all convolution net: 不使用常用的确定性池化，改用步长为2的卷积让模型学到适合自己的下采样。
去掉了全连接层
BN: 在所有层后加BN，会使模型不稳定。去掉G的输出层，和D的输入层后的BN后可以避免该问题。
ReLU in Generator except for the output layer which use Tanh
LeakyReLU in Discriminator
This app can be installed on your PC or mobile device. This will allow this web app to look and behave like any other installed app. You will find it in your app lists and be able to pin it to your home screen, start menus or task bars. This installed web app will also be able to safely interact with other apps and your operating system.
Leonhardt's Blog