SLQ

Caffe implementation of paper "Deep Neural Network Compression with Single and Multiple Level Quantization"

How to use

make the caffe as usual

Download the pretrained AlexNet caffemodel.
In models/bvlc_alexnet fintune alexnet you can have the first quantization iteration of SLQ.

The key code is in the src/caffe/blob.cpp and include/caffe/kmeans.hpp. To have the other iteration you can change the code in kmeans.hpp and finetune the caffemodel of the first iteration.

Details

Three marks "here1","here2","here3"means three palces need to change in the second iteration.

here1

Here determines how many kernels need to update by the means of the weights, we do not update the quantized kernels: During the 5-bit quantization(17 kernels) of Alexnet, 5 iterations, the changes are i=1;i=5;i=9;i=13;i=15 respectively:

  for(int i=1; i<num; i++){//here1
  if(kernel_num[i]>0) kernel[i] = kernel_new[i] / kernel_num[i];}
}

here2

Here is to read the kernel of the last quantization iteration, only the first iteration need to comment this section.

/*std::ifstream fi("/home/xuyuhui/k_0.txt");
        std::ofstream fo;

here3

Here is to partition the clusters and quantize and fix the weights in the chosen clusters.During the 5-bit quantization(17 kernels) of Alexnet, 5 iterations, the changes are T_vec[i]<5;T_vec[i]<9;T_vec[i]<13;T_vec[i]<15;T_vec[i]<17 respectively:

  for(int i=0; i<count; i++){
    if(T_vec[i]<5){//here 3
      data_vec[i] = kernel[int(T_vec[i])];
      dis_vec[i] = 0;
    }
    else dis_vec[i] = 1;
  }

Acceleration

The code of Low-rank decomposition method Trained Rank Pruning is also available.

Citation

@article{xu2018deep,
  title={Deep Neural Network Compression with Single and Multiple Level Quantization},
  author={Xu, Yuhui and Wang, Yongzhuang and Zhou, Aojun and Lin, Weiyao and Xiong, Hongkai},
  journal={arXiv preprint arXiv:1803.03289},
  year={2018}
}

yuhuixu1993/WLQ

SLQ