模型配置文件
YOLO v5的模型配置文件都一样，区别在层深depth_multiple和宽度width_multiple控制不一样。YOLO v5s是最简洁的一个模型，深度为1就是说没有重复模块，因此方便用来分析其结构。模型的具体深度需要跑一下才能看到，或者将depth_multiple与各层 number相乘，按下式计算：
n = max(round(n * gd), 1) if n > 1 else n # depth gain
下面给出了具体的 YOLO v5s 参数配置信息：
from n params module arguments layer cin cout
---------------------------------------------------------------------------------------------------------------------------------------------
0 -1 1 3520 models.common.Focus [3, 32, 3] Focus 3 32
1 -1 1 18560 models.common.Conv [32, 64, 3, 2] Conv 32 64
2 -1 1 19904 models.common.BottleneckCSP [64, 64, 1] BottleneckCSP 64 64
3 -1 1 73984 models.common.Conv [64, 128, 3, 2] Conv 64 128
4 -1 1 161152 models.common.BottleneckCSP [128, 128, 3] BottleneckCSP 128 128
5 -1 1 295424 models.common.Conv [128, 256, 3, 2] Conv 128 256
6 -1 1 641792 models.common.BottleneckCSP [256, 256, 3] BottleneckCSP 256 256
7 -1 1 1180672 models.common.Conv [256, 512, 3, 2] Conv 256 512
8 -1 1 656896 models.common.SPP [512, 512, [5, 9, 13]] SPP 512 512
9 -1 1 1248768 models.common.BottleneckCSP [512, 512, 1, False] BottleneckCSP 512 512
10 -1 1 131584 models.common.Conv [512, 256, 1, 1] Conv 512 256
11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] Upsample 512 256
12 [-1, 6] 1 0 models.common.Concat [1] Concat 512 512
13 -1 1 378624 models.common.BottleneckCSP [512, 256, 1, False] BottleneckCSP 512 256
14 -1 1 33024 models.common.Conv [256, 128, 1, 1] Conv 256 128
15 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] Upsample 256 128
16 [-1, 4] 1 0 models.common.Concat [1] Concat 256 256
17 -1 1 95104 models.common.BottleneckCSP [256, 128, 1, False] BottleneckCSP 256 128
18 -1 1 2322 torch.nn.modules.conv.Conv2d [128, 18, 1, 1] Conv2d 128 255
19 -2 1 147712 models.common.Conv [128, 128, 3, 2] Conv 128 128
20 [-1, 14] 1 0 models.common.Concat [1] Concat 128 256
21 -1 1 313088 models.common.BottleneckCSP [256, 256, 1, False] BottleneckCSP 256 256
22 -1 1 4626 torch.nn.modules.conv.Conv2d [256, 18, 1, 1] Conv2d 256 255
23 -2 1 590336 models.common.Conv [256, 256, 3, 2] Conv 256 256
24 [-1, 10] 1 0 models.common.Concat [1] Concat 256 512
25 -1 1 1248768 models.common.BottleneckCSP [512, 512, 1, False] BottleneckCSP 512 512
26 -1 1 9234 torch.nn.modules.conv.Conv2d [512, 18, 1, 1] Conv2d 512 255
27 [-1, 22, 18] 1 0 Detect [1, anchors Detect 512 255
网络可视化
根据配置文件定义，将网络进行图1划分：
归纳整理得到图2：
搭建网络
根据网络划分和梳理的连接就可以自行搭建网络了。
class YoloModel(nn.Module):
anchors = [[116, 90, 156, 198, 373, 326],
[30, 61, 62, 45, 59, 119],
[10, 13, 16, 30, 33, 23]]
def __init__(self, class_num=1, input_ch=3):
super(YoloModel, self).__init__()
self.build_model(class_num)
# Build strides, anchors
s = 128 # 2x min stride
self.Detect.stride = torch.tensor(
[s / x.shape[-2] for x in self.forward(torch.zeros(1, input_ch, s, s))]) # forward
self.Detect.anchors /= self.Detect.stride.view(-1, 1, 1)
check_anchor_order(self.Detect)
self.stride = self.Detect.stride
# print('Strides: %s' % self.Detect.stride.tolist()) # [8.0, 16.0, 32.0]
print("Input size must be multiple of", self.stride.max().item())
torch_utils.initialize_weights(self)
self._initialize_biases() # only run once
# model_info(self)
def build_model(self, class_num):
# output channels
self.class_num = class_num
self.anchors_num = len(self.anchors[0]) // 2
self.output_ch = self.anchors_num * (5 + class_num)
# backbone
self.Focus = Focus(c1=3, c2=32, k=3, s=1)
self.CBL_1 = self.CBL(c1=32, c2=64, k=3, s=2)
self.CSP_1 = BottleneckCSP(c1=64, c2=64, n=1)
self.CBL_2 = self.CBL(c1=64, c2=128, k=3, s=2)
self.CSP_2 = BottleneckCSP(c1=128, c2=128, n=3)
self.CBL_3 = self.CBL(c1=128, c2=256, k=3, s=2)
self.CSP_3 = BottleneckCSP(c1=256, c2=256, n=3)
self.CBL_4 = self.CBL(c1=256, c2=512, k=3, s=2)
self.SPP = SPP(c1=512, c2=512, k=(5, 9, 13))
# head
self.CSP_4 = BottleneckCSP(c1=512, c2=512, n=1, shortcut=False)
self.CBL_5 = self.CBL(c1=512, c2=256, k=1, s=1)
