devbox
我家自动化
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

YOLOv9大幅度按比例减小模型计算量！加快训练！_github yolov9

作者：我家自动化 | 2024-03-25 22:13:34

赞

踩

github yolov9

 

一、代码及论文链接：

代码链接：GitHub - WongKinYiu/yolov9: Implementation of paper - YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

论文链接：https://github.com/WongKinYiu/yolov9/tree/main

二、 说明

        本文方法并不能直接替代YOLOv9原作者尚未开源的两个小模型，但可以按比例减小模型尺寸。类似YOLOv5、v8等，可以方便测试YOLOv9在数据集上的性能！方法来源于网络。

三、使用步骤

        参照之前的YOLOv9代码，我们运行yolov9-c.yaml的参数量是239 GLOPs。

        我们将以下代码替换掉YOLOv9工程下models包下yolo.py脚本中的代码。

  1. import argparse
  2. import os
  3. import platform
  4. import sys
  5. from copy import deepcopy
  6. from pathlib import Path
  7.  
  8. FILE = Path(__file__).resolve()
  9. ROOT = FILE.parents[1]  # YOLO root directory
  10. if str(ROOT) not in sys.path:
  11.     sys.path.append(str(ROOT))  # add ROOT to PATH
  12. if platform.system() != 'Windows':
  13.     ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
  14.  
  15. from models.common import *
  16. from models.experimental import *
  17. from utils.general import LOGGER, check_version, check_yaml, make_divisible, print_args
  18. from utils.plots import feature_visualization
  19. from utils.torch_utils import (fuse_conv_and_bn, initialize_weights, model_info, profile, scale_img, select_device,
  20.                                time_sync)
  21. from utils.tal.anchor_generator import make_anchors, dist2bbox
  22.  
  23. try:
  24.     import thop  # for FLOPs computation
  25. except ImportError:
  26.     thop = None
  27.  
  28.  
  29. class Detect(nn.Module):
  30.     # YOLO Detect head for detection models
  31.     dynamic = False  # force grid reconstruction
  32.     export = False  # export mode
  33.     shape = None
  34.     anchors = torch.empty(0)  # init
  35.     strides = torch.empty(0)  # init
  36.  
  37.     def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
  38.         super().__init__()
  39.         self.nc = nc  # number of classes
  40.         self.nl = len(ch)  # number of detection layers
  41.         self.reg_max = 16
  42.         self.no = nc + self.reg_max * 4  # number of outputs per anchor
  43.         self.inplace = inplace  # use inplace ops (e.g. slice assignment)
  44.         self.stride = torch.zeros(self.nl)  # strides computed during build
  45.  
  46.         c2, c3 = max((ch[0] // 4, self.reg_max * 4, 16)), max((ch[0], min((self.nc * 2, 128))))  # channels
  47.         self.cv2 = nn.ModuleList(
  48.             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch)
  49.         self.cv3 = nn.ModuleList(
  50.             nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch)
  51.         self.dfl = DFL(self.reg_max) if self.reg_max > 1 else nn.Identity()
  52.  
  53.     def forward(self, x):
  54.         shape = x[0].shape  # BCHW
  55.         for i in range(self.nl):
  56.             x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
  57.         if self.training:
  58.             return x
  59.         elif self.dynamic or self.shape != shape:
  60.             self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
  61.             self.shape = shape
  62.  
  63.         box, cls = torch.cat([xi.view(shape[0], self.no, -1) for xi in x], 2).split((self.reg_max * 4, self.nc), 1)
  64.         dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  65.         y = torch.cat((dbox, cls.sigmoid()), 1)
  66.         return y if self.export else (y, x)
  67.  
  68.     def bias_init(self):
  69.         # Initialize Detect() biases, WARNING: requires stride availability
  70.         m = self  # self.model[-1]  # Detect() module
  71.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
  72.         # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
  73.         for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
  74.             a[-1].bias.data[:] = 1.0  # box
  75.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  76.  
  77.  
  78. class DDetect(nn.Module):
  79.     # YOLO Detect head for detection models
  80.     dynamic = False  # force grid reconstruction
  81.     export = False  # export mode
  82.     shape = None
  83.     anchors = torch.empty(0)  # init
  84.     strides = torch.empty(0)  # init
  85.  
  86.     def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
  87.         super().__init__()
  88.         self.nc = nc  # number of classes
  89.         self.nl = len(ch)  # number of detection layers
  90.         self.reg_max = 16
  91.         self.no = nc + self.reg_max * 4  # number of outputs per anchor
  92.         self.inplace = inplace  # use inplace ops (e.g. slice assignment)
  93.         self.stride = torch.zeros(self.nl)  # strides computed during build
  94.  
  95.         c2, c3 = make_divisible(max((ch[0] // 4, self.reg_max * 4, 16)), 4), max((ch[0], min((self.nc * 2, 128))))  # channels
  96.         self.cv2 = nn.ModuleList(
  97.             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3, g=4), nn.Conv2d(c2, 4 * self.reg_max, 1, groups=4)) for x in ch)
