Skip to content

Latest commit

 

History

History
79 lines (61 loc) · 2.29 KB

README.md

File metadata and controls

79 lines (61 loc) · 2.29 KB

DDSAS: Dynamic and Differentiable Space-Architecture Search (ACML 2021)

license python pytorch

This is an official pytorch implementation for "DDSAS: Dynamic and Differentiable Space-Architecture Search". DDSAS-figure1 DDSAS-figure2

Requirements

  • Python 3.6.8
  • PyTorch 1.4.0

Usage

Data preparation

Download CIFAR10/CIFAR100/ImageNet dataset and place them in .data/ folder.

To obtain the dataset for ImageNet search, run:

python imagenet_split.py

Model Evaluation

Pre-trained checkpoints are released google drive/baiduyun. Place them in the .weights/ folder.

Note: access code for baiduyun is ddsa.

To evaluate a pre-trained DDSAS model on CIFAR10/CIFAR100/ImageNet, run:

bash shell/eval_cifar10.sh
bash shell/eval_cifar100.sh
bash shell/eval_imagenet.sh

Model Search

To search a DDSAS model on CIFAR10/CIFAR100/ImageNet, run:

bash shell/search_cifar10.sh
bash shell/search_cifar100.sh
bash shell/search_imagenet.sh

To search a DDSAS model in a shrinking/expanding search space on CIFAR10/CIFAR100, run:

bash shell/search_cifar10_shrink.sh
bash shell/search_cifar10_expand.sh
bash shell/search_cifar100_shrink.sh
bash shell/search_cifar100_expand.sh

Model Retraining

To retrain a DDSAS on CIFAR10/CIFAR100/ImageNet, run:

bash shell/retrain_cifar10.sh
bash shell/retrain_cifar100.sh
bash shell/retrain_imagenet.sh

Citation

Please cite our paper if you find anything helpful.

@InProceedings{yang21,
      title = {DDSAS: Dynamic and Differentiable Space-Architecture Search},
      author = {Yang, Longxing and Hu, Yu and Lu, Shun and Sun, Zihao and Mei, Jilin and Zeng, Yiming and Shi, Zhiping and Han, Yinhe and Li, Xiaowei},
      booktitle={ACML},
      year={2021}
    }

License

MIT License

Acknowledgement

This code is heavily borrowed from DARTS and SGAS. Great thanks to their contributions.