xGAN

xGAN is highly customizable zero-coding GAN implementation for rapid prototyping. You can use it quickly train your GAN on your dataset.

NOTE: This repo implements DCGAN and does not include recent breakthroughs in the field e.g. Progressive GAN. But I have plan to implement them very soon.

Getting Started

Follow these steps to train your own GAN:

• Clone the repository git clone https://github.com/NaxAlpha/xgan.git
• Install PyTorch and other requirements pip install -r requirements.txt
• Prepare your dataset in dataset-dir like this: dataset-dir/data/abc.png
• Start training: python train.py dataset-dir 100 1024 512 256 128 64
• Duh!

Options

train.py has very diverse set of options available for you to customize. Example usage is:

python train.py <dataset-dir> <network-layers...> [--batch_size=64] [--epochs=100] [--model_dir=None] [--log_iter=10] [--loss_buffer=500] [--n_outputs=3] [--dump_dir=None]

Following is parameter documentation:

• dataset-dir: Path to images you want to train your GAN on
• network-layers: Network architecture from latent space to filters on each layer:
  • Vanilla DCGAN has following parameters: 100 1024 512 256 128 64
  • First layer value is latent space
  • Size of image is determined number of layers e.g. in case of vanilla we have 6 layers: 2^6 => 64
  • An other example architecture would be for image of size 128x128: 100 1024 512 256 128 64 32
• batch_size: Size of single batch
• epochs: Number of epochs for complete iteration of on dataset
• model_dir: Path to directory where to save model (skip if you do not want to save model)
• log_iter: Number of iterations after which to to save model/output
• loss_buffer: Number of values in discriminator/generator loss displayed in output window
• n_outputs: Number of images per row displayed in output window
• dump_dir: Path where to save output of model (skip not to save output)

TODO:

• Sample Jupyter Notebook
• Settings for Training on Colab
• Implement Progressive GAN

Blog Post:

This Icon Does Not Exist — An Application of GANs to Icon Generation