snap-stanford/crust

[NeurIPS 2020] Coresets for Robust Training of Neural Networks against Noisy Labels

crust's Issues

tabular data/ noisy instances

Hi,
thanks for sharing your implementation. I have two questions about it:

1. Does it also work on tabular data?
2. Is it possible to identify the noisy instances (return the noisy IDs or the clean set)?

Thanks!

how to get the V_j

Hi, your work is remarkable and interesting. But I do not find where you generate the V set(according to the original paper, they should be some cluster group in each sub-class, I am right?). Please tell where you implement it in the code. Thanks.

Having trouble replicating table 2 from the paper.

Hi friends, would you mind sharing the commands you used to generate table 2? I am consistently getting way worse results than reported form paper. For instance, I tried:
python robust_cifar_train.py --gpu 0 --use_crust --mislabel-ratio 0.2
and it only gave 84.970.

I also did an ablation study by setting noise ratio to 0, it only gave around 83% accuracy, which was very weird, since with noise labels it could get accuracy to over 90.
python robust_cifar_train.py --gpu 0 --use_crust --mislabel-ratio 0