jsrimr / pathlevel-eas

Path-Level Network Transformation for Efficient Architecture Search, in ICML 2018.

Path-Level Network Transformation for Efficient Architecture Search

Code for the paper Path-Level Network Transformation for Efficient Architecture Search in ICML 2018.

Reference

@article{cai2018path,
  title={Path-Level Network Transformation for Efficient Architecture Search},
  author={Cai, Han and Yang, Jiacheng and Zhang, Weinan and Han, Song and Yu, Yong},
  journal={arXiv preprint arXiv:1806.02639},
  year={2018}
}

Related Projects

• Efficient Architecture Search by Network Transformation in AAAI 2018, Code.

Dependencies

• Python 3.6
• Pytorch 0.3.1

Results

CIFAR-10

Model	Params	Test error (%)
TreeCell-A with DenseNet (N=16, k=48, G=2)	13.1M	3.35
TreeCell-A with PyramidNet (N=18, alpha=84, G=2)	5.7M	3.14
TreeCell-A with PyramidNet (N=18, alpha=84, G=2) + DropPath (600 epochs)	5.7M	2.99
TreeCell-A with PyramidNet (N=18, alpha=84, G=2) + DropPath + Cutout (600 epochs)	5.7M	2.49
TreeCell-A with PyramidNet (N=18, alpha=150, G=2) + DropPath + Cutout (600 epochs)	14.3M	2.30

For checking these networks, please download the corresponding model files and run the following command under the folder of code/CIFAR:

$ python3 run_exp.py --path <nets path>

For example, by running

$ python3 run_exp.py --path ../../Nets/CIFAR10#PyramidTreeCellA#N=18_alpha=150#600#cutout

you will get

test_loss: 0.092100	 test_acc: 97.700000

ImageNet

Model	Multi-Add	Top-1 error (%)
TreeCell-B with CondenseNet (G1=4, G3=8)	594M	25.4

Please refer to the file code/ImageNet/scripts.sh.

Architecture Search

For setting up your environment to run architecture search experiments, please refer to my previous repository.