keras-mobilenet's Introduction

Note: This project is not maintained anymore. Mobilenet implementation is already included in Keras Applications folder. Mobilenet

Keras MobileNet

Google MobileNet Implementation using Keras Framework 2.0

Project Summary

This project is just the implementation of paper from scratch. I don't have the pretrained weights or GPU's to train :)
Separable Convolution is already implemented in both Keras and TF but, there is no BN support after Depthwise layers (Still investigating).
Custom Depthwise Layer is just implemented by changing the source code of Separable Convolution from Keras. Keras: Separable Convolution
There is probably a typo in Table 1 at the last "Conv dw" layer stride should be 1 according to input sizes.
Couldn't find any information about the usage of biases at layers (not used as default).

TODO

Add Custom Depthwise Convolution
Add BN + RELU layers
Check layer shapes
Test Custom Depthwise Convolution
Benchmark training and feedforward pass with both CPU and GPU
Compare with SqueezeNet

Library Versions

Keras v2.0+
Tensorflow 1.0+ (not supporting Theano for now)

References

Licence

MIT License

Note: If you find this project useful, please include reference link in your work.

keras-mobilenet's Issues

Is it faster than Keras built-in SeparableConv2D layer?

Depthwise separable convolutions are notoriously hard to implement efficiently. In particular, SeparableConv2D in Keras (as well as group convolutions in Keras and PyTorch) leads to almost no speedup on CPU, and makes training even slower on GPU, despite the number of parameters dropped by an order of magnitude.

Are you planning to add some sort of benchmarking here? I think fast group convolution and depthwise convolution are the features many people need really badly.