Title: Beyond Filters: Compact Feature Map for Portable Deep Model Authors: Yunhe Wang, Chang Xu, Chao Xu, Dacheng Tao Email: wangyunhe@pku.edu.cn
This code includes the MATLAB file of the compressed network (LeNet.mat) and the demo code (demo.m) for verifying it using MatConvnet.
You should install matconvnet first: http://www.vlfeat.org/matconvnet/, and run(fullfile('matconvnet','matlab','vl_setupnn.m'));
As a result, we obtained a 11.3x compression ratio and a 8.7x speed-up ratio.
Note that, the compressed network generated by exploiting the proposed method is still a regular CNN with 32-bit float values which does not have any decoding or other procedures for online inference.