(paper source) Geometric Self-Supervised Learning for HAR using Inception with Multi-Kernel Convolutions (https://github.com/lky473736/InceptionMK.git)
This repository implements the methodology proposed in the paper Geometric Self-Supervised Learning for HAR using Inception with Multi-Kernel Convolutions.
The system works with four popular HAR datasets.
-
UCI HAR Dataset: Contains data from smartphone sensors for 6 activities.
-
WISDM Dataset: Contains accelerometer data from smartphones for 6 physical activities.
-
PAMAP2 Dataset: Physical Activity Monitoring dataset with data from 18 different physical activities. (At this 12.)
-
mHealth Dataset: Contains data from body-worn sensors for 12 physical activities.
-
Stem Layer: Conv1D + BatchNorm + ReLU to extract initial local features.
-
Inception Block: Parallel depthwise separable convolutions and max pooling to capture multi-scale temporal features.
-
Multi-Kernel Block: Additional depthwise separable convolutions with various kernel sizes (1, 3, 5, 7) for diverse receptive fields.
-
Adaptive Average Pooling + Flatten: Global feature summarization.
-
Embedding Layer: Latent embedding projection for downstream tasks.
-
Dual Heads:
- Activity Classification Head (main task)
- Rotation Angle Classification Head (self-supervised pretext task)
If you use this code in your research, please cite:
@article{Lim2025inceptionmk,
title = {Geometric Self-Supervised Learning for HAR using Inception with Multi-Kernel Convolutions},
author={Gyuyeon Lim and Myung-Kyu Yi},
journal={},
volume={},
pages={},
year={2025}
}
For questions or issues, please contact
- Gyuyeon Lim : lky473736@gmail.com
This project is licensed under the MIT License - see the LICENSE file for details.