深度学习模型之yolov8实例分割模型TesorRT部署-python版本

1 模型转换

从github上下载官方yolov8版本，当前使用的版本是2023年9月份更新的版本，作者一直在更新。官网地址

2 加载模型

模型的训练和测试在官方文档上，有详细的说明，yolov8中文文档这里不做过多说明，v8现在训练是真的方便了，环境部署好之后，几行代码就跑起来了，例如：

from ultralytics import YOLO
from setproctitle import setproctitle
setproctitle("python|yolov8-seg 20231211")

# 载入一个模型
model = YOLO('yolov8n-seg.yaml')  # 从YAML构建一个新模型
# model = YOLO('/data/siping/preweights/yolov8n-seg.pt')    # 载入预训练模型（推荐用于训练）
# model = YOLO('yolov8m-seg.yaml').load('/data/siping/preweights/yolov8m.pt')  # 从YAML构建并传递权重

# 训练模型
model.train(task='segment',
            data='/ultralytics/cfg/datasets/coco128-seg.yaml ',
            epochs=200,
            batch=32,
            imgsz=320,
            device=1,
            translate=0
           )  # 训练模型

model.val()

# Export the model
model.export(format='onnx',imgsz=(320,320),opset=12)

模型训练完成后，对保存下来的best.pt模型转换为onnx模型，由于后期会部署到NX和其他不同类型的显卡上，因此先转换为通用的onnx模型。导出onnx代码：

# Export the model
model.export(format='onnx',imgsz=(384,640),opset=12)

导出的操作可以直接在训练的代码最后面加入一行即可。

onnx模型导出成功后，将模型拷贝到需要运行的机器上，使用tensorrt转换为trt模型，后缀可以是trt或者engine，转换代码这里总结了两个版本的转换代码：
Tensotrt 7.1：

# -*- coding: utf-8 -*- 
import tensorrt as trt
import sys
import os

TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)

def printShape(engine):
    for i in range(engine.num_bindings):
        if engine.binding_is_input(i):
            print("input layer: {}, shape is: {} ".format(i, engine.get_binding_shape(i)))
        else:
            print("output layer: {} shape is: {} ".format(i, engine.get_binding_shape(i)))

def onnx2trt(onnx_path, engine_path):
    with trt.Builder(TRT_LOGGER) as builder, builder.create_network(EXPLICIT_BATCH) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
        builder.max_workspace_size = 1 << 28 # 256MB
        #builder.set_fp16_mode(True)     
        #builder.precision = trt.BuilderFlag.FP16    
        #builder.config.set_flag(trt.BuilderFlag.FP16)           
        with open(onnx_path, 'rb') as model:
            parser.parse(model.read())
        engine = builder.build_cuda_engine(network)

        printShape(engine)            

        with open(engine_path, "wb") as f:
            f.write(engine.serialize())

if __name__ == "__main__":                             
    #onnx文件路径设置
    model_name = "20240115_yolov8Seg_Person"
    input_path = model_name+".onnx"
    #引擎文件保存路径设置
    root_dir = "/ai/good_model/"
    output_path = root_dir + model_name + '.engine'
    onnx2trt(input_path, output_path)

一般是在NX上做模型转换，NX初始化的tensorrt版本较低，这个代码成功导出和部署使用。

tensorrt8.4.0.6版本转换代码：

# 导入必用依赖
import tensorrt as trt
#onnx文件路径设置
onnx_path="./repvgg.onnx"
#引擎文件保存路径设置
engine_path=r"classify.engine"
# 创建logger：日志记录器
logger = trt.Logger(trt.Logger.WARNING)
 
# 创建构建器builder
builder = trt.Builder(logger)
# 预创建网络
network = builder.create_network(1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
# 加载onnx解析器
parser = trt.OnnxParser(network, logger)
success = parser.parse_from_file(onnx_path)
for idx in range(parser.num_errors):
    print(parser.get_error(idx))
if not success:
    print("failed----------------")
    pass  # Error handling code here
    
# builder配置
config = builder.create_builder_config()
# 分配显存作为工作区间，一般建议为显存一半的大小
config.max_workspace_size = 12 << 30  # 1 Mi
builder.max_batch_size = 1  

   
serialized_engine = builder.build_serialized_network(network, config)
# 序列化生成engine文件
with open(engine_path, "wb") as f:
    f.write(serialized_engine)
    print("generate file success!")

这个可以成功导出8.4以上版本的onnx模型，包括CNN的分类模型等。

如果是在windows下做模型导出和测试，可以直接去官网下载一个Tensorrt对应版本的包，直接使用trtexec导出也可以。

3 算法部署

yolov8的分割模型部署代码参考yolov8 github的测试代码

3.1 加载trt模型

加载分割模型的trt模型需要注意的是模型的输入和输出的名称是否对应，可以将onnx模型通过netron查看模型的结构：

3.2 部署代码

部署是加载分割模型的代码：

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import cv2
import os
import torch
import numpy as np
import tensorrt as trt
from utils import ops

def select_device(device='', apex=False, batch_size=None):
    # device = 'cpu' or '0' or '0,1,2,3'
    cpu_request = str(device).lower() == 'cpu'
    if str(device) and not cpu_request:  # if device requested other than 'cpu'
        os.environ['CUDA_VISIBLE_DEVICES'] = str(device)  # set environment variable
        assert torch.cuda.is_available(), 'CUDA unavailable, invalid device %s requested' % device  # check availablity

    cuda = False if cpu_request else torch.cuda.is_available()
    if cuda:
        c = 1024 ** 2  # bytes to MB
        ng = torch.cuda.device_count()
        if ng > 1 and batch_size:  # check that batch_size is compatible with device_count
            assert batch_size % ng == 0, 'batch-size %g not multiple of GPU count %g' % (batch_size, ng)
        x = [torch.cuda.get_device_properties(i) for i in range(ng)]
        s = 'Using CUDA ' + ('Apex ' if apex else '')  # apex for mixed precision https://github.com/NVIDIA/apex
        for i in range(0, ng):
            if i == 1:
                s = ' ' * len(s)
            print("%sdevice%g _CudaDeviceProperties(name='%s', total_memory=%dMB)" %
                  (s, i, x[i].name, x[i].total_memory / c))
    else:
        print('Using CPU')

    print('')  # skip a line
    return torch.device('cuda:0' if cuda else 'cpu')

def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better val mAP)
        r = min(r, 1.0)

