README.md

ITEM

Pytorch implementation of the paper "Debiased Sample Selection for Combating Noisy Labels"

Training

Hyperparameter setup

Our framework mainly contains two hyperparameters, i.e., the number of experts $m$ and the slope parameter $\beta$ in mapping function

We set $m=4$ and $\beta=3$ for all CIFAR experiments.

Run

For CIFAR-10/100 with symmetric or instance-dependent label noise

python Train_cifar.py --dataset ['cifar10', 'cifar100']
                      --batch_size 64
                      --noise_mode sym
                      --r 0.2
                      --cls_num 4
                      --beta 3
                      --gpuid 0

For CIAFAR-10N with worst, random 1/2/3.

python Train_cifarN.py --noise_mode ['worse_label', 'random_label1', 'random_label2', 'random_label3']
                       --cls_num 4
                       --beta 3
                       --gpuid 0