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

class Detect(nn.Module):stride = None  # strides computed during buildonnx_dynamic = False  # ONNX export parameterdef __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layersuper().__init__()self.nc = nc  # number of classesself.no = nc + 5  # number of outputs per anchorself.nl = len(anchors)  # number of detection layersself.na = len(anchors[0]) // 2  # number of anchorsself.grid = [torch.zeros(1)] * self.nl  # init gridself.anchor_grid = [torch.zeros(1)] * self.nl  # init anchor gridself.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 convself.m_box = nn.ModuleList(nn.Conv2d(256, 4 * self.na, 1) for x in ch)  # output convself.m_conf = nn.ModuleList(nn.Conv2d(256, 1 * self.na, 1) for x in ch)  # output convself.m_labels = nn.ModuleList(nn.Conv2d(256, self.nc * self.na, 1) for x in ch)  # output convself.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 = chdef forward(self, x):z = []  # inference outputfor 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_featurem_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:  # inferenceif 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]  # xyy[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # whelse:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  # xywh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # why = 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

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 moduleclass 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 paddingpad = (ksize - 1) // 2self.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

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

class Detect(nn.Module):stride = None  # strides computed during buildonnx_dynamic = False  # ONNX export parameterdef __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layersuper().__init__()self.nc = nc  # number of classesself.no = nc + 5  # number of outputs per anchorself.nl = len(anchors)  # number of detection layersself.na = len(anchors[0]) // 2  # number of anchorsself.grid = [torch.zeros(1)] * self.nl  # init gridself.anchor_grid = [torch.zeros(1)] * self.nl  # init anchor gridself.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 convself.inplace = inplace  # use in-place ops (e.g. slice assignment)def forward(self, x):z = []  # inference outputfor i in range(self.nl):x[i] = self.m[i](x[i])  # convbs, _, 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:  # inferenceif 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]  # xyy[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # whelse:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  # xywh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # why = 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