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    if auto:  # minimum rectangle
        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
    elif scaleFill:  # stretch
        dw, dh = 0.0, 0.0
        new_unpad = (new_shape[1], new_shape[0])
        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios

    dw /= 2  # divide padding into 2 sides
    dh /= 2

    if shape[::-1] != new_unpad:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return im, ratio, (dw, dh)


DEVICE = select_device('0')
SEGLABELS = ['dbseg']
SEGWEIGHT = '20240115_yolov8Seg_lift.engine'

class TrtModelSeg():
    def __init__(self, engine_path, device,shapes=(1, 3, 320, 320)):
        """
        :param engine_path: Trt model path
        :param device: torch.device
        """
        self.device = device
        from collections import OrderedDict, namedtuple
        Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
        logger = trt.Logger(trt.Logger.INFO)
        with open(engine_path, 'rb') as f, trt.Runtime(logger) as runtime:
            model = runtime.deserialize_cuda_engine(f.read())
        self.context = model.create_execution_context()
        self.bindings = OrderedDict()
        dynamic = False
        self.input_name = model.get_binding_name(0)
        self.output_name = model.get_binding_name(1)
        self.output_name1 = model.get_binding_name(2)
        for index in range(model.num_bindings):
            name = model.get_binding_name(index)
            dtype = trt.nptype(model.get_binding_dtype(index))
            # print(model.get_binding_shape(index))
            if -1 in tuple(model.get_binding_shape(index)) and index == 0:
                self.context.set_binding_shape(index, shapes)
            shape = tuple(self.context.get_binding_shape(index))
            data = torch.from_numpy(np.empty(shape, dtype=np.dtype(dtype))).to(select_device(device))
            self.bindings[name] = Binding(name, dtype, shape, data, int(data.data_ptr()))
        self.binding_addrs = OrderedDict((n, d.ptr) for n, d in self.bindings.items())

        self.batch_size = self.bindings[self.input_name].shape[0]

    def __call__(self, img):
        """
        :param img: 输入Tensor 图片
        :return:
        """
        # assert img.shape == self.bindings['input'].shape, (img.shape, self.bindings['input'].shape)
        if img.device == torch.device('cpu'):
            img = img.float().to(select_device(self.device))
        # print('in trt dynamic def----input shape is ----',img.shape)
        self.binding_addrs[self.input_name] = int(img.data_ptr())
        self.context.execute_v2(list(self.binding_addrs.values()))
        return [self.bindings['output0'].data, self.bindings['output1'].data]

# 定义图片预处理函数
def get_input_img(image, img_size=[320, 320]):
    # Padded resize
    img,ratio,(dw,dh) = letterbox(image, img_size, stride=32, auto=False)
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    # img = img.transpose((2, 0, 1))[::-1]
    img = np.ascontiguousarray(img)

    img = torch.from_numpy(img).to(DEVICE)
    img = img.float()  # uint8 to fp16/32
    img /= 255.0  # 0 - 255 to 0.0 - 1.0
    if img.ndimension() == 3:
        img = img.unsqueeze(0)
    return img,ratio,(dw,dh)

#分割模型
class YoloSeg(object):
    def __init__(self,classes=[]):
        self.conf_thr = 0.4  # conf阈值
        self.iou_thr = 0.4  # iou阈值
        self.model = self._load_model()
        try:
            self.labels = self.model.labels
            print("Seg Using meta model ！ ")
        except:
            self.labels = SEGLABELS
            print("Seg Using no meta model ！ ")
        self.classes = self._get_display_class(classes) if classes and isinstance(classes, list) else None
        self._warm_up(size=[320, 320])

    def _load_model(self):
        # Load seg model
        print("Load seg model！ ",SEGWEIGHT)
        self.model = TrtModelSeg(SEGWEIGHT, DEVICE)

        return self.model

    def _get_display_class(self, classes):
        """只计算classes中包含的类别

        Args:
            classes (list): 只显示classes中包含的类别，其他类别将被遗弃
        """
        return [self.labels.index(c) for c in classes if c in self.labels]

    def _warm_up(self, size=[320, 320]):
        x = torch.randn((1, 3, *size)).float().to(DEVICE)
        self.model(x)

    #参考yolov8 segment/predict.py文件
    def detect_image(self, image):
        # 预处理
        img, ratio, (dw, dh) = get_input_img(image)
        # Inference
        preds = self.model(img)

        detpred = preds[0]

        masks = image
        #1.经过置信度和非极大值筛选的数值 NMS
        pred_nms = ops.non_max_suppression(detpred, self.conf_thr, self.iou_thr,nc=len(SEGLABELS))
        retina_masks = True
        proto = preds[1][-1] if len(preds[1]) == 3 else preds[1]  # second output is len 3 if pt, but only 1 if exported
        # print("proto: ",len(preds[1]),proto.shape)
        print("pred_nms: ",len(pred_nms))
        #从预测输出中提取结果
        #pred_nms是掩码信息，如果有多个mask，选择面积最大的mask，因为是吊臂
        for i, pred in enumerate(pred_nms):
            print("len(pred): ", len(pred))

            #获取原图像的大小
            shape = image.shape
            # print("image: ",shape)
            if not len(pred):  # save empty boxes
                masks = None
            elif retina_masks:
                n_class = int(pred[:,5].item())
                print("n_class: ",n_class,SEGLABELS[n_class])
                #如果标签不等于db
                if(n_class!=0):
                    continue
                #对框坐标进行处理。
                pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], shape)
                # print("pred[:, :4]: ",pred[:, :4])
                # 对分割结果进行后处理，输入的是proto、pred:,6:,pred[:,:4]为框坐标img.shape[:2]为（320,320）四个数据
                s_masks = ops.process_mask_native(proto[i], pred[:, 6:], pred[:, :4], shape[:2])  # HWC

