Yolov5-v6.0模型详解

Yolov5-v6.0模型详解

先看看yaml文件:可以知道6.0模型主要构成为conv、bottlenet、cs3、sppf、detect层

yaml文件如下:

# YOLOv5 v6.0 backbone

backbone:

  # [from, number, module, args]

  [[-1, 1, Conv, [64[l1] , 6[l2] , 2[l3] , 2[l4] ]],  # 0-P1/2   

#  来自上层,含瓶颈层数,本层类型,【输出通道,核大小,滑动,?分组】

#含瓶颈层数只对c3有效,其他层均为1

   [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4

   [-1, 3[l5] , C3, [128]],

   [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8

   [-1, 6, C3, [256]],

   [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16

   [-1, 9, C3, [512]],

   [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32

   [-1, 3, C3, [1024]],

   [-1, 1, SPPF, [1024, 5]],  # 9

  ]

# YOLOv5 v6.0 head

head:

  [[-1, 1, Conv, [512, 1, 1]],

   [-1, 1, nn.Upsample, [None, 2, 'nearest']],

   [[-1, 6], 1, Concat, [1]],  # cat backbone P4

   [-1, 3, C3, [512, False]],  # 13

   [-1, 1, Conv, [256, 1, 1]],

   [-1, 1, nn.Upsample, [None, 2, 'nearest']],

   [[-1, 4], 1, Concat, [1]],  # cat backbone P3

   [-1, 3, C3, [256, False]],  # 17 (P3/8-small)

   [-1, 1, Conv, [256, 3, 2]],

   [[-1, 14], 1, Concat, [1]],  # cat head P4

   [-1, 3, C3, [512, False]],  # 20 (P4/16-medium)

   [-1, 1, Conv, [512, 3, 2]],

   [[-1, 10], 1, Concat, [1]],  # cat head P5

   [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)

   [[17, 20, 23], 1, Detect[l6] , [nc, anchors]],  # Detect(P3, P4, P5)

  ]

以下根据图示看源码---------------------------------------------------------

class Conv(nn.Module):

    # Standard convolution

    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups

        super().__init__()

        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)

        self.bn = nn.BatchNorm2d(c2)

        self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())

    def forward(self, x):

        return self.act(self.bn(self.conv(x)))

    def forward_fuse(self, x):

        return self.act(self.conv(x))

class Bottleneck(nn.Module):

    # Standard bottleneck

    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion

        super().__init__()

        c_ = int(c2 * e)  # hidden channels  可以增加通道数,扩充网络宽度

        self.cv1 = Conv(c1, c_, 1, 1)

        self.cv2 = Conv(c_, c2, 3, 1, g=g)  #核是3

        self.add = shortcut and c1 == c2

    def forward(self, x):

        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))

class C3(nn.Module):

    # CSP Bottleneck with 3 convolutions

    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion

        super().__init__()

        c_ = int(c2 * e)  # hidden channels  可以扩充网络宽度

        self.cv1 = Conv(c1, c_, 1, 1)

        self.cv2 = Conv(c1, c_, 1, 1)

        self.cv3 = Conv(2 * c_, c2, 1)  # act=FReLU(c2)  concatconv输入通道要是2 * c_

        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))

#通道数不会变化都是c_

        # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])

    def forward(self, x):

        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))

sppf图示:

  class SPPF(nn.Module):   # SPPF 输入通道c1,输出通道c2,大小不变

    # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher

    def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))  

        super().__init__()

        c_ = c1 // 2  # hidden channels

        self.cv1 = Conv(c1, c_, 1, 1)  #1,滑动1

        self.cv2 = Conv(c_ * 4, c2, 1, 1) #

        self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)  #最大池化,滑动1,此处大小不变

    def forward(self, x):

        x = self.cv1(x)

        with warnings.catch_warnings():

            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning

            y1 = self.m(x)

            y2 = self.m(y1)

            return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))

detect层是最后一层:

如果是训练时,

(1)3个head,经过conv,

(2)每层原来数据为4维:[batch,anchor*(class+5),h,w] 经过处理后变为5维:

[batch,anchor,h,w,class+5]

        见下面绿色代码

如果是预测,3个head预测结果经过(1)(2)

(3)conv后sigmoid变为0~1,然后转化到格子空间。

(4)3个头的数据转化为(batch, _ ,85)然后cat在一起

#-----------------detct源码-------------------------------------------------------

class Detect(nn.Module):#如果是训练,输出3个张量,如果是预测输出【【batch,_,类+5】

    stride = None  # strides computed during build

    onnx_dynamic = False  # ONNX export parameter

    def __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layer

        #ch=()应该是3个头的通道数

        super().__init__()

        self.nc = nc  # number of classes

        self.no = nc + 5  # number of outputs per anchor

        self.nl = len(anchors)  # number of detection layers

        self.na = len(anchors[0]) // 2  # number of anchors,一般为3

        self.grid = [torch.zeros(1)] * self.nl  # init grid

        self.anchor_grid = [torch.zeros(1)] * self.nl  # init anchor grid

        self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2))

 # shape(nl,na,2)

        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv

        #ch 输入的3个头输出的通道【128,256,512】

        self.inplace = inplace  # use in-place ops (e.g. slice assignment)

    def forward(self, x):

        z = []  # inference output

        for i in range(self.nl):  #3个头分别处理 nl是 检测层的数量3

            x[i] = self.m[i](x[i])  # conv

            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)

            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()

            #4维转为5维 x(bs,3,20,20,85)

            if not self.training:  # inference,如果不是训练,把数据处理成(batch,_,85)

                if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:

                    self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)

                y = x[i].sigmoid()  #0~1化

               

                #下面把预测映射到格子空间尺度。预测使用下面结果 y

                if self.inplace:

                    y[..., 0:2] = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  # xy

                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh

                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953

                    xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  # xy

                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh

                    y = torch.cat((xy, wh, y[..., 4:]), -1)

                z.append(y.view(bs, -1, self.no))

        return x if self.training else (torch.cat(z, 1), x)

        #如果是训练直接返回x(3个头),是预测把3个头的数据转化为(batch, _ ,85)然后cat在一起


 [l1]输出通道

 [l2]核大小

 [l3]滑动步长

 [l4]可能是分组

 [l5]C3中残差模块个数

其他层都为1

 [l6]3层预测结果0~1,转为真实的结果输出

  • 14
    点赞
  • 82
    收藏
    觉得还不错? 一键收藏
  • 2
    评论

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论 2
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值