Heptari

STM32 Cube AI Studio 教程

STM32 Cube AI Studio 是 STmicroelectronics 公司推出的一款模型部署和代码生成工具。其可以将 ONNX、tflite 等格式的模型针对 STM32 边缘计算场景进行优化和神经网络结构设计，生成代码到工程中，并提供接口用以进行逻辑编程。

alt text

软件安装

官方链接： STM32Cube.AI

需要注意的是，其依赖数个 STM32 运行库和支持包：

在下载完成后，在 Settings 界面中将以上软件的安装目录设置好：

alt text

回到 Projects 页面，可以看到其允许新建项目，表明安装完成。

项目配置和模型转化

Cube AI Studio 允许将 Cube AI 的部分集成到现有的项目中，也允许新建项目。现在以新建项目为例。

首先，对于一个新的项目，需要新建一个 Runs，即算例，对应的要集成到 MCU 中的模型，如下图所示。

alt text

在右侧，选择模型文件(Model File)，可以为 tflite/onnx/h5 等格式。然后选择在芯片中的存储位置(Memory Pool)、验证数据(Validation Data)等，就可以点击 Run 转化模型。转化完如图所示。

alt text

详情页记录了部分详细信息，如模型名称、目标芯片（开发板）型号、输入输出格式等。该输入输出格式和选择的模型不一定一致，是由 CubeAI 优化过的结果。

该输入输出格式在生成代码后于数据接口中有所体现，即替换一个维数相等的数组作为 input/output

详细信息下方还有内存占用情况、神经网络结构等。

alt text

alt text

需要注意的是，在模型转化过程中，其会将其烧录到单片机中进行验证。所以硬件链接是建议的。

代码生成及部署

检查无误后，即可点击 Generate Code 按钮生成代码，项目类型为 Toolchain 中选择的 CubeIDE/Keil/IAR。其会在项目中生成 Middlewares/ST/AI、AI 两个内容，我们需要关注的部分在 AI/App/app_x-cube-ai.c 中。

alt text

进入 app_x-cube-ai.c 文件，找到函数 acquire_and_process_data() 和 post_process()。前者为获取数据的函数，后者为得到模型预测结果后的处理部分。这两个函数在 main_loop() 中被调用：

/* 
 * Example of main loop function
 */
void main_loop() {
  while (1) {
    /* 1 - Acquire, pre-process and fill the input buffers */
    acquire_and_process_data();

    /* 2 - Call inference engine */
    aiRun();

    /* 3 - Post-process the predictions */
    post_process();
  }
}

而 main_loop() 在 STM32CubeAI_Studio_AI_Process(void) 被调用：

void STM32CubeAI_Studio_AI_Process(void)
{
    main_loop();
}

然后 STM32CubeAI_Studio_AI_Process(void) 在 main() 中被调用。通过这样的方式，我们的模型和数据调用接口被链接到项目中。

下面给出一个利用 mnist-12.onnx 模型进行训练，部署在 NUCLEO-F411RE 开发板上的代码示例：

// 这个变量用来模拟输入的 28x28 矩阵，作为手写数字识别模型的输入。
static const float mnist_test_digit_7[784] = {
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0.3,0.7,0.9,0.9,0.9,0.9,0.9,0.3,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0.1,0.3,0.3,0.3,0.3,0.5,0.9,0.9,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.5,0.9,0.2,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.4,0.9,0.5,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.3,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.2,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.1,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.8,0.9,0.2,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.5,0.9,0.4,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.3,0.9,0.6,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.2,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.1,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.8,0.9,0.3,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.4,0.9,0.6,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.2,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.1,0.9,0.8,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.6,0.9,0.4,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.3,0.9,0.7,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
   0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
 };

int acquire_and_process_data()
{
  if (stai_input[0] != NULL) {
	  memcpy(stai_input[0], mnist_test_digit_4, sizeof(mnist_test_digit_4));
  }
  return 0;
}

int post_process()
{
	float *out_data = (float *)stai_output[0];
	  float max_val = -1.0f;
	  int predicted_digit = -1;

	  // MNIST 输出 10 个类别的概率
	  for (int i = 0; i < 10; i++) {
	      if (out_data[i] > max_val) {
	          max_val = out_data[i];
	          predicted_digit = i;
	      }
	  }

	  /* 通过 USART2 输出结果 */
	  char msg[64];
	  int len = sprintf(msg, "\r\nPredicted Digit: %d \r\n",
	                    predicted_digit);

	  // 假设在 main.c 里的串口句柄是 huart2
	  extern UART_HandleTypeDef huart2;
	  HAL_UART_Transmit(&huart2, (uint8_t*)msg, len, HAL_MAX_DELAY);
  return 0;
}

模型可以在 GitHub 下载: github.com/onnx/models

通过串口输出的预测结果如下：

alt text

Last modified: 2026-05-24