Yolov8n(s)-pose model doesn't work in runtime environment

General Discussion
1

I’ve been developing a product which uses yolov8n-pose model.

This model is not in Hailo Model Zoo, so I parsed, optimized, and compiled it using .yaml file.

I successfully generated the .hef file. However, when I tried to run inference with this model in python, gstreamer library, I encountered the error message shown below.

Running...
terminate called after throwing an instance of 'std::invalid_argument'
  what():  Number should be between 0.0 to 1.0.
Aborted

For detail, when I first run my gstreamer code, it initially runs well, (the camera window appears well), but as soon as a person appears in the camera’s view, this program is immediately crashes and throws the error.

Runtime Environment:

  • raspberri pi cm5 (with cm5 IO board)
  • hailo-8 M.2 AI Acceleration Module
  • Linux Debian
  • I’m using the official postprocess.so file from Hailo · GitHub

Additionally, when I manually parsed, optimized, and compiled yolov8s-pose model which is in hailo model zoo, it doesn’t work too. However, when I tried to use yolov8s-pose.hef model that is provided by hailo model zoo, It works correctly.

I parsed, optimized, and compiled with below codes.

Parsing

hailomz parse --yaml ./yolov8n_pose.yaml --ckpt ./yolov8n-pose.onnx --hw-arch hailo8

Optimizing

hailo optimize --hw-arch hailo8 --use-random-calib-set ./yolov8n_pose.har

Compiling

hailomz compile --yaml ./yolov8n_pose.yaml --hw-arch hailo8 --har ./yolov8n_pose_optimized.har

yolov8n_pose.yaml

base:
- base/yolov8_pose.yaml
network:
  network_name: yolov8n_pose
paths:
  alls_script: /local/shared_with_docker/yolov8n_pose.alls
  network_path:
  - ./yolov8n-pose.onnx
  url: https://hailo-model-zoo.s3.eu-west-2.amazonaws.com/PoseEstimation/yolov8/yolov8m/pretrained/2023-06-11/yolov8m_pose.zip
parser:
  nodes:
  - null
  - - /model.22/cv2.2/cv2.2.2/Conv
    - /model.22/cv3.2/cv3.2.2/Conv
    - /model.22/cv4.2/cv4.2.2/Conv
    - /model.22/cv2.1/cv2.1.2/Conv
    - /model.22/cv3.1/cv3.1.2/Conv
    - /model.22/cv4.1/cv4.1.2/Conv
    - /model.22/cv2.0/cv2.0.2/Conv
    - /model.22/cv3.0/cv3.0.2/Conv
    - /model.22/cv4.0/cv4.0.2/Conv
info:
  task: pose estimation
  input_shape: 640x640x3
  output_shape: 20x20x64, 20x20x1, 20x20x51, 40x40x64, 40x40x1, 40x40x51, 80x80x64,
    80x80x1, 80x80x51
  operations: 9.2G
  parameters: 3.6M
  framework: pytorch
  training_data: coco keypoints train2017
  validation_data: coco keypoints val2017
  eval_metric: mAP
  full_precision_result: 50.4
  source: https://github.com/ultralytics/ultralytics
  license_url: https://github.com/ultralytics/ultralytics/blob/main/LICENSE
  license_name: AGPL-3.0

and I also have yolov8_pose.yaml, yolov8n_pose.alls

2

Hey @Sunghun_Wang,

This could be an issue with either the model or the post-processing. To help identify where the problem is, can you run these commands and share the results?

hailortcli run {model}
hailortcli parse-hef {model}

If both commands work: The model is probably fine, but you might need to adjust the post-processing. One thing I’d recommend is using a better calibration set instead of random data.

If there’s an issue with the model: Try forcing the output range during optimization. For YOLOv8 models with few classes or specific training characteristics, you may need to set a wider dynamic range on the output nodes.