                # 从(batch_size, height, width)的形状转换为(height, width, batch_size)的形状，并将结果存储在NumPy数组中。
                s_masks = (s_masks.cpu().numpy() * 255).astype(np.uint8)
                masks = s_masks.transpose(1, 2, 0)

            else:
                masks = ops.process_mask(proto[i], pred[:, 6:], pred[:, :4], img.shape[2:], upsample=True)  # HWC
                pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], shape)

            print("pred,masks: ", masks.shape)

        #返回预测结果和mask
        return pred_nms,masks

def detect(yolo, image_path):
    image = cv2.imdecode(np.fromfile(image_path, dtype=np.uint8), cv2.IMREAD_COLOR)
    pred_nms, masks = yolo.detect_image(image)
    return pred_nms, masks


if __name__ == '__main__':
    image_path = r"1.jpg"
    model = TrtModelSeg(SEGWEIGHT, DEVICE,shapes=(1, 3, 320, 320))
    yoloseg = YoloSeg(SEGLABELS)
    pred_nms, masks = detect(yoloseg,image_path)
    cv2.imwrite("mask.jpg",masks)

ops.py文件：

# -*- coding: utf-8 -*-
# Ultralytics YOLO 🚀, AGPL-3.0 license

import contextlib
import math
import re
import time

import cv2
import numpy as np
import torch
import torch.nn.functional as F
import torchvision

class Profile(contextlib.ContextDecorator):
    """
    YOLOv8 Profile class. Use as a decorator with @Profile() or as a context manager with 'with Profile():'.

    Example:
        ```python
        from ultralytics.utils.ops import Profile

        with Profile() as dt:
            pass  # slow operation here

        print(dt)  # prints "Elapsed time is 9.5367431640625e-07 s"
        ```
    """

    def __init__(self, t=0.0):
        """
        Initialize the Profile class.

        Args:
            t (float): Initial time. Defaults to 0.0.
        """
        self.t = t
        self.cuda = torch.cuda.is_available()

    def __enter__(self):
        """Start timing."""
        self.start = self.time()
        return self

    def __exit__(self, type, value, traceback):  # noqa
        """Stop timing."""
        self.dt = self.time() - self.start  # delta-time
        self.t += self.dt  # accumulate dt

    def __str__(self):
        """Returns a human-readable string representing the accumulated elapsed time in the profiler."""
        return f'Elapsed time is {self.t} s'

    def time(self):
        """Get current time."""
        if self.cuda:
            torch.cuda.synchronize()
        return time.time()


def segment2box(segment, width=640, height=640):
    """
    Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy).

    Args:
        segment (torch.Tensor): the segment label
        width (int): the width of the image. Defaults to 640
        height (int): The height of the image. Defaults to 640

    Returns:
        (np.ndarray): the minimum and maximum x and y values of the segment.
    """
    # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy)
    x, y = segment.T  # segment xy
    inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height)
    x, y, = x[inside], y[inside]
    return np.array([x.min(), y.min(), x.max(), y.max()], dtype=segment.dtype) if any(x) else np.zeros(
        4, dtype=segment.dtype)  # xyxy


def scale_boxes(img1_shape, boxes, img0_shape, ratio_pad=None, padding=True):
    """
    Rescales bounding boxes (in the format of xyxy) from the shape of the image they were originally specified in
    (img1_shape) to the shape of a different image (img0_shape).

    Args:
        img1_shape (tuple): The shape of the image that the bounding boxes are for, in the format of (height, width).
        boxes (torch.Tensor): the bounding boxes of the objects in the image, in the format of (x1, y1, x2, y2)
        img0_shape (tuple): the shape of the target image, in the format of (height, width).
        ratio_pad (tuple): a tuple of (ratio, pad) for scaling the boxes. If not provided, the ratio and pad will be
            calculated based on the size difference between the two images.
        padding (bool): If True, assuming the boxes is based on image augmented by yolo style. If False then do regular
            rescaling.

    Returns:
        boxes (torch.Tensor): The scaled bounding boxes, in the format of (x1, y1, x2, y2)
    """
    if ratio_pad is None:  # calculate from img0_shape
        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
        pad = round((img1_shape[1] - img0_shape[1] * gain) / 2 - 0.1), round(
            (img1_shape[0] - img0_shape[0] * gain) / 2 - 0.1)  # wh padding
    else:
        gain = ratio_pad[0][0]
        pad = ratio_pad[1]

    if padding:
        boxes[..., [0, 2]] -= pad[0]  # x padding
        boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    return clip_boxes(boxes, img0_shape)


def make_divisible(x, divisor):
    """
    Returns the nearest number that is divisible by the given divisor.

    Args:
        x (int): The number to make divisible.
        divisor (int | torch.Tensor): The divisor.

    Returns:
        (int): The nearest number divisible by the divisor.
    """
    if isinstance(divisor, torch.Tensor):
        divisor = int(divisor.max())  # to int
    return math.ceil(x / divisor) * divisor


#NMS
def non_max_suppression(
        prediction,
        conf_thres=0.25,
        iou_thres=0.45,
        classes=None,
        agnostic=False,
        multi_label=False,
        labels=(),
        max_det=300,
        nc=0,  # number of classes (optional)
        max_time_img=0.05,
        max_nms=30000,
        max_wh=7680,
):
    """
    Perform non-maximum suppression (NMS) on a set of boxes, with support for masks and multiple labels per box.

