SimplePruning

This repository provides a cnn channels pruning demo with tensorflow. You can pruning your own model(support conv2d,depthwise conv2d,pool,fc,concat, add ops and so on) defined in modelsets.py. Have a good time!

•  Author: Haibo Wang
  
•  Email: dasuda2015@163.com
•  Home Page: dasuda.top

Dependencies

• Tensorflow >= 1.10.0
• python >= 3.5
• opencv-python >= 4.1.0
• numpy >= 1.14.5
• matplotlib >= 3.0.3

Getting Started

• Clone the repository

  $ git clone https://github.com/DasudaRunner/SimplePruning.git

• Downdload the Cifar10 dataset, and put into cifar-10-python/

  Url: http://www.cs.toronto.edu/~kriz/cifar.html

• (Optional) Define your model in modesets.py

  You must use add_layer() API defined in pruner.py to set up your model. More details to modelsets.py

• (Optional) Config params in utils/config.py

  e.g. model name, learning rate, pruning rate.

• Train a full model, .ckpt and .pb model file will be saved in ckpt_model/

  $ python full_train.py

• Channel pruning. .ckpt and .pb model file will be saved in channels_pruned_model/

  $ python channels_pruning.py

Supported ops (Tensorflow)

• Conv2d
• FullyConnected
• MaxPooling, AveragePooling
• BatchNormalization
• Activation
• DepthwiseConv2d
• GlobalMaxPooling, GlobalAveragePooling
• Concat
• Add
• Flatten

Evaluation on Cifar10 dataset

Update logs

• 2019.07.24
  • Add support for Add op.
  • Add support for ResNet18/ResNet34 in modelsets.py.
• 2019.07.16
  • Add support for Concat op.
  • Add support for DenseNet40 in modelsets.py.
• 2019.07.14
  • Reconsitution SimplePruning.