Build and run a letter recognition NN model on an STM32L4 Discovery board
Introduction
Prepare development environment
Collect training data
Train the model
Feature extraction
Run the model on development board
Review
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

In this section, you will deploy the model directly on the STM32 board.

Setup STM32CubeMX project

STM32CubeMX is a graphical tool for configuring STM32 microcontrollers.
STM32Cube.AI is an extension of STM32CubeMX to import ML models.

  1. Download and install STM32CubeMX from the STM32CubeMX product page .

  2. Open STM32CubeMX IDE.

  3. Click Access to Board Selector. Find your B-L475E-IOT01A board and click Start Project.

  4. Go to Project Manager. Enter a project name and select the project location where the project will be saved.

  5. Under Pinout & Configuration, expand Pinout menu, and click Clear pinouts.

  6. Expand Software Packs menu, and click Select Components.

  7. Locate X-CUBE-AI, and click its Install button.

  8. Expand the menu and enable X-CUBE-AI > Core. For device application, choose Validation. If the status shows a yellow warning, click Resolve to install any necessary dependencies.

  9. Click OK when done.

Validate the NN model

You will now validate the NN model built in the previous section on the desktop and the development board.

Validate on desktop

  1. On the STM32CubeAI menu, locate X-CUBE-AI configuration from the list of Categories.

  2. click Add network. Browse to the generated model (.h5 file). You can also specify other values, such as validation input, and validation output. Leave as default for now.

  3. Click Analyze, then it will generate a detailed report on the model. You can check the number of parameters, the size of the weights, and the amount of memory to be used.

  4. Click Show graph to visualize your model.

  5. Click Validate on Desktop to validate the model with the provided input and output on your desktop.

Validate on development board

To validate the model on the target, you need to first generate code and install the code on the target.

  1. Click Validate on target, and enable Automatic compilation and download. Select the appropriate toolchain (STM32CubeIDE), and click OK.

Demo with the sample application

A sample application is provided. This uses a feature-based model, trained with the letters O and S written in a horizontal plane.

Notice

There is a bug in the example as supplied. You must edit /Core/Src/main.c before building.

Locate the function call in main() to

  • MX_USART1_UART_Init()

and move to before the call to

  • MX_X_CUBE_AI_Init()

  1. Using a terminal application (such as PuTTY), set up the serial connection to the board. COM number will be as before. The baud rate should be 115200.

  2. Import the application. Click the File tab and click Import. Select Existing project into workspace. Go to tf_stm32 folder and select MCU_Activity_Recognition.

  3. Open X-CUBE-AI/App/app_x-cube-ai.c. Observe that functions are provided for data acquisition from the accelerometer sensor and feature extraction.

  4. Build application and flash to target with Run As.

  5. Observe output on the terminal.

  6. Press the blue button on the board, draw a letter, then press the button again.

Create your own application

You can create your own application from the model you have trained.

  1. Open the list of Software Packages. Change the device application as Application.

  2. Click Generate Code then the project is updated.

  3. Open STM32CubeIDE. Since you re-generated the code again, you need to delete the existing project and import the project again.

  4. Open X-CUBE-AI/App/app_x-cube-ai.c. This is the auto-generated code by STM32CubeMX.

  5. Between USER_CODE_BEGIN and USER_CODE_END, you can write code for your application.

There are two main functions that you should implement:

  • acquire_and_process_data() is for getting data from sensors and processing the data, for example, extracting features.
  • postprocess() is for postprocessing the output
Back
Next