该博客是根据台大李宏毅老师的关于GAN的视频教程所整理的笔记,建议大家可以直接看这个老师的视频(因在youtube,故需翻墙)
文章目录
Improving Generative Adversarial Network
![](https://img-blog.csdnimg.cn/20190415133114798.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2RlZXBpbkM=,size_16,color_FFFFFF,t_70)
f-GAN
总结:除了可以使用 J-S divergence \text{J-S divergence} J-S divergence 外,还可以选择其它任何 f-divergence 来评测两个分布的差异度,即任意 f-divergence 均可套用到 GANs 框架中
f-divergence
P P P 和 Q Q Q 是两个分布。 p ( x ) p(x) p(x) 和 q ( x ) q(x) q(x) 是从这两个分布中采样得到 x x x 的概率,f-divergence 的计算公式为:
D f ( P ∣ Q ) = ∫ x q ( x ) f ( p ( x ) q ( x ) ) d x ≥ 0 , s . t . , f is convex , f ( 1 ) = 0 D_f(P|Q) = \int_x q(x) f(\frac{p(x)}{q(x)}) dx \geq 0, \it{ s.t.}, f \text{ is convex}, f(1)=0 Df(P∣Q)=∫xq(x)f(q(x)p(x))dx≥0,s.t.,f is convex,f(1)=0
该公式用于评测这两个分布之间的差异,值越高,差异越大,反之越小。
(Reverse) KL-divergence 是 f-divergence 的一种特例:
f ( x ) = x log x → D f ( P ∥ Q ) = ∫ x q ( x ) p ( x ) q ( x ) log p ( x ) q ( x ) d x = ∫ x p ( x ) log p ( x ) q ( x ) d x (KL) f ( x ) = − log x → D f ( P ∥ Q ) = ∫ x q ( x ) ( − log p ( x ) q ( x ) ) d x = ∫ x q ( x ) log q ( x ) p ( x ) d x (Reverse KL) f ( x ) = ( x − 1 ) 2 → D f ( P ∥ Q ) = ∫ x q ( x ) ( p ( x ) q ( x ) − 1 ) 2 d x = ∫ x ( p ( x ) − q ( x ) ) 2 q ( x ) d x (Chi Square) \begin{array}{lll} f(x) = x \log x &\to& D_f(P\|Q) = \int_x q(x) \frac{p(x)}{q(x)} \log \frac{p(x)}{q(x)} dx = \int_x p(x) \log \frac{p(x)}{q(x)} dx \ \text{(KL)} \\ f(x) = - \log x &\to& D_f(P\|Q) = \int_x q(x) \left(- \log \frac{p(x)}{q(x)} \right) dx = \int_x q(x) \log \frac{q(x)}{p(x)} dx \ \text{(Reverse KL)} \\ f(x) = (x-1)^2 &\to& D_f(P\|Q) = \int_x q(x) \left( \frac{p(x)}{q(x)} -1 \right)^2 dx = \int_x \frac{(p(x)-q(x))^2}{q(x)} dx \ \text{(Chi Square)} \end{array} f(x)=xlogxf(x)=−logxf(x)=(x−1)2→→→Df(P∥Q)=∫xq(x)q(x)p(x)logq(x)p(x)dx=∫xp(x)logq(x)p(x)dx (KL)Df(P∥Q)=∫xq(x)(−logq(x)p(x))dx=∫xq(x)logp(x)q(x)dx (Reverse KL)Df(P∥Q)=∫xq(x)(q(x)p(x)−1)2dx=∫xq(x)(p(x)−q(x))2dx (Chi Square)
Fenchel Conjugate (共轭函数)
每一个凸函数 f f f 都有一个共轭函数 f ∗ f^* f∗: f ∗ ( t ) = max x t ∈ d o m ( f )