0
点赞
收藏
分享

微信扫一扫

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks


DeepShift: Towards Multiplication-Less Neural Networks
PDF: ​​​https://arxiv.org/pdf/1905.13298.pdf​​​ PyTorch代码: ​​https://github.com/shanglianlm0525/PyTorch-Networks​​

DeepShift 用移位和求反运算代替乘法,可有效缓解计算成本过高的问题.

1 DeepShift Networks

1-1 bit-wise shift (按位移位)

如果输入数字的基本二进制表示形式A为整数或固定点格式,则向左(或右)的逐位移位在数学上等效于将其乘以2的正s(或负s次幂):

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_全连接

1-2 negation operation (求反运算)

求反运算在数学公式 :

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_卷积_02

1-3 DeepShift原理示意图

用按位移位和位取反来代替乘法示意图

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_数据集_03

2 LinearShift Operator (全连接移位)

基于矩阵运算的全连接算子的前向传播 (forward pass)

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_数据集_04


全连接算子的反向传播(backward pass)

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_深度学习_05


全连接移位算子(shift linear operator)

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_深度学习_06

3 ConvShift Operator (卷积移位)

基于矩阵运算的卷积算子的前向传播 (forward pass)

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_卷积_07


卷积算子的反向传播(backward pass)

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_全连接_08


卷积移位算子(shift convolution operator)

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_数据集_09

4 实验对比

4-1 MNIST数据集对比测试结果

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_深度学习_10

4-2 CIFAR10数据集对比测试结果

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_全连接_11

4-3 ImageNet数据集对比测试结果

深度学习论文: DeepShift: Towards Multiplication-Less Neural Networks_卷积_12


举报

相关推荐

0 条评论