Issues with creating Hailo files for custom model

General
1

Hi, as the title suggests I’m trying to convert an onnx file to one Hailo can process (.hef)

I’ve looked through many guides, and I keep running into different issues.

In one of the guides I’m following on YouTube it instructs to use the following command

hailomz compile yolov8s --ckpt=YingCans_Best.onnx --hw-arch hailo8l --calib-path train/images --classes 2 --performance

this results in the error shown in image

I’ve seen other users on the board with similar issues and responses without a through guide.
Running hailomz optimize reports layer yolov8n/conv41 doesn’t have one output layer - Community Projects - Hailo Community

if anyone can help, it would be greatly appreciated

error
but I keep running into issues.

2

In the alls file hailo_model_zoo/hailo_model_zoo/cfg/alls/generic/yolov8s.alls at master · hailo-ai/hailo_model_zoo · GitHub, please remove the part “…/…/postprocess_config/yolov8s_nms_config.json” from nms_postprocess() command.
It should be:
nms_postprocess(meta_arch=yolov8, engine=cpu)

if you are using hailo model zoo prior to version 2.14, you may need to edit the end nodes of the yaml file: hailo_model_zoo/hailo_model_zoo/cfg/networks/yolov8s.yaml at master · hailo-ai/hailo_model_zoo · GitHub
But for version 2.14 and onwards, you don’t need to. The hailomz tool automatically discovers the start and end nodes.

3

Thank you, it got me a bit further but still resulted in an error.

“ValueError: failed to initialize intent(inout) array – expected elsize=8 but got 4”

4

hailo_model_optimization.acceleras.utils.acceleras_exceptions.AccelerasUnsupportedError: layer yolov8n/matmul4 does not support shift delta. To overcome this issue you should force larger range at the inputs of the layer using command quantization_param([layer_name], force_range_in=[range_min, range_max], force_range_index=index) current range of input 0 is [0.001, 0.020] and input 1 is [-3.650, 5.073].You should increase the multiplication of these ranges by a factor of 3.237, e.g. you can apply factor of sqrt(3.237) to both inputs:
quantization_param([yolov8n/matmul4], force_range_in=[0.002, 0.036], force_range_index=0)
quantization_param([yolov8n/matmul4], force_range_in=[-6.567, 9.127], force_range_index=1)

this as well

5

It could be an issue with the calibration dataset. Is your dataset made of RGB images in jpeg, png, format ?

6

yes, they are RGB, png images

7

Can you add these lines to the alls file:

quantization_param([yolov8n/matmul4], force_range_in=[0.002, 0.036], force_range_index=0)
quantization_param([yolov8n/matmul4], force_range_in=[-6.567, 9.127], force_range_index=1)

and try again ?