CPUFronz/sensor_guided_adaptation

Sensor-Guided Adaptive Machine Learning on Resource-Constrained Devices

This is the code for the IoT 2024 paper "Sensor-Guided Adaptive Machine Learning on Resource-Constrained Devices".

Setup

A Conda *.yml file is provided for recreating the development environment. Just run conda env create -f environment.yml to create the required development environment.

How to Run

The files to run the experiments are:

• baseline.py
• cifar10_rotation.py
• cifar10_scale.py
• gtsrb_rotation.py
• gtsrb_scale.py
• inverse.py

Note, that inverse requires a trained baseline model to run, the wandb id of the trained baseline model has to be passed as argument.

The accuracy of the resulting models is automatically evaluated after training. To evaluate the robustness against noisy sensors or measure the runtime, run either run-noisy.sh or run-benchmark.sh. Both Bash scripts call the coresponding Python scripts for evaluation (evaluate_noisy.py and evaluate_runtime.py), for all wandb ids. You have to adjust those ids accordingly. Please note, that you need adb installed and the Android phone connected to evaluate the runtime.

Visualization

To visualize the results, run the corresponding Jupyter Notebooks:

• plot-capacity.ipynb
• plot-header.ipynb
• plot-main.ipynb
• plot-noisy.ipynb
• plot-runtime.ipynb

Data

We use the GTSRB and CIFAR-10 datasets, both are available in PyTorch and will be automatically downloaded.

Citing

If you find our work useful, please cite it using the following BibTex entry:

@inproceedings{papst2024,
  title = {Sensor-Guided Adaptive Machine Learning  on Resource-Constrained Devices},
  author = {Papst, Franz and Kraus, Daniel and Rechberger, Martin and Saukh, Olga},
  booktitle = {14th International Conference on the Internet of Things (IoT 2024)},
  year = {2024},
  doi = {10.1145/3703790.3703801}
}