    Args:
        prediction (torch.Tensor): A tensor of shape (batch_size, num_classes + 4 + num_masks, num_boxes)
            containing the predicted boxes, classes, and masks. The tensor should be in the format
            output by a model, such as YOLO.
        conf_thres (float): The confidence threshold below which boxes will be filtered out.
            Valid values are between 0.0 and 1.0.
        iou_thres (float): The IoU threshold below which boxes will be filtered out during NMS.
            Valid values are between 0.0 and 1.0.
        classes (List[int]): A list of class indices to consider. If None, all classes will be considered.
        agnostic (bool): If True, the model is agnostic to the number of classes, and all
            classes will be considered as one.
        multi_label (bool): If True, each box may have multiple labels.
        labels (List[List[Union[int, float, torch.Tensor]]]): A list of lists, where each inner
            list contains the apriori labels for a given image. The list should be in the format
            output by a dataloader, with each label being a tuple of (class_index, x1, y1, x2, y2).
        max_det (int): The maximum number of boxes to keep after NMS.
        nc (int, optional): The number of classes output by the model. Any indices after this will be considered masks.
        max_time_img (float): The maximum time (seconds) for processing one image.
        max_nms (int): The maximum number of boxes into torchvision.ops.nms().
        max_wh (int): The maximum box width and height in pixels

    Returns:
        (List[torch.Tensor]): A list of length batch_size, where each element is a tensor of
            shape (num_boxes, 6 + num_masks) containing the kept boxes, with columns
            (x1, y1, x2, y2, confidence, class, mask1, mask2, ...).
    """

    # Checks


    #先判断设置的置信度和iou阈值是否在0和1之间，如不在，则报错是不符合要求的阈值。
    assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
    assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'

    #判读那输入的prediction类型，如果之前没有进行提取，则此处再进行提取，如提取过，则直接跳过。

    if isinstance(prediction, (list, tuple)):  # YOLOv8 model in validation model, output = (inference_out, loss_out)
        prediction = prediction[0]  # select only inference output


    bs = prediction.shape[0]  # batch size
    nc = nc or (prediction.shape[1] - 4)  # number of classes 获取类别数目nc，本文是1，其中nm为32是mask的数量
    nm = prediction.shape[1] - nc - 4
    mi = 4 + nc  # mask start index mi为读取mask开始的位置，即数组前边是框、类别置信度，然后是mask
    xc = prediction[:, 4:mi].amax(1) > conf_thres
    # candidates xc为根据置信度分数从结果中筛选结果，即分数大于置信度的为True，小于的为False

    # Settings
    # min_wh = 2  # (pixels) minimum box width and height
    time_limit = 0.5 + max_time_img * bs  # seconds to quit after
    multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)

    prediction = prediction.transpose(-1, -2)  # shape(1,84,6300) to shape(1,6300,84)
    prediction[..., :4] = xywh2xyxy(prediction[..., :4])  # xywh to xyxy

    t = time.time()
    output = [torch.zeros((0, 6 + nm), device=prediction.device)] * bs
    for xi, x in enumerate(prediction):  # image index, image inference
        # Apply constraints
        # x[((x[:, 2:4] < min_wh) | (x[:, 2:4] > max_wh)).any(1), 4] = 0  # width-height
        x = x[xc[xi]]  # confidence

        # Cat apriori labels if autolabelling
        if labels and len(labels[xi]):
            lb = labels[xi]
            v = torch.zeros((len(lb), nc + nm + 4), device=x.device)
            v[:, :4] = xywh2xyxy(lb[:, 1:5])  # box
            v[range(len(lb)), lb[:, 0].long() + 4] = 1.0  # cls
            x = torch.cat((x, v), 0)

        # If none remain process next image
        if not x.shape[0]:
            continue

        # Detections matrix nx6 (xyxy, conf, cls)
        box, cls, mask = x.split((4, nc, nm), 1)

        if multi_label:
            i, j = torch.where(cls > conf_thres)
            x = torch.cat((box[i], x[i, 4 + j, None], j[:, None].float(), mask[i]), 1)
        else:  # best class only
            conf, j = cls.max(1, keepdim=True)
            x = torch.cat((box, conf, j.float(), mask), 1)[conf.view(-1) > conf_thres]

        # Filter by class
        if classes is not None:
            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]

        # Check shape
        n = x.shape[0]  # number of boxes
        if not n:  # no boxes
            continue
        if n > max_nms:  # excess boxes
            x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence and remove excess boxes

        # Batched NMS
        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
        i = i[:max_det]  # limit detections

        # # Experimental
        # merge = False  # use merge-NMS
        # if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
        #     # Update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
        #     from .metrics import box_iou
        #     iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
        #     weights = iou * scores[None]  # box weights
        #     x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
        #     redundant = True  # require redundant detections
        #     if redundant:
        #         i = i[iou.sum(1) > 1]  # require redundancy

        output[xi] = x[i]
        if (time.time() - t) > time_limit:
            print(f'WARNING ?? NMS time limit {time_limit:.3f}s exceeded')
            break  # time limit exceeded
    """
    到此处，此幅图像经过了置信度阈值判断和非极大值抑制处理，确定了图像中的目标类别、分数和框的位置。
    """
    return output


def clip_boxes(boxes, shape):
    """
    Takes a list of bounding boxes and a shape (height, width) and clips the bounding boxes to the shape.

    Args:
        boxes (torch.Tensor): the bounding boxes to clip
        shape (tuple): the shape of the image

    Returns:
        (torch.Tensor | numpy.ndarray): Clipped boxes
    """
    if isinstance(boxes, torch.Tensor):  # faster individually (WARNING: inplace .clamp_() Apple MPS bug)
        boxes[..., 0] = boxes[..., 0].clamp(0, shape[1])  # x1
        boxes[..., 1] = boxes[..., 1].clamp(0, shape[0])  # y1
        boxes[..., 2] = boxes[..., 2].clamp(0, shape[1])  # x2
        boxes[..., 3] = boxes[..., 3].clamp(0, shape[0])  # y2
    else:  # np.array (faster grouped)
        boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1])  # x1, x2
        boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0])  # y1, y2
    return boxes


def clip_coords(coords, shape):
    """
    Clip line coordinates to the image boundaries.

