gmalivenko / cat-gan

GAN for cat images generation

Cat GAN

This is the GAN implementation of cats generator.

What is the GAN? Generative adversarial networks (GANs) are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks contesting with each other in a zero-sum game framework (link). So, it means that we should train 2 networks with different architectures and in different ways.

Architecture

Generator

The generator has very custom but straightforward architecure. ConvTranspose2d -> SELU -> [ConvTranspose2d -> SELU] x N -> ConvTranspose2d -> Tanh

Discriminator

A ResNet-18 with single-neuron output seems compatible for this project.

The generator has very custom but straightforward architecure. ConvTranspose2d -> SELU -> [ConvTranspose2d -> SELU] x N -> ConvTranspose2d -> Tanh

Examples

The output on 5k iteration (for default configuration): 5 minute after: 2 hours after:

Installation

pip install -r requirements.txt

For the PyTorch installation, please follow this guide.

Training

To run the training script, make changes in the configuration (example is in config/train.yml).

python train.py --config config/train.yml

Testing

To get radom cat (fill restore.generator in config/train.yml):

python test.py --config config/train.yml

References

Inspired by https://arxiv.org/abs/1406.2661

License

This software is covered by MIT License.