PENG TinyML Final

Data

You can download data here: https://drive.google.com/drive/u/1/folders/1XXiJoPyWChSsvOFa0eVxZ4AdnfQh_lOA

File Description

【Model_Training_Design_Repo】: This file contains the 5 layer CNN model with focal loss using Pytorch frameworks.

【C_Sourch_Code_Repo】: This file contains the C code generated by STM32CubeMX.

【Quant_fft_model】: This file Contains a quanzied CNN model with 2 channel input (time and frequency).

【Trained_Model_Weights】: The training model weights file in onnx.

【Validation】:File you can used to test our model once you download the C code your board.

【requirements.txt】: The required version you need to install.

How to run the code

1. Connect your STM 32 using usb2micro usb cable to your computer and double-check the port number via the Device Manager.
2. Open \PENG\C_Sourch_Code_Repo\MDK_ARM\Peng_model.uvprojx in KeiluVision5, and download it on your STM32.
3. Run following code in your terminal.

  pip install requirements.txt

4. Open Validation as a project in you python IDE (eg: Pycharm) and change com4 in validation.py to the one you checked in the first step.
5. Run the validation.py and press reset button.

What's new about our model

1. We use train_validation_split when we train the model
2. We change loss function to Focal Loss
3. We process the original data and cange 1 channel time data to 2 channel input data (time + frequence)
4. We use static quantization in pytorch to quant final model