• YOLOv5、YOLOv8改进：Decoupled Head解耦头

    目录

     

    1.Decoupled Head介绍

     2.Yolov5加入Decoupled_Detect

    2.1 DecoupledHead加入common.py中：

    2.2 Decoupled_Detect加入yolo.py中：

    2.3修改yolov5s_decoupled.yaml 

    1.Decoupled Head介绍

    Decoupled Head是一种图像分割任务中常用的网络结构，用于提取图像特征并预测每个像素的类别。传统的图像分割网络通常将特征提取和像素预测过程集成在同一个网络中，而Decoupled Head则将这两个过程进行解耦，分别处理。

    Decoupled Head的核心思想是通过引入额外的分支网络来进行像素级的预测。这个分支网络通常被称为“头”（head），因此得名Decoupled Head。具体而言，Decoupled Head网络在主干网络的特征图上添加一个或多个额外的分支，用于预测像素的类别。

    Decoupled Head的优势在于可以更好地处理不同尺度和精细度的语义信息。通过将像素级的预测与特征提取分开，可以更好地利用底层和高层特征之间的语义信息，从而提高分割的准确性和细节保留能力。

    Decoupled Head的优点：

    1. 分离特征提取和像素预测：Decoupled Head将特征提取和像素级预测分离开来，使得网络可以更加灵活地处理不同尺度和语义信息。

    2. 多尺度特征融合：通过在主干网络的不同层级添加分支，Decoupled Head可以融合来自不同尺度的特征信息，从而提高对多尺度目标的分割能力。

    3. 更好的像素级预测：由于Decoupled Head将像素级的预测作为独立的任务进行处理，可以更好地保留细节和边缘信息，提高分割的精确性。

    4. 可扩展性：Decoupled Head结构可以根据需要进行扩展和修改，例如添加更多的分支或调整分支的结构，以适应不同的任务和数据集需求。

     YOLOv6 采用了解耦检测头（Decoupled Head）结构，同时综合考虑到相关算子表征能力和硬件上计算开销这两者的平衡，采用 Hybrid Channels 策略重新设计了一个更高效的解耦头结构，在维持精度的同时降低了延时，缓解了解耦头中 3x3 卷积带来的额外延时开销。

