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.
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-dirlike this:dataset-dir/data/abc.png - Start training:
python train.py dataset-dir 100 1024 512 256 128 64 - Duh!
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 onnetwork-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
- Vanilla DCGAN has following parameters:
batch_size: Size of single batchepochs: Number of epochs for complete iteration of on datasetmodel_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/outputloss_buffer: Number of values in discriminator/generator loss displayed in output windown_outputs: Number of images per row displayed in output windowdump_dir: Path where to save output of model (skip not to save output)
- Sample Jupyter Notebook
- Settings for Training on Colab
- Implement Progressive GAN