Skip to content

Latest commit

 

History

History

MobileNet

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
V1
V1
 
 
V2
V2
 
 
imgs
imgs
 
 
README.md
README.md
 
 

README.md

MobileNets Implementation

Implementing MobileNets in pytorch.
Welcome any advice with widely open arms.



Todo

  • MobileNet version 1
  • MobileNet version 2
  • MobileNet version 3
  • Training MobileNets on ImageNet Dataset...



Version 1

  • MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
  • Authors
    • [Andrew G. Howard | Menglong Zhu | Bo Chen | Dmitry Kalenichenko | Weijun Wang | Tobias Weyand | Marco Andreetto | Hartwig Adam]

  • [Paper] | [Code]
  • Introduce a Depthwise convolution which is consist of two layers, depthwise and pointwise convolutions. It had similar performance to normal(original?) convolution, however it does have lower computation cost.





Version 2

  • MobileNetV2: Inverted Residuals and Linear Bottlenecks
  • Authors
    • [Mark Sandler | Andrew Howard | Menglong Zhu | Andrey Zhmoginov | Liang-Chieh Chen]

  • [Paper] | [Code]
  • Introduce a Inverted Residual and Linear Bottlenecks Relu is capable of preserving complete information about the input manifold, but only if the input manifold lies in a low-dimensional subspace. So, in MobileNet version 2, we are gonna use Inverted Residual block, which using relu in a lower dimension and expanding it into a higher dimension following a linear transformation. Using linear layers is crucial as it prevents non-linearities from destroying too much information.