编辑：OAK中国
首发：oakchina.cn
喜欢的话，请多多👍⭐️✍
内容可能会不定期更新，官网内容都是最新的，请查看首发地址链接。

Hello，大家好，这里是OAK中国，我是Ashely。

专注科技，专注分享。

最近真的很忙，已经好久不发博客了。这个月有朋友问怎么在OAK相机上部署yolov9，正好给大家出个教程。

1.其他Yolo转换及使用教程请参考
2.检测类的yolo模型建议使用在线转换（地址），如果在线转换不成功，你再根据本教程来做本地转换。

▌.pt 转换为 .onnx

使用下列脚本(将脚本放到 YOLOv9 根目录中)将 pytorch 模型转换为 onnx 模型，若已安装 openvino_dev，则可进一步转换为 OpenVINO 模型：

示例用法：

python export_onnx.py -w <path_to_model>.pt -imgsz 640 

export_onnx.py :

#!/usr/bin/env python3
# -*- coding:utf-8 -*-
import argparse
import json
import logging
import math
import os
import platform
import sys
import time
import warnings
from io import BytesIO
from pathlib import Path

import torch
from torch import nn

warnings.filterwarnings("ignore")

FILE = Path(__file__).resolve()
ROOT = FILE.parents[0]  # YOLO root directory
if str(ROOT) not in sys.path:
    sys.path.append(str(ROOT))  # add ROOT to PATH
if platform.system() != "Windows":
    ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative

from models.experimental import attempt_load
from models.yolo import DDetect, Detect, DualDDetect, DualDetect, TripleDDetect, TripleDetect
from utils.torch_utils import select_device

try:
    from rich import print
    from rich.logging import RichHandler

    logging.basicConfig(
        level="INFO",
        format="%(message)s",
        datefmt="[%X]",
        handlers=[
            RichHandler(
                rich_tracebacks=False,
                show_path=False,
            )
        ],
    )
except ImportError:
    logging.basicConfig(
        level="INFO",
        format="%(asctime)s\t%(levelname)s\t%(message)s",
        datefmt="[%X]",
    )


class DetectV9(nn.Module):
    """YOLOv9 Detect head for detection models"""

    dynamic = False  # force grid reconstruction
    export = False  # export mode
    shape = None
    anchors = torch.empty(0)  # init
    strides = torch.empty(0)  # init

    def __init__(self, old_detect):
        super().__init__()
        self.nc = old_detect.nc  # number of classes
        self.nl = old_detect.nl  # number of detection layers
        self.reg_max = old_detect.reg_max  # DFL channels (ch[0] // 16 to scale 4/8/12/16/20 for n/s/m/l/x)
        self.no = old_detect.no  # number of outputs per anchor
        self.stride = old_detect.stride  # strides computed during build

        self.cv2 = old_detect.cv2
        self.cv3 = old_detect.cv3
        self.dfl = old_detect.dfl
        self.f = old_detect.f
        self.i = old_detect.i

    def forward(self, x):
        shape = x[0].shape  # BCHW

        d1 = [torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1) for i in range(self.nl)]

        box, cls = torch.cat([xi.view(shape[0], self.no, -1) for xi in d1], 2).split((self.reg_max * 4, self.nc), 1)
        box = self.dfl(box)
        cls_output = cls.sigmoid()
        # Get the max
        conf, _ = cls_output.max(1, keepdim=True)
        # Concat
        y = torch.cat([box, conf, cls_output], dim=1)
        # Split to 3 channels
        outputs = []
        start, end = 0, 0
        for xi in x:
            end += xi.shape[-2] * xi.shape[-1]
            outputs.append(y[:, :, start:end].view(xi.shape[0], -1, xi.shape[-2], xi.shape[-1]))
            start += xi.shape[-2] * xi.shape[-1]

        return outputs

    def bias_init(self):
        # Initialize Detect() biases, WARNING: requires stride availability
        m = self  # self.model[-1]  # Detect() module

        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (.01 objects, 80 classes, 640 img)


class DualDetectV9(DetectV9):
    def __init__(self, old_detect):
        super().__init__(old_detect)

        self.cv4 = old_detect.cv4
        self.cv5 = old_detect.cv5
        self.dfl2 = old_detect.dfl2

    def forward(self, x):
        shape = x[0].shape  # BCHW

        d2 = [torch.cat((self.cv4[i](x[self.nl + i]), self.cv5[i](x[self.nl + i])), 1) for i in range(self.nl)]

        box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
        box2 = self.dfl2(box2)
        cls_output2 = cls2.sigmoid()
        # Get the max
        conf2, _ = cls_output2.max(1, keepdim=True)
        # Concat
        y2 = torch.cat([box2, conf2, cls_output2], dim=1)