    原始 YOLOv5 的检测头是通过分类和回归分支融合共享的方式来实现的，因此加入 Decoupled Head。

    为什么要用到解耦头？

    因为分类和定位的关注点不同；
    分类更关注目标的纹理内容；
    定位更关注目标的边缘信息

     2.Yolov5加入Decoupled_Detect

    2.1 DecoupledHead加入common.py中：

    1. #======================= 解耦头=============================#
    2. class DecoupledHead(nn.Module):
    3.     def __init__(self, ch=256, nc=80,  anchors=()):
    4.         super().__init__()
    5.         self.nc = nc  # number of classes
    6.         self.nl = len(anchors)  # number of detection layers
    7.         self.na = len(anchors[0]) // 2  # number of anchors
    8.         self.merge = Conv(ch, 256 , 1, 1)
    9.         self.cls_convs1 = Conv(256 , 256 , 3, 1, 1)
    10.         self.cls_convs2 = Conv(256 , 256 , 3, 1, 1)
    11.         self.reg_convs1 = Conv(256 , 256 , 3, 1, 1)
    12.         self.reg_convs2 = Conv(256 , 256 , 3, 1, 1)
    13.         self.cls_preds = nn.Conv2d(256 , self.nc * self.na, 1) # 一个1x1的卷积，把通道数变成类别数，比如coco 80类（主要对目标框的类别，预测分数）
    14.         self.reg_preds = nn.Conv2d(256 , 4 * self.na, 1)       # 一个1x1的卷积，把通道数变成4通道，因为位置是xywh
    15.         self.obj_preds = nn.Conv2d(256 , 1 * self.na, 1)       # 一个1x1的卷积，把通道数变成1通道，通过一个值即可判断有无目标（置信度）
    16.  
    17.     def forward(self, x):
    18.         x = self.merge(x)
    19.         x1 = self.cls_convs1(x)
    20.         x1 = self.cls_convs2(x1)
    21.         x1 = self.cls_preds(x1)
    22.         x2 = self.reg_convs1(x)
    23.         x2 = self.reg_convs2(x2)
    24.         x21 = self.reg_preds(x2)
    25.         x22 = self.obj_preds(x2)
    26.         out = torch.cat([x21, x22, x1], 1) # 把分类和回归结果按channel维度，即dim=1拼接
    27.         return out
    28.  
    29. class Decoupled_Detect(nn.Module):
    30.     stride = None  # strides computed during build
    31.     onnx_dynamic = False  # ONNX export parameter
    32.     export = False  # export mode
    33.  
    34.     def __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layer
    35.         super().__init__()
    36.  
    37.         self.nc = nc  # number of classes
    38.         self.no = nc + 5  # number of outputs per anchor
    39.         self.nl = len(anchors)  # number of detection layers
    40.         self.na = len(anchors[0]) // 2  # number of anchors
    41.         self.grid = [torch.zeros(1)] * self.nl  # init grid
    42.         self.anchor_grid = [torch.zeros(1)] * self.nl  # init anchor grid
    43.         self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2))  # shape(nl,na,2)
    44.         self.m = nn.ModuleList(DecoupledHead(x, nc, anchors) for x in ch)
    45.         self.inplace = inplace  # use in-place ops (e.g. slice assignment)
    46.  
    47.     def forward(self, x):
    48.         z = []  # inference output
    49.         for i in range(self.nl):
    50.             x[i] = self.m[i](x[i])  # conv
    51.             bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
    52.             x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
    53.  
    54.             if not self.training:  # inference
    55.                 if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
    56.                     self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
    57.  
    58.                 y = x[i].sigmoid()
    59.                 if self.inplace:
    60.                     y[..., 0:2] = (y[..., 0:2] * 2 + self.grid[i]) * self.stride[i]  # xy
    61.                     y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
    62.                 else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
    63.                     xy, wh, conf = y.split((2, 2, self.nc + 1), 4)  # y.tensor_split((2, 4, 5), 4)  # torch 1.8.0
    64.                     xy = (xy * 2 + self.grid[i]) * self.stride[i]  # xy
    65.                     wh = (wh * 2) ** 2 * self.anchor_grid[i]  # wh
    66.                     y = torch.cat((xy, wh, conf), 4)
    67.                 z.append(y.view(bs, -1, self.no))
    68.  
    69.         return x if self.training else (torch.cat(z, 1),) if self.export else (torch.cat(z, 1), x)
    70.  
    71.     def _make_grid(self, nx=20, ny=20, i=0):
    72.         d = self.anchors[i].device
    73.         t = self.anchors[i].dtype
    74.         shape = 1, self.na, ny, nx, 2  # grid shape
    75.         y, x = torch.arange(ny, device=d, dtype=t), torch.arange(nx, device=d, dtype=t)
    76.         if check_version(torch.__version__, '1.10.0'):  # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
    77.             yv, xv = torch.meshgrid(y, x, indexing='ij')
    78.         else:
    79.             yv, xv = torch.meshgrid(y, x)
    80.         grid = torch.stack((xv, yv), 2).expand(shape) - 0.5  # add grid offset, i.e. y = 2.0 * x - 0.5
    81.         anchor_grid = (self.anchors[i] * self.stride[i]).view((1, self.na, 1, 1, 2)).expand(shape)
    82.         return grid, anchor_grid

    2.2 Decoupled_Detect加入yolo.py中：

    class BaseModel(nn.Module):

    1.     def _apply(self, fn):
    2.         # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers
    3.         self = super()._apply(fn)
    4.         m = self.model[-1]  # Detect()
    5.         if isinstance(m, (Detect, Segment,Decoupled_Detect)):
    6.             m.stride = fn(m.stride)
    7.             m.grid = list(map(fn, m.grid))
    8.             if isinstance(m.anchor_grid, list):
    9.                 m.anchor_grid = list(map(fn, m.anchor_grid))
    10.         return self

    class DetectionModel(BaseModel):

