Index page

Mobile Net

[5] Mobile Net MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications (2017), Andrew G. Howard et al. Paper

Notes

• All layers except first(normal one for that) :nxm Depthwise Conv->BN->ReLU->1x1 Conv->BN->ReLU
• Depthwise separable filters : Add group = no of input channels in Conv2d
• Splits it into layers
  • filtering layer
  • combining layer
• use groups when going from lower channels to higher
• Depthwise convolutions + pointwise convolutions
• use 1x1 convs
• Features:
• Smaller net
• More Efficient convs
• Model size
• AvgPool at the end -RMSProp
• Very little or no weight decay for Depthwise filters

  In effect-> Width, resolution multiplier