YOLOv5介绍
YOLOv5为兼顾速度与性能的目标检测算法。笔者将在近期更新一系列YOLOv5的代码导读博客。YOLOv5为2021.1.5日发布的4.0版本。
YOLOv5开源项目github网址
源代码导读汇总网址
本博客导读的代码为utils文件夹下的
activations.py
最后更新日期为2021.1.7
激活函数-数学表达
首先介绍YOLOv5激活函数中涉及到的函数其具体形式:
Sigmoid激活函数:
S
i
g
m
o
i
d
(
x
)
=
1
1
+
e
−
x
=
e
x
e
x
+
1
Sigmoid(x) = \frac{1}{1+e^{-x} } = \frac{e^{x}}{e^{x}+1 }
S
i
g
m
o
i
d
(
x
)
=
1
+
e
−
x
1
=
e
x
+
1
e
x
Swish激活函数:
S
w
i
s
h
(
x
)
=
x
1
+
e
−
β
x
Swish(x) = \frac{x}{1+e^{-\beta x}}
S
w
i
s
h
(
x
)
=
1
+
e
−
β
x
x
ReLU激活函数:
R
e
L
U
(
x
)
=
{
x
,
if
x
>
0
0
,
if
x
≤
0
=
m
a
x
(
0
,
x
)
ReLU(x)= \begin{cases} x, & \text {if $x$ $>$ 0} \\ 0, & \text{if $x$ $\leq$ 0} \end{cases} = max(0,x)
R
e
L
U
(
x
)
=
{
x
,
0
,
if
x
>
0
if
x
≤
0
=
m
a
x
(
0
,
x
)
ReLU6限制上述函数输出在6以内:
R
e
L
U
6
(
x
)
=
m
i
n
(
6
,
m
a
x
(
0
,
x
)
)
ReLU6(x) = min(6,max(0,x))
R
e
L
U
6
(
x
)
=
m
i
n
(
6
,
m
a
x
(
0
,
x
)
)
Hard-swish激活函数:
H
a
r
d
−
s
w
i
s
h
(
x
)
=
x
∗
R
e
L
U
6
(
x
+
3
)
6
Hard-swish(x) = \frac{x \ast ReLU6(x+3)}{6 }
H
a
r
d
−
s
w
i
s
h
(
x
)
=
6
x
∗
R
e
L
U
6
(
x
+
3
)
Swish激活函数求导过程(
取
β
=
1
取\beta=1
取
β
=
1
):
S
w
i
s
h
′
(
x
)
=
1
+
e
−
x
+
x
e
−
x
(
1
+
e
−
x
)
2
=
1
1
+
e
−
x
+
x
e
−
x
(
1
+
e
−
x
)
2
Swish^{‘}(x) = \frac{1+e^{-x}+xe^{-x}}{(1+e^{-x})^2} = \frac{1}{1+e^{-x} } + \frac{xe^{-x}}{(1+e^{-x})^2}
S
w
i
s
h
′
(
x
)
=
(
1
+
e
−
x
)
2
1
+
e
−
x
+
x
e
−
x
=
1
+
e
−
x
1
+
(
1
+
e
−
x
)
2
x
e
−
x
=
1
1
+
e
−
x
+
x
1
+
e
−
x
∗
(
1
−
1
1
+
e
−
x
)
=\frac{1}{1+e^{-x} } + \frac{x}{1+e^{-x} } \ast (1 – \frac{1}{1+e^{-x}} )
=
1
+
e
−
x
1
+
1
+
e
−
x
x
∗
(
1
−
1
+
e
−
x
1
)
tanh函数:
t
a
n
h
(
x
)
=
e
x
−
e
−
x
e
x
+
e
−
x
tanh(x) = \frac{e^{x}-e^{-x}}{e^{x}+e^{-x} }
t
a
n
h
(
x
)
=
e
x
+
e
−
x
e
x
−
e
−
x
softplus函数(ReLU函数的平滑版):
s
o
f
t
p
l
u
s
(
x
)
=
l
n
(
1
+
e
x
)
softplus(x) = ln({1+e^{x} })
s
o
f
t
p
l
u
s
(
x
)
=
l
n
(
1
+
e
x
)
Mish激活函数:
M
i
s
h
(
x
)
=
x
∗
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
Mish(x) = x \ast tanh(softplus(x))
M
i
s
h
(
x
)
=
x
∗
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
Mish激活函数求导过程:
M
i
s
h
′
(
x
)
=
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
+
x
∗
(
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
)
′
Mish^{‘}(x) = tanh(softplus(x))+x \ast (tanh(softplus(x)))^{‘}
M
i
s
h
′
(
x
)
=
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
+
x
∗
(
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
)
′
设
f
x
=
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
设fx = tanh(softplus(x))
设
f
x
=
t
a
n
h
(
s
o
f
t
p
l
u
s
(
x
)
)
M
i
s
h
′
(
x
)
=
f
x
+
x
∗
s
o
f
t
p
l
u
s
′
(
x
)
∗
t
a
n
h
′
(
s
o
f
t
p
l
u
s
(
x
)
)
Mish^{‘}(x) = fx + x \ast softplus^{‘}(x) \ast tanh^{‘}(softplus(x))
M
i
s
h
′
(
x
)
=
f
x
+
x
∗
s
o
f
t
p
l
u
s
′
(
x
)
∗
t
a
n
h
′
(
s
o
f
t
p
l
u
s
(
x
