Hi, thanks for sharing this great project! Since there is no clean labels of training set and no noisy labels of test set, I am wondering the noise rate of Chaoyang dataset. This noise rate of real-world dataset is important to design the synthetic label noise settiing.

Also, I found the script step1.py defines args.noise_rate = 0.15 for Chaoyang datset. Is this value computed with clean labels, or estimated by other algorithm?

May I ask why there are no weight parameters saved in the code, how can we get the same best accuracy in the next time?

Hello, I am replicating your code now, because I had a problem when downloading your dataset, I want to replace it with my own dataset, please ask the "Almost clean dataset" file named "chaoyang_15_STEP1.p" generated by step1.py. What's in it? Thank you very much!

When I use the DM and NF-Net methods to reproduce on the Chaoyang dataset,
the former's results(acc = 39.50) are very poor but the latter's results are much higher(acc = 75.06) than the results in your paper.
I am confused about this. Could you give me some ideas? Or maybe the hyperparameter in my settings is not right?
Thanks!

I just use the method sklearn.metrics.roc_auc_score to calculate AUC, but i don't know the parameter 'multi_class' should be 'ovo' or 'ovr'?? Then i find the result is close to your paper when use 'ovr', but worse when use 'ovo'. It's my appreciate if you could tell me which one i should choose.

1、Why the value of noise_rate is 0.15 in line 66 step1.py?
2、Also in step1.py, why the filter_ratio is the product of noise_rate and 1.5 in line 216, why 1.5?
3、Why you set the value of forget_rate 0.01 when running NSHE.py?
Sorry to disturb you again

I just want to run the comparative experiment on Digestpath2019, i set the batch-size to 16, because the Co-teaching takes about 30G graphics memory！a batch size of 96 may be too large for me! I want to know what is your GPU device like? Thanks！