The accuracy of the custom-trained Yolov8m model is low

General
1

Hello,

I trained a Yolov8m model on custom data, but the compiled result shows lower detection accuracy compared to the original pt file.

  1. yolo detect train data=/workspace/ultralytics/ultralytics/datasets/custom-data.yaml model=yolov8m.pt name=retrain_yolov8m_batch8 epochs=100 batch=8 cache=True
  2. yolo export model=/workspace/ultralytics/runs/detect/retrain_yolov8m_batch8/weights/best.pt imgsz=640 format=onnx opset=11
  3. hailomz compile --ckpt /local/workspace/hailo_model_zoo/best.onnx --calib-path /local/shared_with_docker/dataset/cal --yaml /local/workspace/hailo_model_zoo/hailo_model_zoo/cfg/networks/yolov8m.yaml --classes 1 --hw-arch hailo8

I used the above commands.

The calibration data path is set to a directory with around 1000 images, and I did not modify the yolov8m.yaml file.

Do the .har and .hef files need to be in the same directory?

Are there any important settings or modifications I should make?

2

Hi @sidedoor.kim,

We have a section for debugging accuracy in our guide.

Particularly for yolov8, I would recommend implementing this solution too.