝐾𝐿(𝑞2 ∥𝑈 )).* But when i read the code, i find it is calculated by**" L_e = - (torch.mean(torch.sum(sm(logits1) * lsm(logits1), dim=1)) + torch.mean(torch.sum(sm(logits1) * lsm(logits1), dim=1))) * 0.5 "**. It seems ignore the " -sum(q1logU)+sum(q2logU)"???

Hello and thank your for your research and code.
I have some question on the notebook provided here.

1. Under the 'Client' cell, inside model training to get logits1 and logits2 outputs from the model are multiplied by 3, can you comment on that?
2. The self entropy regularization L_e is it mentioned on the paper? and maybe a typo logits1 is used in both sum operations.

Thanks :)

I used your code for training on Cifar100 data and I also replicated your code myself for training on Cifar100. I find that both converge very slowly. I am guessing that there is something wrong with my hyperparameter settings, do you have any suggestions? Also if possible I would like to know your hyperparameter settings with Clothing1M.