How to implement a multi class neural network with STM32F103?

Implementing a multi-class neural network on an STM32F103 (Cortex-M3, 72 MHz, no FPU, limited RAM/Flash) is challenging due to its limited resources, but it is possible for small networks and low-dimensional inputs (e.g., 2D/3D sensor inputs).

Here's how you can do it step-by-step:

 1. Design a Small Neural Network on PC

Use Python with TensorFlow/Keras to design and train a model:

 2. Convert Model to C (Quantized if Needed)

Options:

• STM32Cube.AI (official ST tool)

  • Converts Keras/TensorFlow models to C code optimized for STM32

  • Integrates with STM32CubeMX

• uTensor, CMSIS-NN, or TensorFlow Lite for Microcontrollers (TFLM)

  • Lightweight inference engines for ARM Cortex-M cores

 On STM32F103, use integer quantization (8-bit) to reduce memory and increase speed:

 3. Integrate with STM32 Project

A. Using STM32Cube.AI:

1. Install STM32CubeMX + STM32Cube.AI plugin

2. Import .h5 model

3. Enable X-CUBE-AI middleware

4. Generate code

5. Use aiRun() to pass input data and get output class

B. Manual (for tiny models):

If doing it manually:

• Extract weights and biases

• Implement forward pass in C:

 4. Run and Test

• Flash to STM32 using STM32CubeIDE

• Send input via UART or use ADC/sensors

• Print the predicted class

 Optimization Tips

• Use fixed-point arithmetic (Q7, Q15 types)

• Use CMSIS-DSP for matrix operations

• Use CMSIS-NN if available — optimized for Cortex-M

• Keep model size tiny: e.g., 2–1–3 or 2–4–3 architecture

 Resources

• STM32Cube.AI documentation

• CMSIS-NN

• TFLite for Microcontrollers