At cifar10 0.2 sym noise rate setting, cann't repeat the results reported at paper.

LiJunnan1992 / DivideMix

Code for paper: DivideMix: Learning with Noisy Labels as Semi-supervised Learning

MIT License

529 stars 83 forks source link

At cifar10 0.2 sym noise rate setting, cann't repeat the results reported at paper. #50

Open sillytheif opened 1 year ago

sillytheif commented 1 year ago

The cifar10 0.2 noise rate, I can only get about 91.5% acc useing the code, and I can't find the reason

kovelxyz commented 1 year ago

I found a similar question and got the same precision as yours.

xhl-video commented 1 year ago

I can only get ~92% performance as well.

xhl-video commented 1 year ago

I have found the solution. We should follow Table 7 in the appendix to assign λu. It should be 0 for the Cifar10 20% dataset, which is pretty different for other noise ratios.

ljzoe commented 2 days ago

How long do you run this code?It takes me 1 hour to train one iteration!Do you have the same situation?