TensorFlow VAEs and GANs

TensorFlow implementation of various deep generative networks such as VAE and GAN.

Models

Standard models

• Variational autoencoder (VAE) [Kingma et al. 2013]
• Generative adversarial network (GAN or DCGAN) [Goodfellow et al. 2014]

Conditional models

• Conditional variational autoencoder [Kingma et al. 2014]
• CVAE-GAN [Bao et al. 2017]

Usage

Prepare datasets

MNIST and SVHN

MNIST and SVHN datasets are automatically downloaded from their websites.

CelebA

First, download img_align_celeba.zip and list_attr_celeba.txt from CelebA webpage. Then, place these files to datasets and run create_database.py on databsets directory.

Training

# Both standard and conditional models are available!
python train.py --model=dcgan --epoch=200 --batchsize=100 --output=output

TensorBoard is also available with the following script.

tensorboard --logdir="output/dcgan/log"

Results

DCGAN (for SVHN 50 epochs)

CVAE-GAN (for SVHN 50 epochs)

References

• Kingma et al., "Auto-Encoding Variational Bayes", arXiv preprint 2013.
• Goodfellow et al., "Generative adversarial nets", NIPS 2014.

• Kingma et al., "Semi-supervised learning with deep generative models", NIPS 2014.
• Bao et al., "CVAE-GAN: Fine-Grained Image Generation through Asymmetric Training", arXiv preprint 2017.