  98.         self.cv3 = nn.ModuleList(
  99.             nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch)
  100.         self.dfl = DFL(self.reg_max) if self.reg_max > 1 else nn.Identity()
  101.  
  102.     def forward(self, x):
  103.         shape = x[0].shape  # BCHW
  104.         for i in range(self.nl):
  105.             x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
  106.         if self.training:
  107.             return x
  108.         elif self.dynamic or self.shape != shape:
  109.             self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
  110.             self.shape = shape
  111.  
  112.         box, cls = torch.cat([xi.view(shape[0], self.no, -1) for xi in x], 2).split((self.reg_max * 4, self.nc), 1)
  113.         dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  114.         y = torch.cat((dbox, cls.sigmoid()), 1)
  115.         return y if self.export else (y, x)
  116.  
  117.     def bias_init(self):
  118.         # Initialize Detect() biases, WARNING: requires stride availability
  119.         m = self  # self.model[-1]  # Detect() module
  120.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
  121.         # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
  122.         for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
  123.             a[-1].bias.data[:] = 1.0  # box
  124.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  125.  
  126.  
  127. class DualDetect(nn.Module):
  128.     # YOLO Detect head for detection models
  129.     dynamic = False  # force grid reconstruction
  130.     export = False  # export mode
  131.     shape = None
  132.     anchors = torch.empty(0)  # init
  133.     strides = torch.empty(0)  # init
  134.  
  135.     def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
  136.         super().__init__()
  137.         self.nc = nc  # number of classes
  138.         self.nl = len(ch) // 2  # number of detection layers
  139.         self.reg_max = 16
  140.         self.no = nc + self.reg_max * 4  # number of outputs per anchor
  141.         self.inplace = inplace  # use inplace ops (e.g. slice assignment)
  142.         self.stride = torch.zeros(self.nl)  # strides computed during build
  143.  
  144.         c2, c3 = max((ch[0] // 4, self.reg_max * 4, 16)), max((ch[0], min((self.nc * 2, 128))))  # channels
  145.         c4, c5 = max((ch[self.nl] // 4, self.reg_max * 4, 16)), max((ch[self.nl], min((self.nc * 2, 128))))  # channels
  146.         self.cv2 = nn.ModuleList(
  147.             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch[:self.nl])
  148.         self.cv3 = nn.ModuleList(
  149.             nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
  150.         self.cv4 = nn.ModuleList(
  151.             nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, 4 * self.reg_max, 1)) for x in ch[self.nl:])
  152.         self.cv5 = nn.ModuleList(
  153.             nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:])
  154.         self.dfl = DFL(self.reg_max)
  155.         self.dfl2 = DFL(self.reg_max)
  156.  
  157.     def forward(self, x):
  158.         shape = x[0].shape  # BCHW
  159.         d1 = []
  160.         d2 = []
  161.         for i in range(self.nl):
  162.             d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
  163.             d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
  164.         if self.training:
  165.             return [d1, d2]
  166.         elif self.dynamic or self.shape != shape:
  167.             self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
  168.             self.shape = shape
  169.  
  170.         box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
  171.         dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  172.         box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
  173.         dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  174.         y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1)]
  175.         return y if self.export else (y, [d1, d2])
  176.  
  177.     def bias_init(self):
  178.         # Initialize Detect() biases, WARNING: requires stride availability
  179.         m = self  # self.model[-1]  # Detect() module
  180.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
  181.         # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
  182.         for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
  183.             a[-1].bias.data[:] = 1.0  # box
  184.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  185.         for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
  186.             a[-1].bias.data[:] = 1.0  # box
  187.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  188.  
  189.  
  190. class DualDDetect(nn.Module):
  191.     # YOLO Detect head for detection models
  192.     dynamic = False  # force grid reconstruction
  193.     export = False  # export mode
  194.     shape = None
  195.     anchors = torch.empty(0)  # init
  196.     strides = torch.empty(0)  # init
  197.  
