CNN网络量化 - Quantized Convolutional Neural Networks for Mobile Devices

Quantized Convolutional Neural Networks for Mobile Devices

CVPR2016

GitHub code: https://github.com/jiaxiang-wu/quantized-cnn

本文主要是通过对CNN网络的量化，达到压缩模型大小及加快速度的目的，牺牲的准确率比较小。

CNN网络在 test phase ， 运算时间主要消耗在卷积层，CNN网络的参数主要集中在全连接层。主要在卷积层需要加速，在全连接层压缩参数空间。这里我们采用了 Product quantization 技术。

3 Quantized CNN
3.1. Quantizing the Fully-connected Layer

3.2. Quantizing the Convolutional Layer
Similar to the fully-connectedlayer, we pre-compute the look-up tables of inner products with the input feature maps

3.3. Quantization with Error Correction
对每层量化进行误差纠正，避免累计误差太大。

3.4. Computation Complexity

the reduction in the computation and storage overhead largely depends on two hyper-parameters, M (number of subspaces) and K (number of sub-codewords in each subspace)

5.1. Results on MNIST

5.2. Results on ILSVRC-12

5.3. Results on Mobile Devices

5.4. Theoretical vs. Realistic Speed-up