README.md

Efficient Architecture Search by Network Transformation

Code for the paper Efficient Architecture Search by Network Transformation in AAAI 2018.

Reference

@inproceedings{cai2018efficient,
  title={Efficient Architecture Search by Network Transformation},
  author={Cai, Han and Chen, Tianyao and Zhang, Weinan and Yu, Yong and Wang, Jun},
  booktitle={AAAI},
  year={2018}
}

Related Projects

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

Dependencies

• Python 3.6
• Tensorflow 1.3.0

Top Nets

nets	test accuracy (%)	Dataset
C10+_Conv_Depth_20	95.77	C10+
C10+_DenseNet_Depth_76	96.56	C10+
C10_DenseNet_Depth_70	95.34	C10
SVHN_Conv_Depth_20	98.27	SVHN

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

$ python3 main.py --test --path=<nets path>

For example, by running

$ python3 main.py --test --path=../final_nets/C10+_Conv_Depth_20

you will get

Testing...
mean cross_entropy: 0.210500, mean accuracy: 0.957700
test performance: 0.9577

Acknowledgement

The DenseNet part of this code is based on the repository by Illarion. Many thanks to Illarion.