YOLOV5更换轻量级的backbone：mobilenetV2

简洁概要：

MobileNetV2主要采用了深度可分离卷积，在MobileNetv1的基础上引用了残差模块以及Relu6的激活函数，用1n，n1的思想代替了n*n的矩阵，计算量会更小。

修改主干网络：

一：添加自己主干网络

yolov5 6.1的版本中，在models/common中添加MobilenetV2作为backbone

class ConvBNReLU(nn.Sequential):  # 该函数主要做卷积 池化 ReLU6激活操作
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
        padding = (kernel_size - 1) // 2  # 池化 = （步长-1）整除2
        super(ConvBNReLU, self).__init__(  # 调用ConvBNReLU父类添加模块
            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, bias=False, groups=groups),  # bias默认为False
            nn.BatchNorm2d(out_planes),
            nn.ReLU6(inplace=True))

class InvertedResidual(nn.Module):  # 该模块主要实现了倒残差模块
    def __init__(self, inp, oup, stride, expand_ratio):  # inp 输入 oup 输出 stride步长 exoand_ratio 按比例扩张
        super(InvertedResidual, self).__init__()
        self.stride = stride
        assert stride in [1, 2]
        hidden_dim = int(round(inp * expand_ratio))  # 由于有到残差模块有1*1,3*3的卷积模块，所以可以靠expand_rarton来进行升维
        self.use_res_connect = self.stride == 1 and inp == oup  # 残差连接的判断条件：当步长=1且输入矩阵与输出矩阵的shape相同时进行
        layers = []
        if expand_ratio != 1:  # 如果expand_ratio不等于1，要做升维操作，对应图中的绿色模块
            # pw
            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))  # 这里添加的是1*1的卷积操作
        layers.extend([
            # dw
            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
            # 这里做3*3的卷积操作，步长可能是1也可能是2,groups=hidden_dim表示这里使用了分组卷积的操作，对应图上的蓝色模块

            # pw-linear
            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),  # 对应图中的黄色模块
            nn.BatchNorm2d(oup),
        ])
        self.conv = nn.Sequential(*layers)  # 将layers列表中的元素解开依次传入nn.Sequential

    def forward(self, x):
        if self.use_res_connect:  # 如果使用了残差连接，就会进行一个x+的操作
            return x + self.conv(x)
        else:
            return self.conv(x)  # 否则不做操作

二：在yolo.py中添加common中的两个函数

if m in (Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv,
                 BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x,
                 ConvBNReLU, InvertedResidual):  # 添加 common中新加的两个模块 ConvBNReLU和InvertedResidual

三：制作mobilenetv2的yaml配置文件

# Parameters
nc: 1  # number of classes

depth_multiple: 1.0  # model depth multiple
width_multiple: 1.0  # layer channel multiple

anchors:
  - [ 10,13, 16,30, 33,23 ]  # P3/8
  - [ 30,61, 62,45, 59,119 ]  # P4/16
  - [ 116,90, 156,198, 373,326 ]  # P5/32

# YOLOv5 v6.0 backbone
backbone:
  # [from, number, module, args]
  [ [ -1, 1, Conv, [ 32, 3, 2 ] ],  # 0-P1/2 32x320x320
    [ -1, 1, InvertedResidual, [ 16, 1, 1 ] ],  # 1        16x320x320
    [ -1, 1, InvertedResidual, [ 24, 2, 6 ] ],  # 2-P2/4   24x160x160
    [ -1, 1, InvertedResidual, [ 24, 1, 6 ] ],  # 3-P2/4   24x160x160
    [ -1, 1, InvertedResidual, [ 32, 2, 6 ] ],  # 4-P3/8   32x80x80
    [ -1, 1, InvertedResidual, [ 32, 1, 6 ] ],  # 5-P3/8   32x80x80
    [ -1, 1, InvertedResidual, [ 32, 1, 6 ] ],  # 6-P3/8   32x80x80
    [ -1, 1, InvertedResidual, [ 64, 2, 6 ] ],  # 7-P4/16  64x40x40
    [ -1, 1, InvertedResidual, [ 64, 1, 6 ] ],  # 8-P4/16  64x40x40
    [ -1, 1, InvertedResidual, [ 64, 1, 6 ] ],  # 9-P4/16  64x40x40
    [ -1, 1, InvertedResidual, [ 64, 1, 6 ] ],  # 10-P4/16 64x40x40
    [ -1, 1, InvertedResidual, [ 96, 1, 6 ] ],  # 11       96X40X40
    [ -1, 1, InvertedResidual, [ 96, 1, 6 ] ],  # 12       96X40X40
    [ -1, 1, InvertedResidual, [ 96, 1, 6 ] ],  # 13       96X40X40
    [ -1, 1, InvertedResidual, [ 160, 2, 6 ] ], # 14-P5/32  160X20X20
    [ -1, 1, InvertedResidual, [ 160, 1, 6 ] ], # 15-P5/32  160X20X20
    [ -1, 1, InvertedResidual, [ 160, 1, 6 ] ], # 16-P5/32  160X20X20
    [ -1, 1, InvertedResidual, [ 320, 1, 6 ] ],  # 17       320X20X20
  ]