        # Split to 3 channels
        outputs2 = []
        start2, end2 = 0, 0
        for _i, xi in enumerate(x[3:]):
            end2 += xi.shape[-2] * xi.shape[-1]
            outputs2.append(y2[:, :, start2:end2].view(xi.shape[0], -1, xi.shape[-2], xi.shape[-1]))
            start2 += xi.shape[-2] * xi.shape[-1]

        return outputs2

    def bias_init(self):
        # Initialize Detect() biases, WARNING: requires stride availability
        m = self  # self.model[-1]  # Detect() module

        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(
                5 / m.nc / (640 / s) ** 2
            )  # cls (5 objects and 80 classes per 640 image)
        for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(
                5 / m.nc / (640 / s) ** 2
            )  # cls (5 objects and 80 classes per 640 image)


class TripleDetectV9(DualDetectV9):
    def __init__(self, old_detect):
        super().__init__(old_detect)

        self.cv6 = old_detect.cv6
        self.cv7 = old_detect.cv7
        self.dfl3 = old_detect.dfl3

    def forward(self, x):
        shape = x[0].shape  # BCHW

        d3 = [
            torch.cat(
                (self.cv6[i](x[self.nl * 2 + i]), self.cv7[i](x[self.nl * 2 + i])),
                1,
            )
            for i in range(self.nl)
        ]

        box3, cls3 = torch.cat([di.view(shape[0], self.no, -1) for di in d3], 2).split((self.reg_max * 4, self.nc), 1)
        box3 = self.dfl3(box3)
        cls_output3 = cls3.sigmoid()
        # Get the max
        conf3, _ = cls_output3.max(1, keepdim=True)
        # Concat
        y3 = torch.cat([box3, conf3, cls_output3], dim=1)

        # Split to 3 channels
        outputs3 = []
        start3, end3 = 0, 0
        for _i, xi in enumerate(x[6:]):
            end3 += xi.shape[-2] * xi.shape[-1]
            outputs3.append(y3[:, :, start3:end3].view(xi.shape[0], -1, xi.shape[-2], xi.shape[-1]))
            start3 += xi.shape[-2] * xi.shape[-1]

        return outputs3

    def bias_init(self):
        # Initialize Detect() biases, WARNING: requires stride availability
        m = self  # self.model[-1]  # Detect() module

        for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(
                5 / m.nc / (640 / s) ** 2
            )  # cls (5 objects and 80 classes per 640 image)
        for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(
                5 / m.nc / (640 / s) ** 2
            )  # cls (5 objects and 80 classes per 640 image)
        for a, b, s in zip(m.cv6, m.cv7, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(
                5 / m.nc / (640 / s) ** 2
            )  # cls (5 objects and 80 classes per 640 image)


def parse_args():
    parser = argparse.ArgumentParser(
        description="Tool for converting Yolov9 models to the blob format used by OAK",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
    )
    parser.add_argument(
        "-m",
        "-i",
        "-w",
        "--input_model",
        type=Path,
        required=True,
        help="weights path",
    )
    parser.add_argument(
        "-imgsz",
        "--img-size",
        nargs="+",
        type=int,
        default=[640, 640],
        help="image size",
    )  # height, width
    parser.add_argument("-op", "--opset", type=int, default=12, help="opset version")

    parser.add_argument(
        "-n",
        "--name",
        type=str,
        help="The name of the model to be saved, none means using the same name as the input model",
    )
    parser.add_argument(
        "-o",
        "--output_dir",
        type=Path,
        help="Directory for saving files, none means using the same path as the input model",
    )
    parser.add_argument(
        "-b",
        "--blob",
        action="store_true",
        help="OAK Blob export",
    )
    parser.add_argument(
        "-s",
        "--spatial_detection",
        action="store_true",
        help="Inference with depth information",
    )
    parser.add_argument(
        "-sh",
        "--shaves",
        type=int,
        help="Inference with depth information",
    )
    parser.add_argument(
        "-t",
        "--convert_tool",
        type=str,
        help="Which tool is used to convert, docker: should already have docker (https://docs.docker.com/get-docker/) and docker-py (pip install docker) installed; blobconverter: uses an online server to convert the model and should already have blobconverter (pip install blobconverter); local: use openvino-dev (pip install openvino-dev) and openvino 2022.1 ( https://docs.oakchina.cn/en/latest /pages/Advanced/Neural_networks/local_convert_openvino.html#id2) to convert",
        default="blobconverter",
        choices=["docker", "blobconverter", "local"],
    )

    args = parser.parse_args()
    args.input_model = args.input_model.resolve().absolute()
    if args.name is None:
        args.name = args.input_model.stem

    if args.output_dir is None:
        args.output_dir = args.input_model.parent

    args.img_size *= 2 if len(args.img_size) == 1 else 1  # expand

    if args.shaves is None:
        args.shaves = 5 if args.spatial_detection else 6

    return args


def export(input_model, img_size, output_model, opset, **kwargs):
    t = time.time()