  198.     def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
  199.         super().__init__()
  200.         self.nc = nc  # number of classes
  201.         self.nl = len(ch) // 2  # number of detection layers
  202.         self.reg_max = 16
  203.         self.no = nc + self.reg_max * 4  # number of outputs per anchor
  204.         self.inplace = inplace  # use inplace ops (e.g. slice assignment)
  205.         self.stride = torch.zeros(self.nl)  # strides computed during build
  206.  
  207.         c2, c3 = make_divisible(max((ch[0] // 4, self.reg_max * 4, 16)), 4), max((ch[0], min((self.nc * 2, 128))))  # channels
  208.         c4, c5 = make_divisible(max((ch[self.nl] // 4, self.reg_max * 4, 16)), 4), max((ch[self.nl], min((self.nc * 2, 128))))  # channels
  209.         self.cv2 = nn.ModuleList(
  210.             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3, g=4), nn.Conv2d(c2, 4 * self.reg_max, 1, groups=4)) for x in ch[:self.nl])
  211.         self.cv3 = nn.ModuleList(
  212.             nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
  213.         self.cv4 = nn.ModuleList(
  214.             nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3, g=4), nn.Conv2d(c4, 4 * self.reg_max, 1, groups=4)) for x in ch[self.nl:])
  215.         self.cv5 = nn.ModuleList(
  216.             nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:])
  217.         self.dfl = DFL(self.reg_max)
  218.         self.dfl2 = DFL(self.reg_max)
  219.  
  220.     def forward(self, x):
  221.         shape = x[0].shape  # BCHW
  222.         d1 = []
  223.         d2 = []
  224.         for i in range(self.nl):
  225.             d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
  226.             d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
  227.         if self.training:
  228.             return [d1, d2]
  229.         elif self.dynamic or self.shape != shape:
  230.             self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
  231.             self.shape = shape
  232.  
  233.         box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
  234.         dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  235.         box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
  236.         dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  237.         y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1)]
  238.         return y if self.export else (y, [d1, d2])
  239.         #y = torch.cat((dbox2, cls2.sigmoid()), 1)
  240.         #return y if self.export else (y, d2)
  241.         #y1 = torch.cat((dbox, cls.sigmoid()), 1)
  242.         #y2 = torch.cat((dbox2, cls2.sigmoid()), 1)
  243.         #return [y1, y2] if self.export else [(y1, d1), (y2, d2)]
  244.         #return [y1, y2] if self.export else [(y1, y2), (d1, d2)]
  245.  
  246.     def bias_init(self):
  247.         # Initialize Detect() biases, WARNING: requires stride availability
  248.         m = self  # self.model[-1]  # Detect() module
  249.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
  250.         # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
  251.         for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
  252.             a[-1].bias.data[:] = 1.0  # box
  253.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  254.         for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
  255.             a[-1].bias.data[:] = 1.0  # box
  256.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  257.  
  258.  
  259. class TripleDetect(nn.Module):
  260.     # YOLO Detect head for detection models
  261.     dynamic = False  # force grid reconstruction
  262.     export = False  # export mode
  263.     shape = None
  264.     anchors = torch.empty(0)  # init
  265.     strides = torch.empty(0)  # init
  266.  
  267.     def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
  268.         super().__init__()
  269.         self.nc = nc  # number of classes
  270.         self.nl = len(ch) // 3  # number of detection layers
  271.         self.reg_max = 16
  272.         self.no = nc + self.reg_max * 4  # number of outputs per anchor
  273.         self.inplace = inplace  # use inplace ops (e.g. slice assignment)
  274.         self.stride = torch.zeros(self.nl)  # strides computed during build
  275.  
  276.         c2, c3 = max((ch[0] // 4, self.reg_max * 4, 16)), max((ch[0], min((self.nc * 2, 128))))  # channels
  277.         c4, c5 = max((ch[self.nl] // 4, self.reg_max * 4, 16)), max((ch[self.nl], min((self.nc * 2, 128))))  # channels
  278.         c6, c7 = max((ch[self.nl * 2] // 4, self.reg_max * 4, 16)), max((ch[self.nl * 2], min((self.nc * 2, 128))))  # channels
  279.         self.cv2 = nn.ModuleList(
  280.             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch[:self.nl])
  281.         self.cv3 = nn.ModuleList(
  282.             nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
  283.         self.cv4 = nn.ModuleList(
  284.             nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, 4 * self.reg_max, 1)) for x in ch[self.nl:self.nl*2])
