This is the official repository for the paper Pseudo-Label Correction for Instance-Dependent Noise using Teacher-Student Framework. The following sections provide information on replicating the experiments. We test our method on three different datasets MNIST, FashionMNIST, and SVHN. Please download all files in the repository.
We adopted the IDN generation proposed by Chen et al.'s Beyond Class-Conditional Assumption: A Primary Attempt to Combat Instance-Dependent Label Noise. Please refer this paper to generate noisy labels with desired noise level.
We use 22%, 32%, 42%, and 52% noise in our method, but 20%, 30%, 40%, and 50% noise for existing methods. This is done to accommodate for methods that do not need a small set of clean data and ensure fair comparisons.
All datasets must be downloaded through Pytorch. We test each dataset on all noise levels {0.22, 0.32, 0.42, 0.52} on three different seeds {0, 1, 2}.
The following script must be run to reproduce results for MNIST.
python3 train_mnist.py --seed 0 --noise_rate 0.22 --epochs_retrain 50
The following script must be run to reproduce results for Fashion-MNIST.
python3 train_fmnist.py --seed 0 --noise_rate 0.22 --epochs_retrain 50
The following script must be run to reproduce results for SVHN.
python3 train_svhn.py --seed 0 --noise_rate 0.22 --epochs_retrain 50