转自:https://blog.csdn.net/weixin_36714575/article/details/114238645
前言
在上次的文章中我们解析了backbone网络的构建源码,在这篇中我们针对model.py剩余的部分进行debug解析。如果没看过之前文章的小伙伴,推荐先查看这个系列的第一篇和第二篇。下面贴上传送门:
今天我们继续对model.py里的Detect类进行解析,这部分对应yolov5的检查头部分。
detect类在model.py里,这部分代码如下:
class Detect(nn.Module):
stride = None # strides computed during build
export = False # onnx export
def __init__(self, nc=80, anchors=(), ch=()): # detection layer
super(Detect, self).__init__()
self.nc = nc # number of classes
self.no = nc + 5 # number of outputs per anchor 85 for coco
self.nl = len(anchors) # number of detection layers 3
self.na = len(anchors[0]) // 2 # number of anchors 3
self.grid = [torch.zeros(1)] * self.nl # init grid
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
self.register_buffer('anchors', a) # shape(nl,na,2)
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv 128=>255/256=>255/512=>255
def forward(self, x):
# x = x.copy() # for profiling
z = [] # inference output
self.training |= self.export
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]:
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
y = x[i].sigmoid()
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
z.append(y.view(bs, -1, self.no))
return x if self.training else (torch.cat(z, 1), x)
@staticmethod
def _make_grid(nx=20, ny=20):
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
我们首先来看这个类的__init__()函数:
def __init__(self, nc=80, anchors=(), ch=()): # detection layer
super(Detect, self).__init__()
self.nc = nc # number of classes
self.no = nc + 5 # number of outputs per anchor 85 for coco
self.nl = len(anchors) # number of detection layers 3
self.na = len(anchors[0]) // 2 # number of anchors 3
self.grid = [torch.zeros(1)] * self.nl # init grid
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
self.register_buffer('anchors', a) # shape(nl,na,2)
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv 128=>255/256=>255/512=>255
yolov5的检测头仍为FPN结构,所以self.m为3个输出卷积。这三个输出卷积模块的channel变化分别为128=>255|256=>255|512=>255。
self.no为每个anchor位置的输出channel维度,每个位置都预测80个类(coco)+ 4个位置坐标xywh + 1个confidence score。所以输出channel为85。每个尺度下有3个anchor位置,所以输出85*3=255个channel。
下面我们再来看下head部分的forward()函数:
def forward(self, x):
# x = x.copy() # for profiling
z = [] # inference output
self.training |= self.export
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]:
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
y = x[i].sigmoid()
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
z.append(y.view(bs, -1, self.no))
return x if self.training else (torch.cat(z, 1), x)
x是一个列表的形式,分别对应着3个head的输入。它们的shape分别为:
- [B, 128, 32, 32]
- [B, 256, 16, 16]
- [B, 512, 8, 8]
三个输入先后被送入了3个卷积,得到输出结果。
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
这里将x进行变换从:
- x[0]:(bs,255,32,32) => x(bs,3,32,32,85)
- x[1]:(bs,255,32,32) => x(bs,3,16,16,85)
- x[2]:(bs,255,32,32) => x(bs,3,8,8,85)
if not self.training: # inference
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
def _make_grid(nx=20, ny=20):
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
这里的_make_grid()函数是准备好格点。所有的预测的单位长度都是基于grid层面的而不是原图。注意每一层的grid的尺寸都是不一样的,和每一层输出的尺寸w,h是一样的。
y = x[i].sigmoid()
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
z.append(y.view(bs, -1, self.no))
这里是inference的核心代码,我们要好好剖析一下,相比于yolov3,yolov5有一些变化:
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i]
这里可以明显发现box center的x,y的预测被乘以2并减去了0.5,所以这里的值域从yolov3里的(0,1)注意是开区间,变成了(-0.5, 1.5)。
这样改的原因目前还未知,从表面理解是可以跨半个格点预测了,这样应该能提高一些召回。当然还有一个好处就是也解决了yolov3中因为sigmoid开区间而导致中心无法到达边界处的问题。这里是我分析的观点,如果读者有其他的思路欢迎留言点拨。
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]
这里是预测boundingbox的wh。先回顾下yolov3里的预测:
pred_boxes[..., 0] = x.data + self.grid_x
pred_boxes[..., 1] = y.data + self.grid_y
pred_boxes[..., 2] = torch.exp(w.data) * self.anchor_w
pred_boxes[..., 3] = torch.exp(h.data) * self.anchor_h
是一个基于框的w,h的e指数函数。而在yolov5中这里变成了:(2*w_pred/h_pred) ^2。
值域从原来的(1,e)变成了(0,4)。这里我的理解是这个预测的框范围变得更大了,不仅可以预测到4倍以内的大物体,而且可以预测到比anchor小的boundingbox。和上面一样,这里是我分析的观点,如果读者有其他的思路欢迎留言点拨。
到这里我们就分析完了Detect类里面的所有代码。下面我回到Model类里面,最后分析它的前向传播过程,这里有两个函数forward()和forward_once()两个函数:
def forward_once(self, x, profile=False):
y, dt = [], [] # outputs
for m in self.model:
if m.f != -1: # if not from previous layer
x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers
if profile:
o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPS
t = time_synchronized()
for _ in range(10):
_ = m(x)
dt.append((time_synchronized() - t) * 100)
print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type))
x = m(x) # run
y.append(x if m.i in self.save else None) # save output
if profile:
print('%.1fms total' % sum(dt))
return x
self.foward_once()就是前向执行一次model里的所有module,得到结果。profile参数打开会记录每个模块的平均执行时长和flops用于分析模型的瓶颈,提高模型的执行速度和降低显存占用。
def forward(self, x, augment=False, profile=False):
if augment:
img_size = x.shape[-2:] # height, width
s = [1, 0.83, 0.67] # scales
f = [None, 3, None] # flips (2-ud, 3-lr)
y = [] # outputs
for si, fi in zip(s, f):
xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
yi = self.forward_once(xi)[0] # forward
# cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save
yi[..., :4] /= si # de-scale
if fi == 2:
yi[..., 1] = img_size[0] - 1 - yi[..., 1] # de-flip ud
elif fi == 3:
yi[..., 0] = img_size[1] - 1 - yi[..., 0] # de-flip lr
y.append(yi)
return torch.cat(y, 1), None # augmented inference, train
else:
return self.forward_once(x, profile) # single-scale inference, train
self.forward()函数里面augment可以理解为控制TTA,如果打开会对图片进行scale和flip。默认是关闭的。
def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416)
# scales img(bs,3,y,x) by ratio constrained to gs-multiple
if ratio == 1.0:
return img
else:
h, w = img.shape[2:]
s = (int(h * ratio), int(w * ratio)) # new size
img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize
if not same_shape: # pad/crop img
h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)]
return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean
scale_img的源码如上,就是通过普通的双线性插值实现,根据ratio来控制图片的缩放比例,最后通过pad 0补齐到原图的尺寸。
至此整个yolov5 head的前向传播和inference的源码就分析完了。整个实现debug下来感觉也是比较通俗易懂的,整体上yolov3的差距也不是特别大。