Run Time Adaptive Network Slimming for Mobile Environments

Authors: Hong-Ming Chiu, Kuan-Chih Lin, Tian Sheuan Chang

Link to Paper

Code for training resnet models referenced from: https://github.com/kuangliu/pytorch-cifar

Cite

Please cite our paper if you use this code in your own work:

@inproceedings{chiu2019,
  author={Hong-Ming Chiu and Kuan-Chih Lin and Tian Sheuan Chang,
  booktitle={2019 IEEE International Symposium on Circuits and Systems (ISCAS)}, 
  title={Run Time Adaptive Network Slimming for Mobile Environments}, 
  year={2019},
  pages={1-4},
  doi={10.1109/ISCAS.2019.8701884}
  }

Table of Contents

• Intorduction
• Directory Tree
• Setup and Usage
• Reference
• Author

Intorduction

This Python program contains the code for the paper "Run Time Adaptive Network Slimming for Mobile Environments" and the code for training the ResNet [2] model. This program performs the interface stage pruning and conculates the pruning rate based on the saving in floating point operations (FLOP).

Directory Tree

• /:
  • Adapted_Network.py: main function for Adaptive Network Slimming. [1]
• model_pkl/:
  • <model name>.pkl: pretrained model file.
• train_model/
  • main.py: main function for training model.
• result/
  • resnet18_uniform.png: result of pruning ResNet18 with uniform pruning rate.
  • resnet18_nonuniform.png: result of pruning ResNet18 with non-uniform pruning rate.

Setup and Usage

Package Version

• Python 3.6
• PyTorch 1.10
• Torchvision 0.3.0
• CUDA 10.0

Hardware and Operation System

• CPU: Intel i7-8700 3.2GHz
• Graphics Cards: GeForce RTX 2080 Ti
• OS: Ubuntu 18.04

Environment setup

1. Set up the environment using Anaconda

conda create -n myenv python=3.6
conda install --name myenv pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=10.0 -c pytorch

2. Activate conda environment

conda activate myenv

Usage

1. Run network slimming model with uniform pruning rate. Use -net to select ResNet with different depth, available options are resnet18, resnet34 and resnet50.

python3 Adapted_Network.py -net resnet18

Use --nonuniform to run network slimming model with non-uniform pruning rate.

python3 Adapted_Network.py -net resnet18 --nonuniform

Reference

[1] Hong-Ming Chiu, Kuan-Chih Lin and Tian Sheuan Chang, "Run Time Adaptive Network Slimming for Mobile Environments," 2019 IEEE International Symposium on Circuits and Systems (ISCAS).

[2] Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun, "Deep Residual Learning for Image Recognition," 2015.

Author

Name : Hong-Ming Chiu

Email : hmchiu2 [at] illinois.edu

Website : https://hong-ming.github.io

License

MIT License