DU-DARTS: Decreasing the Uncertainty of Differentiable Architecture Search (BMVC 2021)

This is an official pytorch implementation for "DU-DARTS: Decreasing the Uncertainty of Differentiable Architecture Search ".

Requirements

• Python 3.7.10
• PyTorch 1.7.1

Usage

Model Evaluation

Pre-trained checkpoints and experimental logs are released here.

To evaluate a pre-trained DU-DARTS model on CIFAR-10, run:

python evaluate_cifar.py --arch du_darts_c10_s0 --dataset cifar10 --pretrained_ckpt ./pretrained/c10_s0_best_weights.pt

To evaluate a pre-trained DU-DARTS model on CIFAR-100, run:

python evaluate_cifar.py --arch du_darts_c100_s0 --dataset cifar100 --pretrained_ckpt ./pretrained/c100_s0_best_weights.pt

To evaluate a pre-trained DU-DARTS model on ImageNet, run:

python evaluate_imagenet.py --arch du_darts_c100_s0 --pretrained_ckpt ./pretrained/c10_imagenet_transfer_best_model.pth

Model Search

To search a DU-DARTS model on CIFAR-10, run:

python train_search.py  --arch_name du_darts_c10_s0 --layers 8 --loss_type entropy --dataset cifar10

To search a DU-DARTS model on CIFAR-100, run:

python train_search.py  --arch_name du_darts_c100_s0 --layers 5 --loss_type entropy --dataset cifar100

Model Retraining

To retrain a DU-DARTS model on CIFAR-10, run:

python retrain_cifar.py  --arch du_darts_c10_s0 --dataset cifar10 --auxiliary --cutout

To retrain a DU-DARTS model on CIFAR-100, run:

python retrain_cifar.py  --arch du_darts_c100_s0 --dataset cifar100 --auxiliary --cutout

To retrain a DU-DARTS model on ImageNet, run:

python -u retrain_imagenet.py --arch du_darts_c100_s0 --auxiliary

Citation

Please cite our paper if you find anything helpful.

@inproceedings{lu2021dudarts,
        title={DU-DARTS: Decreasing the Uncertainty of Differentiable Architecture Search},
        author={Lu, Shun and Hu, Yu and Yang, Longxing and Sun, Zihao and Mei, Jilin and Zeng Yiming and Li, Xiaowei },
        booktitle={BMVC},
        year={2021}
}

License

MIT License

Acknowledgement

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