Skip to content

README.md

Latest commit

 

History

History
86 lines (62 loc) · 2.16 KB

README.md

File metadata and controls

86 lines (62 loc) · 2.16 KB
Raw

Practise

This is the official Pytorch implementation for Practical Network Acceleration with Tiny Sets.

In this project, we also implement MiR which is Compressing Models with Few Samples: Mimicking then Replacing.

Requirements

  • Python3
  • pytorch
  • pandas

Datasets

Please prepare datasets first, and then modify path in dataset.py.

Usage

Compute the recoverability of one block

python main.py --num_sample 500 --seed 2021 --epoch 1000 --practise one --rm_blocks layer1.1 --gpu_id 0

Compute recoverabilities of all blocks

python main.py --num_sample 500 --seed 2021 --epoch 1000 --practise all --rm_blocks 2 --gpu_id 0

Finetune the pruned network

python main.py --num_sample 500 --seed 2021 --epoch 2000 --FT MiR --rm_blocks layer1.1 --gpu_id 0

ResNet

For ResNet-34 and ResNet-50, the removable blocks are

1: layer1.1,layer1.2
2: layer2.1,layer2.2,layer2.3
3: layer3.1,layer3.2,layer3.3,layer3.4,layer3.5
4: layer4.1,layer4.2

MobileNet V2

For MobileNetV2, the removable blocks are

1: 24->24: features.3
2: 32->32: features.5,features.6
3: 64->64: features.8,features.9,features.10
4: 96->96: features.12,features.13
5: 160->160: features.15,features.16

Test the latency

python speed.py --model mobilenet_v2 --cudnn-benchmark --rm_blocks features.9

Results in our paper

We provide all shells to reproduce all results in our paper. Please check shells in the exp folder.

Citation

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

@inproceedings{wang2023practical,
  title={Practical Network Acceleration with Tiny Sets},
  author={Wang, Guo-Hua and Wu, Jianxin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}

@inproceedings{wang2022compressing,
  title={Compressing models with few samples: Mimicking then replacing},
  author={Wang, Huanyu and Liu, Junjie and Ma, Xin and Yong, Yang and Chai, Zhenhua and Wu, Jianxin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={701--710},
  year={2022}
}