We propose a new graph-based framework, namely Noisy Graph Clean- ing (NGC), which collects clean samples by leveraging ge- ometric structure of data and model predictive confidence.
Hyper parameters can be found in configs folder.
We provide all necessary datasets to reproduce results declared in paper. Please refer Baidu(code:a7d4).
- Install requirements by
cd papers/ICCV2021-NGC
pip install -r requirements.txt
- Build c++ library
cd impls
python setup.py build_ext -i
cd ..
- Run
python run.py --config configs/cifar10_sym.py
- Change random seed
python run.py --config config/cifar10_sym.py --work_dir work_cifar10_sym --seed 1001
If you find NGC useful for your research, please consider citing the paper as follows:
@article{2021NGC,
title={NGC: A Unified Framework for Learning with Open-World Noisy Data},
author={ Wu, Zhi Fan and Wei, Tong and Jiang, Jianwen and Mao, Chaojie and Tang, Mingqian and Li, Yu Feng },
year={2021},
}