2 BACKGROUND
2.2 GENERATIVE ADVERSARIAL NETS
GAN的min-max优化目标可以抽象为
min
θ
D
max
θ
G
E
x
∼
p
(
x
)
[
h
D
(
D
(
x
)
)
]
+
E
z
∼
p
(
z
)
[
h
G
(
D
(
G
(
z
)
)
)
]
(
1
)
\underset{\theta_D}{\min}\ \underset{\theta_G}{\max}\ \mathbb{E}_{x\sim p(x)}\left [ h_D\left ( D(x) \right ) \right ]+\mathbb{E}_{z\sim p(z)}\left [ h_G\left ( D\left ( G(z) \right ) \right ) \right ] \qquad(1)
θDmin θGmax Ex∼p(x)[hD(D(x))]+Ez∼p(z)[hG(D(G(z)))](1)
不同版本的GAN使用不同的
h
(
⋅
)
h(\cdot)
h(⋅)
【原始版本GAN】
min
θ
G
max
θ
D
E
x
∼
p
(
x
)
log
D
(
x
)
+
E
z
∼
p
(
z
)
log
(
1
−
D
(
G
(
z
)
)
)
\underset{\theta_G}{\min}\ \underset{\theta_D}{\max}\ \mathbb{E}_{x\sim p(x)}\log D(x)+\mathbb{E}_{z\sim p(z)}\log\left ( 1-D\left ( G(z) \right ) \right )
θGmin θDmax Ex∼p(x)logD(x)+Ez∼p(z)log(1−D(G(z)))
将
min
θ
G
max
θ
D
\underset{\theta_G}{\min}\ \underset{\theta_D}{\max}
θGmin θDmax变为
min
θ
D
max
θ
G
\underset{\theta_D}{\min}\ \underset{\theta_G}{\max}
θDmin θGmax,有
min
θ
D
max
θ
G
E
x
∼
p
(
x
)
−
log
D
(
x
)
+
E
z
∼
p
(
z
)
−
log
(
1
−
D
(
G
(
z
)
)
)
\underset{\theta_D}{\min}\ \underset{\theta_G}{\max}\ \mathbb{E}_{x\sim p(x)}-\log D(x)+\mathbb{E}_{z\sim p(z)}-\log\left ( 1-D\left ( G(z) \right ) \right )
θDmin θGmax Ex∼p(x)−logD(x)+Ez∼p(z)−log(1−D(G(z)))
于是可以看出
h
D
(
t
)
=
−
log
(
t
)
h_D(t) = -\log(t)
hD(t)=−log(t),
h
G
(
t
)
=
−
log
(
1
−
t
)
h_G(t) = -\log(1-t)
hG(t)=−log(1−t)
【WGAN】
对于
G
G
G:
max
θ
G
E
z
∼
p
(
z
)
D
(
G
(
z
)
)
\underset{\theta_G}{\max}\ \mathbb{E}_{z\sim p(z)}D(G(z))
θGmax Ez∼p(z)D(G(z))
对于
D
D
D:
min
θ
D
E
z
∼
p
(
z
)
D
(
G
(
z
)
)
−
E
x
∼
p
(
x
)
D
(
x
)
\underset{\theta_D}{\min}\ \mathbb{E}_{z\sim p(z)}D(G(z)) - \mathbb{E}_{x\sim p(x)}D(x)
θDmin Ez∼p(z)D(G(z))−Ex∼p(x)D(x)
合并起来有
min
θ
G
max
θ
D
E
x
∼
p
(
x
)
−
D
(
x
)
+
E
z
∼
p
(
z
)
D
(
G
(
z
)
)
\underset{\theta_G}{\min}\ \underset{\theta_D}{\max}\ \mathbb{E}_{x\sim p(x)}-D(x) + \mathbb{E}_{z\sim p(z)}D(G(z))
θGmin θDmax Ex∼p(x)−D(x)+Ez∼p(z)D(G(z))
于是可以看出
h
D
(
t
)
=
−
t
h_D(t)=-t
hD(t)=−t,
h
G
(
t
)
=
t
h_G(t)=t
hG(t)=t
为了简化公式,令
f
(
z
)
=
D
(
G
(
z
)
)
f(z)=D(G(z))
f(z)=D(G(z)),则公式(1)简化为
min
θ
D
max
θ
G
E
x
∼
p
(
x
)
[
h
D
(
D
(
x
)
)
]
+
E
z
∼
p
(
z
)
[
h
G
(
f
(
z
)
)
]
\underset{\theta_D}{\min}\ \underset{\theta_G}{\max}\ \mathbb{E}_{x\sim p(x)}\left [ h_D\left ( D(x) \right ) \right ]+\mathbb{E}_{z\sim p(z)}\left [ h_G\left ( f(z) \right ) \right ]
θDmin θGmax Ex∼p(x)[hD(D(x))]+Ez∼p(z)[hG(f(z))]