YOLOv5的head详解
在前两篇文章中我们对YOLO的backbone和neck进行了详尽的解读，如果有小伙伴没看这里贴一下传送门：
YOLOv5的Backbone设计
YOLOv5的Neck端设计
在这篇文章中，我们将针对YOLOv5的head进行解读，head虽然在网络中占比最少，但这却是YOLO最核心的内容，话不多说，进入正题。
1 YOLOv5s网络结构总览
要了解head，就不能将其与前两部分割裂开。head中的主体部分就是三个Detect检测器，即利用基于网格的anchor在不同尺度的特征图上进行目标检测的过程。由下面的网络结构图可以很清楚的看出：当输入为640*640时，三个尺度上的特征图分别为：80x80、40x40、20x20。现在问题的关键变为，Detect的过程细节是怎样的？如何在多个检测框中选择效果最好的？
2 YOLO核心：Detect
首先看一下yolo中Detect的源码组成：
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.grid[i].shape[2:4] != x[i].shape[2:4] or self.onnx_dynamic:
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)
def _make_grid(self, nx=20, ny=20, i=0):
d = self.anchors[i].device
yv, xv = torch.meshgrid([torch.arange(ny).to(d), torch.arange(nx).to(d)])
grid = torch.stack((xv, yv), 2).expand((1, self.na, ny, nx, 2)).float()
anchor_grid = (self.anchors[i].clone() * self.stride[i]) \
.view((1, self.na, 1, 1, 2)).expand((1, self.na, ny, nx, 2)).float()
return grid, anchor_grid
Detect很重要，但是内容不多，那我们就将其解剖开来，一部分一部分地看。
2.1 initial部分
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)
self.anchor=anchors
initial部分定义了Detect过程中的重要参数
1. nc:类别数目
2. no:每个anchor的输出，包含类别数nc+置信度1+xywh4，故nc+5
3. nl:检测器的个数。以上图为例，我们有3个不同尺度上的检测器：[[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]]，故检测器个数为3。
4. na:每个检测器中anchor的数量，个数为3。由于anchor是w h连续排列的，所以需要被2整除。
5. grid:检测器Detect的初始网格
6. anchor_grid:anchor的初始网格
7. m：每个检测器的最终输出，即检测器中anchor的输出no×anchor的个数nl。打印出来很好理解（60是因为我的数据集nc为15，coco是80）：
ModuleList(
(0): Conv2d(128, 60, kernel_size=(1, 1), stride=(1, 1))
(1): Conv2d(256, 60, kernel_size=(1, 1), stride=(1, 1))
(2): Conv2d(512, 60, kernel_size=(1, 1), stride=(1, 1))
)
2.2 forward
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.grid[i].shape[2:4] != x[i].shape[2:4] or self.onnx_dynamic:
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)
在forward操作中，网络接收3个不同尺度的特征图，如下图所示：
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()
网络的for loop次数为3，也就是依次在这3个特征图上进行网格化预测，利用卷积操作得到通道数为no×nl的特征输出。拿128x80x80举例，在nc=15的情况下经过卷积得到60x80x80的特征图，这个特征图就是后续用于格点检测的特征图。
if not self.training: # inference
if self.grid[i].shape[2:4] != x[i].shape[2:4] or self.onnx_dynamic:
self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
def _make_grid(self, nx=20, ny=20, i=0):
d = self.anchors[i].device
yv, xv = torch.meshgrid([torch.arange(ny).to(d), torch.arange(nx).to(d)])
grid = torch.stack((xv, yv), 2).expand((1, self.na, ny, nx, 2)).float()
anchor_grid = (self.anchors[i].clone() * self.stride[i]) \
.view((1, self.na, 1, 1, 2)).expand((1, self.na, ny, nx, 2)).float()
return grid, anchor_grid
随后就是基于经过检测器卷积后的特征图划分网格，网格的尺寸是与输入尺寸相同的，如20x20的特征图会变成20x20的网格，那么一个网格对应到原图中就是32x32像素；40x40的一个网格就会对应到原图的16x16像素，以此类推。
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))
这里其实就是预测偏移的主体部分了。
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
这一句是对x和y进行预测。x、y在输入网络前都是已经归一好的（0,1），乘以2再减去0.5就是（-0.5，1.5），也就是让x、y的预测能够跨网格进行。后边self.grid[i]) * self.stride[i]就是将相对位置转为网格中的绝对位置了。
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
这里对宽和高进行预测，没啥好说的。
z.append(y.view(bs, -1, self.no))
最后将结果填入z