Metropolis-Hastings Generative Adversarial Networks Ryan Turner, - - PowerPoint PPT Presentation

▶
metropolis hastings generative adversarial networks
SMART_READER_LITE
LIVE PREVIEW

Metropolis-Hastings Generative Adversarial Networks Ryan Turner, - - PowerPoint PPT Presentation

Jan 28, 2023 205 likes •492 views

Metropolis-Hastings Generative Adversarial Networks Ryan Turner, Jane Hung, Eric Frank, Yunus Saatci, Jason Yosinski Poster #201 Typical GAN training Typical GAN training D tries to get here G tries to get here Typical GAN training D

slide-1
SLIDE 1

Metropolis-Hastings Generative Adversarial Networks

Ryan Turner, Jane Hung, Eric Frank, Yunus Saatci, Jason Yosinski Poster #201

slide-2
SLIDE 2

Typical GAN training

slide-3
SLIDE 3

Typical GAN training

D tries to get here G tries to get here

slide-4
SLIDE 4

Typical GAN training

D tries to get here G tries to get here

slide-5
SLIDE 5

Typical GAN training

D tries to get here G tries to get here

slide-6
SLIDE 6

Typical GAN training

D tries to get here G tries to get here

slide-7
SLIDE 7

Typical GAN training … gets stuck

D tries to get here G tries to get here

slide-8
SLIDE 8

MH-GAN helps you reach the star

D tries to get here G tries to get here

slide-9
SLIDE 9

MH-GAN helps you reach the star

  • Wrap G and D to

build better G' Metropolis-Hastings Selector MH-GAN GAN D tries to get here G tries to get here

slide-10
SLIDE 10

MH recovers the true data distribution

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019.

slide-11
SLIDE 11

MH recovers the true data distribution

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019.

slide-12
SLIDE 12

MH recovers the true data distribution

Dropped modes!

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019.

slide-13
SLIDE 13

MH recovers the true data distribution

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019.

slide-14
SLIDE 14

MH recovers the true data distribution

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019.

slide-15
SLIDE 15

MH recovers the true data distribution

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019. Dropped modes!

slide-16
SLIDE 16

MH recovers the true data distribution

“Mixture of Gaussians” dataset [1]

[1] Azadi et al. ICLR 2019.

slide-17
SLIDE 17

Motivation for Metropolis-Hastings

  • Use MCMC independence sampler: sample pD from G
  • Given a perfect D and imperfect G, still obtain exact samples from true data distribution!
  • Avoid densities in MCMC, just need density ratios:

Metropolis-Hastings Selector MH-GAN GAN

slide-18
SLIDE 18

Metropolis-Hastings as a post-processing step for generators

Metropolis-Hastings Selector MH-GAN GAN

slide-19
SLIDE 19

MH recovers the correct score distribution

slide-20
SLIDE 20

MH recovers the correct score distribution

Discriminator gives different scores to fakes

slide-21
SLIDE 21

MH recovers the correct score distribution

Score distribution now matches real data Discriminator gives different scores to fakes

slide-22
SLIDE 22

Also… sample images

slide-23
SLIDE 23

Progressive GAN (base)

slide-24
SLIDE 24

Progressive GAN (base)

slide-25
SLIDE 25

Progressive GAN (base) PGAN + DRS (calibrated)

slide-26
SLIDE 26

Progressive GAN (base) PGAN + DRS (calibrated) PGAN + MH-GAN (calibrated)

slide-27
SLIDE 27

Metropolis-Hastings GAN Poster #201

ne Hung, Eric Frank, Yunus Saatci, Jason Yosinski Ryan Turner, Jane Hung, Eric Frank, Yunus Saatci, Jason Yosinski https://github.com/uber-research/metropolis-hastings-gans

slide-28
SLIDE 28

MH recovers the true data distribution

1) 1D mixture of 4 Gaussians, missing

  • ne mixture