• AutoPruner: An End-to-End Trainable Filter Pruning Method for Efficient Deep Model Inference, Pattern Recognition, In Press.
• [DOI] [Manuscript]

PyTorch environment:

• Python 3.6.9
• PyTorch 1.2.0 (MobileNetv2 codes)
• torchsummaryX

1. clone this repository.
2. download the ImageNet dataset and organize this dataset with train and val folders.
3. select subfolder (e.g., MobileNetv2):

   cd MobileNetv2
   

4. modify your configuration path and start pruning:
   • modify 1_pruning/main.py and run run_this.sh
   • if you want to get a smaller model, you should use a smaller compression_rate in 1_pruning/main.py
5. fine-tune the pruned model:

   cd ../2_fine_tune
   ./run_this.sh
   

6. test the accuray of pruned model:

   cd ../released_model
   run_this.sh
   

• MobileNetV2
  • MobileNetV2 is conducted using PyTorch 1.2.0.
  • the training log files and pretrained weights are also provided in corresponding folders.
• ResNet and VGG
  • unfortunately, the experiments of ResNet and VGG are conducted 2 years ago, I have forgotten the PyTorch version.
  • the training log files are also provided, but the pretrained weighs are missing.

We prune the MobileNetV2 model on ImageNet dataset:

The results on VGG16 and ResNet50 (PyTorch official models):

Note that:

• FLOPs = 2*MACs
• The latency is tested on a M40 GPU with 32 batch size.

If you find this work is useful for your research, please cite:

@article{AutoPruner,
  author = {Jian-Hao Luo, Jianxin Wu},
  title = {AutoPruner: An End-to-End Trainable Filter Pruning Method for Efficient Deep Model Inference},
  journal = {Pattern Recognition},
  year = {2020},
}

Feel free to contact me if you have any question (Jian-Hao Luo [email protected] or [email protected]).