Training Generative Adversarial Networks with Limited Data (2020-06-01T00:00:00.000000Z)

TL;DR

It is demonstrated, 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, and is expected to open up new application domains for GANs.

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.

Authors

S. Laine

8 papers

Timo Aila

8 papers

Tero Karras

5 papers

TL;DR

Abstract

Authors

References56 items

Differentiable Augmentation for Data-Efficient GAN Training

On Data Augmentation for GAN Training

Image Augmentations for GAN Training

Feature Quantization Improves GAN Training

A U-Net Based Discriminator for Generative Adversarial Networks

Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs

Improved Consistency Regularization for GANs

Unsupervised multi-modal Styled Content Generation

MineGAN: Effective Knowledge Transfer From GANs to Target Domains With Few Images

cGANs with Multi-Hinge Loss

StarGAN v2: Diverse Image Synthesis for Multiple Domains

Analyzing and Improving the Image Quality of StyleGAN

Consistency Regularization for Generative Adversarial Networks

Randaugment: Practical automated data augmentation with a reduced search space

AutoGAN: Neural Architecture Search for Generative Adversarial Networks

A survey on Image Data Augmentation for Deep Learning

E-LPIPS: Robust Perceptual Image Similarity via Random Transformation Ensembles

Improved Precision and Recall Metric for Assessing Generative Models

Image Generation From Small Datasets via Batch Statistics Adaptation

BreCaHAD: a dataset for breast cancer histopathological annotation and diagnosis

A Style-Based Generator Architecture for Generative Adversarial Networks

Bag of Tricks for Image Classification with Convolutional Neural Networks

Self-Supervised GANs via Auxiliary Rotation Loss

Large Scale GAN Training for High Fidelity Natural Image Synthesis

PA-GAN: Improving GAN Training by Progressive Augmentation

Generative Adversarial Network in Medical Imaging: A Review

Assessing Generative Models via Precision and Recall

AutoAugment: Learning Augmentation Policies from Data

Self-Attention Generative Adversarial Networks

Transferring GANs: generating images from limited data

Spectral Normalization for Generative Adversarial Networks

AmbientGAN: Generative models from lossy measurements

cGANs with Projection Discriminator

Which Training Methods for GANs do actually Converge?

Demystifying MMD GANs

Progressive Growing of GANs for Improved Quality, Stability, and Variation

Improved Regularization of Convolutional Neural Networks with Cutout

GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

DeLiGAN: Generative Adversarial Networks for Diverse and Limited Data

Megapixel Size Image Creation using Generative Adversarial Networks

Improved Training of Wasserstein GANs

Towards Principled Methods for Training Generative Adversarial Networks

Least Squares Generative Adversarial Networks

Amortised MAP Inference for Image Super-resolution

Temporal Ensembling for Semi-Supervised Learning

Regularization With Stochastic Transformations and Perturbations for Deep Semi-Supervised Learning

Improved Techniques for Training GANs

Adam: A Method for Stochastic Optimization

Ten Lectures on Wavelets

The distribution of second order moment statistics in a normal system

Power measurement tutorial for the Green500 list

Dropout: a simple way to prevent neural networks from overfitting

I and J

Learning Multiple Layers of Features from Tiny Images

Ten Lectures on Wavelets Volume 921

I. Daubechies

Field of Study

Journal Information

Name

Volume

Venue Information

Name

Type

URL

Alternate Names