hitcszx / alfs Goto Github PK

Dear authors,

How are you?

I read your great work about asymmetric loss functions: Asymmetric Loss Functions for Learning with Noisy Labels, ICML 2021.

However, we have two pieces of work which worked on this aspect too, which I believe to be highly relevant.

(1) In DERIVATIVE MANIPULATION FOR GENERAL EXAMPLE WEIGHTING (https://arxiv.org/pdf/1905.11233.pdf), we mentioned :

(2) In IMAE for Noise-Robust Learning: Mean Absolute Error Does Not Treat Examples Equally and Gradient Magnitude’s Variance Matters, we mentioned:

Both papers have been included in my PhD thesis: Example weighting for deep representation learning, Xinshao Wang, 2020

Therefore, could I ask you kindly to cite our two papers? If you could update your arXiv version by citing our two papers, I would appreciate it a lot.

Thanks very much.
Kind regards,

Hi,
Thanks for this great implementation.
There are some hyper-parameters in the proposed ALFs. And the test accuracy corresponding to each hyper-parameter with different values is shown in the paper. However, there is no information on how to adjust the hyper-parameters in the training phase. So I wonder how you choose the value of these hyper-parameters. Is there a validation set with noisy labels? If yes, how to evaluate the performance on the noisy validation set? Looking forward to your reply.

Best Regards,

Hi,

Thanks for this great implementation. When I ran this code, I found that the implementation of the NLNL loss might exist problem that will cause loss to be Nan. It might be caused by this line:

. This line will make all labels to be -100. Would you mind checking this issue?

Best Regards,
Hongxin