Deep Bayes: Normalizing flow

Introduction

这篇记录了Flow-based 生成模型。转载请注明。
Ref:
1.Deep Bayes slides
2.李宏毅machine learning

Change of variables formula

1d-case:
Given

Nd-case:

Coupling layers

Forward:

Inverse

Jacobian

NFs

1. Expressive parametric model
2. Tractable computation/optimization of an exact
3. Fast and tractable sampling

Multi-scale Architecture

Masked Coupling layers

Summary

Flow是一个比较可控的生成模型，但是它的生成能力可能还是不如GAN那么强，但是比较可以解释。