[ICCV 2023] MADAug: When to Learn What: Model-Adaptive Data Augmentation Curriculum

Introduction

Model-Adaptive Data Augmentation (MADAug) that jointly trains an augmentation policy network to teach the model “when to learn what”. In this paper, we study two fundamental problems towards developing a data-and-model-adaptive data augmentation policy that determines a curriculum of “when to learn what” to train a model: (1) when to apply data augmentation in training? (2) what data augmentations should be applied to each training sample at different training stages?

Getting Started

Code supports Python 3.

Install requirements

pip install -r requirements.txt

Run data augmentation

Script to search for the dynamic augmentation policy and train task model for is located in main.sh. Pass the dataset name as the argument to call the script.

For example, to use the dynamic augmentation policy for classifying the reduced_cifar10 dataset:

bash main.sh reduced_cifar10

References & Opensources

Part of our implementation is adopted from the Fast AutoAugment and DADA repositories.

• Fast AutoAugment (https://github.com/kakaobrain/fast-autoaugment)
• DADA (https://github.com/VDIGPKU/DADA)
• AdaAug (https://github.com/jamestszhim/adaptive_augment)

Citation

If you find MADAug helpful in your research, please consider citing:

@article{hou2023learn,
  title={When to Learn What: Model-Adaptive Data Augmentation Curriculum},
  author={Hou, Chengkai and Zhang, Jieyu and Zhou, Tianyi},
  journal={arXiv preprint arXiv:2309.04747},
  year={2023}
}