    Args:
        coords (torch.Tensor | numpy.ndarray): A list of line coordinates.
        shape (tuple): A tuple of integers representing the size of the image in the format (height, width).

    Returns:
        (torch.Tensor | numpy.ndarray): Clipped coordinates
    """
    if isinstance(coords, torch.Tensor):  # faster individually (WARNING: inplace .clamp_() Apple MPS bug)
        coords[..., 0] = coords[..., 0].clamp(0, shape[1])  # x
        coords[..., 1] = coords[..., 1].clamp(0, shape[0])  # y
    else:  # np.array (faster grouped)
        coords[..., 0] = coords[..., 0].clip(0, shape[1])  # x
        coords[..., 1] = coords[..., 1].clip(0, shape[0])  # y
    return coords


def scale_image(masks, im0_shape, ratio_pad=None):
    """
    Takes a mask, and resizes it to the original image size.

    Args:
        masks (np.ndarray): resized and padded masks/images, [h, w, num]/[h, w, 3].
        im0_shape (tuple): the original image shape
        ratio_pad (tuple): the ratio of the padding to the original image.

    Returns:
        masks (torch.Tensor): The masks that are being returned.
    """
    # Rescale coordinates (xyxy) from im1_shape to im0_shape
    im1_shape = masks.shape
    if im1_shape[:2] == im0_shape[:2]:
        return masks
    if ratio_pad is None:  # calculate from im0_shape
        gain = min(im1_shape[0] / im0_shape[0], im1_shape[1] / im0_shape[1])  # gain  = old / new
        pad = (im1_shape[1] - im0_shape[1] * gain) / 2, (im1_shape[0] - im0_shape[0] * gain) / 2  # wh padding
    else:
        gain = ratio_pad[0][0]
        pad = ratio_pad[1]
    top, left = (int(round(pad[1] - 0.1)), int(round(pad[0] - 0.1)))  # y, x
    bottom, right = (int(round(im1_shape[0] - pad[1] + 0.1)), int(round(im1_shape[1] - pad[0] + 0.1)))

    if len(masks.shape) < 2:
        raise ValueError(f'"len of masks shape" should be 2 or 3, but got {len(masks.shape)}')
    masks = masks[top:bottom, left:right]
    masks = cv2.resize(masks, (im0_shape[1], im0_shape[0]))
    if len(masks.shape) == 2:
        masks = masks[:, :, None]

    return masks


def xyxy2xywh(x):
    """
    Convert bounding box coordinates from (x1, y1, x2, y2) format to (x, y, width, height) format where (x1, y1) is the
    top-left corner and (x2, y2) is the bottom-right corner.

    Args:
        x (np.ndarray | torch.Tensor): The input bounding box coordinates in (x1, y1, x2, y2) format.

    Returns:
        y (np.ndarray | torch.Tensor): The bounding box coordinates in (x, y, width, height) format.
    """
    assert x.shape[-1] == 4, f'input shape last dimension expected 4 but input shape is {x.shape}'
    y = torch.empty_like(x) if isinstance(x, torch.Tensor) else np.empty_like(x)  # faster than clone/copy
    y[..., 0] = (x[..., 0] + x[..., 2]) / 2  # x center
    y[..., 1] = (x[..., 1] + x[..., 3]) / 2  # y center
    y[..., 2] = x[..., 2] - x[..., 0]  # width
    y[..., 3] = x[..., 3] - x[..., 1]  # height
    return y


def xywh2xyxy(x):
    """
    Convert bounding box coordinates from (x, y, width, height) format to (x1, y1, x2, y2) format where (x1, y1) is the
    top-left corner and (x2, y2) is the bottom-right corner.

    Args:
        x (np.ndarray | torch.Tensor): The input bounding box coordinates in (x, y, width, height) format.

    Returns:
        y (np.ndarray | torch.Tensor): The bounding box coordinates in (x1, y1, x2, y2) format.
    """
    assert x.shape[-1] == 4, f'input shape last dimension expected 4 but input shape is {x.shape}'
    y = torch.empty_like(x) if isinstance(x, torch.Tensor) else np.empty_like(x)  # faster than clone/copy
    dw = x[..., 2] / 2  # half-width
    dh = x[..., 3] / 2  # half-height
    y[..., 0] = x[..., 0] - dw  # top left x
    y[..., 1] = x[..., 1] - dh  # top left y
    y[..., 2] = x[..., 0] + dw  # bottom right x
    y[..., 3] = x[..., 1] + dh  # bottom right y
    return y


def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0):
    """
    Convert normalized bounding box coordinates to pixel coordinates.

    Args:
        x (np.ndarray | torch.Tensor): The bounding box coordinates.
        w (int): Width of the image. Defaults to 640
        h (int): Height of the image. Defaults to 640
        padw (int): Padding width. Defaults to 0
        padh (int): Padding height. Defaults to 0
    Returns:
        y (np.ndarray | torch.Tensor): The coordinates of the bounding box in the format [x1, y1, x2, y2] where
            x1,y1 is the top-left corner, x2,y2 is the bottom-right corner of the bounding box.
    """
    assert x.shape[-1] == 4, f'input shape last dimension expected 4 but input shape is {x.shape}'
    y = torch.empty_like(x) if isinstance(x, torch.Tensor) else np.empty_like(x)  # faster than clone/copy
    y[..., 0] = w * (x[..., 0] - x[..., 2] / 2) + padw  # top left x
    y[..., 1] = h * (x[..., 1] - x[..., 3] / 2) + padh  # top left y
    y[..., 2] = w * (x[..., 0] + x[..., 2] / 2) + padw  # bottom right x
    y[..., 3] = h * (x[..., 1] + x[..., 3] / 2) + padh  # bottom right y
    return y


def xyxy2xywhn(x, w=640, h=640, clip=False, eps=0.0):
    """
    Convert bounding box coordinates from (x1, y1, x2, y2) format to (x, y, width, height, normalized) format. x, y,
    width and height are normalized to image dimensions.

