Go to index

Mobile Net

Reading time : ~1 min

by Subhaditya Mukherjee

• Notes

Paper notes for the paper

[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

Related posts:  FP16  AI Superpowers Kai Fu Lee  Digital Minimalism Cal Newport  More Deep Learning, Less Crying - A guide  Super resolution  Federated Learning  Taking Batchnorm For Granted  A murder mystery and Adversarial attack  Thank you and a rain check  Pruning