  285.         self.cv5 = nn.ModuleList(
  286.             nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:self.nl*2])
  287.         self.cv6 = nn.ModuleList(
  288.             nn.Sequential(Conv(x, c6, 3), Conv(c6, c6, 3), nn.Conv2d(c6, 4 * self.reg_max, 1)) for x in ch[self.nl*2:self.nl*3])
  289.         self.cv7 = nn.ModuleList(
  290.             nn.Sequential(Conv(x, c7, 3), Conv(c7, c7, 3), nn.Conv2d(c7, self.nc, 1)) for x in ch[self.nl*2:self.nl*3])
  291.         self.dfl = DFL(self.reg_max)
  292.         self.dfl2 = DFL(self.reg_max)
  293.         self.dfl3 = DFL(self.reg_max)
  294.  
  295.     def forward(self, x):
  296.         shape = x[0].shape  # BCHW
  297.         d1 = []
  298.         d2 = []
  299.         d3 = []
  300.         for i in range(self.nl):
  301.             d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
  302.             d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
  303.             d3.append(torch.cat((self.cv6[i](x[self.nl*2+i]), self.cv7[i](x[self.nl*2+i])), 1))
  304.         if self.training:
  305.             return [d1, d2, d3]
  306.         elif self.dynamic or self.shape != shape:
  307.             self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
  308.             self.shape = shape
  309.  
  310.         box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
  311.         dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  312.         box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
  313.         dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  314.         box3, cls3 = torch.cat([di.view(shape[0], self.no, -1) for di in d3], 2).split((self.reg_max * 4, self.nc), 1)
  315.         dbox3 = dist2bbox(self.dfl3(box3), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  316.         y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1), torch.cat((dbox3, cls3.sigmoid()), 1)]
  317.         return y if self.export else (y, [d1, d2, d3])
  318.  
  319.     def bias_init(self):
  320.         # Initialize Detect() biases, WARNING: requires stride availability
  321.         m = self  # self.model[-1]  # Detect() module
  322.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
  323.         # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
  324.         for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
  325.             a[-1].bias.data[:] = 1.0  # box
  326.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  327.         for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
  328.             a[-1].bias.data[:] = 1.0  # box
  329.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  330.         for a, b, s in zip(m.cv6, m.cv7, m.stride):  # from
  331.             a[-1].bias.data[:] = 1.0  # box
  332.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  333.  
  334.  
  335. class TripleDDetect(nn.Module):
  336.     # YOLO Detect head for detection models
  337.     dynamic = False  # force grid reconstruction
  338.     export = False  # export mode
  339.     shape = None
  340.     anchors = torch.empty(0)  # init
  341.     strides = torch.empty(0)  # init
  342.  
  343.     def __init__(self, nc=80, ch=(), inplace=True):  # detection layer
  344.         super().__init__()
  345.         self.nc = nc  # number of classes
  346.         self.nl = len(ch) // 3  # number of detection layers
  347.         self.reg_max = 16
  348.         self.no = nc + self.reg_max * 4  # number of outputs per anchor
  349.         self.inplace = inplace  # use inplace ops (e.g. slice assignment)
  350.         self.stride = torch.zeros(self.nl)  # strides computed during build
  351.  
  352.         c2, c3 = make_divisible(max((ch[0] // 4, self.reg_max * 4, 16)), 4), \
  353.                                 max((ch[0], min((self.nc * 2, 128))))  # channels
  354.         c4, c5 = make_divisible(max((ch[self.nl] // 4, self.reg_max * 4, 16)), 4), \
  355.                                 max((ch[self.nl], min((self.nc * 2, 128))))  # channels
  356.         c6, c7 = make_divisible(max((ch[self.nl * 2] // 4, self.reg_max * 4, 16)), 4), \
  357.                                 max((ch[self.nl * 2], min((self.nc * 2, 128))))  # channels
  358.         self.cv2 = nn.ModuleList(
  359.             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3, g=4), 
  360.                           nn.Conv2d(c2, 4 * self.reg_max, 1, groups=4)) for x in ch[:self.nl])
  361.         self.cv3 = nn.ModuleList(
  362.             nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch[:self.nl])
  363.         self.cv4 = nn.ModuleList(
  364.             nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3, g=4), 
  365.                           nn.Conv2d(c4, 4 * self.reg_max, 1, groups=4)) for x in ch[self.nl:self.nl*2])
  366.         self.cv5 = nn.ModuleList(
  367.             nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nc, 1)) for x in ch[self.nl:self.nl*2])
  368.         self.cv6 = nn.ModuleList(
  369.             nn.Sequential(Conv(x, c6, 3), Conv(c6, c6, 3, g=4), 
