Awesome

ENC: Efficient Neural Network Compression

"Efficient Neural Network Compression", CVPR 2019, by Hyeji Kim, Muhammad Umar Karim Khan, Chong-Min Kyung.
[Paper] [Supp] [Poster]

Overall Framework

ENC is the fast network compression platform

• network decomposition based on truncated SVD
• determination of the number of filters for each layer (convolutional, fully-connected)
• reduction of convolution filters and fully-connected nodes

This repository contains

• ./compress : proposed network compression methods
• ./init : pre-processing scripts for a new network & network decomposition
• ./base_models : caffemodel & prototxt files
• ./caffe : modified caffe.cpp for Caffe framework

Download the Compressed Networks

• https://github.com/Hyeji-Kim/ENC-models

Requirements

• Caffe
• MATLAB
• google.protobuf: sudo pip install google sudo pip install protobuf
• python-tk: sudo apt-get install python-tk
• scipy: sudo apt-get install libblas-dev sudo apt-get install liblapack-dev sudo pip install scipy
• skimage: sudo apt-get install python-skimage
• realpath: sudo apt-get realpath

Quick Start (ex. ResNet-56)

1. copy caffe/caffe.cpp to ${your caffe home}/tools/. and rebuild :

cp caffe/caffe.cpp ${your caffe home}/tools/.
make -j20 all
make -j20 pycaffe

2. modify the dataset path in below prototxt files :

./base_model/res56/train_val_test.prototxt
./base_model/res56/train_val_val.prototxt

./init/decomp_val/res56/models_res56/train_val.prototxt
./init/decomp_val/res56/comp_res56/train_val/train_val_conv_rmax.prototxt

./init/decomp_test/res56/models_res56/train_val.prototxt
./init/decomp_test/res56/comp_res56/train_val/train_val_conv_rmax.prototxt

3. modify the symbolic link of caffe to ${your caffe home}/:

./init/decomp_val/caffe
./init/decomp_test/caffe
./compress/ENC-Inf/res56/caffe
./compress/ENC-Map/res56/caffe

NOTE: symbolic link of caffe directory should be the absolute path. (ex. /home/hjkim/ENC/caffe/)

4. set the constraints in RUN.sh.

cd ./compress/ENC-Map/res56
vi RUN.sh # set variables:

TAR_COMP="{target complexity}" # which is (1-compression rate), range: 0.0 ~ 1.0, ex) 0.5
GPU_IDX="{avaiable gpu indice}" # ex) GPU_IDX="0" or GPU_IDX="0,2,3"

5. run the script :

./RUN.sh

For VGG-16 Compression

Download [orig-vgg16] and [decomp-vgg16], then :

mv vgg16.caffemodel ./base_models/vgg16/vgg16.caffemodel # original vgg16
mv vgg16_svd_1.caffemodel ./init/decomp_init/vgg16/comp_vgg16/model/vgg16_svd_1.caffemodel # decomposed vgg16

Citation

@CONFERENCE{ENC_CVPR19,
  author={Hyeji Kim, Muhammad Umar Karim Khan, Chong-Min Kyung},
  title={Efficient Neural Network Compression},
  booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2019},
}

Reference

• Network decomposition : "https://github.com/chengtaipu/lowrankcnn"
• ResNet-56 for Cifar10 : "https://github.com/yihui-he/resnet-cifar10-caffe"
• VBMF : "https://sites.google.com/site/shinnkj23/downloads"

Contact

hyejikim7833@gmail.com, Feel free to contact me. :-)