/Collaborative_HAR

Primary LanguagePython

Towards Better Accuracy-Efficiency Trade-Offs: Dynamic Activity Inference via Collaborative Learning from Various Width-Resolution Configurations

Lutong Qin, Lei Zhang*, Chengrun Li, Chaoda Song, Dongzhou Cheng, Shuoyuan Wang, Hao Wu, Aiguo Song

Nanjing Normal University   |   Shanghai Jiao Tong University   |  Case Western Reserve University   |   University of Macau   |  Yunnan University   |  Southeast University

Towards Better Accuracy-Efficiency Trade-Offs: Dynamic Activity Inference via Collaborative Learning from Various Width-Resolution Configurations



📝 Introduction

This paper introduces a revolutionary approach: Dynamic Activity Inference via Collaborative Learning. Traditional methods for human activity recognition (HAR) on mobile devices are rigid and lack adaptability to changing resource constraints. Our method breaks this mold by training subnetworks at different widths and resolutions, enabling dynamic accuracy-latency trade-offs at runtime. Unlike conventional approaches requiring multiple networks, ours seamlessly switches between configurations without retraining, slashing costs. Extensive experiments validate its efficacy.

✨Compared to existing works, Dynamic Activity Inference via Collaborative Learning achieves:

  • Generalizability: It adapts to a wide range of real-world scenarios without the need for human-specific fine-tuning. Previous methods often specialize in specific domains, limiting their applicability.

  • State-of-the-art Results: Our approach achieves current state-of-the-art performance in human activity recognition.

🌟With this project, you can:

  • Utilize our easy-to-use codebase for re-implementation and further development.

  • Experience our model or deploy the model for inference locally.



🚀 Getting Started

Requirements

  • Python 3.8
  • PyTorch 1.11.0
  • Numpy 1.23.4
  • pyyaml 5.1

Training

UniMiB-SHAR dataset

python train.py --model_name='unimib'                      # CNN
python train.py --model_name='resnet_unimib'               # ResNet

WISDM dataset

python train.py --model_name='wisdm'                       # CNN
python train.py --model_name='resnet_wisdm'                # ResNet

UCI-HAR dataset

Get UCI dataset from UCI Machine Learning Repository(http://archive.ics.uci.edu/ml/index.php), do data pre-processing by sliding window strategy and split the data into training and test sets

python train.py --model_name='uci'                         # CNN
python train.py --model_name='resnet_uci'                  # ResNet

Results

Comparing with both baselines tuning either width or resolution independently

Alt text Alt text

Comparing with independently trained networks

Alt text Alt text

Boosting the Performance of a Single Network

Dataset WISDM UCI-HAR UniMiB-SHAR
CNN 97.10% & 11.21M 95.95% & 13.06M 75.49% & 15.49M
Ours 98.71% & 11.21M 97.00% & 13.06M 78.13% & 15.49M
------------ ---------------------- ------------------------- ------------------------
ResNet 97.52% & 32.22M 96.51% & 43.5M 76.29% & 48.97M
Ours 99.12% & 32.22M 97.64% & 43.5M 78.91% & 48.97M
------------ ---------------------- ------------------------- ------------------------

Practical On-device Activity Inference

A practical use case for on-device activity inference at various width-resolution configurations

Alt text

The runtime under static computing environment

The runtime under time-varying workloads

Contact

For any questions on this project, please contact Lutong Qin (211843003@njnu.edu.cn)