Skip to content

Getting Started with EdgeVolution

This guide will help you set up and run EdgeVolution on your system.

Setup

Cloning the repository

git clone https://github.com/ankilab/EdgeVolution

Docker Build Targets

The Dockerfile provides two build targets. The embedded image extends ml — it contains everything in ml plus the hardware toolchain:

┌──────────────────────────────────────────────────────────────┐
│  embedded   (docker build --target embedded .)               │
│                                                              │
│  nRF CLI tools, Segger J-Link, Zephyr SDK, west workspace   │
│                                                              │
│  ┌────────────────────────────────────────────────────────┐  │
│  │  ml   (docker build .)              ← DEFAULT target   │  │
│  │                                                        │  │
│  │  TensorFlow 2.9.1-gpu · Python deps · project code    │  │
│  └────────────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────────────┘

ML/NAS only (default) — architecture search, training, evaluation: 

docker build -t edgevolution .

Embedded — everything above plus cross-compilation and flashing to MCUs: 

docker build --target embedded -t edgevolution-embedded .

The embedded build downloads the Zephyr SDK and modules (~1.5 GB). Docker caches this layer, so subsequent builds are fast.

Running the Container

ML/NAS (GPU-accelerated): 

docker run -it --rm --gpus all -v $(pwd):/EdgeVolution edgevolution

Embedded (with USB passthrough for J-Link): 

docker run -it --rm --privileged --gpus all -v $(pwd):/EdgeVolution edgevolution-embedded

Configuration

Defining hyperparameters

Before 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.

  1. Search space setup
    Search Space

  2. Hyperparameters
    Hyperparameters --> Important to update results path!

  3. Microcontroller setup
    Microcontrollers

  4. Dataloader definition
    Dataloader Definition

Execution

Running an experiment

python main.py +hyperparameters=<your_hyperparams> +search_space=<your_search_space> +boards=nrf52840dk

Visualization

The EvoVis Dashboard can be used to visualize and interpret the results of an EdgeVolution experiment run.