Caffe Implementation of ThiNet

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]

Requirements

Python 2.6 & Caffe environment:

Python2.6
Caffe & Caffe's Python interface

Usage

Clone the ThiNet repository.
select ThiNet_ICCV or ThiNet_TPAMI subfolder:
```
cd ThiNet_ICCV
```
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.
Run the pruning demo:
```
./run_this.sh
```

Other Toolkits

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.

Results

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

Method	Top-1 Acc.	Top-5 Acc.	#Param.	#FLOPs
original VGG16	71.50%	90.01%	138.24M	30.94B
ThiNet_ICCV	69.80%	89.53%	131.44M	9.58B
ThiNet_TPAMI	69.74%	89.41%	131.44M	9.58B

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

Method	Top-1 Acc.	Top-5 Acc.	#Param.	#FLOPs
original ResNet50	75.30%	92.20%	25.56M	7.72B
ThiNet_ICCV	72.04%	90.67%	16.94M	4.88B
ThiNet_TPAMI	74.03%	92.11%	16.94M	4.88B

Citation

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},
}

Contact

Feel free to contact me if you have any question (Jian-Hao Luo luojh@lamda.nju.edu.cn or jianhao920@gmail.com).