-
Wasserstein GAN
- the GAN training can also be interpreted as the minimization of the Jensen-Shannon (JS) divergence
- K-Lipschitz constraint $$ |D(x_1)-D(x_2)|\le K|x_1-x_2| $$
- if D satisfies K-Lipschitz constraint, minimax game of WGAN can be represented as $$ \min_{\theta_G}\max_{\theta_D}\mathbb E_x[D(x)]-E_z[D(G(z))] $$
- To make the discriminator be the K-Lipschitz: clamps all the weights in the discriminator to a fixed box denoted as
$w\in [-c, c]$
为了K-Lipschitz,D中的参数被clamp
-
Temporal GAN
$G_0: z_0\in \mathbb R^{k_0}\to [z_1^1,...,z_1^T]\in\mathbb R^{T\times K_1}$
$T$ 为时间,$[z_1^1,...,z_1^T]$ 为latent variables-
Generated video:
$[G_1(z_0,z_1^1),...,G_1(z_0,z_1^T)]$ -
objective $$ \min_{\theta_{G_0},\theta_{G_1}}\max_{\theta_{D}}\mathbb E_{[x^1,...,x^T]}[D([x^1,...,x^T])]-\mathbb E_{z_0}[D([G_1(z_0,z_1^1),...,G_1(z_0,z_1^T)])$$
-
Network configuration
G:Temporal generator + image generator D: 3D conv Training: Wasserstein GAN