README.md

Meta Label Correction for Noisy Label Learning

This repository contains the source code for the AAAI paper "Meta Label Correction for Noisy Label Learning".

Data

The code will download automatically the CIFAR data set; for Clothing1M, please contact the original creator for access. Put the obtained Clothing1M data set under directory data/clothing1M. Then execute cd CLOTHING1M; python3 load_cloth1m_data.py to generate necessary folders for training.

Example runs

On CIFAR-10 run MLC with UNIF noise and a noise level of 0.4 by executing

python3 main.py --dataset cifar10 --optimizer sgd --bs 100 --corruption_type unif --corruption_level 0.4 --gold_fraction 0.02 --epochs 120 --main_lr 0.1 --meta_lr 3e-4 --runid cifar10_run  --cls_dim 128

On CIFAR-100, run MLC with FLIP noise and a noise level of 0.6 by executing

python3 main.py --dataset cifar100 --optimizer sgd --bs 100 --corruption_type flip --corruption_level 0.6 --gold_fraction 0.02 --epochs 120 --main_lr 0.1 --meta_lr 3e-4 --runid cifar100_run  --cls_dim 128

On Clothing1M, run MLC as

python3 main.py --dataset clothing1m --optimizer sgd --bs 32 --corruption_type unif --corruption_level 0.1 --gold_fraction 0.1 --epochs 15 --main_lr 0.1 --meta_lr 0.003 --runid clothing1m_run --cls_dim 128 --skip --gradient_steps 5

(Note that for clothing1m, corruption_type, corruption_level, and gold_fraction have no effect as the original dataset comes with actual noisy labels and clean/noisy data splits.)

Refer to python3 main.py --help for a detailed explanations of all applicable arguments.

Citation

If you find MLC useful, please cite the following paper

@inproceedings{zheng2021mlc,
  title={Meta Label Correction for Noisy Label Learning},
  author={Zheng, Guoqing and Awadallah, Ahmed Hassan and Dumais, Susan},  
  journal={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
  volume={35},
  year={2021},
}

For any questions, please submit an issue or contact zheng@microsoft.com.

This repository is released under MIT License. (See LICENSE)