Training Generative Adversarial Networks with Limited Data

Tero Karras; Miika Aittala; Janne Hellsten; Samuli Laine; Jaakko Lehtinen; Timo Aila

Training Generative Adversarial Networks with Limited Data

Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, Timo Aila

NVIDIA Corporation

Research output: Chapter in Book/Report/Conference proceeding › Conference article in proceedings › Scientific › peer-review

Abstract

Training generative adversarial networks (GAN) using too little data typically leads to discriminator overfitting, causing training to diverge. We propose an adaptive discriminator augmentation mechanism that significantly stabilizes training in limited data regimes. The approach does not require changes to loss functions or network architectures, and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. We demonstrate, on several datasets, that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. We expect this to open up new application domains for GANs. We also find that the widely used CIFAR-10 is, in fact, a limited data benchmark, and improve the record FID from 5.59 to 2.42.

Original language	English
Title of host publication	Thirty-fourth Conference on Neural Information Processing Systems
Publisher	Morgan Kaufmann Publishers
Number of pages	11
Publication status	Published - 2020
MoE publication type	A4 Conference publication
Event	Conference on Neural Information Processing Systems - Virtual, Vancouver, Canada Duration: 6 Dec 2020 → 12 Dec 2020 Conference number: 34

Publication series

Name	Advances in neural information processing systems
Publisher	Morgan Kaufmann Publishers
Volume	33
ISSN (Print)	1049-5258

Conference

Conference	Conference on Neural Information Processing Systems
Abbreviated title	NeurIPS
Country/Territory	Canada
City	Vancouver
Period	06/12/2020 → 12/12/2020

Access to Document

https://papers.nips.cc/paper/2020/hash/8d30aa96e72440759f74bd2306c1fa3d-Abstract.html

Cite this

Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., & Aila, T. (2020). Training Generative Adversarial Networks with Limited Data. In Thirty-fourth Conference on Neural Information Processing Systems (Advances in neural information processing systems; Vol. 33). Morgan Kaufmann Publishers. https://papers.nips.cc/paper/2020/hash/8d30aa96e72440759f74bd2306c1fa3d-Abstract.html

@inproceedings{2841c569649042fba8259838f3907c6d,

title = "Training Generative Adversarial Networks with Limited Data",

abstract = "Training generative adversarial networks (GAN) using too little data typically leads to discriminator overfitting, causing training to diverge. We propose an adaptive discriminator augmentation mechanism that significantly stabilizes training in limited data regimes. The approach does not require changes to loss functions or network architectures, and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. We demonstrate, on several datasets, that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. We expect this to open up new application domains for GANs. We also find that the widely used CIFAR-10 is, in fact, a limited data benchmark, and improve the record FID from 5.59 to 2.42.",

author = "Tero Karras and Miika Aittala and Janne Hellsten and Samuli Laine and Jaakko Lehtinen and Timo Aila",

year = "2020",

language = "English",

series = "Advances in neural information processing systems",

publisher = "Morgan Kaufmann Publishers",

booktitle = "Thirty-fourth Conference on Neural Information Processing Systems",

address = "United States",

note = "Conference on Neural Information Processing Systems, NeurIPS ; Conference date: 06-12-2020 Through 12-12-2020",

}

Karras, T, Aittala, M, Hellsten, J, Laine, S, Lehtinen, J & Aila, T 2020, Training Generative Adversarial Networks with Limited Data. in Thirty-fourth Conference on Neural Information Processing Systems. Advances in neural information processing systems, vol. 33, Morgan Kaufmann Publishers, Conference on Neural Information Processing Systems, Vancouver, Canada, 06/12/2020. <https://papers.nips.cc/paper/2020/hash/8d30aa96e72440759f74bd2306c1fa3d-Abstract.html>

Training Generative Adversarial Networks with Limited Data. / Karras, Tero; Aittala, Miika; Hellsten, Janne et al.
Thirty-fourth Conference on Neural Information Processing Systems. Morgan Kaufmann Publishers, 2020. (Advances in neural information processing systems; Vol. 33).

Research output: Chapter in Book/Report/Conference proceeding › Conference article in proceedings › Scientific › peer-review

TY - GEN

T1 - Training Generative Adversarial Networks with Limited Data

AU - Karras, Tero

AU - Aittala, Miika

AU - Hellsten, Janne

AU - Laine, Samuli

AU - Lehtinen, Jaakko

AU - Aila, Timo

N1 - Conference code: 34

PY - 2020

Y1 - 2020

N2 - Training generative adversarial networks (GAN) using too little data typically leads to discriminator overfitting, causing training to diverge. We propose an adaptive discriminator augmentation mechanism that significantly stabilizes training in limited data regimes. The approach does not require changes to loss functions or network architectures, and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. We demonstrate, on several datasets, that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. We expect this to open up new application domains for GANs. We also find that the widely used CIFAR-10 is, in fact, a limited data benchmark, and improve the record FID from 5.59 to 2.42.

AB - Training generative adversarial networks (GAN) using too little data typically leads to discriminator overfitting, causing training to diverge. We propose an adaptive discriminator augmentation mechanism that significantly stabilizes training in limited data regimes. The approach does not require changes to loss functions or network architectures, and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. We demonstrate, on several datasets, that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. We expect this to open up new application domains for GANs. We also find that the widely used CIFAR-10 is, in fact, a limited data benchmark, and improve the record FID from 5.59 to 2.42.

UR - https://github.com/NVlabs/stylegan2-ada

M3 - Conference article in proceedings

T3 - Advances in neural information processing systems

BT - Thirty-fourth Conference on Neural Information Processing Systems

PB - Morgan Kaufmann Publishers

T2 - Conference on Neural Information Processing Systems

Y2 - 6 December 2020 through 12 December 2020

ER -

Training Generative Adversarial Networks with Limited Data

Abstract

Publication series

Conference

Access to Document

Other files and links

Fingerprint

Cite this