  370.                           nn.Conv2d(c6, 4 * self.reg_max, 1, groups=4)) for x in ch[self.nl*2:self.nl*3])
  371.         self.cv7 = nn.ModuleList(
  372.             nn.Sequential(Conv(x, c7, 3), Conv(c7, c7, 3), nn.Conv2d(c7, self.nc, 1)) for x in ch[self.nl*2:self.nl*3])
  373.         self.dfl = DFL(self.reg_max)
  374.         self.dfl2 = DFL(self.reg_max)
  375.         self.dfl3 = DFL(self.reg_max)
  376.  
  377.     def forward(self, x):
  378.         shape = x[0].shape  # BCHW
  379.         d1 = []
  380.         d2 = []
  381.         d3 = []
  382.         for i in range(self.nl):
  383.             d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))
  384.             d2.append(torch.cat((self.cv4[i](x[self.nl+i]), self.cv5[i](x[self.nl+i])), 1))
  385.             d3.append(torch.cat((self.cv6[i](x[self.nl*2+i]), self.cv7[i](x[self.nl*2+i])), 1))
  386.         if self.training:
  387.             return [d1, d2, d3]
  388.         elif self.dynamic or self.shape != shape:
  389.             self.anchors, self.strides = (d1.transpose(0, 1) for d1 in make_anchors(d1, self.stride, 0.5))
  390.             self.shape = shape
  391.  
  392.         box, cls = torch.cat([di.view(shape[0], self.no, -1) for di in d1], 2).split((self.reg_max * 4, self.nc), 1)
  393.         dbox = dist2bbox(self.dfl(box), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  394.         box2, cls2 = torch.cat([di.view(shape[0], self.no, -1) for di in d2], 2).split((self.reg_max * 4, self.nc), 1)
  395.         dbox2 = dist2bbox(self.dfl2(box2), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  396.         box3, cls3 = torch.cat([di.view(shape[0], self.no, -1) for di in d3], 2).split((self.reg_max * 4, self.nc), 1)
  397.         dbox3 = dist2bbox(self.dfl3(box3), self.anchors.unsqueeze(0), xywh=True, dim=1) * self.strides
  398.         #y = [torch.cat((dbox, cls.sigmoid()), 1), torch.cat((dbox2, cls2.sigmoid()), 1), torch.cat((dbox3, cls3.sigmoid()), 1)]
  399.         #return y if self.export else (y, [d1, d2, d3])
  400.         y = torch.cat((dbox3, cls3.sigmoid()), 1)
  401.         return y if self.export else (y, d3)
  402.  
  403.     def bias_init(self):
  404.         # Initialize Detect() biases, WARNING: requires stride availability
  405.         m = self  # self.model[-1]  # Detect() module
  406.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
  407.         # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
  408.         for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
  409.             a[-1].bias.data[:] = 1.0  # box
  410.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  411.         for a, b, s in zip(m.cv4, m.cv5, m.stride):  # from
  412.             a[-1].bias.data[:] = 1.0  # box
  413.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  414.         for a, b, s in zip(m.cv6, m.cv7, m.stride):  # from
  415.             a[-1].bias.data[:] = 1.0  # box
  416.             b[-1].bias.data[:m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (5 objects and 80 classes per 640 image)
  417.  
  418.  
  419. class Segment(Detect):
  420.     # YOLO Segment head for segmentation models
  421.     def __init__(self, nc=80, nm=32, npr=256, ch=(), inplace=True):
  422.         super().__init__(nc, ch, inplace)
  423.         self.nm = nm  # number of masks
  424.         self.npr = npr  # number of protos
  425.         self.proto = Proto(ch[0], self.npr, self.nm)  # protos
  426.         self.detect = Detect.forward
  427.  
  428.         c4 = max(ch[0] // 4, self.nm)
  429.         self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch)
  430.  
  431.     def forward(self, x):
  432.         p = self.proto(x[0])
  433.         bs = p.shape[0]
  434.  
  435.         mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
  436.         x = self.detect(self, x)
  437.         if self.training:
  438.             return x, mc, p
  439.         return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))
  440.  
  441.  
  442. class Panoptic(Detect):
  443.     # YOLO Panoptic head for panoptic segmentation models
  444.     def __init__(self, nc=80, sem_nc=93, nm=32, npr=256, ch=(), inplace=True):
  445.         super().__init__(nc, ch, inplace)
  446.         self.sem_nc = sem_nc
  447.         self.nm = nm  # number of masks
  448.         self.npr = npr  # number of protos
  449.         self.proto = Proto(ch[0], self.npr, self.nm)  # protos
  450.         self.uconv = UConv(ch[0], ch[0]//4, self.sem_nc+self.nc)
  451.         self.detect = Detect.forward
  452.  
  453.         c4 = max(ch[0] // 4, self.nm)
  454.         self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch)
  455.  
  456.  
  457.     def forward(self, x):
  458.         p = self.proto(x[0])
  459.         s = self.uconv(x[0])
  460.         bs = p.shape[0]
  461.  
  462.         mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
  463.         x = self.detect(self, x)
  464.         if self.training:
  465.             return x, mc, p, s
