MobileNets are based on streamlined architecture that uses depthwise separable convolutions to build light weight deep nueral network.
思想
- 基于depthwise separable convolution(来源于FactorizedNet)来实现
- depth-wise convolution:每一层分别作卷积
- point-wise convolution :1x1卷积,把各个层连接起来
注:左侧图片来自FactorizedNet,右侧来则mobileNet
作者讨论了,采用separable depthwise convolution与传统的卷积的计算量进行了比较:
Standard Convolution Filters:
DK∗Dk∗M∗N∗DG∗DG
Depthwise Concolution Filters:
DK∗DK∗M∗DG∗DG+M∗N∗DG∗