OnlineAugment (Accepted at ECCV 2020)

Official OnlineAugment implementation in PyTorch

• More automatic than AutoAugment and related
  • Towards fully automatic (STN and VAE, No need to specify the image primitives).
  • Broad domains (natural, medical images, etc).
  • Diverse tasks (classification, segmentation, etc).
• Easy to use
  • One-stage training (user-friendly).
  • Simple code (single GPU training, no need for parallel optimization).
• Orthogonal to AutoAugment and related
  • Online v.s. Offline (Joint optimization, no expensive offline policy searching).
  • State-of-the-art performance (in combination with AutoAugment).

(In this implementation, we disable the meta-gradient for efficient training. The code is also refactored accordingly, achieving comparable performance. Especially for reduced CIFARs, we observe higher accuracy than reported in the paper.)

Visualization on CIFAR-10

A-STN

D-VAE

P-VAE

Run

We conducted experiments in

• python 3.7
• pytorch 1.2, torchvision 0.4.0, cuda10

The searching of policies and training of target model is optimized jointly.

For example, training wide-resnet using STN on reduced CIFAR-10, using the script in r-cifar10-wrn-scripts

./run-aug-stn.sh

Citation

If this code is helpful for your research, please cite:

@article{tang2020onlineaugment,
  title={OnlineAugment: Online Data Augmentation with Less Domain Knowledge},
  author={Tang, Zhiqiang and Gao, Yunhe and Karlinsky, Leonid and Sattigeri, Prasanna and Feris, Rogerio and Metaxas, Dimitris},
  journal={arXiv preprint arXiv:2007.09271},
  year={2020}
}