yolov5中head修改为decouple head详解

 现成的YOLOv5代码真的很香,不管口碑怎么样,我用着反正是挺爽的,下面这篇文章主要给大家介绍了关于yolov5中head修改为decouple head的相关资料,需要的朋友可以参考下

 

目录

• yolox的decoupled head结构
• 对于decouple head的改进
• 特点
• 疑问
• 总结

yolov5的head修改为decouple head

yolox的decoupled head结构

 

本来想将yolov5的head修改为decoupled head，与yolox的decouple head对齐，但是没注意，该成了如下结构：

 

感谢少年肩上杨柳依依的指出，如还有问题欢迎指出

 

1.修改models下的yolo.py文件中的Detect

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64

class Detect(nn.Module):     stride = None  # strides computed during build     onnx_dynamic = False  # ONNX export parameter       def __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layer         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         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         self.m_box = nn.ModuleList(nn.Conv2d(256, 4 * self.na, 1) for x in ch)  # output conv         self.m_conf = nn.ModuleList(nn.Conv2d(256, 1 * self.na, 1) for x in ch)  # output conv         self.m_labels = nn.ModuleList(nn.Conv2d(256, self.nc * self.na, 1) for x in ch)  # output conv         self.base_conv = nn.ModuleList(BaseConv(in_channels = x, out_channels = 256, ksize = 1, stride = 1) for x in ch)         self.cls_convs = nn.ModuleList(BaseConv(in_channels = 256, out_channels = 256, ksize = 3, stride = 1) for x in ch)         self.reg_convs = nn.ModuleList(BaseConv(in_channels = 256, out_channels = 256, ksize = 3, stride = 1) for x in ch)                   # self.m = nn.ModuleList(nn.Conv2d(x, 4 * self.na, 1) for x in ch, nn.Conv2d(x, 1 * self.na, 1) for x in ch,nn.Conv2d(x, self.nc * self.na, 1) for x in ch)         self.inplace = inplace  # use in-place ops (e.g. slice assignment)self.ch = ch       def forward(self, x):         z = []  # inference output         for i in range(self.nl):             # # x[i] = self.m[i](x[i])  # convs             # print("&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&", i)             # print(x[i].shape)             # print(self.base_conv[i])             # print("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")                                                       x_feature = self.base_conv[i](x[i])             # x_feature = x[i]                           cls_feature = self.cls_convs[i](x_feature)             reg_feature = self.reg_convs[i](x_feature)             # reg_feature = x_feature                           m_box = self.m_box[i](reg_feature)             m_conf = self.m_conf[i](reg_feature)             m_labels = self.m_labels[i](cls_feature)             x[i] = torch.cat((m_box,m_conf, m_labels),1)             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()               if not self.training:  # inference                 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()                 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)

 

2.在yolo.py中添加

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41

def get_activation(name="silu", inplace=True):     if name == "silu":         module = nn.SiLU(inplace=inplace)     elif name == "relu":         module = nn.ReLU(inplace=inplace)     elif name == "lrelu":         module = nn.LeakyReLU(0.1, inplace=inplace)     else:         raise AttributeError("Unsupported act type: {}".format(name))     return module       class BaseConv(nn.Module):     """A Conv2d -> Batchnorm -> silu/leaky relu block"""       def __init__(         self, in_channels, out_channels, ksize, stride, groups=1, bias=False, act="silu"     ):         super().__init__()         # same padding         pad = (ksize - 1) // 2         self.conv = nn.Conv2d(             in_channels,             out_channels,             kernel_size=ksize,             stride=stride,             padding=pad,             groups=groups,             bias=bias,         )         self.bn = nn.BatchNorm2d(out_channels)         self.act = get_activation(act, inplace=True)       def forward(self, x):         # print(self.bn(self.conv(x)).shape)         return self.act(self.bn(self.conv(x)))         # return self.bn(self.conv(x))       def fuseforward(self, x):         return self.act(self.conv(x))

 

decouple head的特点：

由于训练模型时，应该是channels = 256的地方改成了channels = x(失误)，所以在decoupled head的部分参数量比yolox要大一些，以下的结果是在channels= x的情况下得出

比yolov5s参数多，计算量大，在我自己的2.5万的数据量下map提升了3%多

1.模型给出的目标cls较高，需要将conf的阈值设置较大（0.5），不然准确率较低

?

1	parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold')

 

2.对于少样本的检测效果较好，召回率的提升比准确率多

3.在conf设置为0.25时，召回率比yolov5s高，但是准确率低；在conf设置为0.5时，召回率与准确率比yolov5s高

4.比yolov5s参数多，计算量大，在2.5万的数据量下map提升了3%多

对于decouple head的改进

 

改进：

1.将红色框中的conv去掉，缩小参数量和计算量；

2.channels =256 ，512 ，1024是考虑不增加参数，不进行featuremap的信息压缩

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

class Detect(nn.Module):     stride = None  # strides computed during build     onnx_dynamic = False  # ONNX export parameter       def __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layer         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         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         self.inplace = inplace  # use in-place ops (e.g. slice assignment)       def forward(self, x):         z = []  # inference output         for i in range(self.nl):             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()               if not self.training:  # inference                 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()                 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)

 

特点

1.模型给出的目标cls较高，需要将conf的阈值设置较大（0.4），不然准确率较低

2.对于少样本的检测效果较好，准确率的提升比召回率多

3. 准确率的提升比召回率多，

该改进不如上面的模型提升多，但是参数量小，计算量小少9Gflop，占用显存少

decoupled head指标提升的原因：由于yolov5s原本的head不能完全的提取featuremap中的信息，decoupled head能够较为充分的提取featuremap的信息；

疑问

为什么decoupled head目标的cls会比较高，没想明白

为什么去掉base_conv，召回率要比准确率提升少

 

 

原文链接：https://blog.csdn.net/qq_34496674/article/details/124828868