    # Load PyTorch model
    device = select_device("cpu")
    # load FP32 model
    model = attempt_load(input_model, device=device, inplace=True, fuse=True)
    labels = model.module.names if hasattr(model, "module") else model.names  # get class names
    labels = labels if isinstance(labels, list) else list(labels.values())

    # check num classes and labels
    assert model.nc == len(labels), f"Model class count {model.nc} != len(names) {len(labels)}"

    # Replace with the custom Detection Head
    if isinstance(model.model[-1], (Detect, DDetect)):
        logging.info("Replacing model.model[-1] with DetectV9")
        model.model[-1] = DetectV9(model.model[-1])
    elif isinstance(model.model[-1], (DualDetect, DualDDetect)):
        logging.info("Replacing model.model[-1] with DualDetectV9")
        model.model[-1] = DualDetectV9(model.model[-1])
    elif isinstance(model.model[-1], (TripleDetect, TripleDDetect)):
        logging.info("Replacing model.model[-1] with TripleDetectV9")
        model.model[-1] = TripleDetectV9(model.model[-1])

    num_branches = model.model[-1].nl

    # Input
    img = torch.zeros(1, 3, *img_size).to(device)  # image size(1,3,320,320) Detection

    model.eval()

    model(img)  # dry runs

    # ONNX export
    try:
        import onnx

        print()
        logging.info(f"Starting ONNX export with onnx {onnx.__version__}...")
        output_list = ["output%s_yolov6r2" % (i + 1) for i in range(num_branches)]
        with BytesIO() as f:
            torch.onnx.export(
                model,
                img,
                f,
                verbose=False,
                opset_version=opset,
                input_names=["images"],
                output_names=output_list,
            )

            # Checks
            onnx_model = onnx.load_from_string(f.getvalue())  # load onnx model
            onnx.checker.check_model(onnx_model)  # check onnx model

        try:
            import onnxsim

            logging.info("Starting to simplify ONNX...")
            onnx_model, check = onnxsim.simplify(onnx_model)
            assert check, "assert check failed"

        except ImportError:
            logging.warning(
                "onnxsim is not found, if you want to simplify the onnx, "
                + "you should install it:\n\t"
                + "pip install -U onnxsim onnxruntime\n"
                + "then use:\n\t"
                + f'python -m onnxsim "{output_model}" "{output_model}"'
            )
        except Exception:
            logging.exception("Simplifier failure")

        onnx.save(onnx_model, output_model)
        logging.info(f"ONNX export success, saved as:\n\t{output_model}")

    except Exception:
        logging.exception("ONNX export failure")

    # generate anchors and sides
    anchors = []

    # generate masks
    masks = {}

    logging.info(f"anchors:\n\t{anchors}")
    logging.info(f"anchor_masks:\n\t{masks}")
    export_json = output_model.with_suffix(".json")
    export_json.write_text(
        json.dumps(
            {
                "nn_config": {
                    "output_format": "detection",
                    "NN_family": "YOLO",
                    "input_size": f"{img_size[0]}x{img_size[1]}",
                    "NN_specific_metadata": {
                        "classes": model.nc,
                        "coordinates": 4,
                        "anchors": anchors,
                        "anchor_masks": masks,
                        "iou_threshold": 0.3,
                        "confidence_threshold": 0.5,
                    },
                },
                "mappings": {"labels": labels},
            },
            indent=4,
        )
    )
    logging.info(f"Anchors data export success, saved as:\n\t{export_json}")

