Code Generation for Vision Model

This section provides step-by-step information on generating, optimizing and validating the code for object detection models trained in DEEPCRAFT™ Studio.

Prerequisites

Before you begin, set up the following hardware and software requirements:

Hardware:

• PSOC™ Edge board for deploying the generated code
• Supported camera
• Supported display

For detailed information on boards, compatible displays and cameras, refer to this section (opens in a new tab).(Link to CE)

Software:

• DEEPCRAFT™ Studio 5.6 or later (for training and downloading the trained Object Detection models)
• Eclipse IDE for ModusToolbox™ 2025.8 or later (recommended this section demonstrates the flow using Eclipse IDE)
• ModusToolbox™ Tools Package 3.6.0 or later
• ModusToolbox™ Machine Learning Pack 3.0.0 or later

ModusToolbox™ and ModusToolbox™ Machine Learning Pack can be installed through ModusToolbox™ Setup. Download the setup application from here (opens in a new tab).

Retrieving the trained Object Detection model

The quickest way to retrieve a trained Object Detection model for deployment is through the New Project-wizard in Studio. This is accessed through 'File' > 'New Project...' or through the quick pane under the File menu.

Note: You have the choice to create a ModusToolbox™ project when creating the Studio project. You can do this at this stage or later, but for the purpose of following the flow of this guide, you should do it later.

Get the model file

The model file is located under the \Models directory in the newly created project. You have to use the '_int8.tflite' version for compatibility with the rest of this guide.

If you retrained this model or another Object Detection from scratch you will have to download the model from your training job to retrieve it. To know how to download the Model file, refer to Download the Model and Model Predictions (opens in a new tab).

Creating Object Detection compatible ModusToolbox™ project

1. Open ModusToolbox™ > Eclipse IDE for ModusToolbox™. The Eclipse IDE for ModusToolbox window appears.

2. Select New Application from the Quick Panel or navigate to File> New> Modus Toolbox™ Application to open the Project Creator Tool.

3. Expand the PSOC™ Edge BSPs and select the BSPs (Board Support Packages) that matches your hardware and click Next. The Select Application window appears.

4. In Template Application, expand Machine Learning and select the PSOC Edge Machine Learning DEEPCRAFT Deploy Vision deploy code example and click Create. The project creation might take a while as the required libraries and configuration scripts are downloaded and executed.

5. After you create the project, open Project Explorer, expand 'proj_cm55' directory, and double‑click design.mtbl file to launch ML Configurator with optimized, compatible code‑generation settings. For some projects, where design.mtbl is missing, select your ModusToolbox project directory, navigate to Quick panel> Library Configurators and click ML Configurator 3.0.

Code Generation

1. Under the Model window, in the Pretrained model field, click the folder icon to browse and select the model_int8.tflite downloaded earlier.

2. In Output folder, click the folder icon to browse and select the ModusToolbox™ project directory to generate the code in the 'proj_cm55/mth_ml_gen' folder.

3. Click Generate Source to generate the code.

Optional: Configuring the code generation from scratch

If you want to create a valid code generation configuration from scratch, below are the required settings for Object detection models.

Parameter Name	Recommended Value	Description
Output file prefix	YOLOV5N	This prefix name will generate code that can be used directly with the provided code example project
Output folder	mtb_ml_gen	The folder will allow the generated code to directly replace the existing files and project structure for the provided code example project
Target device	PSOC Edge - Arm Cortex M55 + Ethos U55	The provided code example project utilizes the CM55 core
Pretrained model	(path)/model_int8.tflite	Browse to the location of the .tflite model from the DEEPCRAFT Studio model directory or location of downloaded model file
Imported tflite model quantization	int8x8	Selection to align with model quantization
Type	Classification	This object detection model is a classification type model
System config	PSE8x_U55_400Mhz_SOCMEM_200Mhz_QUAD_XIP or PSE8x_U55_400Mhz_SOCMEM_300Mhz_QUAD_XIP	-
Dataset structure	Random	-
Sample count	>= 10	Recommended to reduce this value to less than 10. A very low value might cause issues during validation
Arena cache size	>= 4000000	Recommended to increase this value to greater than 1000000

After you have generated the code, refer to the Deploy Vision Model on PSOC™ Edge boards to know the deployment steps.