StyleGAN that produces anime faces. This is an implementation I made as a learning project, there may be some mistakes as this is my first GAN!

The dataset has been taken from kaggle. Thanks to splcher for providing this funny dataset.

Link to the WandB project can be found here.

After cloning the repo, you will have do download the dataset. It can be found here.

Then you can run the initialization script:

cd path/to/repo/AnimeStyleGAN
python3 prepare.py [path_to_dataset.zip]

This script will extract the images and put them in the right directory.

To train the model: python3 launch_training.py

The configuration is stored in the scr/train.py. The best configuration found is the following:

config = {
    # Global params
    'dim_image': 64,
    'batch_size': 256,
    'epochs': 50,
    'dropout': 0.3,
    'device': 'cuda' if torch.cuda.is_available() else 'cpu',
    'seed': 0,
    'n_iter_log': 10,

    # StyleGAN params
    'n_channels': 512,
    'dim_z': 196,
    'n_layers_z': 4,
    'n_layers_block': 3,
    'n_noise': 10,
    'lr_g': 1e-4,
    'betas_g': (0.5, 0.5),
    'milestones_g': [15],
    'gamma_g': 0.1,
    'running_avg_factor_G': 0.9,
    'weight_avg_factor_g': 0.5,

    # Discriminator params
    'n_first_channels': 12,
    'n_layers_d_block': 5,
    'lr_d': 1e-4,
    'betas_d': (0.5, 0.99),
    'milestones_d': [15],
    'gamma_d': 0.1,
    'running_avg_factor_D': 0.9,
    'weight_avg_factor_d': 0.5,
}

It is a pretty big model though (15M parameters for both the generator and the discriminator).

To make the training easier, there are many tricks:

• Style mixing (from the StyleGAN paper)
• Label smoothing
• Historical averaging
• Spectral norm
• Two-timescale update rule
• Noisy input to the discriminator

Many of those tricks can be found in this paper: Improved Techniques for Training GANs, Salimans et al.

I tried the GP-WGAN but it did not perform very well. Another improvement would be to do experience replay for the discriminator.

• The difficulty of finding the right hyperparameters for GAN training
• BatchNorm are dangerous as we should avoid sharing information among samples
• Take care of the scales when using noises
• Some tutos may be wrong, reading the original paper is better (Improved Techniques for Training GANs looks usually badly explained)
• The usefulness of Paper With Code
• The importance of betas parameters for Adam optimizer
• The importance of all the tiny tricks that can greatly help convergence
• The tiny ommitted implementation details that are crucial
• Do alternate the training of D and G (between minibatchs)!!

• StyleGAN paper: https://arxiv.org/abs/1812.04948v3
• Improved techniques paper: https://arxiv.org/abs/1606.03498v1
• Dataset: https://www.kaggle.com/splcher/animefacedataset