Compilation fails for YOLO11s ONNX model on Hailo8

General
1

Hi Hailo Team

I successfully trained a YOLO11s model using Ultralytics YOLO. I was able to export it to ONNX and run the parse and optimize steps without errors.

However, when I run the compilation step:

hailomz compile yolov11s --hw-arch hailo8 --calib-path ../images --resize 640 640 --classes 1 --har yolov11s.har --performance

I get the following error:

[error] Failed to produce compiled graph
[error] BackendAllocatorException: Compilation failed: map::at

ONNX model was exported with:

  • opset=13

  • imgsz=640

  • simplify=True

  • dynamic=False

I have ~220 calibration images. Parsing and optimization steps complete successfully.

Could you advise how to successfully compile this model for Hailo8?

Thanks in advance!

2

Hi @Yogeesh_Hegde

Welcome to the Hailo community. At DeGirum (a SW partner of Hailo), we developed a cloud compiler to help users convert YOLO checkpoints to hef: Early Access to DeGirum Cloud Compiler. Please see if you find it helpful for your use case.

3

Hi sashi, DeGirum should implement nv12 to rgb and other conversions to be added in the hef.

4

Hi @Frenki

Thanks for the suggestion. However, since such conversion can be done in preprocessor, we are unlikely to add such an option to our cloud compiler.