Synthetic Data Generation

User friendly library to generate datasets using GANs and VAEs.

Paper supporting this library: Synthetic Data Generation To Better Perceive Classifier Predictions

Installation

Use git to install generate_datasets.

git clone https://github.com/Mr-Vicente/generate_datasets.git

Usage

GAN

from generate_datasets import data_access
from generate_datasets.WGAN_GP.wgan_gp_class_generic import WGAN_GP

EPOCHS = 500
OUTPUT_DIR = "imgs_gen"
params = {
    'lr': 0.0001,
    'beta1': 0,
    'batch_size': 64,
    'latent_dim': 128,
    'image_size': 152
}

class_names = ['Blond-yellow','Yellow','Orange','Orange-Brown','Blond',
                'Light-Brown','Brown','Black','Gray','White']


Number_Dataset_Classes = len(class_names)

gan = WGAN_GP(params)
gan.load_dataset(data_access.prepare_data('gan'),Number_Dataset_Classes)
gan.train_model(EPOCHS)
gan.generate_images(10,OUTPUT_DIR,class_names)

If no classifier is provided to GANs, the gan model will behave like a normal gan. The interest of one providing a classifier to our GAN model might be to better understand how it is operating, or even to generate images of a specific class.

VAE

from generate_datasets import data_access
from generate_datasets.VAE.generic_vannila_vae import VAE
from generate_datasets.Processing import process_cartoon

EPOCHS = 500
n_images = 10
OUTPUT_DIR = "imgs_gen"

params = {
    'lr': 0.0001,
    'beta1': 0,
    'batch_size': 64,
    'latent_dim': 128,
    'image_size': 152
}

class_names = ['Blond-yellow','Yellow','Orange','Orange-Brown','Blond',
                'Light-Brown','Brown','Black','Gray','White']

def load_cartoon_data():
    images, labels = process_cartoon.decode_data_cartoon()
    return images, labels

tf.keras.backend.clear_session()

model = VAE()
model.load_dataset(data_access.prepare_dataset('vae',load_cartoon_data(),image_size=(128,128)))
model.train_model(epochs)
model.generate_images(n_images,OUTPUT_DIR)

Library Info

This library has several generative models at your despose:

• GAN (Generative Adversarial Network)
  • Status: Working
  • Paper: GAN
  • Official Implementation: None
• VAE (Variational Autoencoder)
  • Status: Working
  • Paper: VAE
  • Official Implementation: None
• WGAN (Wasserstein GAN)
  • Status: Working
  • Paper: WGAN
  • Official Implementation: None
• WGAN-GP (Wasserstein GAN with Gradient Penalty)
  • Status: Working
  • Paper: WGAN-GP
  • Official Implementation: None, but check this one WGAN-GP Implementation
• PGGAN (Progressive Growing GAN)
  • Status: Not working
  • Paper: PGGAN
  • Official Implementation: PGGAN Implementation
• IntroVAE (Introspective Variational Autoencoder)
  • Status: Not giving proper results
  • Paper: IntroVAE
  • Official Implementation: None

To learn more about these Generative models visit the referenced papers/implementations.

Generated Images (examples)

Epochs	Datasets	Original Images	WGAN-GP	VAE	IntroVAE
~20	Fashion MNIST				-
~200	Cartoon Dataset				-
~200	Train Dataset			-
~200	LSUN Bedroom Dataset			-	-

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

Project status

On going!

Authors and acknowledgment

Frederico Vicente & Ludwig Krippahl

License

MIT