This project implements an image classifier using the 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, micromlgen and m2cgen and they have been tested on Arduino Mega and ESP32-CAM boards.
- Download the dataset.
- Install the python dependencies listed in
requirements.txt. - Install the toolchain packages required by RIOT for each board. Consult RIOT's documentation for how to do it.
# 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 <image-path>
$ ./driver features <float|double> 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 for creating new apps is also recommended.
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 |