search

LiJunnan1992 / DivideMix

Code for paper: DivideMix: Learning with Noisy Labels as Semi-supervised Learning
MIT License
543 stars 84 forks source link

some question about when noise_mode = asym #13

Closed Wongcheukwai closed 4 years ago

Wongcheukwai commented 4 years ago

Hi Junnan,

I really like your paper and am running your code. But I have some question regarding how you deal with asymmetric noise.

  1. In line 24 in dataloader_cifar.py, did you just match similar class manually? Because I checked the Cifar official website, it seems you just match similar classes like cats and dogs, deers and horses, birds and planes. May I ask why you generate asymmetric data like this?

  2. I didn't find the asymmetric class transition for Cifar100 in your code and it is interesting that you didn't report asymmetric noise accuracy in you paper in Table 5. So can you tell me how you generate asymmetric data for Cifar100?

Looking forward to your reply!

LiJunnan1992 commented 4 years ago

Hi,

  1. Yes for asymmetric noise only visually similar class labels are exchanged. The similar classes were defined by previous papers and I followed their definition.
  2. I didn't generate asymmetric noise for cifar100. As far as I know, there hasn't been a commonly-followed definition of what classes are visually similar for cifar100.
Wongcheukwai commented 4 years ago

thank you for your reply. As in previous paper, do you mean Unsupervised label noise modelling and loss correction?

LiJunnan1992 commented 4 years ago

The F-correction, Joint-Optim, Meta-Learning, and P-correction paper use the same asymmetric noise.

Wongcheukwai commented 4 years ago

thanks Junnan I will close this one. Just wanted to say your paper inspired me a lot.