Flow++: Improving Flow-Based Generative Models with Variational - - PowerPoint PPT Presentation

▶
flow improving flow based generative models with
SMART_READER_LITE
LIVE PREVIEW

Flow++: Improving Flow-Based Generative Models with Variational - - PowerPoint PPT Presentation

Jun 26, 2023 245 likes •365 views

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design Jonathan Ho*, Xi Chen*, Aravind Srinivas, Yan Duan, Pieter Abbeel Overview - Goal: likelihood-based model with Fast sampling and training -

slide-1
SLIDE 1

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Jonathan Ho*, Xi Chen*, Aravind Srinivas, Yan Duan, Pieter Abbeel

slide-2
SLIDE 2

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Overview

  • Goal: likelihood-based model with
  • Fast sampling and training
  • Good samples and density estimation performance
  • Our strategy: improve flow models
  • Uniform dequantization -> variational dequantization
  • Affine coupling -> mixture of logistics coupling
  • Convolutions -> convolutions + self-attention
slide-3
SLIDE 3

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Continuous flows for discrete data

■ A problem arises when fitting continuous density models to

discrete data: degeneracy

■

When the data are 3-bit pixel values,

■

What density does a model assign to values between bins like 0.4, 0.42…?

■ Correct semantics: we want the integral of probability density

within a discrete interval to approximate discrete probability mass

3

slide-4
SLIDE 4

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Continuous flows for discrete data

■ Solution: Dequantization. Add noise to data.

■ ■

We draw noise u uniformly from

4

[Theis, Oord, Bethge, 2016]

slide-5
SLIDE 5

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Variational Dequantization

5

■ Variational Dequantization. Add a learnable noise q to data. [Ho et al., 2019]

slide-6
SLIDE 6

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Coupling layers

RealNVP Ours: logistic mixture CDF convolutions convolutions & self-attention

slide-7
SLIDE 7

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Ablation on CIFAR

7

slide-8
SLIDE 8

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Results

slide-9
SLIDE 9

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Samples (CIFAR10, ImageNet 64x64)

slide-10
SLIDE 10

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

Samples (CelebA 5-bit)

slide-11
SLIDE 11

Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

■ Slides adapted from Berkeley CS294-158 Deep Unsupervised

Learning class:

https://sites.google.com/view/berkeley-cs294-158-sp19/home

■

Want to learn more about foundation of Deep Generative Models & Self-Supervised learning methods?

■

All lecture videos are available on youtube, featuring guest speakers: Ilya Sutskever, Alyosha Efros, Alec Radford, Aaron van den Oord