    # Finish
    logging.info("Export complete (%.2fs).\n" % (time.time() - t))


def convert(convert_tool, output_model, shaves, output_dir, name, **kwargs):
    t = time.time()

    export_dir: Path = output_dir.joinpath(name + "_openvino")
    export_dir.mkdir(parents=True, exist_ok=True)

    export_xml = export_dir.joinpath(name + ".xml")
    export_blob = export_dir.joinpath(name + ".blob")

    if convert_tool == "blobconverter":
        import blobconverter

        blobconverter.from_onnx(
            model=str(output_model),
            data_type="FP16",
            shaves=shaves,
            use_cache=False,
            version="2021.4",
            output_dir=export_dir,
            optimizer_params=[
                "--scale=255",
                "--reverse_input_channel",
                # "--use_new_frontend",
            ],
            # download_ir=True,
        )
        """
        with ZipFile(blob_path, "r", ZIP_LZMA) as zip_obj:
            for name in zip_obj.namelist():
                zip_obj.extract(
                    name,
                    export_dir,
                )
        blob_path.unlink()
        """
    elif convert_tool == "docker":
        import docker

        export_dir = Path("/io").joinpath(export_dir.name)
        export_xml = export_dir.joinpath(name + ".xml")
        export_blob = export_dir.joinpath(name + ".blob")

        client = docker.from_env()
        image = client.images.pull("openvino/ubuntu20_dev", tag="2022.3.1")
        docker_output = client.containers.run(
            image=image.tags[0],
            command=f"bash -c \"mo -m {name}.onnx -n {name} -o {export_dir} --static_shape --reverse_input_channels --scale=255 --use_new_frontend && echo 'MYRIAD_ENABLE_MX_BOOT NO' | tee /tmp/myriad.conf >> /dev/null && /opt/intel/openvino/tools/compile_tool/compile_tool -m {export_xml} -o {export_blob} -ip U8 -VPU_NUMBER_OF_SHAVES {shaves} -VPU_NUMBER_OF_CMX_SLICES {shaves} -d MYRIAD -c /tmp/myriad.conf\"",
            remove=True,
            volumes=[
                f"{output_dir}:/io",
            ],
            working_dir="/io",
        )
        logging.info(docker_output.decode("utf8"))
    else:
        import subprocess as sp

        # OpenVINO export
        logging.info("Starting to export OpenVINO...")
        OpenVINO_cmd = f"mo --input_model {output_model} --output_dir {export_dir} --data_type FP16 --scale 255 --reverse_input_channel"
        try:
            sp.check_output(OpenVINO_cmd, shell=True)
            logging.info(f"OpenVINO export success, saved as {export_dir}")
        except sp.CalledProcessError:
            logging.exception("")
            logging.warning("OpenVINO export failure!")
            logging.warning(f"By the way, you can try to export OpenVINO use:\n\t{OpenVINO_cmd}")

        # OAK Blob export
        logging.info("Then you can try to export blob use:")
        blob_cmd = (
            "echo 'MYRIAD_ENABLE_MX_BOOT ON' | tee /tmp/myriad.conf"
            + f"compile_tool -m {export_xml} -o {export_blob} -ip U8 -d MYRIAD -VPU_NUMBER_OF_SHAVES {shaves} -VPU_NUMBER_OF_CMX_SLICES {shaves} -c /tmp/myriad.conf"
        )
        logging.info(f"{blob_cmd}")

        logging.info(
            "compile_tool maybe in the path: /opt/intel/openvino/tools/compile_tool/compile_tool, if you install openvino 2022.1 with apt"
        )

    logging.info("Convert complete (%.2fs).\n" % (time.time() - t))


if __name__ == "__main__":
    args = parse_args()
    logging.info(args)
    print()
    output_model = args.output_dir / (args.name + ".onnx")

    export(output_model=output_model, **vars(args))
    if args.blob:
        convert(output_model=output_model, **vars(args))

可以使用 Netron 查看模型结构：

▌转换

openvino 本地转换

onnx -> openvino

mo 是 openvino_dev 2022.1 中脚本，安装命令为 pip install openvino-dev

mo --input_model yolov9-c.onnx --scale=255 --reverse_input_channel

openvino -> blob
compile_tool 是 OpenVINO Runtime 中脚本

<path>/compile_tool -m yolov9-c.xml 
-ip U8 -d MYRIAD 
-VPU_NUMBER_OF_SHAVES 6 
-VPU_NUMBER_OF_CMX_SLICES 6

在线转换

blobconvert 网页 http://blobconverter.luxonis.com/

• 进入网页，按下图指示操作：
  

• 修改参数，转换模型：
  

1. 选择 onnx 模型
2. 修改 optimizer_params 为 --data_type=FP16 --scale=255 --reverse_input_channel
3. 修改 shaves 为 6
4. 转换

blobconverter python 代码:

blobconverter.from_onnx(
            "yolov9-c.onnx",	
            optimizer_params=[
                "--scale=255",
                "--reverse_input_channel",
            ],
            shaves=6,
        )

blobconvert cli

blobconverter --onnx yolov9-c.onnx -sh 6 -o . --optimizer-params "scale=255 --reverse_input_channel"

