Home

Awesome

Group Fisher Pruning for Practical Network Compression (ICML2021)

By Liyang Liu*, Shilong Zhang*, Zhanghui Kuang, Jing-Hao Xue, Aojun Zhou, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, Wayne Zhang

Updates

NOTES

All models about detection has been released. The classification models will be released later, because we want to refactor all our code into a Hook , so that it can become a more general tool for all tasks in OpenMMLab.

We will continue to improve this method and apply it to more other tasks, such as segmentation and pose.

The layer grouping algorithm is implemtated based on the AutoGrad of Pytorch, If you are not familiar with this feature and you can read Chinese, then these materials may be helpful to you.

  1. AutoGrad in Pytorch

  2. Hook of MMCV

Introduction

1. Compare with state-of-the-arts.

<div align="center"> <img src="resources/sota.png" width="800" height="400"/> </div>

2. Can be applied to various complicated structures and various tasks.

<div align="center"> <img src="resources/structures.png" width="900"/> </div>

3. Boosting inference speed on GPU under same flops.

<div align="center"> <img src="resources/speed_up.png" width="600", height="300"/> </div>

Get Started

1. Creat a basic environment with pytorch 1.3.0 and mmcv-full

Due to the frequent changes of the autograd interface, we only guarantee the code works well in pytorch==1.3.0.

  1. Creat the environment
conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab
  1. Install PyTorch 1.3.0 and corresponding torchvision.
conda install pytorch=1.3.0 cudatoolkit=10.0 torchvision=0.2.2 -c pytorch
  1. Build the mmcv-full from source with pytorch 1.3.0 and cuda 10.0

Please use gcc-5.4 and nvcc 10.0

 git clone https://github.com/open-mmlab/mmcv.git
 cd mmcv
 MMCV_WITH_OPS=1 pip install -e .

2. Install the corresponding codebase in OpenMMLab.

e.g. MMdetection

pip install mmdet==2.13.0

3. Pruning the model.

e.g. Detection

cd detection

Modify the load_from as the path to the baseline model in of xxxx_pruning.py

# for slurm train
sh tools/slurm_train.sh PATITION_NAME JOB_NAME configs/retina/retina_pruning.py work_dir
# for slurm_test
sh tools/slurm_test.sh PATITION_NAME JOB_NAME configs/retina/retina_pruning.py PATH_CKPT --eval bbox
# for torch.dist
# sh tools/dist_train.sh configs/retina/retina_pruning.py 8

4. Finetune the model.

e.g. Detection

cd detection

Modify the deploy_from as the path to the pruned model in custom_hooks of xxxx_finetune.py

# for slurm train
sh tools/slurm_train.sh PATITION_NAME JOB_NAME configs/retina/retina_finetune.py work_dir
# for slurm test
sh tools/slurm_test.sh PATITION_NAME JOB_NAME configs/retina/retina_fintune.py PATH_CKPT --eval bbox
# for torch.dist
# sh tools/dist_train.sh configs/retina/retina_finetune.py 8

Models

Detection

MethodBackboneBaseline(mAP)Finetuned(mAP)Download
RetinaNetR-50-FPN36.536.5Baseline/Pruned/Finetuned
ATSS*R-50-FPN38.137.9Baseline/Pruned/Finetuned
PAA*R-50-FPN39.039.4Baseline/Pruned/Finetuned
FSAFR-50-FPN37.437.4Baseline/Pruned/Finetuned

* indicate with no Group Normalization in heads.

Classification

Coming soon.

Please cite our paper in your publications if it helps your research.

@inproceedings{liu2021group,
  title={Group fisher pruning for practical network compression},
  author={Liu, Liyang and Zhang, Shilong and Kuang, Zhanghui and Zhou, Aojun and Xue, Jing-Hao and Wang, Xinjiang and Chen, Yimin and Yang, Wenming and Liao, Qingmin and Zhang, Wayne},
  booktitle={International Conference on Machine Learning},
  pages={7021--7032},
  year={2021},
  organization={PMLR}
}