EdgeVolution is a software-hardware end-to-end pipeline that allows to optimize preprocessing and neural network architectures for microcontrollers in an end-to-end fashion. It features different building blocks that can be customized individually.
Welcome to EdgeVolution! You can follow this simple guide here to get a high-level overview. You can find a more comprehensive documentation here.
git clone https://github.com/ankilab/EdgeVolution
EdgeVolution provides two Docker build targets:
ML/NAS target (default) — For running neural architecture search and training:
docker build -t edgevolution .Embedded target — For flashing optimized models to MCU hardware (includes nRF tools, J-Link, Zephyr SDK):
docker build --target embedded -t edgevolution-embedded .Note: The embedded target downloads the Zephyr SDK and modules (~1.5 GB) during the build. This is cached for subsequent builds.
ML/NAS (GPU-accelerated):
docker run -it --rm --gpus all -v $(pwd):/EdgeVolution edgevolutionEmbedded (with USB device passthrough for J-Link):
docker run -it --rm --privileged --gpus all -v $(pwd):/EdgeVolution edgevolution-embeddedInside the container:
python3 -m pytest tests/ -v -p no:dashBefore the EdgeVolution optimization run is started, some configurations must be made. These include the definition of hyperparameters, the search space and the boards that are to be used for evaluating the candidates on the microcontroller.
-
Search space setup search space setup
-
Hyperparameters hyperparameters --> Important to update results path!
-
Microcontroller setup microcontroller boards
python main.py +hyperparameters=speech_commands +search_space=speech_commands +boards=none
EdgeVolution supports multiple search strategies via a pluggable ask/tell interface. Select a strategy with the search_strategy= flag:
| Strategy | Flag | Description |
|---|---|---|
| Genetic Algorithm | search_strategy=genetic_algorithm |
Standard GA with crossover and mutation (default) |
| Random Search | search_strategy=random |
Random baseline — no selection or crossover |
| Regularized Evolution | search_strategy=regularized_evolution |
Tournament selection with aging (Real et al. 2019) |
| PyMOO (NSGA-II) | search_strategy=pymoo |
Multi-objective optimization (requires pymoo) |
| Ax (Bayesian Opt.) | search_strategy=ax |
Bayesian optimization via BoTorch (requires ax-platform) |
Example:
python main.py +hyperparameters=speech_commands +search_space=speech_commands +boards=none search_strategy=regularized_evolutionFor more examples and common commands, see the Usage Guide.
Community support is provided via a Discord Server for real-time community support.
To contribute to EdgeVolution, follow these steps:
- Fork this repository.
- Create a branch:
git checkout -b <branch_name>. - Make your changes and commit them:
git commit -m '<commit_message>' - Push to the original branch:
git push origin <project_name>/<location> - Create the pull request.
Alternatively see the GitHub documentation on creating a pull request.
If you want to contact me you can create a github issue or reach out on Discord.
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
tbd