self.Upsample_5 = nn.Upsample(size=None, scale_factor=2, mode="nearest")
self.Concat_5 = Concat(dimension=1)
self.CSP_5 = BottleneckCSP(c1=512, c2=256, n=1, shortcut=False)
self.CBL_6 = self.CBL(c1=256, c2=128, k=1, s=1)
self.Upsample_6 = nn.Upsample(size=None, scale_factor=2, mode="nearest")
self.Concat_6 = Concat(dimension=1)
self.CSP_6 = BottleneckCSP(c1=256, c2=128, n=1, shortcut=False)
self.Conv_6 = nn.Conv2d(in_channels=128, out_channels=self.output_ch, kernel_size=1, stride=1)
self.CBL_7 = self.CBL(c1=128, c2=128, k=3, s=2)
self.Concat_7 = Concat(dimension=1)
self.CSP_7 = BottleneckCSP(c1=256, c2=256, n=1, shortcut=False)
self.Conv_7 = nn.Conv2d(in_channels=256, out_channels=self.output_ch, kernel_size=1, stride=1)
self.CBL_8 = self.CBL(c1=256, c2=256, k=3, s=2)
self.Concat_8 = Concat(dimension=1)
self.CSP_8 = BottleneckCSP(c1=512, c2=512, n=1, shortcut=False)
self.Conv_8 = nn.Conv2d(in_channels=512, out_channels=self.output_ch, kernel_size=1, stride=1)
# detection
self.Detect = Detect(nc=self.class_num, anchors=self.anchors)
def forward(self, x):
# backbone
x = self.Focus(x) # 0
x = self.CBL_1(x)
x = self.CSP_1(x)
x = self.CBL_2(x)
y1 = self.CSP_2(x) # 4
x = self.CBL_3(y1)
y2 = self.CSP_3(x) # 6
x = self.CBL_4(y2)
x = self.SPP(x)
# head
x = self.CSP_4(x)
y3 = self.CBL_5(x) # 10
x = self.Upsample_5(y3)
x = self.Concat_5([x, y2])
x = self.CSP_5(x)
y4 = self.CBL_6(x) # 14
x = self.Upsample_6(y4)
x = self.Concat_6([x, y1])
y5 = self.CSP_6(x) # 17
output_1 = self.Conv_6(y5) # 18 output_1
x = self.CBL_7(y5)
x = self.Concat_7([x, y4])
y6 = self.CSP_7(x) # 21
output_2 = self.Conv_7(y6) # 22 output_2
x = self.CBL_8(y6)
x = self.Concat_8([x, y3])
x = self.CSP_8(x)
output_3 = self.Conv_8(x) # 26 output_3
output = self.Detect([output_1, output_2, output_3])
return output
@staticmethod
def CBL(c1, c2, k, s):
return nn.Sequential(
nn.Conv2d(c1, c2, k, s, autopad(k), bias=False),
nn.BatchNorm2d(c2),
nn.LeakyReLU(0.1, inplace=True),
)
def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
conv_layers = [self.Conv_6, self.Conv_7, self.Conv_8]
for conv_layer, s in zip(conv_layers, self.Detect.stride):
bias = conv_layer.bias.view(self.anchors_num, -1)
bias[:, 4] += math.log(8 / (640 / s) ** 2) # initialize confidence
bias[:, 5:] += math.log(0.6 / (self.class_num - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls
conv_layer.bias = torch.nn.Parameter(bias.view(-1), requires_grad=True)
检测模块
关于上图中的 Detect 模块需要指出的是，在ONNX中被转化成了 reshape + transpose，这是因为模型在导入ONNX时设置了参数self.Detect.export = True，根据检测端的源码可知，检测端在训练和模型导出时直接输出的是三个预测张量，其shape = (bs, na, H, W, no)，其中na*no=255，即图2中输出张量的通道数。这一变换过程对应源码：
bs, _, ny, nx = x[i].shape # x(bs,na×no,20,20) to x(bs,na,20,20,no)
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
变换结果：
input.shape = torch.Size([1, 3, 640, 640]) # NCHW
torch.Size([1, 3, 80, 80, 85])
torch.Size([1, 3, 40, 40, 85])
torch.Size([1, 3, 20, 20, 85])
而在Python端进行推理预测时，输出则是tuple(torch.cat(z, 1), x)，直接对第一项进行处理即可：共计25200个预测框，每个预测框包含了80个类的预测概率、4个边框坐标和1个置信度。就是说，在推理过程中，多进行了归纳合并这一步。
torch.Size([1, 25200, 85])
( 80 × 80 + 40 × 40 + 20 × 20 ) × 3 = 25200 (80 \times 80 + 40 \times 40 + 20 \times 20 ) \times3 = 25200 (80×80+40×40+20×20)×3=25200
下面是完整的Detect模块定义：
class Detect(nn.Module):
def __init__(self, nc=80, anchors=()): # detection layer
super(Detect, self).__init__()
self.stride = None # strides computed during build
self.nc = nc # number of classes
self.no = nc + 5 # channels of output tensor
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
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.export = False # model export
def forward(self, x):
# x = x.copy() # for profiling
z = [] # inference output
self.training |= self.export
for i in range(self.nl):
bs, _, ny, nx = x[i].shape # x(bs,na×no,20,20) to x(bs,na,20,20,no)
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)
7018
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
