Deep Compression on AlexNet

This is a demo of Deep Compression compressing AlexNet from 233MB to 8.9MB without loss of accuracy. It only differs from the paper that Huffman coding is not applied. Deep Compression's video from ICLR'16 best paper award presentation is available.

Related Papers

Learning both Weights and Connections for Efficient Neural Network (NIPS'15)

Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding (ICLR'16, best paper award)

EIE: Efficient Inference Engine on Compressed Deep Neural Network (ISCA'16)

If you find Deep Compression useful in your research, please consider citing the paper:

@inproceedings{han2015learning,
  title={Learning both Weights and Connections for Efficient Neural Network},
  author={Han, Song and Pool, Jeff and Tran, John and Dally, William},
  booktitle={Advances in Neural Information Processing Systems (NIPS)},
  pages={1135--1143},
  year={2015}
}


@article{han2015deep_compression,
  title={Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding},
  author={Han, Song and Mao, Huizi and Dally, William J},
  journal={International Conference on Learning Representations (ICLR)},
  year={2016}
}

A hardware accelerator working directly on the deep compressed model:

@article{han2016eie,
  title={EIE: Efficient Inference Engine on Compressed Deep Neural Network},
  author={Han, Song and Liu, Xingyu and Mao, Huizi and Pu, Jing and Pedram, Ardavan and Horowitz, Mark A and Dally, William J},
  journal={International Conference on Computer Architecture (ISCA)},
  year={2016}
}

Usage:

export CAFFE_ROOT=$your caffe root$

python decode.py bvlc_alexnet_deploy.prototxt AlexNet_compressed.net $CAFFE_ROOT/alexnet.caffemodel 

cd $CAFFE_ROOT

./build/tools/caffe test --model=models/bvlc_alexnet/train_val.prototxt --weights=alexnet.caffemodel --iterations=1000 --gpu 0

Test Result:

I1022 20:18:58.336736 13182 caffe.cpp:198] accuracy_top1 = 0.57074
I1022 20:18:58.336745 13182 caffe.cpp:198] accuracy_top5 = 0.80254