Dear sir, can you give me the hyperparameter setting for DivideMix on this dataset?

[LanXiaoPang613]

Dear sir, can you give me the hyperparameter setting for DivideMix on this dataset?

[LanXiaoPang613]

When i reproduce the experiments of dmix on cifar10N, the results for aggregate based on ResNet-34 and default setting of divideMix are only nearly 84%. So if possible, pls give me the hyper-parameter setting for the divideMix

[weijiaheng]

Thanks for your interest in our dataset:

We fixed the parameter lambda=25 for DivideMix in all our experiments with different noise levels. We contacted the authors of DivideMix and found out that a smaller lambda is required for low noise (this is mentioned in Appendix B of the open review version of DivideMix). Therefore, in the camera-ready version, we re-adjusted the parameters for the low noise situation of DivideMix. You might want to try and see how it works with lambda=0 instead of lambda=0. Feel free to contact us if you have further questions.