Extracting detections from yolov8s HAR model

General
1

Hi everyone!

I’m following the ‘Quick Optimization Tutorial‘ from the ‘hailo_dataflow_compiler_v3.27.0_user_guide.pdf‘ to evaluate/improve my model accuracy.

My code (adapted from the tutorial’s code) uses ClientRunner to get inferences from my HAR model, as follows:

model_name = ‘yolov8s.har’
runner = ClientRunner(har=model_name)
with runner.infer_context(InferenceContext.SDK_NATIVE) as ctx:
native_res = runner.infer(ctx, image_dataset[:IMAGES_TO_VISUALIZE, :,:, :])

Checking native_res, I realized that the model outputs the state of the penultimate layers [‘yolov8s/conv41’,’yolov8s/conv42’, ‘yolov8s/conv52’, ‘yolov8s/conv53’,’yolov8s/conv62’,’yolov8s/conv63’] instead of the last layer ‘yolov8s/yolov8_nms_postprocess‘.

This is the mentioned .har drawn on https://netron.app/:

image
image1037×317 33.8 KB

This is the content of yolov8s.alls:

normalization1 = normalization([0.0, 0.0, 0.0], [255.0, 255.0, 255.0])
change_output_activation(conv42, sigmoid)
change_output_activation(conv53, sigmoid)
change_output_activation(conv63, sigmoid)
nms_postprocess(“../../postprocess_config/yolov8s_nms_config.json”, meta_arch=yolov8, engine=cpu)

I found this post, where it is said that we need to implement the postprocess layer to obtain the proper detections. It is not clear to me why, because the .har model already has the postprocess layer.

Can someone help me extract the proper detections from the ClientRunner.inferoutput using a yolov8 model?

2

Hey @Luiz_doleron,

You’re getting those head outputs because infer() just gives you whatever comes directly out of the neural network itself.

That yolov8_nms_postprocess layer you mentioned in the .alls file isn’t actually part of the network - it’s a separate post-processing step that runs on the CPU after the main inference is done. The Model-Zoo runner handles this automatically, but when you call ClientRunner.infer(...), you’re only getting the raw outputs from the YOLOv8 head layers (those conv41/42, conv52/53, conv62/63 tensors).

This is actually the expected behavior. To get your final detection results, you’ll need to run that post-processing step that’s defined in your .alls file on these raw outputs.

So your workflow should be: preprocessing → inference → post-processing.