youzhonghui / gate-decorator-pruning

Code for the NuerIPS'19 paper "Gate Decorator: Global Filter Pruning Method for Accelerating Deep Convolutional Neural Networks"

Gate Decorator (NeurIPS 2019)

This repo contains required scripts to reproduce results from paper:

Gate Decorator: Global Filter Pruning Method for Accelerating Deep Convolutional Neural Networks

---

Requirements

python 3.6+ and PyTorch 1.0+

Installation

1. clone the code
2. pip install --upgrade git+https://github.com/youzhonghui/pytorch-OpCounter.git
3. pip install tqdm

How to use

(1). Notebook (ResNet-56)

In the run/resnet-56 folder, we provide an example which reduces the FLOPs of resnet-56 by 70%, but still maintains 93.15% accuracy on CIFAR-10:

1. The run/resnet-56/resnet56_prune.ipynb prunes the network with Tick-Tock framework.
2. The run/resnet-56/finetune.ipynb shows how to finetune the pruned network to get better results.

If you want to run the demo code, you may need to install jupyter notebook

(2). Command line (VGG-16)

In the run/vgg16 folder, we provide an example executed by command line, which reduces the FLOPs of VGG-16 by 90% (98% parameters), and keep 92.07% accuracy on CIFAR-10.

The instructions can be found here

(3). Save and load the pruned model

In the run/load_pruned_model/ folder, we provide an example shows how to save and load a pruned model (VGG-16 with only 0.3M float parameters).

Todo

• Basic running example.
• PyTorch 1.2 compatibility test.
• The command-line execution demo.
• Save and load the pruned model.
• ResNet-50 pruned model.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{zhonghui2019gate,
  title={Gate Decorator: Global Filter Pruning Method for Accelerating Deep Convolutional Neural Networks},
  author={Zhonghui You and
          Kun Yan and
          Jinmian Ye and
          Meng Ma and
          Ping Wang},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2019}
}