Compiling custom YOLO model on Hailo 8.

General
1

Hello,

I am using the Raspberry Pi 5 with the AI Hat+ and am doing a custom YOLO model implementation. I have a couple questions regarding what is possible with the Hailo 8.

  1. Looking at examples in the hailo_model_zoo/training folder on the github, I see that in order to compile and genrate a .HEF, we need to “Choose the corresponding YAML from our networks configuration directory, i.e. hailo_model_zoo/cfg/networks/yolov5s.yaml, and run compilation using the model zoo”(README in the yolov5 implementation). Opening the networks/yolov5s.yaml file, I see that this is different than a classic .yaml file for yolov5s. How does one adapt a truly custom model architecture to this?

  2. Also, I noticed that the in the benchmarks for the YOLO models in the model zoo, the largest input resolution I saw was 1280x1280. Does anyone know the maximum resolution that the Hailo 8 AI Hat has supported? Something tells me it isn’t 1280x1280, because they were able to run it with batch size 8 at 11fps. Has anyone got it to work with a higher resolution than this?

2

Welcome to the Hailo Community!

If you’re working with a fully custom model, it’s best to use the Hailo Dataflow Compiler directly rather than using the Model Zoo. This gives you greater flexibility and control over model conversion and integration into your workflow.

I recommend going through the built-in tutorials included in the Hailo AI Software Suite Docker. You can launch a Jupyter notebook server with step-by-step workflow examples using the following command:

hailo tutorial

Can you clarify your concern? Most models operate on input sizes smaller than the raw resolution of camera sensors.

If you’re working with high-resolution images and need to detect small objects, consider using tiling. We provide tiling examples in our GitHub Tappas repository:

GitHub - Tappas - Tiling