AGNAS: Attention-Guided Micro- and Macro-Architecture Search ()

This is an official pytorch implementation for (AGNAS: Attention-Guided Micro- and Macro-Architecture Search) by Zihao Sun, and et. al.

Requirements

• Python 3.7.10
• PyTorch >= 1.1.0

Dataset

CIFAR-10 can be automatically downloaded using this code by torchvision, and place them in the folder ./dataset/cifar10

ImageNet needs to be manually downloaded and following the instructions here.

Usage

AGNAS in Darts_Search_Space

To evaluate the AGNAS searched model on CIFAR-10

Pre-trained checkpoints are released in google drive. Download and place them in the ./Darts_Search_Space/eval_retrained_model, and run

cd Darts_Search_Space
python test.py

To search AGNAS model on CIFAR-10, run:

cd Darts_Search_Space
python train_search.py

To retrain AGNAS searched model on CIFAR-10, run:

cd Darts_Search_Space
python train.py

AGNAS in ProxylessNAS_Search_Space

To evaluate the AGNAS searched model on ImageNet

Pre-trained checkpoints are released in google drive. Download and place them in the ./ProxylessNAS_Search_Space/retrain_architecture/eval_retrained_model, and run

cd ProxylessNAS_Search_Space/retrain_architecture
python valid.py

To search AGNAS model on ImageNet, run:

cd ProxylessNAS_Search_Space/attention_search
bash run_train.sh

To retrain AGNAS model on ImageNet, run:

cd ProxylessNAS_Search_Space/retrain_architecture
bash run_retrain.sh

Citation

Please cite our paper if you find anything helpful.

@inproceedings{sun2022agnas,
  title={AGNAS: Attention-Guided Micro and Macro-Architecture Search},
  author={Sun, Zihao and Hu, Yu and Lu, Shun and Yang, Longxing and Mei, Jilin and Han, Yinhe and Li, Xiaowei},
  booktitle={International Conference on Machine Learning},
  pages={20777--20789},
  year={2022},
  organization={PMLR}
}