Skip to content

A new training strategy of GANs where the latent space dimension is grown progressively during training

Notifications You must be signed in to change notification settings

eriksandstroem/Latent-Space-Growing-of-GANs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Latent Space Growing of Generative Adversarial Networks

This project was conducted as a master thesis by Erik Sandström during the spring of 2019 at Lund University.

This README-file contains extra material for Section 7.2.1 and 7.2.2 not covered in the thesis report.

Below, the random walk interpolations are presented for each model when restricting the random walks for dimensions {1-8,9-16,17-32,33-64,65-128,129-256}, one at a time.

clgGAN

Restricted to dimensions: 1-8

Restricted to dimensions: 9-16

Restricted to dimensions: 17-32

Restricted to dimensions: 33-64

Restricted to dimensions: 65-128

Restricted to dimensions: 129-256

cgGAN

Restricted to dimensions: 1-8

Restricted to dimensions: 9-16

Restricted to dimensions: 17-32

Restricted to dimensions: 33-64

Restricted to dimensions: 65-128

Restricted to dimensions: 129-256

bgGAN

Restricted to dimensions: 1-8

Restricted to dimensions: 9-16

Restricted to dimensions: 17-32

Restricted to dimensions: 33-64

Restricted to dimensions: 65-128

Restricted to dimensions: 129-256

lgGAN

Restricted to dimensions: 1-8

Restricted to dimensions: 9-16

Restricted to dimensions: 17-32

Restricted to dimensions: 33-64

Restricted to dimensions: 65-128

Restricted to dimensions: 129-256

bGAN

Restricted to dimensions: 1-8

Restricted to dimensions: 9-16

Restricted to dimensions: 17-32

Restricted to dimensions: 33-64

Restricted to dimensions: 65-128

Restricted to dimensions: 129-256

Below, the random walk interpolations along the coordinate axis for the clgGAN are presented when restricting random walks along the coordinate axis dimensions {1,2,5,90,195,250}, one at a time.

Coordinate walk along dimension: 1

Coordinate walk along dimension: 2

Coordinate walk along dimension: 5

Coordinate walk along dimension: 90

Coordinate walk along dimension: 195

Coordinate walk along dimension: 250

Setup

Place a directory called "celebA_dataset" where the model directory, e.g., "clgGAN" is located. In the "celebA_dataset" directory, place another folder called "celebA" containing the images belonging to the dataset. Use the cropped and aligned version of CelebA

About

A new training strategy of GANs where the latent space dimension is grown progressively during training

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages