OAK相机如何将yoloV5lite模型转换成blob格式？

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。

▌前言

Hello，大家好，这里是OAK中国，我是助手君。

最近咱社群里有几个朋友在将yolox转换成blob的过程有点不清楚，所以我就写了这篇博客。（请夸我贴心！咱的原则：合理要求，有求必应！）

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 YOLOv5 lite 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 320 

export_onnx.py :

# coding=utf-8
import argparse
from io import BytesIO
import json
import logging
import sys
import time
import warnings
from pathlib import Path

warnings.filterwarnings("ignore")

FILE = Path(__file__).resolve()
ROOT = FILE.parents[0]
if str(ROOT) not in sys.path:
    sys.path.append(str(ROOT))  # add ROOT to PATH

import torch
import torch.nn as nn

from models.common import Conv
from models.experimental import attempt_load
from utils.activations import Hardswish, SiLU
from utils.general import check_img_size
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=True,
                show_path=False,
            )
        ],
    )
except ImportError:
    logging.basicConfig(
        level="INFO",
        format="%(message)s",
        datefmt="[%X]",
    )


class DetectV5(nn.Module):
    # YOLOv5 Detect head for detection models
    dynamic = False  # force grid reconstruction
    export = True  # export mode

    def __init__(self, old_detect):  # detection layer
        super().__init__()
        self.nc = old_detect.nc  # number of classes
        self.no = old_detect.no  # number of outputs per anchor
        self.nl = old_detect.nl  # number of detection layers
        self.na = old_detect.na
        self.anchors = old_detect.anchors
        self.grid = old_detect.grid  # [torch.zeros(1)] * self.nl
        self.anchor_grid = old_detect.anchor_grid  # anchor grid

        self.stride = old_detect.stride
        if hasattr(old_detect, "inplace"):
            self.inplace = old_detect.inplace

        self.f = old_detect.f
        self.i = old_detect.i
        self.m = old_detect.m

    def forward(self, x):
        for i in range(self.nl):
            x[i] = self.m[i](x[i])  # conv
            x[i] = x[i].sigmoid()
        return x


def parse_args():
    parser = argparse.ArgumentParser(
        description="Tool for converting YOLOv5-Lite 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")
    model = attempt_load(input_model, map_location=device)  # load FP32 model
    labels = model.names
    labels = labels if isinstance(labels, list) else list(labels.values())

    # Checks
    gs = int(max(model.stride))  # grid size (max stride)
    img_size = [
        check_img_size(x, gs) for x in img_size
    ]  # verify img_size are gs-multiples

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

    # Update model
    model.eval()
    for k, m in model.named_modules():
        if isinstance(m, Conv):  # assign export-friendly activations
            m._non_persistent_buffers_set = set()  # torch 1.6.0 compatibility
            if isinstance(m.act, nn.SiLU):
                m.act = SiLU()
            if isinstance(m.act, nn.Hardswish):
                m.act = Hardswish()
        elif isinstance(m, nn.Upsample):
            m.recompute_scale_factor = None  # torch 1.11.0 compatibility

    model.model[-1] = DetectV5(model.model[-1])

    m = model.module.model[-1] if hasattr(model, "module") else model.model[-1]
    num_branches = len(m.anchor_grid)

    y = model(img)  # dry runs

    # ONNX export
    try:
        import onnx

        print()
        logging.info("Starting ONNX export with onnx %s..." % onnx.__version__)
        output_list = ["output%s_yolov5" % (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("ONNX export success, saved as:\n\t%s" % output_model)

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

    # generate anchors and sides
    anchors, sides = [], []
    m = model.module.model[-1] if hasattr(model, "module") else model.model[-1]
    for i in range(num_branches):
        sides.append(int(img_size[0] // m.stride[i]))
        for j in range(m.anchor_grid[i].size()[1]):
            anchors.extend(m.anchor_grid[i][0, j, 0, 0].numpy())
    anchors = [float(x) for x in anchors]

    # generate masks
    masks = dict()
    # for i, num in enumerate(sides[::-1]):
    for i, num in enumerate(sides):
        masks[f"side{num}"] = list(range(i * 3, i * 3 + 3))

    logging.info("anchors:\n\t%s" % anchors)
    logging.info("anchor_masks:\n\t%s" % 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": 0.5,
                        "confidence_threshold": 0.5,
                    },
                },
                "mappings": {"labels": labels},
            },
            indent=4,
        )
    )
    logging.info("Anchors data export success, saved as:\n\t%s" % export_json)

    # Finish
    logging.info("Export complete (%.2fs)." % (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
        from zipfile import ZIP_LZMA, ZipFile

        blob_path = blobconverter.from_onnx(
            model=str(output_model),
            data_type="FP16",
            shaves=shaves,
            use_cache=False,
            version="2022.1",
            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,
                    output_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="2022.1.0")
        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 = (
            "mo --input_model %s --output_dir %s --data_type FP16 --scale 255 --reverse_input_channel"
            % (output_model, export_dir)
        )
        try:
            sp.check_output(OpenVINO_cmd, shell=True)
            logging.info("OpenVINO export success, saved as %s" % export_dir)
        except Exception:
            logging.exception("")
            logging.warning("OpenVINO export failure!")
            logging.warning(
                "By the way, you can try to export OpenVINO use:\n\t%s" % 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"
            + "compile_tool -m %s -o %s -ip U8 -d MYRIAD -VPU_NUMBER_OF_SHAVES %s -VPU_NUMBER_OF_CMX_SLICES %s -c /tmp/myriad.conf"
            % (export_xml, export_blob, shaves, shaves)
        )
        logging.info("%s" % 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()
    print(args)
    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 v5lite.onnx  --scale 255 --reverse_input_channel

openvino -> blob

<path>/compile_tool -m v5lite.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(
            "v5lite.onnx",	
            optimizer_params=[
                " --scale 255",
                "--reverse_input_channel",
            ],
            shaves=6,
        )

blobconvert cli

blobconverter --onnx v5lite.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：

使用 export_onnx.py 转换模型时会将相关参数写入 json 文件中，可根据 json 文件中数据添加下列参数

• 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

• setAnchors - yolo 锚点

• setAnchorMasks - 锚掩码

Anchors:

训练模型时 cfg 中的 anchors，例如:

[10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326]

是从 v5Lite-e.yaml 中 获取

AnchorMasks :

如果使用不同的输入宽度，还应该重新设置sideX , sideY, sideZ, 其中 X = width/8 , Y = width/16，Z = width/32 。如果您使用的是微型（tiny）模型，那么只要设置sideX , sideY，其中 X = width/16 , Y = width/32。

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

numClasses = 80
model = dai.OpenVINO.Blob("v5lite.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(
    [
    10,13, 16,30, 33,23,
    30,61, 62,45, 59,119,
    116,90, 156,198, 373,326
    ]
)
detectionNetwork.setAnchorMasks(
    {
        "side%s" % (W // 8): [0, 1, 2],
        "side%s" % (W // 16): [3, 4, 5],
        "side%s" % (W // 32): [6, 7, 8],
    }
)
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↗↗