Large Scale GAN Training for High Fidelity Natural Image Synthesis (2018-09-27T00:00:00.000000Z)

TL;DR

It is found that applying orthogonal regularization to the generator renders it amenable to a simple "truncation trick," allowing fine control over the trade-off between sample fidelity and variety by reducing the variance of the Generator's input.

Abstract

Despite recent progress in generative image modeling, successfully generating high-resolution, diverse samples from complex datasets such as ImageNet remains an elusive goal. To this end, we train Generative Adversarial Networks at the largest scale yet attempted, and study the instabilities specific to such scale. We find that applying orthogonal regularization to the generator renders it amenable to a simple "truncation trick," allowing fine control over the trade-off between sample fidelity and variety by reducing the variance of the Generator's input. Our modifications lead to models which set the new state of the art in class-conditional image synthesis. When trained on ImageNet at 128x128 resolution, our models (BigGANs) achieve an Inception Score (IS) of 166.5 and Frechet Inception Distance (FID) of 7.4, improving over the previous best IS of 52.52 and FID of 18.6.

Authors

Jeff Donahue

8 papers

K. Simonyan

25 papers

Andrew Brock

4 papers

TL;DR

Abstract

Authors

References60 items

The Unusual Effectiveness of Averaging in GAN Training

Self-Attention Generative Adversarial Networks

Comparing Generative Adversarial Network Techniques for Image Creation and Modification

Is Generator Conditioning Causally Related to GAN Performance?

Spectral Normalization for Generative Adversarial Networks

cGANs with Projection Discriminator

Improving GANs Using Optimal Transport

On Convergence and Stability of GANs

Which Training Methods for GANs do actually Converge?

A Note on the Inception Score

Demystifying MMD GANs

Non-local Neural Networks

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

Many Paths to Equilibrium: GANs Do Not Need to Decrease a Divergence At Every Step

FiLM: Visual Reasoning with a General Conditioning Layer

Wasserstein Generative Adversarial Networks

Revisiting Unreasonable Effectiveness of Data in Deep Learning Era

Modulating early visual processing by language

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

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

Megapixel Size Image Creation using Generative Adversarial Networks

The Cramer Distance as a Solution to Biased Wasserstein Gradients

Geometric GAN

Improved Training of Wasserstein GANs

Hierarchical Implicit Models and Likelihood-Free Variational Inference

Least Squares Generative Adversarial Networks

On the Quantitative Analysis of Decoder-Based Generative Models

Conditional Image Synthesis with Auxiliary Classifier GANs

A Learned Representation For Artistic Style

Deconvolution and Checkerboard Artifacts

Amortised MAP Inference for Image Super-resolution

Neural Photo Editing with Introspective Adversarial Networks

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Improved Techniques for Training GANs

f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization

Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks

Deep Residual Learning for Image Recognition

Rethinking the Inception Architecture for Computer Vision

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

A note on the evaluation of generative models

Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Adam: A Method for Stochastic Optimization

Conditional Generative Adversarial Nets

Very Deep Convolutional Networks for Large-Scale Image Recognition

ImageNet Large Scale Visual Recognition Challenge

Exact solutions to the nonlinear dynamics of learning in deep linear neural networks

Understanding the difficulty of training deep feedforward neural networks

Eigenvalue computation in the 20th century

GENERATIVE ADVERSARIAL NETS

2018), which employ a z ∈ R512 latent space on a highly constrained dataset with 30,000 images) is improper

We use a single shared class embedding in G, and skip connections

2016), and is chosen to allow some factors of variation to be discrete, while others are continuous. This latent outperforms N (0, I) by around 5

Dropout: a simple way to prevent neural networks from overfitting

Learning Multiple Layers of Features from Tiny Images

This Paper Is Included in the Proceedings of the 12th Usenix Symposium on Operating Systems Design and Implementation (osdi '16). Tensorflow: a System for Large-scale Machine Learning Tensorflow: a System for Large-scale Machine Learning

• We swept the strength of the modified Orthogonal Regularization penalty in G through

We find that the strength of the penalty correlates negatively with performance, but that settings above 0.5 impart training stability

• We swept the R1 gradient penalty strength through

We find that DropOut has a similar stabilizing effect to R1 but also degrades performance. regularization effect similar to DropOut, but not to significantly improve results

Field of Study

Journal Information

Name

Volume

Venue Information

Name

Type

URL

Alternate Names