GANs 2.0: Generative Adversarial Networks in TensorFlow 2.0

Project aim

The main aim of this project is to speed up a process of building deep learning pipelines that are based on Generative Adversarial Networks and simplify prototyping of various generator/discriminator models. This library provides several GAN trainers that can be used as off-the-shelf features such us:

• Vanilla GAN
• Conditional GAN
• Cycle GAN
• Wasserstein GAN
• Progressive GAN (WIP)

Examples

Function modeling

Vanilla GAN (Gaussian function)	Vanilla GAN (sigmoid function)

Image generation

Vanilla GAN (MNIST)	Conditional GAN (MNIST)

Vanilla GAN (FASHION_MNIST)	Conditional GAN (FASHION_MNIST)

Vanilla GAN (CIFAR10)	Conditional GAN (CIFAR10)

Image translation

Cycle GAN (SUMMER2WINTER)	Cycle GAN (WINTER2SUMMER)

Installation

Installs with GPU support

pip install gans2[tensorflow_gpu]

Installs with CPU support

pip install gans2[tensorflow]

Running training pipeline code examples for Vanilla GAN for MNIST digit generation

Pre-defined models

import tensorflow as tf
from easydict import EasyDict as edict

from gans.datasets import mnist
from gans.models.discriminators import discriminator
from gans.models.generators.latent_to_image import latent_to_image
from gans.trainers import optimizers
from gans.trainers import vanilla_gan_trainer

model_parameters = edict({
    'img_height':                  28,
    'img_width':                   28,
    'num_channels':                1,
    'batch_size':                  16,
    'num_epochs':                  10,
    'buffer_size':                 1000,
    'latent_size':                 100,
    'learning_rate_generator':     0.0001,
    'learning_rate_discriminator': 0.0001,
    'save_images_every_n_steps':   10
})

dataset = mnist.MnistDataset(model_parameters)

generator = latent_to_image.LatentToImageGenerator(model_parameters)
discriminator = discriminator.Discriminator(model_parameters)

generator_optimizer = optimizers.Adam(
    learning_rate=model_parameters.learning_rate_generator,
    beta_1=0.5,
)
discriminator_optimizer = optimizers.Adam(
    learning_rate=model_parameters.learning_rate_discriminator,
    beta_1=0.5,
)

gan_trainer = vanilla_gan_trainer.VanillaGANTrainer(
    batch_size=model_parameters.batch_size,
    generator=generator,
    discriminator=discriminator,
    training_name='VANILLA_GAN_MNIST',
    generator_optimizer=generator_optimizer,
    discriminator_optimizer=discriminator_optimizer,
    latent_size=model_parameters.latent_size,
    continue_training=False,
    save_images_every_n_steps=model_parameters.save_images_every_n_steps,
    visualization_type='image',
)

gan_trainer.train(
    dataset=dataset,
    num_epochs=model_parameters.num_epochs,
)

Custom models

import tensorflow as tf
from easydict import EasyDict as edict
from tensorflow.python import keras
from tensorflow.python.keras import layers

from gans.datasets import mnist
from gans.models import sequential
from gans.trainers import optimizers
from gans.trainers import vanilla_gan_trainer

model_parameters = edict({
    'img_height':                  28,
    'img_width':                   28,
    'num_channels':                1,
    'batch_size':                  16,
    'num_epochs':                  10,
    'buffer_size':                 1000,
    'latent_size':                 100,
    'learning_rate_generator':     0.0001,
    'learning_rate_discriminator': 0.0001,
    'save_images_every_n_steps':   10
})

dataset = mnist.MnistDataset(model_parameters)

generator = sequential.SequentialModel(
    layers=[
        keras.Input(shape=[model_parameters.latent_size]),
        layers.Dense(units=7 * 7 * 256, use_bias=False),
        layers.BatchNormalization(),
        layers.LeakyReLU(),

        layers.Reshape((7, 7, 256)),
        layers.Conv2DTranspose(128, (5, 5), strides=(1, 1), padding='same', use_bias=False),
        layers.BatchNormalization(),
        layers.LeakyReLU(),

        layers.Conv2DTranspose(64, (5, 5), strides=(2, 2), padding='same', use_bias=False),
        layers.BatchNormalization(),
        layers.LeakyReLU(),

        layers.Conv2DTranspose(1, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh')
    ]
)

discriminator = sequential.SequentialModel(
    [
        keras.Input(
            shape=[
                model_parameters.img_height,
                model_parameters.img_width,
                model_parameters.num_channels,
            ]),
        layers.Conv2D(filters=64, kernel_size=(5, 5), strides=(2, 2), padding='same'),
        layers.LeakyReLU(),
        layers.Dropout(0.3),

        layers.Conv2D(filters=128, kernel_size=(5, 5), strides=(2, 2), padding='same'),
        layers.LeakyReLU(),
        layers.Dropout(rate=0.3),

        layers.Flatten(),
        layers.Dense(units=1),
    ]
)

generator_optimizer = optimizers.Adam(
    learning_rate=model_parameters.learning_rate_generator,
    beta_1=0.5,
)
discriminator_optimizer = optimizers.Adam(
    learning_rate=model_parameters.learning_rate_discriminator,
    beta_1=0.5,
)

gan_trainer = vanilla_gan_trainer.VanillaGANTrainer(
    batch_size=model_parameters.batch_size,
    generator=generator,
    discriminator=discriminator,
    training_name='VANILLA_GAN_MNIST_CUSTOM_MODELS',
    generator_optimizer=generator_optimizer,
    discriminator_optimizer=discriminator_optimizer,
    latent_size=model_parameters.latent_size,
    continue_training=False,
    save_images_every_n_steps=model_parameters.save_images_every_n_steps,
    visualization_type='image',
)

gan_trainer.train(
    dataset=dataset,
    num_epochs=model_parameters.num_epochs,
)

More code examples

Vanilla GAN for Gaussian function modeling

Vanilla GAN for sigmoid function modeling

Conditional GAN for MNIST digit generation

Cycle GAN for summer2winter/winter2summer style transfer

Wasserstein GAN for MNIST digit generatio

Monitoring model training

In order to visualize a training process (loss values, generated outputs) run the following command in the project directory:

tensorboard --logdir outputs

To follow the training process go to the default browser and type the following address http://your-workstation-name:6006/

The below picture presents the TensorBoard view lunched for two experiments: VANILLA_MNIST and VANILLA_FASION_MNIST.

References

1. Deep Convolutional Generative Adversarial Network Tutorial in TensorFlow
2. Cycle GAN Tutorial in TensorFlow
3. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks paper