▌DepthAI 示例

正确解码需要可配置的网络相关参数：

• setNumClasses – YOLO 检测类别的数量
• setIouThreshold – iou 阈值
• setConfidenceThreshold – 置信度阈值，低于该阈值的对象将被过滤掉

# coding=utf-8
import cv2
import depthai as dai
import numpy as np

numClasses = 80
model = dai.OpenVINO.Blob("yolov9-c.blob")
dim = next(iter(model.networkInputs.values())).dims
W, H = dim[:2]

output_name, output_tenser = next(iter(model.networkOutputs.items()))
if "yolov6" in output_name:
    numClasses = output_tenser.dims[2] - 5
else:
    numClasses = output_tenser.dims[2] // 3 - 5

labelMap = [
    # "class_1","class_2","..."
    "class_%s" % i
    for i in range(numClasses)
]

# Create pipeline
pipeline = dai.Pipeline()

# Define sources and outputs
camRgb = pipeline.create(dai.node.ColorCamera)
detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)
xoutRgb = pipeline.create(dai.node.XLinkOut)
xoutNN = pipeline.create(dai.node.XLinkOut)

xoutRgb.setStreamName("image")
xoutNN.setStreamName("nn")

# Properties
camRgb.setPreviewSize(W, H)
camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
camRgb.setInterleaved(False)
camRgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.BGR)

# Network specific settings
detectionNetwork.setBlob(model)
detectionNetwork.setConfidenceThreshold(0.5)

# Yolo specific parameters
detectionNetwork.setNumClasses(numClasses)
detectionNetwork.setCoordinateSize(4)
detectionNetwork.setAnchors([])
detectionNetwork.setAnchorMasks({})
detectionNetwork.setIouThreshold(0.5)

# Linking
camRgb.preview.link(detectionNetwork.input)
camRgb.preview.link(xoutRgb.input)
detectionNetwork.out.link(xoutNN.input)

# Connect to device and start pipeline
with dai.Device(pipeline) as device:
    # Output queues will be used to get the rgb frames and nn data from the outputs defined above
    imageQueue = device.getOutputQueue(name="image", maxSize=4, blocking=False)
    detectQueue = device.getOutputQueue(name="nn", maxSize=4, blocking=False)

    frame = None
    detections = []

    # nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/height
    def frameNorm(frame, bbox):
        normVals = np.full(len(bbox), frame.shape[0])
        normVals[::2] = frame.shape[1]
        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)

    def drawText(frame, text, org, color=(255, 255, 255), thickness=1):
        cv2.putText(
            frame, text, org, cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), thickness + 3, cv2.LINE_AA
        )
        cv2.putText(
            frame, text, org, cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, thickness, cv2.LINE_AA
        )

    def drawRect(frame, topLeft, bottomRight, color=(255, 255, 255), thickness=1):
        cv2.rectangle(frame, topLeft, bottomRight, (0, 0, 0), thickness + 3)
        cv2.rectangle(frame, topLeft, bottomRight, color, thickness)

    def displayFrame(name, frame):
        color = (128, 128, 128)
        for detection in detections:
            bbox = frameNorm(
                frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax)
            )
            drawText(
                frame=frame,
                text=labelMap[detection.label],
                org=(bbox[0] + 10, bbox[1] + 20),
            )
            drawText(
                frame=frame,
                text=f"{detection.confidence:.2%}",
                org=(bbox[0] + 10, bbox[1] + 35),
            )
            drawRect(
                frame=frame,
                topLeft=(bbox[0], bbox[1]),
                bottomRight=(bbox[2], bbox[3]),
                color=color,
            )
        # Show the frame
        cv2.imshow(name, frame)

    while True:
        imageQueueData = imageQueue.tryGet()
        detectQueueData = detectQueue.tryGet()

        if imageQueueData is not None:
            frame = imageQueueData.getCvFrame()

        if detectQueueData is not None:
            detections = detectQueueData.detections

        if frame is not None:
            displayFrame("rgb", frame)

        if cv2.waitKey(1) == ord("q"):
            break

▌参考资料

https://docs.oakchina.cn/en/latest/
https://www.oakchina.cn/selection-guide/

---

OAK中国
| OpenCV AI Kit在中国区的官方代理商和技术服务商
| 追踪AI技术和产品新动态

戳「+关注」获取最新资讯↗↗