This repo contains required scripts to reproduce results from paper:

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

python 3.6+ and PyTorch 1.0+

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

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

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

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).

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

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}
}