    Args:
        x (np.ndarray | torch.Tensor): The input bounding box coordinates in (x1, y1, x2, y2) format.
        w (int): The width of the image. Defaults to 640
        h (int): The height of the image. Defaults to 640
        clip (bool): If True, the boxes will be clipped to the image boundaries. Defaults to False
        eps (float): The minimum value of the box's width and height. Defaults to 0.0

    Returns:
        y (np.ndarray | torch.Tensor): The bounding box coordinates in (x, y, width, height, normalized) format
    """
    if clip:
        x = clip_boxes(x, (h - eps, w - eps))
    assert x.shape[-1] == 4, f'input shape last dimension expected 4 but input shape is {x.shape}'
    y = torch.empty_like(x) if isinstance(x, torch.Tensor) else np.empty_like(x)  # faster than clone/copy
    y[..., 0] = ((x[..., 0] + x[..., 2]) / 2) / w  # x center
    y[..., 1] = ((x[..., 1] + x[..., 3]) / 2) / h  # y center
    y[..., 2] = (x[..., 2] - x[..., 0]) / w  # width
    y[..., 3] = (x[..., 3] - x[..., 1]) / h  # height
    return y


def xywh2ltwh(x):
    """
    Convert the bounding box format from [x, y, w, h] to [x1, y1, w, h], where x1, y1 are the top-left coordinates.

    Args:
        x (np.ndarray | torch.Tensor): The input tensor with the bounding box coordinates in the xywh format

    Returns:
        y (np.ndarray | torch.Tensor): The bounding box coordinates in the xyltwh format
    """
    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
    y[..., 0] = x[..., 0] - x[..., 2] / 2  # top left x
    y[..., 1] = x[..., 1] - x[..., 3] / 2  # top left y
    return y


def xyxy2ltwh(x):
    """
    Convert nx4 bounding boxes from [x1, y1, x2, y2] to [x1, y1, w, h], where xy1=top-left, xy2=bottom-right.

    Args:
        x (np.ndarray | torch.Tensor): The input tensor with the bounding boxes coordinates in the xyxy format

    Returns:
        y (np.ndarray | torch.Tensor): The bounding box coordinates in the xyltwh format.
    """
    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
    y[..., 2] = x[..., 2] - x[..., 0]  # width
    y[..., 3] = x[..., 3] - x[..., 1]  # height
    return y


def ltwh2xywh(x):
    """
    Convert nx4 boxes from [x1, y1, w, h] to [x, y, w, h] where xy1=top-left, xy=center.

    Args:
        x (torch.Tensor): the input tensor

    Returns:
        y (np.ndarray | torch.Tensor): The bounding box coordinates in the xywh format.
    """
    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
    y[..., 0] = x[..., 0] + x[..., 2] / 2  # center x
    y[..., 1] = x[..., 1] + x[..., 3] / 2  # center y
    return y


def xyxyxyxy2xywhr(corners):
    """
    Convert batched Oriented Bounding Boxes (OBB) from [xy1, xy2, xy3, xy4] to [xywh, rotation].

    Args:
        corners (numpy.ndarray | torch.Tensor): Input corners of shape (n, 8).

    Returns:
        (numpy.ndarray | torch.Tensor): Converted data in [cx, cy, w, h, rotation] format of shape (n, 5).
    """
    is_numpy = isinstance(corners, np.ndarray)
    atan2, sqrt = (np.arctan2, np.sqrt) if is_numpy else (torch.atan2, torch.sqrt)

    x1, y1, x2, y2, x3, y3, x4, y4 = corners.T
    cx = (x1 + x3) / 2
    cy = (y1 + y3) / 2
    dx21 = x2 - x1
    dy21 = y2 - y1

    w = sqrt(dx21 ** 2 + dy21 ** 2)
    h = sqrt((x2 - x3) ** 2 + (y2 - y3) ** 2)

    rotation = atan2(-dy21, dx21)
    rotation *= 180.0 / math.pi  # radians to degrees

    return np.vstack((cx, cy, w, h, rotation)).T if is_numpy else torch.stack((cx, cy, w, h, rotation), dim=1)


def xywhr2xyxyxyxy(center):
    """
    Convert batched Oriented Bounding Boxes (OBB) from [xywh, rotation] to [xy1, xy2, xy3, xy4].

    Args:
        center (numpy.ndarray | torch.Tensor): Input data in [cx, cy, w, h, rotation] format of shape (n, 5).

    Returns:
        (numpy.ndarray | torch.Tensor): Converted corner points of shape (n, 8).
    """
    is_numpy = isinstance(center, np.ndarray)
    cos, sin = (np.cos, np.sin) if is_numpy else (torch.cos, torch.sin)

    cx, cy, w, h, rotation = center.T
    rotation *= math.pi / 180.0  # degrees to radians

    dx = w / 2
    dy = h / 2

    cos_rot = cos(rotation)
    sin_rot = sin(rotation)
    dx_cos_rot = dx * cos_rot
    dx_sin_rot = dx * sin_rot
    dy_cos_rot = dy * cos_rot
    dy_sin_rot = dy * sin_rot

    x1 = cx - dx_cos_rot - dy_sin_rot
    y1 = cy + dx_sin_rot - dy_cos_rot
    x2 = cx + dx_cos_rot - dy_sin_rot
    y2 = cy - dx_sin_rot - dy_cos_rot
    x3 = cx + dx_cos_rot + dy_sin_rot
    y3 = cy - dx_sin_rot + dy_cos_rot
    x4 = cx - dx_cos_rot + dy_sin_rot
    y4 = cy + dx_sin_rot + dy_cos_rot

    return np.vstack((x1, y1, x2, y2, x3, y3, x4, y4)).T if is_numpy else torch.stack(
        (x1, y1, x2, y2, x3, y3, x4, y4), dim=1)


def ltwh2xyxy(x):
    """
    It converts the bounding box from [x1, y1, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right.

