• ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression, ICCV 2017.
• ThiNet: Pruning CNN Filters for a Thinner Net, TPAMI 2018.
• [ICCV Project Page] [Pretrained Models]

Python 2.6 & Caffe environment:

• Python2.6
• Caffe & Caffe's Python interface

1. Clone the ThiNet repository.
2. select ThiNet_ICCV or ThiNet_TPAMI subfolder:

   cd ThiNet_ICCV
   

3. modify your configuration path:
   • modify the caffe path (caffe_root) at the beginning of net_generator.py and compress_model.py
   • modify ImageNet lmdb file path in line 212 and line 217 of net_generator.py
   • modify ImageNet dataset path in line 54, 55, 60 of compress_model.py
   • modify line 2 and 4 in run_this.sh with correct file path.
4. Run the pruning demo:

   ./run_this.sh
   

• Image Resize:

  • Note that there are two different strategies to organize ImageNet dataset:
    1. fixed size: each image is firstly resized to 256×256, then center-cropped to obtain a 224×224 regin;
    2. keep aspect ratio: each image is firstly resized with shorter side=256, then center-cropped;
  • The default caffe create_lmdb.sh file will convert images into 256x256. If you want to keep the original ratio:
    1. replace caffe/src/caffe/util/io.cpp with toolkit/caffe_lmdb_keep_ratio/io.cpp
    2. rebuild caffe
    3. use the provided script toolkit/caffe_lmdb_keep_ratio/create_lmdb.sh to create the lmdb file
    4. and, do not forget to modify the configuration path of this script.

• FLOPs Calculation:

  cd toolkit
  modify the caffe_root at the beginning of FLOPs_and_size.py file.
  python FLOPs_and_size.py [the path of *.prototxt file]
  

  NOTE: we regard the vector multiplication as TWO float-point operations (multiplication and addition). In some paper, it is calculated as ONE operation. Do not be confused if the result is twice larger.

We prune the VGG_ILSVRC_16_layers model on ImageNet dataset with ratio=0.5:

There are no difference on VGG16, but ThiNet_TPAMI is much better on ResNet50:

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

@CONFERENCE{ThiNet_ICCV17,
  author={Jian-Hao Luo, Jianxin Wu, and Weiyao Lin},
  title={ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression},
  booktitle={ICCV},
  year = {2017},
  pages={5058-5066},
}

@article{ThiNet_TPAMI,
  author = {Jian-Hao Luo, Hao Zhang, Hong-Yu Zhou, Chen-Wei Xie, Jianxin Wu, and Weiyao Lin},
  title = {ThiNet: Pruning CNN Filters for a Thinner Net},
  journal = {IEEE Trans. on Pattern Analysis and Machine Intelligence},
  year = {2008},
}

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