PixelCNN

Mask type A 和 Mask type B  

所有PixelCNN类型 Autoregressive models — PixelCNN Modelling coloured images PixelCNN’s blind spot in the receptive field Fixing the blind spot — Gated PixelCNN Conditional generation with Gated PixelCNN Gated PixelCNN with cropped convolutions Improving performance — PixelCNN++ Improving sampling time — Fast PixelCNN++ Using attention mechanisms — PixelSNAIL Generating Diverse High-Fidelity Images — VQ-VAE 2

最早，由Google的Deepmind团队宣布通过其最新的Wavenet模型获得了良好的音频生成效果。 PixelCNN的基本网络结构如下图： 其中的残差块Residual Blocks如下：

Fully visible belief network Explicit density model Use chain rule to decompose likelihood of an image x into product of 1D  distributions: Then maximize likelihood of training data Complex distribution over pixel values => Express using a neural  network! Will need to define  ordering of “previous  pixels”. Generate image pixels starting from corner Dependency on previous pixels modeled  using an RNN (LSTM) Drawback: sequential generation is slow!