[ICLR 2016] Unsupervised representation learning with DCGANs
目录
- Introduction
- Model Architecture
- References
Introduction
- 作者提出了用 CNN 搭建 GAN，使得 GAN 训练更加稳定的一系列准则，并将满足这些设计理念的模型称为 DCGANs (Deep Convolutional GANs). 此外，作者将 trained discriminators 用于图像分类任务，相比于其他无监督算法取得了较好的效果，这证明模型通过无监督学习学得了很好的图像特征表示
Model Architecture
- (1) Replace any pooling layers with strided convolutions (discriminator) (下采样) and fractional-strided convolutions (generator) (上采样).
- (2) Remove fully connected hidden layers for deeper architectures. generator 的输入层为全连接层，输入 uniform noise distribution