    1.     def _initialize_dh_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
    2.         # https://arxiv.org/abs/1708.02002 section 3.3
    3.         # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
    4.         m = self.model[-1]  # Detect() module
    5.         for mi, s in zip(m.m, m.stride):  # from
    6.             # reg_bias = mi.reg_preds.bias.view(m.na, -1).detach()
    7.             # reg_bias += math.log(8 / (640 / s) ** 2)
    8.             # mi.reg_preds.bias = torch.nn.Parameter(reg_bias.view(-1), requires_grad=True)
    9.  
    10.             # cls_bias = mi.cls_preds.bias.view(m.na, -1).detach()
    11.             # cls_bias += math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # cls
    12.             # mi.cls_preds.bias = torch.nn.Parameter(cls_bias.view(-1), requires_grad=True)
    13.             b = mi.b3.bias.view(m.na, -1)
    14.             b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
    15.             mi.b3.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
    16.             b = mi.c3.bias.data
    17.             b += math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # cls
    18.             mi.c3.bias = torch.nn.Parameter(b, requires_grad=True)
    1.   if isinstance(m, (Detect, Segment,ASFF_Detect)):
    2.             s = 256  # 2x min stride
    3.             m.inplace = self.inplace
    4.             forward = lambda x: self.forward(x)[0] if isinstance(m, Segment) else self.forward(x)
    5.             m.stride = torch.tensor([s / x.shape[-2] for x in forward(torch.zeros(1, ch, s, s))])  # forward
    6.             check_anchor_order(m)
    7.             m.anchors /= m.stride.view(-1, 1, 1)
    8.             self.stride = m.stride
    9.             self._initialize_biases()  # only run once
    10.         elif isinstance(m, Decoupled_Detect):
    11.             s = 256  # 2x min stride
    12.             m.inplace = self.inplace
    13.             m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
    14.             check_anchor_order(m)  # must be in pixel-space (not grid-space)
    15.             m.anchors /= m.stride.view(-1, 1, 1)
    16.             self.stride = m.stride
    17.             self._initialize_dh_biases()  # only run once

    def parse_model(d, ch):  # model_dict, input_channels(3)

    1.         elif m in {Detect, Segment,Decoupled_Detect}:
    2.             args.append([ch[x] for x in f])
    3.             if isinstance(args[1], int):  # number of anchors
    4.                 args[1] = [list(range(args[1] * 2))] * len(f)
    5.             if m is Segment:
    6.                 args[3] = make_divisible(args[3] * gw, 8)

    2.3修改yolov5s_decoupled.yaml 

    1. # YOLOv5 🚀 by Ultralytics, GPL-3.0 license
    2.  
    3. # Parameters
    4. nc: 1  # number of classes
    5. depth_multiple: 0.33  # model depth multiple
    6. width_multiple: 0.50  # layer channel multiple
    7. anchors:
    8.   - [10,13, 16,30, 33,23]  # P3/8
    9.   - [30,61, 62,45, 59,119]  # P4/16
    10.   - [116,90, 156,198, 373,326]  # P5/32
    11.  
    12. # YOLOv5 v6.0 backbone
    13. backbone:
    14.   # [from, number, module, args]
    15.   [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
    16.    [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
    17.    [-1, 3, C3, [128]],
    18.    [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    19.    [-1, 6, C3, [256]],
    20.    [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    21.    [-1, 9, C3, [512]],
    22.    [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
    23.    [-1, 3, C3, [1024]],
    24.    [-1, 1, SPPF, [1024, 5]],  # 9
    25.   ]
    26.  
    27. # YOLOv5 v6.0 head
    28. head:
    29.   [[-1, 1, Conv, [512, 1, 1]],
    30.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
    31.    [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    32.    [-1, 3, C3, [512, False]],  # 13
    33.  
    34.    [-1, 1, Conv, [256, 1, 1]],
    35.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
    36.    [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    37.    [-1, 3, C3, [256, False]],  # 17 (P3/8-small)
    38.  
    39.    [-1, 1, Conv, [256, 3, 2]],
    40.    [[-1, 14], 1, Concat, [1]],  # cat head P4
    41.    [-1, 3, C3, [512, False]],  # 20 (P4/16-medium)
    42.  
    43.    [-1, 1, Conv, [512, 3, 2]],
    44.    [[-1, 10], 1, Concat, [1]],  # cat head P5
    45.    [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)
    46.  
    47.    [[17, 20, 23], 1, Decoupled_Detect, [nc, anchors]],  # Detect(P3, P4, P5),解耦
    48.   ]

  • 相关阅读:
    Dubbo分布式日志链路追踪
    vue computed计算属性及应用
    实在智能应邀出席中国移动科技工作者论坛，分享基于大模型+Agent智能体的“企业大脑”
    原生m1/m2 Sketch v90 for Mac最新中文版Sketch已更新，已解决闪退打不开问题
    ffmpeg在windows的安装、合并、切片、.m4s、.m3u8处理
    77-Mybatis-Plus详解
    esbuild中文文档-输出内容配置项（Output contents - Banner、Charset、Footer）
    springboot整合其他项目
    【数据结构】带头节点双向循环链表
    用MiniPC搭建个人服务器
  • 原文地址：https://blog.csdn.net/weixin_45303602/article/details/133185486