DualCrossTAP Knowledge Distillation Code

This repository contains the code for training and evaluation of the teacher model (DualCrossTAP) and student models (CNN-S1 & CNN-S2) from Composite human activity recognition utilizing knowledge distillation and sensor fusion focusing on resource constrained microcontrollers.

Usage

• First run run_preprocessing.py to download and prepare the dataset
• To train models with KD, run run_kd_pipeline.py
• To perform cross validation on the teacher model, run run_cross_validation.py
• To calculate the floating point operations (FLOPS) of each model, run run_flops_analysis.py
• To generate the C header files to run student models on a microcontroller, run run_deployment_prep.py

Structure

├── config.py  # Hyperparameters and other configuration values
├── data_pipeline
│   ├── downloader.py  # Helper script to download the CAPP dataset
│   ├── processor.py  # Helper script to process the CAPP dataset
├── models
│   ├── analysis.py  # Analysis utilities
│   ├── deployment_prep_utils.py  # Generate C header files from trained student models
│   ├── knowledge_distillation.py  # Train student model with KD
│   ├── student.py  # Train student model without KD
│   ├── teacher.py  # Train teacher model
│   └── utils.py  # Training utilities
├── README.md
├── run_cross_validation.py  # Perform k-fold cross validation on DualCrossTAP
├── run_deployment_prep.py  # Convert student models to C header files
├── run_flops_analysis.py  # Calculate model MFLOPS
├── run_kd_pipeline.py  # Train DualCrossTAP and students (CNN-S1 & CNN-S2) with and without KD
├── run_preprocessing.py  # Prepare the CAPP dataset

Cite

Consider citing our paper published in Expert Systems with Applications (ESWA) if you use the codebase in your work:

@article{dual_cross_tap,
    title = {Composite human activity recognition utilizing knowledge distillation and sensor fusion focusing on resource constrained microcontrollers},
    journal = {Expert Systems with Applications},
    volume = {298},
    pages = {129652},
    year = {2026},
    issn = {0957-4174},
    doi = {10.1016/j.eswa.2025.129652},
    url = {https://doi.org/10.1016/j.eswa.2025.129652},
    author = {Athar Noor Mohammad Rafee and John Clear and Jannatun Noor}
}

 

Copyright © 2025-present Athar Noor Mohammad Rafee and John Clear IV