mldl / propmix

Code from PropMix, accepted at BMVC'21

PropMix: Hard Sample Filtering and Proportional MixUp for Learning with Noisy Labels

This repository is the official implementation of Hard Sample Filtering and Proportional MixUp for Learning with Noisy Labels (BMVC 2021).

Authors: Filipe R. Cordeiro; Vasileios Belagiannis, Ian Reid and Gustavo Carneiro

Illustration

Requirements

• This codebase is written for python3.
• To install necessary python packages, run pip install -r requirements.txt.

Training and Evaluating

The pipeline for training with PropMix is the following:

1. Self-supervised pretrain.

In our paper we use SimCLR for most of the datasets. We use moco-v2 to pre-train an InceptionResNetV2 on Webvision. Other self-supervised methods can be used as well.

If you use SimCLR, run:

python simclr.py --config_env configs/env.yml --config_exp configs/pretext/<config_file.yml> --cudaid 0

2. Clustering

python scan.py --config_env configs/env.yml --config_exp configs/scan/<config_file.yml> --cudaid 0

3. Train the model (using the pretraining from steps 1 and 2)

For CIFAR-10/CIFAR-100:

python propmix.py --r [0.2/0.5/0.8/0.9] --noise_mode [sym/asym] --config_env configs/env.yml --config_exp configs/propmix/<config_file.yml> --cudaid 0

Add --nopt if you wish to train from scratch, without the self-supervised pretrain, from steps 1 and 2. Add --strong_aug to use strong augmentation. Recommended for high noise rates.

License and Contributing

• This README is formatted based on paperswithcode.
• This project is licensed under the terms of the MIT license.
• Feel free to post issues via Github.

Cite PropMix
If you find the code useful in your research, please consider citing our paper:

@article{cordeiroPropMix2021,
  title={PropMix: Hard Sample Filtering and Proportional MixUp for Learning with Noisy Labels},
  author={Cordeiro, F. R. and Belagiannis, Vasileios and Reid, Ian and Carneiro, Gustavo},
  journal={The 32nd British Machine Vision Conference},
  volume={?},
  year={2021}
}

Contact

Please contact [email protected] if you have any question on the codes.