NEW RELEASE: we released our new, engineered and user-friendly DNAS library named PLiNIO. PLiNIO has been used as the main optimization engine to build the proposed classifier. We highly suggest to try PLiNIO for your experiments!
This repository contains four folders:
python code: folder with the python training code for the proposed classifier.c_code_l432kc: folder with the c code generated using STM32CubeMX with X-Cube-AI and Keil uVision for the nucleo-l432kc board.c_code_l4a6zg: folder with the c code generated using STM32CubeMX with X-Cube-AI and Keil uVision for the nucleo-l4a6zg board.trained_model: folder with the trained and deployed onnx model.
The code has been tested on Ubuntu 18.04.06 LTS with Bash 4.4 Shell
You can run the code contained in python_code by first installing the python requirements:
pip install requirements.txt
Then, downloading data:
chmod +x download_data.sh
./download_data.sh
N.B., you may require to install unrar within your system (e.g., sudo apt install unrar) to run the download_data.sh script
and then, running:
python train.py DiscIEGMNet_cbr --epoch 30 --lr 1e-2 --augment
The outputs of this execution will be saved in the path: saved_models/DiscIEGMNet_cbr/<timestamp>, where <timestamp> contains information about the date and time of when the code is executed.
In particular, the trained model's weights will be saved as model.pkl within this path.
You can firstly convert the model to onnx format by running
python pkl2onnx.py DiscIEGMNet_cbr --model-path <path-where-model.pkl-is-stored>
Once we obtain the onnx model file, we could deploy the model on the board by following the provided instructions using XCube-AI and Keil uVision.
N.B., we propose two different C code folders (c_code_l432kc and c_code_l4a6zg) because we tested our code on a slightly different nucleo-L4 board. The main difference between the two implementations regards the UART pheripheral. In the evaluation phase please refer to the contained in c_code_l432kc.