Create a custom .hef

Community Projects
,
1

Hi everyone,

I’m looking for a stereo depth model that can run in real-time (100–200 ms) on a Raspberry Pi 5 with a Hailo-8*accelerator.

I tried the stereo model from the hailo_model_zoo, but inference takes about 800 ms which is too slow for my use case.

So I started testing other stereo depth models like FastACVNet, CREStereo, and HitNet using the following conversion code:

import os
import onnx
from onnxsim import simplify
import numpy as np
from hailo_sdk_client import ClientRunner
from PIL import Image

def parse_stereo_onnx(onnx_path: str, left_dir: str, right_dir: str, net_name: str, hw_arch: str, target_size=(640, 480)):
    # -------------------------------
    # Step 1. Simplify the ONNX model
    # -------------------------------
    model = onnx.load(onnx_path)
    model_simp, check = simplify(model)
    if not check:
        raise RuntimeError("Simplified ONNX model validation failed.")
        
    simplified_onnx_path = os.path.splitext(onnx_path)[0] + "_simplified.onnx"
    onnx.save(model_simp, simplified_onnx_path)
    print(f"[✓] Simplified ONNX model saved to: {simplified_onnx_path}")

    # -----------------------------------------------------
    # Step 2. Translate the simplified ONNX model to Hailo format
    # -----------------------------------------------------
    runner = ClientRunner(hw_arch=hw_arch)

    # Précise les noms des entrées et sorties comme dans le YAML
    input_names = ["left_image", "right_image"]
    output_names = ["output"]

    hn, params = runner.translate_onnx_model(simplified_onnx_path, net_name,
                                             input_names=input_names,
                                             outputs_names=output_names)
    print("[✓] Model translation to Hailo format completed.")

    # -----------------------------------------------------
    # Step 3. Load stereo calibration dataset
    # -----------------------------------------------------
    calib_dataset = load_stereo_calib_dataset(left_dir, right_dir, target_size)
    print("[✓] Calibration dataset prepared.")

    # -----------------------------------------------------
    # Step 4. Optimize the model (quantization)
    # -----------------------------------------------------
    runner.optimize(calib_dataset)
    print("[✓] Model quantization complete.")

    # -----------------------------------------------------
    # Step 5. Save HAR file
    # -----------------------------------------------------
    har_file = f"{net_name}_quantized.har"
    runner.save_har(har_file)
    print(f"[✓] HAR file saved to: {har_file}")

def load_stereo_calib_dataset(left_dir, right_dir, target_size=(640, 480)):
    """
    Charge et redimensionne les images gauche/droite pour la calibration.
    Retourne un dict: { "left_image": [...], "right_image": [...] }
    """
    left_imgs = sorted([f for f in os.listdir(left_dir) if f.lower().endswith(('.jpg', '.png'))])
    right_imgs = sorted([f for f in os.listdir(right_dir) if f.lower().endswith(('.jpg', '.png'))])

    assert len(left_imgs) == len(right_imgs), "Nombre d'images gauche/droite différent"
    images_left = []
    images_right = []

    for l_img, r_img in zip(left_imgs, right_imgs):
        l_path = os.path.join(left_dir, l_img)
        r_path = os.path.join(right_dir, r_img)

        left = Image.open(l_path).convert("RGB").resize(target_size, Image.BILINEAR)
        right = Image.open(r_path).convert("RGB").resize(target_size, Image.BILINEAR)

        images_left.append(np.array(left, dtype=np.float32))
        images_right.append(np.array(right, dtype=np.float32))

    return {
        "left_image": np.stack(images_left),
        "right_image": np.stack(images_right)
    }

def main():
    onnx_path = "./model/crestereo_combined_iter10_480x640.onnx"
    calib_left = "./calibration/left"
    calib_right = "./calibration/right"
    net_name = "stereo_depth_net"
    hw_arch = "hailo8l"

    parse_stereo_onnx(onnx_path, calib_left, calib_right, net_name, hw_arch)

if __name__ == "__main__":
    main()

and i have some issue like

[warning] This model has non-default (reflective/edge) padding layers which are not supported currently, and were replaced with zero padding.

or

ValueError: axes don't match array

This seems to occur during model translation. Has anyone encountered and resolved this when working with stereo ONNX models on Hailo?
I try

hailo tutorial

to get some informations but i get :slight_smile:

[C 2025-06-11 10:26:16.209 ServerApp] Running as root is not recommended. Use --allow-root to bypass.

After this, the Jupyter server crashes. Is there a clean way to launch the tutorial environment inside Docker?

Any suggestions, tips, or shared experiences would be greatly appreciated!

Thanks in advance!