    Args:
        x (np.ndarray | torch.Tensor): the input image

    Returns:
        y (np.ndarray | torch.Tensor): the xyxy coordinates of the bounding boxes.
    """
    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
    y[..., 2] = x[..., 2] + x[..., 0]  # width
    y[..., 3] = x[..., 3] + x[..., 1]  # height
    return y


def segments2boxes(segments):
    """
    It converts segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh)

    Args:
        segments (list): list of segments, each segment is a list of points, each point is a list of x, y coordinates

    Returns:
        (np.ndarray): the xywh coordinates of the bounding boxes.
    """
    boxes = []
    for s in segments:
        x, y = s.T  # segment xy
        boxes.append([x.min(), y.min(), x.max(), y.max()])  # cls, xyxy
    return xyxy2xywh(np.array(boxes))  # cls, xywh


def resample_segments(segments, n=1000):
    """
    Inputs a list of segments (n,2) and returns a list of segments (n,2) up-sampled to n points each.

    Args:
        segments (list): a list of (n,2) arrays, where n is the number of points in the segment.
        n (int): number of points to resample the segment to. Defaults to 1000

    Returns:
        segments (list): the resampled segments.
    """
    for i, s in enumerate(segments):
        s = np.concatenate((s, s[0:1, :]), axis=0)
        x = np.linspace(0, len(s) - 1, n)
        xp = np.arange(len(s))
        segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)],
                                     dtype=np.float32).reshape(2, -1).T  # segment xy
    return segments


#裁剪掩码mask 输入的是masks和downsampled_bboxes
def crop_mask(masks, boxes):
    """
    It takes a mask and a bounding box, and returns a mask that is cropped to the bounding box.

    Args:
        masks (torch.Tensor): [n, h, w] tensor of masks
        boxes (torch.Tensor): [n, 4] tensor of bbox coordinates in relative point form

    Returns:
        (torch.Tensor): The masks are being cropped to the bounding box.



    确保masks的值在图像大小范围之内。

    masks.gt_（0.5）判断masks中的值是否大于0.5，大于则为true。
    """

    n, h, w = masks.shape
    # print("n, h, w :",masks.shape)
    #然后对boxes的内容进行拆分，torch.chunk（），拆分为四个，x1, y1, x2, y2
    x1, y1, x2, y2 = torch.chunk(boxes[:, :, None], 4, 1)  # x1 shape(n,1,1)
    r = torch.arange(w, device=masks.device, dtype=x1.dtype)[None, None, :]  # rows shape(1,1,w)
    c = torch.arange(h, device=masks.device, dtype=x1.dtype)[None, :, None]  # cols shape(1,h,1)

    return masks * ((r >= x1) * (r < x2) * (c >= y1) * (c < y2))


def process_mask_upsample(protos, masks_in, bboxes, shape):
    """
    Takes the output of the mask head, and applies the mask to the bounding boxes. This produces masks of higher quality
    but is slower.

    Args:
        protos (torch.Tensor): [mask_dim, mask_h, mask_w]
        masks_in (torch.Tensor): [n, mask_dim], n is number of masks after nms
        bboxes (torch.Tensor): [n, 4], n is number of masks after nms
        shape (tuple): the size of the input image (h,w)

    Returns:
        (torch.Tensor): The upsampled masks.
    """
    c, mh, mw = protos.shape  # CHW
    masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)
    masks = F.interpolate(masks[None], shape, mode='bilinear', align_corners=False)[0]  # CHW
    masks = crop_mask(masks, bboxes)  # CHW

    #判断masks中的值是否大于0.5，大于则为true
    return masks.gt_(0.5)


def process_mask(protos, masks_in, bboxes, shape, upsample=False):
    """
    Apply masks to bounding boxes using the output of the mask head.

    Args:
        protos (torch.Tensor): A tensor of shape [mask_dim, mask_h, mask_w].
        masks_in (torch.Tensor): A tensor of shape [n, mask_dim], where n is the number of masks after NMS.
        bboxes (torch.Tensor): A tensor of shape [n, 4], where n is the number of masks after NMS.
        shape (tuple): A tuple of integers representing the size of the input image in the format (h, w).
        upsample (bool): A flag to indicate whether to upsample the mask to the original image size. Default is False.

    Returns:
        (torch.Tensor): A binary mask tensor of shape [n, h, w], where n is the number of masks after NMS, and h and w
            are the height and width of the input image. The mask is applied to the bounding boxes.
    """

    c, mh, mw = protos.shape  # CHW
    ih, iw = shape
    masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)  # CHW

    downsampled_bboxes = bboxes.clone()
    downsampled_bboxes[:, 0] *= mw / iw
    downsampled_bboxes[:, 2] *= mw / iw
    downsampled_bboxes[:, 3] *= mh / ih
    downsampled_bboxes[:, 1] *= mh / ih

    masks = crop_mask(masks, downsampled_bboxes)  # CHW
    if upsample:
        masks = F.interpolate(masks[None], shape, mode='bilinear', align_corners=False)[0]  # CHW
    return masks.gt_(0.5)


def process_mask_native(protos, masks_in, bboxes, shape):
    """
    It takes the output of the mask head, and crops it after upsampling to the bounding boxes.

    Args:
        protos (torch.Tensor): [mask_dim, mask_h, mask_w]
        masks_in (torch.Tensor): [n, mask_dim], n is number of masks after nms
        bboxes (torch.Tensor): [n, 4], n is number of masks after nms
        shape (tuple): the size of the input image (h,w)

    Returns:
        masks (torch.Tensor): The returned masks with dimensions [h, w, n]



    """
    c, mh, mw = protos.shape  # CHW
    masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)
    masks = scale_masks(masks[None], shape)[0]  # CHW
    masks = crop_mask(masks, bboxes)  # CHW
    return masks.gt_(0.5)


def scale_masks(masks, shape, padding=True):
    """
    Rescale segment masks to shape.