)
)
s
o
f
t
p
l
u
s
′
(
x
)
=
l
n
(
1
+
e
x
)
′
=
e
x
e
x
+
1
=
s
i
g
m
o
i
d
(
x
)
softplus^{‘}(x) = ln({1+e^{x} })^{‘} = \frac{e^{x}}{e^{x}+1} = sigmoid(x)
s
o
f
t
p
l
u
s
′
(
x
)
=
l
n
(
1
+
e
x
)
′
=
e
x
+
1
e
x
=
s
i
g
m
o
i
d
(
x
)
t
a
n
h
′
(
x
)
=
1
−
t
a
n
h
2
(
x
)
tanh^{‘}(x) = 1 – tanh^{2}(x)
t
a
n
h
′
(
x
)
=
1
−
t
a
n
h
2
(
x
)
则:
M
i
s
h
′
(
x
)
=
f
x
+
x
∗
s
i
g
m
o
i
d
(
x
)
∗
(
1
−
(
f
x
)
2
)
Mish^{‘}(x) = fx + x \ast sigmoid(x) \ast (1 – (fx)^{2})
M
i
s
h
′
(
x
)
=
f
x
+
x
∗
s
i
g
m
o
i
d
(
x
)
∗
(
1
−
(
f
x
)
2
)
activations.py
该文件定义了Yolov5中的各种激活函数。各种激活函数的表达形式及求导过程均总结如上。
# Activation functions
#定义了网络架构中使用的激活函数
import torch
import torch.nn as nn
import torch.nn.functional as F
swish类定义了Swish激活函数
# Swish https://arxiv.org/pdf/1905.02244.pdf ---
class Swish(nn.Module): #Swish激活函数
@staticmethod
def forward(x): # 此处beta定为1
return x * torch.sigmoid(x)
hardswish类定义了hardswish激活函数
class Hardswish(nn.Module): # nn.Hardswish()的输出友好版本
@staticmethod
def forward(x): #数学形式见上文
# return x * F.hardsigmoid(x) # for torchscript and CoreML
return x * F.hardtanh(x + 3, 0., 6.) / 6. # for torchscript, CoreML and ONNX
高效应用swish激活函数的类
class MemoryEfficientSwish(nn.Module): #节省内存的Swish 不采用自动求导
class F(torch.autograd.Function):
@staticmethod
def forward(ctx, x):
#save_for_backward会保留x的全部信息(一个完整的外挂Autograd Function的Variable),
#并提供避免in-place操作导致的input在backward被修改的情况.
#in-place操作指不通过中间变量计算的变量间的操作。
ctx.save_for_backward(x)
return x * torch.sigmoid(x)
@staticmethod
def backward(ctx, grad_output):
#此处saved_tensors[0] 作用同上文 save_for_backward
x = ctx.saved_tensors[0]
sx = torch.sigmoid(x)
# 返回该激活函数求导之后的结果 求导过程见上文
return grad_output * (sx * (1 + x * (1 - sx)))
def forward(self, x): #应用前向传播方法
return self.F.apply(x)
Mish类定义了Mish激活函数
# Mish https://github.com/digantamisra98/Mish ----
class Mish(nn.Module): #这里定义了Mish激活函数
@staticmethod #数学公式见上
def forward(x):
return x * F.softplus(x).tanh()
高效应用Mish激活函数的类
class MemoryEfficientMish(nn.Module): # 节省内存版
class F(torch.autograd.Function): #该类形式内容同上一个类
@staticmethod
def forward(ctx, x):
ctx.save_for_backward(x)
return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x)))
@staticmethod
def backward(ctx, grad_output):
x = ctx.saved_tensors[0]
sx = torch.sigmoid(x) # sx = sigmoid(x)
fx = F.softplus(x).tanh() # fx = tanh(softplus(x))
return grad_output * (fx + x * sx * (1 - fx * fx))
def forward(self, x):
return self.F.apply(x)
定义了FReLU激活函数
# FReLU https://arxiv.org/abs/2007.11824 ---
# 旷视研究院2020ECCV 提出的激活函数 (形式真的简单)
class FReLU(nn.Module):
def __init__(self, c1, k=3): # ch_in, kernel
super().__init__()
self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1)
self.bn = nn.BatchNorm2d(c1)
def forward(self, x):
return torch.max(x, self.bn(self.conv(x)))