Paper Reading:MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications(arXiv:1704.04861v1)
Motivation
提出来用在手机或者边缘视觉应用上的轻量化模型,可以有效的做精度和延迟之间的trade off。
Architecture
-
Depthwise Separable Convolution
使用DC和PC代替标准卷积
input :
output:
kernel K of size:
-
standard convolution cost
-
Depthwise convolution cost
-
Depthwise Separable Convolution = Depthwise convolution + Pointwise convolution
最终计算量减少的比例
-
Network Structure and Training
Depthwise Separable Convolution模块的构建方法:
各个子模块计算量的百分比: -
Width Multiplier: Thinner Models Although
width multiplier α, the number of input channels M becomes αM and the number of output channels N becomes αN,使用随机采样输入的通道实现?是属于[0-1]的。
计算量减为:
-
Resolution Multiplier: Reduced Representation
将网络输入的分辨率使用缩放因子改小,resize?
计算量减为:
$DK ·DK · αM· ρDF · ρDF +αM· αN · ρDF · ρDF $
Experiments
- classification
- detection