二值网络--TBN: Convolutional Neural Network with Ternary Inputs and Binary Weights

TBN: Convolutional Neural Network with Ternary Inputs and Binary Weights
ECCV2018

本文的思路就是： 对 weight 进行二值量化+ scaling factor α ， 对于 网络层的输入 进行 ternary value {−1,0,1} without the scaling factor

3 Ternary-Binary Networks
3.1 Convolution with Matrix Multiplication
use matrix multiplication to implement the convolution layer ⟨I,W,∗⟩
ten2mat 将一组滤波器 weight filters W can be reshaped to a matrix
ten2mat reshape the matrix back to output tensor C

3.2 Binary Weights

3.3 Ternary Inputs
the scaling factors of the binary inputs does not affect the performance of XNOR-Networks
这里我们不采用 scaling factor 了



3.4 Ternary-Binary Dot Product
这样卷积就没有乘法运算了，只有 位运算和 加减

量化梯度求导采用：

provides ∼ 32× memory savings and 40× faster convolutional operations

11