T_Revision

NeurIPS‘19: Are Anchor Points Really Indispensable in Label-Noise Learning? (PyTorch implementation).

This is the code for the paper: Are Anchor Points Really Indispensable in Label-Noise Learning?
Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, Masashi Sugiyama.

If you find this code useful in your research, please cite

@inproceedings{xia2019t_revision,
  title={Are Anchor Points Really Indispensable in Label-Noise Learning?},
  author={Xia, Xiaobo and Liu, Tongliang and Wang, Nannan and Han, Bo and Gong, Chen and Niu, Gang and Sugiyama, Masashi},
  booktitle={NeurIPS},
  year={2019}
}

Dependencies

we implement our methods by PyTorch on NVIDIA Tesla V100. The environment is as bellow:

• Ubuntu 16.04 Desktop
• PyTorch, version >= 0.4.1
• CUDA, version >= 9.0
• Anaconda3

Install PyTorch and Torchvision (Conda):

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

Install PyTorch and Torchvision (Pip3):

pip3 install torch torchvision

Experiments

We verify the effectiveness of T_revision on three synthetic noisy datasets (MNIST, CIFAR-10, CIFAR-100), and one real-world noisy dataset (clothing1M). And We provide datasets (the images and labels have been processed to .npy format).
Here is an example:

python main.py --dataset cifar10 --noise_rate 0.5