Estoy utilizando el Hailo AI Software Suite versión 2025.04.

General
,
1

I’m trying to compile a custom YOLOv8m model (trained with Ultralytics) in HEF format.
I exported my PyTorch model (best.pt) to ONNX (best.onnx) using Ultralytics with imgsz=640 and opset=11 (simplify=True, dynamic=False), as suggested in their GitHub guide ‘YOLOv8 Retraining’.

My goal is to obtain a best.hef with three classes.

I’ve tried the hailo parser → hailo optimize → hailo compile script and the hailomz compile command, following their documentation and making logical inferences based on the CLI usage prompts.

The persistent errors are:

Hailo optimize fails with: hailo_sdk_client.sdk_backend.sdk_backend_exceptions.AllocatorScriptParserException: The given layers best/conv71 do not exist in the HN (when using /local/workspace/hailo_model_zoo/hailo_model_zoo/cfg/alls/generic/yolov8m.alls).

Hailomz compile fails with: hailomz compile: error: argument model_name: invalid option: ‘/local/workspace/best.hef’ (when I try to specify the input ONNX and output HEF, even though their GitHub guide shows it).

I need the exact command-line syntax or an example script (Python or .alls) that is 100% compatible with version 2025.04 of my Hailo AI Software Suite to compile a custom 3-class YOLOv8m model. Could you please provide me with the direct solution for this use case?

2

Welcome to the Hailo Community!

If you want everything easy train the YOLOv8 model using the Model Zoo instructions:

GitHub - Hailo Model Zoo - Training - Yolov8

This will create an ONNX file that is exactly as the hailomz command expects it.

Otherwise, I you will need to convert the model with the model standard model conversion flow. I recommend starting with the tutorials build into the Hailo AI Software Suite Docker. Run the following command:

hailo tutorial

This will start a Jupyter Notebook server explaining you each step in the process. To pick the right start- and end-nodes you can use Netron to view the ONNX or TFLite files.

Netron.app