caffe-quantization

Added quantization layer into caffe (support a coarse level fixed point simulation). The working flow is to insert the newly-built QuantizationLayer between computation data path. It will round the output of the previous layer using the specified round method and strategy under a pre-determined bit width.

Installation

Only 3 files needed to be added to the original Caffe framework (https://github.com/BVLC/caffe).

• quantization_layer.hpp: header file for QuantizationLayer, add to $CAFFE_ROOT/include/caffe/layers
• QuantizationLayer.cpp: source file for QuantizationLayer, add to $CAFFE_ROOT/src/caffe/layers
• caffe.proto: Added id = 147 QuantizationParameter for QuantizationLayer, replace the original one at $CAFFE_ROOT/src/caffe/proto

Re-Compile the Caffe framework and PyCaffe/MatCaffe if you use the Python/Matlab Interface.

Usage

Three IPython Notebooks are provided to demonstrate how to use PyCaffe for manipulating the QuantizationLayer.