SampleCaffeModelCompression

This is a really simple code for compressing caffemodel (no algorithm and technique). This code only copy array from pre-trained mode to new model (not all layer).

Prerequisite:

1. you have caffe installed in your pc.

How to do compression in terminal ?

Quick demo:

bash run.sh

or:

1. export CAFFE_ROOT = your_caffe_path

2. python sample_compress.py --before_prototxt $CAFFE_ROOT/models/bvlc_alexnet/deploy.prototxt --after_prototxt $CAFFE_ROOT/models/bvlc_alexnet/deploy_compress.prototxt --net $CAFFE_ROOT/models/bvlc_alexnet/bvlc_alexnet_beforecaffemodel --output_path $CAFFE_ROOT/models/bvlc_alexnet/bvlc_alexnet_compressed.caffemodel

To test the model:

examples$:python 00-classification.py

Some example results:

The testing picture:

Original bvlc_reference_caffenet model:
model_def = caffe_root + 'models/bvlc_reference_caffenet/deploy.prototxt'
model_weights = caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'

mean-subtracted values: [('B', 104.0069879317889), ('G', 116.66876761696767), ('R', 122.6789143406786)] predicted class is: 281 output label: n02123045 tabby, tabby cat probabilities and labels:

Original bvlc_alexnet model:
model_def = caffe_root + 'models/bvlc_alexnet/deploy.prototxt'
model_weights = caffe_root + 'models/bvlc_alexnet/bvlc_alexnet.caffemodel'

mean-subtracted values: [('B', 104.0069879317889), ('G', 116.66876761696767), ('R', 122.6789143406786)] predicted class is: 285 output label: n02124075 Egyptian cat probabilities and labels:

Compressed bvlc_alexnet model:
model_def = caffe_root + 'models/bvlc_alexnet/deploy_compress.prototxt'
model_weights = caffe_root + 'models/bvlc_alexnet/bvlc_alexnet_compressed.caffemodel'

mean-subtracted values: [('B', 104.0069879317889), ('G', 116.66876761696767), ('R', 122.6789143406786)] predicted class is: 254 output label: n02110958 pug, pug-dog probabilities and labels:

Size comparison:

cpchung:examples$ ls ../models/bvlc_alexnet/ -lrth

\total 291M
-rw-rw-r-- 1 cpchung cpchung 233M Aug 22 2014 bvlc_alexnet.caffemodel
-rw-rw-r-- 1 cpchung cpchung 3.6K Feb 22 22:18 deploy.prototxt
-rw-rw-r-- 1 cpchung cpchung 3.9K Feb 22 22:18 deploy_compress.prototxt
-rw-rw-r-- 1 cpchung cpchung 59M Feb 22 23:05 bvlc_alexnet_compressed.caffemodel

Thank you very much for all references below:

1. https://github.com/BVLC/caffe
2. https://github.com/yuanyuanli85/CaffeModelCompression
3. https://github.com/songhan/Deep-Compression-AlexNet
4. https://github.com/rbgirshick/fast-rcnn