DHA: End-to-End Joint Optimization of Data Augmentation Policy, Hyper-parameter and Architecture

Paper Link

We propose DHA, which achieves joint optimization of Data augmentation policy, Hyper-parameter and Architecture.

🌟 End-toE-End Traning: First to efficiently and jointly realize Data Augmentation, Hyper-parameter Optimization, and Neural Architecture Search in an end-to-end manner without retraining.

🌟 State-of-the-art: State-of-the-art accuracy on ImageNet with both cell-based and Mobilenet-like architecture search space

🌟 Findings: Demonstrate the advantages of doing joint-training over optimizing each AutoML component in sequence

💾 Dataset Preparation

[Cifar10] [Cifar100] [IMAGENET] [SPORTS8] [MIT67] [FLOWERS102]

💻 Usage

Training with flowers102 dataset

Training normal two-stage ISTA Model

python search_darts.py --conf_path conf/ista/flowers102_ista.json --Running ista_nor

Training normal single-stage ISTA Model

python search_darts.py --conf_path conf/ista/flowers102_ista_single.json --Running ista_single_nor

Training DHA Model

python search_darts.py --conf_path conf/ista/flowers102_ista_single.json --Running ista_single_doal

📄 Citation

If you find our work useful or interesting, please cite our paper:

@article{zhou2021dha,
  title={Dha: End-to-end joint optimization of data augmentation policy, hyper-parameter and architecture},
  author={Zhou, Kaichen and Hong, Lanqing and Hu, Shoukang and Zhou, Fengwei and Ru, Binxin and Feng, Jiashi and Li, Zhenguo},
  journal={arXiv preprint arXiv:2109.05765},
  year={2021}
}