    Args:
        masks (torch.Tensor): (N, C, H, W).
        shape (tuple): Height and width.
        padding (bool): If True, assuming the boxes is based on image augmented by yolo style. If False then do regular
            rescaling.
    """
    mh, mw = masks.shape[2:]
    gain = min(mh / shape[0], mw / shape[1])  # gain  = old / new
    pad = [mw - shape[1] * gain, mh - shape[0] * gain]  # wh padding
    if padding:
        pad[0] /= 2
        pad[1] /= 2
    top, left = (int(round(pad[1] - 0.1)), int(round(pad[0] - 0.1))) if padding else (0, 0)  # y, x
    bottom, right = (int(round(mh - pad[1] + 0.1)), int(round(mw - pad[0] + 0.1)))
    masks = masks[..., top:bottom, left:right]

    masks = F.interpolate(masks, shape, mode='bilinear', align_corners=False)  # NCHW
    return masks


def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None, normalize=False, padding=True):
    """
    Rescale segment coordinates (xy) from img1_shape to img0_shape.

    Args:
        img1_shape (tuple): The shape of the image that the coords are from.
        coords (torch.Tensor): the coords to be scaled of shape n,2.
        img0_shape (tuple): the shape of the image that the segmentation is being applied to.
        ratio_pad (tuple): the ratio of the image size to the padded image size.
        normalize (bool): If True, the coordinates will be normalized to the range [0, 1]. Defaults to False.
        padding (bool): If True, assuming the boxes is based on image augmented by yolo style. If False then do regular
            rescaling.

    Returns:
        coords (torch.Tensor): The scaled coordinates.
    """
    if ratio_pad is None:  # calculate from img0_shape
        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
    else:
        gain = ratio_pad[0][0]
        pad = ratio_pad[1]

    if padding:
        coords[..., 0] -= pad[0]  # x padding
        coords[..., 1] -= pad[1]  # y padding
    coords[..., 0] /= gain
    coords[..., 1] /= gain
    coords = clip_coords(coords, img0_shape)
    if normalize:
        coords[..., 0] /= img0_shape[1]  # width
        coords[..., 1] /= img0_shape[0]  # height
    return coords


def masks2segments(masks, strategy='largest'):
    """
    It takes a list of masks(n,h,w) and returns a list of segments(n,xy)

    Args:
        masks (torch.Tensor): the output of the model, which is a tensor of shape (batch_size, 160, 160)
        strategy (str): 'concat' or 'largest'. Defaults to largest

    Returns:
        segments (List): list of segment masks
    """
    segments = []
    for x in masks.int().cpu().numpy().astype('uint8'):
        c = cv2.findContours(x, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
        if c:
            if strategy == 'concat':  # concatenate all segments
                c = np.concatenate([x.reshape(-1, 2) for x in c])
            elif strategy == 'largest':  # select largest segment
                c = np.array(c[np.array([len(x) for x in c]).argmax()]).reshape(-1, 2)
        else:
            c = np.zeros((0, 2))  # no segments found
        segments.append(c.astype('float32'))
    return segments


def convert_torch2numpy_batch(batch: torch.Tensor) -> np.ndarray:
    """
    Convert a batch of FP32 torch tensors (0.0-1.0) to a NumPy uint8 array (0-255), changing from BCHW to BHWC layout.

    Args:
        batch (torch.Tensor): Input tensor batch of shape (Batch, Channels, Height, Width) and dtype torch.float32.

    Returns:
        (np.ndarray): Output NumPy array batch of shape (Batch, Height, Width, Channels) and dtype uint8.
    """
    return (batch.permute(0, 2, 3, 1).contiguous() * 255).clamp(0, 255).to(torch.uint8).cpu().numpy()


def clean_str(s):
    """
    Cleans a string by replacing special characters with underscore _

    Args:
        s (str): a string needing special characters replaced

    Returns:
        (str): a string with special characters replaced by an underscore _
    """
    return re.sub(pattern='[|@#!?·$€%&()=??^*;:,¨′><+]', repl='_', string=s)


#mask处理
def masks(img,masks, colors, im_gpu, alpha=0.5, retina_masks=False):
    """save masks at once.
    Args:
        masks (tensor): predicted masks on cuda, shape: [n, h, w]
        colors (List[List[Int]]): colors for predicted masks, [[r, g, b] * n]
        im_gpu (tensor): img is in cuda, shape: [3, h, w], range: [0, 1]
        alpha (float): mask transparency: 0.0 fully transparent, 1.0 opaque
    """
    if len(masks) == 0:
       img = im_gpu.permute(1, 2, 0).contiguous().cpu().numpy() * 255
    colors = torch.tensor(colors, device=im_gpu.device, dtype=torch.float32) / 255.0
    colors = colors[:, None, None]  # shape(n,1,1,3)
    masks = masks.unsqueeze(3)  # shape(n,h,w,1)
    masks_color = masks * (colors * alpha)  # shape(n,h,w,3)

    inv_alph_masks = (1 - masks * alpha).cumprod(0)  # shape(n,h,w,1)
    mcs = (masks_color * inv_alph_masks).sum(0) * 2  # mask color summand shape(n,h,w,3)

    im_gpu = im_gpu.flip(dims=[0])  # flip channel
    im_gpu = im_gpu.permute(1, 2, 0).contiguous()  # shape(h,w,3)
    im_gpu = im_gpu * inv_alph_masks[-1] + mcs
    im_mask = (im_gpu * 255)
    im_mask_np = im_mask.byte().cpu().numpy()
    img = im_mask_np if retina_masks else scale_image(im_gpu.shape, im_mask_np, img.shape)

其中上面代码中的opt是官方代码，在models/utils/ops.py中，直接拿过来用就行了，以上的推理代码也是参考的训练模型的代码中的models/segments/predict.py函数，最终完成模型的部署。

4 测试结果