  466.         return (torch.cat([x, mc], 1), p, s) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p, s))
  467.     
  468.  
  469. class BaseModel(nn.Module):
  470.     # YOLO base model
  471.     def forward(self, x, profile=False, visualize=False):
  472.         return self._forward_once(x, profile, visualize)  # single-scale inference, train
  473.  
  474.     def _forward_once(self, x, profile=False, visualize=False):
  475.         y, dt = [], []  # outputs
  476.         for m in self.model:
  477.             if m.f != -1:  # if not from previous layer
  478.                 x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
  479.             if profile:
  480.                 self._profile_one_layer(m, x, dt)
  481.             x = m(x)  # run
  482.             y.append(x if m.i in self.save else None)  # save output
  483.             if visualize:
  484.                 feature_visualization(x, m.type, m.i, save_dir=visualize)
  485.         return x
  486.  
  487.     def _profile_one_layer(self, m, x, dt):
  488.         c = m == self.model[-1]  # is final layer, copy input as inplace fix
  489.         o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0  # FLOPs
  490.         t = time_sync()
  491.         for _ in range(10):
  492.             m(x.copy() if c else x)
  493.         dt.append((time_sync() - t) * 100)
  494.         if m == self.model[0]:
  495.             LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s}  module")
  496.         LOGGER.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f}  {m.type}')
  497.         if c:
  498.             LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s}  Total")
  499.  
  500.     def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
  501.         LOGGER.info('Fusing layers... ')
  502.         for m in self.model.modules():
  503.             if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'):
  504.                 m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
  505.                 delattr(m, 'bn')  # remove batchnorm
  506.                 m.forward = m.forward_fuse  # update forward
  507.         self.info()
  508.         return self
  509.  
  510.     def info(self, verbose=False, img_size=640):  # print model information
  511.         model_info(self, verbose, img_size)
  512.  
  513.     def _apply(self, fn):
  514.         # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers
  515.         self = super()._apply(fn)
  516.         m = self.model[-1]  # Detect()
  517.         if isinstance(m, (Detect, DualDetect, TripleDetect, DDetect, DualDDetect, TripleDDetect, Segment)):
  518.             m.stride = fn(m.stride)
  519.             m.anchors = fn(m.anchors)
  520.             m.strides = fn(m.strides)
  521.             # m.grid = list(map(fn, m.grid))
  522.         return self
  523.  
  524.  
  525. class DetectionModel(BaseModel):
  526.     # YOLO detection model
  527.     def __init__(self, cfg='yolo.yaml', ch=3, nc=None, anchors=None):  # model, input channels, number of classes
  528.         super().__init__()
  529.         if isinstance(cfg, dict):
  530.             self.yaml = cfg  # model dict
  531.         else:  # is *.yaml
  532.             import yaml  # for torch hub
  533.             self.yaml_file = Path(cfg).name
  534.             with open(cfg, encoding='ascii', errors='ignore') as f:
  535.                 self.yaml = yaml.safe_load(f)  # model dict
  536.  
  537.         # Define model
  538.         ch = self.yaml['ch'] = self.yaml.get('ch', ch)  # input channels
  539.         if nc and nc != self.yaml['nc']:
  540.             LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
  541.             self.yaml['nc'] = nc  # override yaml value
  542.         if anchors:
  543.             LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}')
  544.             self.yaml['anchors'] = round(anchors)  # override yaml value
  545.         self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
  546.         self.names = [str(i) for i in range(self.yaml['nc'])]  # default names
  547.         self.inplace = self.yaml.get('inplace', True)
  548.  
  549.         # Build strides, anchors
  550.         m = self.model[-1]  # Detect()
  551.         if isinstance(m, (Detect, DDetect, Segment)):
  552.             s = 256  # 2x min stride
  553.             m.inplace = self.inplace
  554.             forward = lambda x: self.forward(x)[0] if isinstance(m, (Segment)) else self.forward(x)
  555.             m.stride = torch.tensor([s / x.shape[-2] for x in forward(torch.zeros(1, ch, s, s))])  # forward
  556.             # check_anchor_order(m)
  557.             # m.anchors /= m.stride.view(-1, 1, 1)
  558.             self.stride = m.stride
  559.             m.bias_init()  # only run once
  560.         if isinstance(m, (DualDetect, TripleDetect, DualDDetect, TripleDDetect)):
  561.             s = 256  # 2x min stride
  562.             m.inplace = self.inplace
  563.             #forward = lambda x: self.forward(x)[0][0] if isinstance(m, (DualSegment)) else self.forward(x)[0]
