SSPQ

Spatial Shift Point-Wise Quantization code open

introduction

This repository is related with the paper Spatial Shift Point-Wise Quantization. The main idea is

• substitute the spatial operation to active shift operation
• apply inverted residual block structure as introduced by the paper all you need is a few shifts - https://arxiv.org/abs/1903.05285 .
• quantize 1x1 point-wise convolution part by using dorefa-net.
• keep the quantization bit for activation over 4-bit since quantization under 4-bit occur unstablility of convergence.

our experimental Results for SSPQ 50

This figure shows the SSPQ 50 light-weight model performance in the metric of accuracy and model size.

pre-requisite

Note that this code is tested only in the environment decribed below. Mismatched versions does not guarantee correct execution.

• Ubuntu kernel ver. 4.15.0-117-generic #118~16.04.1

• Cuda 10.0

• cudnn 7.6.5

• Tensorflow 1.15.3

• Tensorpack

• g++ 7.5.0

• python 3.7

• install https://github.com/Eunhui-Kim/custom-op-for-shiftNet

  After the test works fine, at first

• put the shift2d.py in the tensorpack models path. in my case, the tensorpack models is $HOME/.local/lib/python3.7/site-packages/tensorpack/models/

Testing

sspq50 model for imagenet dataset,

1. put 'imagenet-resnet-QresIBShift.py, resnet_model2.py, with imagenet_utils.py and dorefa.py of tensorpack file together in the tensorpack example path
2. run python imagenet-resnet-QresIBShift.py --mode QresIBShift --data $Imagenet_Path --batch 256 --depth 50