# YOLOv5 v6.0 head
head:
  [ [ -1, 1, Conv, [ 160, 1, 1 ] ],
    [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ],
    [ [ -1, 13 ], 1, Concat, [ 1 ] ],  # cat backbone P4
    [ -1, 1, C3, [ 160, False ] ],  # 21

    [ -1, 1, Conv, [ 80, 1, 1 ] ],
    [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ],
    [ [ -1, 6 ], 1, Concat, [ 1 ] ],  # cat backbone P3
    [ -1, 1, C3, [ 80, False ] ],  # 25 (P3/8-small)

    [ -1, 1, Conv, [ 80, 3, 2 ] ],
    [ [ -1, 22 ], 1, Concat, [ 1 ] ],  # cat head P4
    [ -1, 1, C3, [ 160, False ] ],  # 28 (P4/16-medium)

    [ -1, 1, Conv, [ 160, 3, 2 ] ],
    [ [ -1, 18 ], 1, Concat, [ 1 ] ],  # cat head P5
    [ -1, 1, C3, [ 320, False ] ],  # 31 (P5/32-large)

    [ [ 25, 28, 31 ], 1, Detect, [ nc, anchors ] ],  # Detect(P3, P4, P5)
  ]

四：制作数据集VOC的yaml配置文件

# YOLOv5  by Ultralytics, GPL-3.0 license
# PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC by University of Oxford
# Example usage: python train.py --data VOC.yaml
# parent
# ├── yolov5
# └── datasets
#     └── VOC  ← downloads here (2.8 GB)

# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: E:\yolov5-6.1\VOCdevkit
train: # train images (relative to 'path')  16551 images
  - images/train
val: # val images (relative to 'path')  4952 images
  - images/val
test: # test images (optional)

# Classes
nc: 1  # number of classes
names: [ 'ball' ]  # class names

# Download script/URL (optional) ---------------------------------------------------------------------------------------
download: |
  import xml.etree.ElementTree as ET

  from tqdm import tqdm
  from utils.general import download, Path

  def convert_label(path, lb_path, year, image_id):
      def convert_box(size, box):
          dw, dh = 1. / size[0], 1. / size[1]
          x, y, w, h = (box[0] + box[1]) / 2.0 - 1, (box[2] + box[3]) / 2.0 - 1, box[1] - box[0], box[3] - box[2]
          return x * dw, y * dh, w * dw, h * dh

      in_file = open(path / f'VOC{year}/Annotations/{image_id}.xml')
      out_file = open(lb_path, 'w')
      tree = ET.parse(in_file)
      root = tree.getroot()
      size = root.find('size')
      w = int(size.find('width').text)
      h = int(size.find('height').text)

      for obj in root.iter('object'):
          cls = obj.find('name').text
          if cls in yaml['names'] and not int(obj.find('difficult').text) == 1:
              xmlbox = obj.find('bndbox')
              bb = convert_box((w, h), [float(xmlbox.find(x).text) for x in ('xmin', 'xmax', 'ymin', 'ymax')])
              cls_id = yaml['names'].index(cls)  # class id
              out_file.write(" ".join([str(a) for a in (cls_id, *bb)]) + '\n')

  # Download
  dir = Path(yaml['path'])  # dataset root dir
  url = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/'
  urls = [f'{url}VOCtrainval_06-Nov-2007.zip',  # 446MB, 5012 images
          f'{url}VOCtest_06-Nov-2007.zip',  # 438MB, 4953 images
          f'{url}VOCtrainval_11-May-2012.zip']  # 1.95GB, 17126 images
  download(urls, dir=dir / 'images', delete=False, curl=True, threads=3)

  # Convert
  path = dir / 'images/VOCdevkit'
  for year, image_set in ('2007', 'train'), ('2007', 'val'), ('2007', 'test'):
      imgs_path = dir / 'images' / f'{image_set}{year}'
      lbs_path = dir / 'labels' / f'{image_set}{year}'
      imgs_path.mkdir(exist_ok=True, parents=True)
      lbs_path.mkdir(exist_ok=True, parents=True)

      with open(path / f'VOC{year}/ImageSets/Main/{image_set}.txt') as f:
          image_ids = f.read().strip().split()
      for id in tqdm(image_ids, desc=f'{image_set}{year}'):
          f = path / f'VOC{year}/JPEGImages/{id}.jpg'  # old img path
          lb_path = (lbs_path / f.name).with_suffix('.txt')  # new label path
          f.rename(imgs_path / f.name)  # move image
          convert_label(path, lb_path, year, id)  # convert labels to YOLO format

五：启用训练

由于修改了网络所以不能加载预训练模型进行

预训练模型的作用：加快模型训练初期的超参数训练时间

weights修改为空

cfg修改为自己网络模型的配置文件

data修改为自己VOC数据集的配置文件

六：性能检测

修改val.py的参数，与上一步一致

这里分别用了V5s,V5n,以及mobilenetV2分别做了150批次训练来对比

mobilenetV2

V5s

V5n

对比可以发现 V5n与mobilenetV2的相差并不大，相比较于这两个模型，V5s的精度稍微高一些，但是它模型的复杂度会略微大一丢丢，推理时间略大一点。