  564.             forward = lambda x: self.forward(x)[0]
  565.             m.stride = torch.tensor([s / x.shape[-2] for x in forward(torch.zeros(1, ch, s, s))])  # forward
  566.             # check_anchor_order(m)
  567.             # m.anchors /= m.stride.view(-1, 1, 1)
  568.             self.stride = m.stride
  569.             m.bias_init()  # only run once
  570.  
  571.         # Init weights, biases
  572.         initialize_weights(self)
  573.         self.info()
  574.         LOGGER.info('')
  575.  
  576.     def forward(self, x, augment=False, profile=False, visualize=False):
  577.         if augment:
  578.             return self._forward_augment(x)  # augmented inference, None
  579.         return self._forward_once(x, profile, visualize)  # single-scale inference, train
  580.  
  581.     def _forward_augment(self, x):
  582.         img_size = x.shape[-2:]  # height, width
  583.         s = [1, 0.83, 0.67]  # scales
  584.         f = [None, 3, None]  # flips (2-ud, 3-lr)
  585.         y = []  # outputs
  586.         for si, fi in zip(s, f):
  587.             xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
  588.             yi = self._forward_once(xi)[0]  # forward
  589.             # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
  590.             yi = self._descale_pred(yi, fi, si, img_size)
  591.             y.append(yi)
  592.         y = self._clip_augmented(y)  # clip augmented tails
  593.         return torch.cat(y, 1), None  # augmented inference, train
  594.  
  595.     def _descale_pred(self, p, flips, scale, img_size):
  596.         # de-scale predictions following augmented inference (inverse operation)
  597.         if self.inplace:
  598.             p[..., :4] /= scale  # de-scale
  599.             if flips == 2:
  600.                 p[..., 1] = img_size[0] - p[..., 1]  # de-flip ud
  601.             elif flips == 3:
  602.                 p[..., 0] = img_size[1] - p[..., 0]  # de-flip lr
  603.         else:
  604.             x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale  # de-scale
  605.             if flips == 2:
  606.                 y = img_size[0] - y  # de-flip ud
  607.             elif flips == 3:
  608.                 x = img_size[1] - x  # de-flip lr
  609.             p = torch.cat((x, y, wh, p[..., 4:]), -1)
  610.         return p
  611.  
  612.     def _clip_augmented(self, y):
  613.         # Clip YOLO augmented inference tails
  614.         nl = self.model[-1].nl  # number of detection layers (P3-P5)
  615.         g = sum(4 ** x for x in range(nl))  # grid points
  616.         e = 1  # exclude layer count
  617.         i = (y[0].shape[1] // g) * sum(4 ** x for x in range(e))  # indices
  618.         y[0] = y[0][:, :-i]  # large
  619.         i = (y[-1].shape[1] // g) * sum(4 ** (nl - 1 - x) for x in range(e))  # indices
  620.         y[-1] = y[-1][:, i:]  # small
  621.         return y
  622.  
  623.  
  624. Model = DetectionModel  # retain YOLO 'Model' class for backwards compatibility
  625.  
  626.  
  627. class SegmentationModel(DetectionModel):
  628.     # YOLO segmentation model
  629.     def __init__(self, cfg='yolo-seg.yaml', ch=3, nc=None, anchors=None):
  630.         super().__init__(cfg, ch, nc, anchors)
  631.  
  632.  
  633. class ClassificationModel(BaseModel):
  634.     # YOLO classification model
  635.     def __init__(self, cfg=None, model=None, nc=1000, cutoff=10):  # yaml, model, number of classes, cutoff index
  636.         super().__init__()
  637.         self._from_detection_model(model, nc, cutoff) if model is not None else self._from_yaml(cfg)
  638.  
  639.     def _from_detection_model(self, model, nc=1000, cutoff=10):
  640.         # Create a YOLO classification model from a YOLO detection model
  641.         if isinstance(model, DetectMultiBackend):
  642.             model = model.model  # unwrap DetectMultiBackend
  643.         model.model = model.model[:cutoff]  # backbone
  644.         m = model.model[-1]  # last layer
  645.         ch = m.conv.in_channels if hasattr(m, 'conv') else m.cv1.conv.in_channels  # ch into module
  646.         c = Classify(ch, nc)  # Classify()
  647.         c.i, c.f, c.type = m.i, m.f, 'models.common.Classify'  # index, from, type
  648.         model.model[-1] = c  # replace
  649.         self.model = model.model
  650.         self.stride = model.stride
  651.         self.save = []
  652.         self.nc = nc
  653.  
  654.     def _from_yaml(self, cfg):
  655.         # Create a YOLO classification model from a *.yaml file
  656.         self.model = None
  657.  
  658.  
  659. def parse_model(d, ch):  # model_dict, input_channels(3)
  660.     # Parse a YOLO model.yaml dictionary
  661.     LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10}  {'module':<40}{'arguments':<30}")
  662.     anchors, nc, gd, gw, act = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'], d.get('activation')
  663.     if act:
  664.         Conv.default_act = eval(act)  # redefine default activation, i.e. Conv.default_act = nn.SiLU()
