DivAug: Plug-in Automated Data Augmentation with Explicit Diversity Maximization

This is the official project repository for DivAug: Plug-in Automated Data Augmentation with Explicit Diversity Maximization [ICCV 2021].

Zirui Liu*, Haifeng Jin*, Ting-Hsiang Wang, Kaixiong Zhou, and Xia Hu.

TL; DR. DivAug is a unsupervised automated data augmentation method without requiring a separate search process.

Implementation

The ./cpp_extension directory contains the C++ implementation of the parallel KMEAN++ algorithm based on OpenMP. Our implementation of KMEAN++ is based on KMEAN module in scikit-learn. We also provide the vanilla python implementation of KMEAN++ in ./common/utils:line 48.

Install

• Requirements

torch>=1.8.0
torchvision>=0.9.0
warmup_scheduler (https://github.com/ildoonet/pytorch-gradual-warmup-lr)

• Build

cd cpp_extension
pip install -v -e .

Reproduce results

See ./scripts. For example, if you want to train wide-resnet-28-10 with DivAug on CIFAR-100, just run

bash scripts/wresnet_28x10divaug_cifar100.sh

Citation

@misc{liu2021divaug,
      title={DivAug: Plug-in Automated Data Augmentation with Explicit Diversity Maximization}, 
      author={Zirui Liu and Haifeng Jin and Ting-Hsiang Wang and Kaixiong Zhou and Xia Hu},
      year={2021},
      eprint={2103.14545},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

Our codebase is mainly based on

https://github.com/kakaobrain/fast-autoaugment

https://github.com/ildoonet/unsupervised-data-augmentation

https://github.com/SenWu/dauphin