Help with compile of Custom yolov8s_pose

General Discussion
,
1

Hello, lately im trying to create a HEF model using my yolov8s_pose PT, im using this ALL file:
normalization1 = normalization([0.0, 0.0, 0.0], [255.0, 255.0, 255.0])

#change_output_activation(conv71, sigmoid)
#change_output_activation(conv58, sigmoid)
#change_output_activation(conv44, sigmoid)

pre_quantization_optimization(equalization, policy=enabled)

quantization_param(conv71, precision_mode = a16_w16)
quantization_param(conv44, precision_mode = a16_w16)

quantization_param(conv45, precision_mode=a16_w16)
quantization_param(conv59, precision_mode=a16_w16)
quantization_param(conv72, precision_mode=a16_w16)

model_optimization_flavor(optimization_level=3)

post_quantization_optimization(finetune, policy=enabled, learning_rate=0.00015, epochs=20)

Also using this YAML:

base:

  • base/yolov8_pose.yaml
    network:
    network_name: yolov8s_pose
    paths:
    alls_script: yolov8s_pose.alls
    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: 480x480x3
        output_shape: 15x15x64, 15x15x1, 15x15x15, 30x30x64, 30x30x1, 30x30x15, 60x60x64, 60x60x1, 60x60x15

        My calib files are my training data, can be a imagem with no objects.

        But when i test the compiled HEF, i got poor results, often not even detecting any keypoints, im using the code from Hailo-Application-Code in github, just modifiyng everything to be 5 keypoints instead of the default 17, how can i discover if the problem is in this code, or let my HEF file be as close as possible from my PT file.
        Thanks in advance

2

Hi @Thiago_Medeiros thanks for sharing details. If you want to skip local toolchain setup, the DeGirum Cloud Compiler can compile your Ultralytics YOLO PyTorch checkpoint (.pt) for {{Hailo-8 or Hailo-8L}}). It’s a fast way to validate performance and iterate. It’s currently in early access but you can request access here: https://forms.degirum.com/cloud-compiler