  665.         LOGGER.info(f"{colorstr('activation:')} {act}")  # print
  666.     na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
  667.     no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
  668.  
  669.     layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
  670.     for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
  671.         m = eval(m) if isinstance(m, str) else m  # eval strings
  672.         for j, a in enumerate(args):
  673.             with contextlib.suppress(NameError):
  674.                 args[j] = eval(a) if isinstance(a, str) else a  # eval strings
  675.  
  676.         n = n_ = max(round(n * gd), 1) if n > 1 else n  # depth gain
  677.         if m in {
  678.             Conv, AConv, ConvTranspose, 
  679.             Bottleneck, SPP, SPPF, DWConv, BottleneckCSP, nn.ConvTranspose2d, DWConvTranspose2d, SPPCSPC, ADown,
  680.             RepNCSPELAN4, SPPELAN}:
  681.             c1, c2 = ch[f], args[0]
  682.             if c2 != no:  # if not output
  683.                 c2 = make_divisible(c2 * gw, 8)
  684.             if m in (RepNCSPELAN4, ):
  685.                 args[1] = make_divisible(args[1] * gw, 8)
  686.                 args[2] = make_divisible(args[2] * gw, 8)
  687.  
  688.             args = [c1, c2, *args[1:]]
  689.             if m in {BottleneckCSP, SPPCSPC}:
  690.                 args.insert(2, n)  # number of repeats
  691.                 n = 1
  692.         elif m is nn.BatchNorm2d:
  693.             args = [ch[f]]
  694.         elif m is Concat:
  695.             c2 = sum(ch[x] for x in f)
  696.         elif m is Shortcut:
  697.             c2 = ch[f[0]]
  698.         elif m is ReOrg:
  699.             c2 = ch[f] * 4
  700.         elif m is CBLinear:
  701.             c2 = [make_divisible(i * gw, 8) for i in args[0]]
  702.             c1 = ch[f]
  703.             args = [c1, c2, *args[1:]]
  704.         elif m is CBFuse:
  705.             c2 = ch[f[-1]]
  706.         # TODO: channel, gw, gd
  707.         elif m in {Detect, DualDetect, TripleDetect, DDetect, DualDDetect, TripleDDetect, Segment}:
  708.             args.append([ch[x] for x in f])
  709.             # if isinstance(args[1], int):  # number of anchors
  710.             #     args[1] = [list(range(args[1] * 2))] * len(f)
  711.             if m in {Segment}:
  712.                 args[2] = make_divisible(args[2] * gw, 8)
  713.         elif m is Contract:
  714.             c2 = ch[f] * args[0] ** 2
  715.         elif m is Expand:
  716.             c2 = ch[f] // args[0] ** 2
  717.         else:
  718.             c2 = ch[f]
  719.  
  720.         m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
  721.         t = str(m)[8:-2].replace('__main__.', '')  # module type
  722.         np = sum(x.numel() for x in m_.parameters())  # number params
  723.         m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
  724.         LOGGER.info(f'{i:>3}{str(f):>18}{n_:>3}{np:10.0f}  {t:<40}{str(args):<30}')  # print
  725.         save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
  726.         layers.append(m_)
  727.         if i == 0:
  728.             ch = []
  729.         ch.append(c2)
  730.     return nn.Sequential(*layers), sorted(save)
  731.  
  732.  
  733. if __name__ == '__main__':
  734.     parser = argparse.ArgumentParser()
  735.     parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml')
  736.     parser.add_argument('--batch-size', type=int, default=1, help='total batch size for all GPUs')
  737.     parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
  738.     parser.add_argument('--profile', action='store_true', help='profile model speed')
  739.     parser.add_argument('--line-profile', action='store_true', help='profile model speed layer by layer')
  740.     parser.add_argument('--test', action='store_true', help='test all yolo*.yaml')
  741.     opt = parser.parse_args()
  742.     opt.cfg = check_yaml(opt.cfg)  # check YAML
  743.     print_args(vars(opt))
  744.     device = select_device(opt.device)
  745.  
  746.     # Create model
  747.     im = torch.rand(opt.batch_size, 3, 640, 640).to(device)
  748.     model = Model(opt.cfg).to(device)
  749.     model.eval()
  750.  
  751.     # Options
  752.     if opt.line_profile:  # profile layer by layer
  753.         model(im, profile=True)
  754.  
  755.     elif opt.profile:  # profile forward-backward
  756.         results = profile(input=im, ops=[model], n=3)
  757.  
  758.     elif opt.test:  # test all models
  759.         for cfg in Path(ROOT / 'models').rglob('yolo*.yaml'):
  760.             try:
  761.                 _ = Model(cfg)
  762.             except Exception as e:
  763.                 print(f'Error in {cfg}: {e}')
  764.  
  765.     else:  # report fused model summary
  766.         model.fuse()

        最后，修改模型配置文件的深度与宽度，运行即可。运行报错的可以看一下我之前的文章，或者评论区提问。

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/我家自动化/article/detail/312799
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号