https://github.com/dipa09/riot_imgclf
Multi-class image classifier for RIOT-OS
https://github.com/dipa09/riot_imgclf
arduino-mega-2560 emlearn esp32-cam m2cgen micromlgen riot-os scikit-learn tinyml
Last synced: about 1 month ago
JSON representation
Multi-class image classifier for RIOT-OS
- Host: GitHub
- URL: https://github.com/dipa09/riot_imgclf
- Owner: dipa09
- License: mit
- Created: 2024-07-24T12:55:21.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2024-09-18T10:06:19.000Z (over 1 year ago)
- Last Synced: 2025-03-25T15:51:42.278Z (about 1 year ago)
- Topics: arduino-mega-2560, emlearn, esp32-cam, m2cgen, micromlgen, riot-os, scikit-learn, tinyml
- Language: Python
- Homepage:
- Size: 39.1 KB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
This project implements an image classifier using the [DeepWeeds](https://github.com/AlexOlsen/DeepWeeds) dataset,
which contains more that 8000 images of 8 different weeds' species.
The goal is to evaluate the performance of some TinyML libraries on RIOT OS.
The libraries that have been selected are:
[emlearn](https://github.com/emlearn/emlearn),
[micromlgen](https://github.com/eloquentarduino/micromlgen) and
[m2cgen](https://github.com/BayesWitnesses/m2cgen)
and they have been tested on Arduino Mega and ESP32-CAM boards.
## Dependencies
1. Download the [dataset](https://github.com/AlexOlsen/DeepWeeds).
2. Install the python dependencies listed in `requirements.txt`.
3. Install the toolchain packages required by RIOT for each board.
Consult [RIOT's documentation](https://api.riot-os.org/getting-started.html) for how to do it.
## Usage
```
# Build tools needed for the traning phase.
$ make driver liblbp.so
# Create a custom dataset.
$ python3 ./prepare.py
# Extract all features.
$ python3 ./extract.py
# Train the model.
$ python3 ./train.py
# Export the model to C.
$ python3 ./generate.py [-l library]
# Extract features for a test image
$ python3 ./extract.py -i
$ ./driver features features.h
# Build for Linux (optional)
$ make main [LIB=LIBRARY]
# Copy the project to the RIOT application (edit 'Makefile' or set RIOT_DIR)
$ make sync
# From the RIOT application directory. Build the app for the selected board.
$ make [BOARD=...]
# Connect the board and run
$ make flash [BOARD=...]
```
The preapration, extraction and training steps can be customized, that's why the workflow is
so fragmented. Please use `--help` for more information on how to use the various tools.
Reading the [RIOT documentation](https://api.riot-os.org/creating-an-application.html) for creating
new apps is also recommended.
## Operations
## Tested version
This project has been tested using the following packages' versions.
If you have issues try to install the specific package version.
| Package | Version |
|---------------|-----------|
| emlearn | 0.20.4 |
| h5py | 3.11.0 |
| joblib | 1.4.2 |
| m2cgen | 0.10.0 |
| mahotas | 1.4.15 |
| matplotlib | 3.9.0 |
| micromlgen | 1.1.28 |
| numpy | 1.26.4 |
| opencv-python | 4.10.0.84 |
| scikit-learn | 1.5.0 |