README.md

ADAM-ADMM

A Unified, Systematic Framework of Structured Weight Pruning for DNNs

Content

caffe-admm/ Customized Caffe for non-structured pruning (structured pruning version will be released soon)

prototxt/ Sample prototxt for non-structured pruning

Results

1. CaffeNet non-structured pruing

compression ratio	top-1	top-5	download link
conv2-5 16.1x	57.50%	80.28%	http://bit.ly/2NUgpn4
conv2-5 40.5x	55.38%	78.64%	http://bit.ly/2vm6jnl

2. CaffeNet structured pruing

compression ratio	top-1	top-5	download link
conv2-5 1.5x	59.04%	81.65%	http://bit.ly/2vlgWXD
conv2-5 4.8x	57.47%	80.15%	http://bit.ly/2NYF1Lv
conv2-5 13.2x	55.33%	78.82%	http://bit.ly/2NYqzmB

Our paper

"ADAM-ADMM: A Unified, Systematic Framework of Structured Weight Pruning for DNNs" (https://arxiv.org/abs/1807.11091)

Citation

If you use these models or code in your research, please cite:

@article{zhang2018admm,
  title={ADAM-ADMM: A Unified, Systematic Framework of Structured Weight Pruning for DNNs},
  author={Zhang, Tianyun and Zhang, Kaiqi and Ye, Shaokai and Li, Jiayu and Tang, Jian and Wen, Wujie and Fardad, Makan and Wang, Yanzhi},
  journal={arXiv preprint arXiv:1807.11091},
  year={2018}
}