Creating a project in STM32Cube AI Studio

This message will disappear after all relevant tasks have been resolved.

Semantic MediaWiki

There are 1 incomplete or pending task to finish installation of Semantic MediaWiki. An administrator or user with sufficient rights can complete it. This should be done before adding new data to avoid inconsistencies.

STM32Cube AI Studio is the next-generation tool for deploying AI models on STM32 microcontrollers.

This guide provides a comprehensive, step-by-step walkthrough for creating a new project, inspired by best practices from the legacy plugin X-CUBE-AI.

1. Introduction

STM32Cube AI Studio enables:

Import and validation of AI models (supported types are: ONNX, TFLite, Keras)
Selection and configuration of STM32 hardware targets
Analysis, optimization, and visualization of models
Management of memory pools and hardware accelerators
Export of ready-to-deploy code and binaries.

2. Project workflow

2.1. Launch and Prepare

Open STM32Cube AI Studio from the host PC desktop or start menu.
Ensure the latest version for new features and device support are available on the host PC.

2.2. Create a new project

Start a project

Click the Project button on the dashboard or sidebar.

Enter a Project Name (for example, gesture_recognition) and define the Workspace Path.

Select a target (for example: STM32F746G-DISCO)

Select a toolchain (for example: STM32CubeIDE labelled as "IDE")

Project Structure

Once the project is started, a structure is set up, the project includes:

Model files
Configuration files
Optimization results
Exported code and binaries

For a generated project, these files are directly included in theproject code structure, in:

Middlewares/ST/AI: Includes ST Edge AI Core middleware and libraries
AI/App: Application code

Note

ST Edge AI Core outputs are stored in a .ai folder which is also used as a temporary folder / backup folder.

2.3. Start a Run

The Run button generate a report summarizing model performance.
Before running, configure the following fields:

Target

Configure target-specific options, such as:

activating the Neural-ART NPU, or other specific hardware
selecting optimization profiles
defining custom runtime arguments

Model

Click to import a model. Supported formats are:

ONNX (.onnx)
TensorFlow Lite (.tflite)
Keras (.h5, .keras, .hdf5).

Memory pools

Define where the model activations will be allocated in RAM, either automatically or manually (customize pool priority, size, and offset).

Below is an example on NUCLEO-H743ZI using default memory pools, where DTCMRAM will be filled first, then RAM_D1, then RAM_D2, and so on. Add a new memory pool row manually if needed, with custom offset and size, for example when using external RAM.

It is also possible to export model weights as a binary file (.bin) or to copy weights to RAM at startup.

For more information about memory pools, see this section.

Validation

This section defines:

which data will be used for model validation
how many samples will be used
where validation will be performed: on host (PC) or on target (STM32 board)

Note

Run on host relies on the ST Edge AI Core validate command, or analyze when not supported.
Run on target relies on STM32CubeMX (building a validation project), ST Edge AI core (generating model C code via the generate command), and STM32CubeProgrammer (flashing validation app on target).

For more information, refer to the ST Edge AI Core CLI command workflow.

2.4. Generate code

The Generate code button is used to:

generate optimized C code (.c and .h) for the imported model
generates a full project for the selected IDE

The generated project can be:

Application: a hello_world project template to get started creating your own application
Validation: a specialized project to run on-target validation using ST Edge AI Core

Both options rely on the ST Edge AI Core generate command.

3. Best Practices & Troubleshooting

Model compatibility

Use supported layers and formats for best results.
Refer to the ST Edge AI Core supported layers and operators list for guidance on supported layers and formats.

Memory optimization

If memory errors occur, reduce the model size or optimize memory pools.
Select a device with more resources if needed.

4. Example: Creating a gesture recognition project

Launch STM32Cube AI Studio
Create a new project named gesture_recognition
Import a model, such as, gesture_model.onnx
Select the target board, such as, STM32H747I-DISCO
Execute a run, selecting to run either on the target or on the host
Visualize the model architecture
Generate a STM32project
Integrate into the application

Creating a project in STM32Cube AI Studio

Contents