Information-theoretic stochastic contrastive conditional GAN: InfoSCC-GAN

This repos contains official Pytorch implementation of the paper: Information-theoretic stochastic contrastive conditional GAN: InfoSCC-GAN, accepted at NeurIPS 2021 Workshop Bayesian Deep Learning

Paper Poster

Citation

@inproceedings { kinakh2021informationtheoretic,
    title = { Information-theoretic stochastic contrastive conditional GAN: InfoSCC-GAN },
    author = { Kinakh, Vitaliy and Drozdova, Mariia and Quétant, Guillaume and Golling, Tobias and Voloshynovskiy, Slava },
    booktitle = { Bayesian Deep Learning NeurIPS workshop },
    year = { 2021 }
}

Demos

CelebA demo with 10 attributes | CelebA demo with 15 attributes | AFHQ demo

Contents

1. Animal Faces High Quality experiments
   1.1 Animal Faces Demo
   1.2 EigenGAN experiment
   1.3 InfoSCC-GAN experiments
   
   • Oneclass global discriminator, Hinge loss, each 2-nd iter classification regularization
     
   • Oneclass global discriminator, Non saturating loss, each 2-nd iter classification regularization
   • Multiclass global discriminator, Hinge loss, each 2-nd iteration classification regularization
     
   • Oneclass patch discriminator, Hinge loss, each 2-nd iteration classification regularization
     
   • Oneclass patch discriminator, Non-saturating loss, each 2-nd iteration classification regularization
     
   • Oneclass path discriminator, LSGAN loss, each 2-nd iteration classification regularization
2. CelebA experiments
   2.1 CelebA Demo with 10 attributes
   2.2 CelebA Demo with 15 attributes
   
3. Run experiment

Animal Faces High Quality experiments

AFHQ demo

EigenGAN exploration

InfoSCC-GAN

Oneclass global discriminator, Hinge loss, each 2-nd iter classification regularization

latent variables exploration

Explore

Each row: fix input label and latent variables , randomly change

Explore and

Each row: fix input label , randomly change latent variables and

Explore

Fix and , explore all

Oneclass global discriminator, Non saturating loss, each 2-nd iter classification regularization

latent variables exploration

Explore

Each row: fix input label and latent variables , randomly change

Explore and

Each row: fix input label , randomly change latent variables and

Explore

Fix and , explore all

Multiclass global discriminator, Hinge loss, each 2-nd iter classification regularization

latent variables exploration

Explore

Each row: fix input label and latent variables , randomly change

Explore and

Each row: fix input label , randomly change latent variables and

Explore

Fix and , explore all

One class patch discriminator, Hinge loss, each 2-nd iter classification regularization

latent variables exploration

Explore

Each row: fix input label and latent variables , randomly change

Explore and

Each row: fix input label , randomly change latent variables and

Explore

Fix and , explore all

One class patch discriminator, Non-saturating loss, each 2-nd iter classification regularization

latent variables exploration

Explore

Each row: fix input label and latent variables , randomly change

Explore and

Each row: fix input label , randomly change latent variables and

Explore

Fix and , explore all

One class patch discriminator, LSGAN loss, each 2-nd iter classification regularization

latent variables exploration

Explore

Each row: fix input label and latent variables , randomly change

Explore and

Each row: fix input label , randomly change latent variables and

Explore

Fix and , explore all

CelebA experiments

CelebA demo with 10 attributes

CelebA demo with 15 attributes