README.md

Testing PSNR Loss with Vanilla GAN

Simulate experiments with a different loss functions on the Vanilla GAN architecture.

Setting up the Vanilla GAN Experiments package

• Setup the environment -

make setup

• Allow the .envrc to activate vitual environment -

direnv allow

• Install the requirements specified in requirements.txt -

make install

Testing PSNR Loss with Vanilla GAN against other losses (MSE, BCE)

For every experiment run:

• Clean the directory structure -

make clean

• Reset the directory structure -

make reset

• Define config parameters for package (each key in config.py - config dictionary must have value)

For each experiment run:

• Train the model -

make experiments

• Visualize model training - inferred and true data distributions, losses [Generator, Discriminator, Discriminator (Real Samples), Discriminator (Fake Sampless)] and generated samples (Uses Tensorboard) -

make visualize

Example For Using API

• Training from scratch -

def train_from_scratch_example() -> None:
    '''
    Train a model from scratch.
    Returns:
        None
    '''
    # Create directory structure for the experiment
    create_directory_structure = DirectoryStructure(home_dir=config['device']['home directory'])
    create_directory_structure.create_directory_structure()

    # Create the experiments
    experiments = Experiments(config=config)

    # Train the model
    experiments.train(verbose=True, checkpoint=None)

• Training from checkpoint -

def train_from_checkpoint_example() -> None:
    '''
    Train a model from a checkpoint.
    Returns:
        None
    '''
    # Create the experiments
    experiments = Experiments(config=config)

    checkpoint = {
        'generator': './weights/generator_epoch_0_loss_0.pt',
        'discriminator': './weights/discriminator_epoch_0_loss_0.pt'
    }

    # Train the model
    experiments.train(verbose=True, checkpoint=checkpoint)

References:

The Generate Adversarial Network (GAN) training algorithm can be found here -

@misc{https://doi.org/10.48550/arxiv.1406.2661,
  doi = {10.48550/ARXIV.1406.2661},
  url = {https://arxiv.org/abs/1406.2661},
  author = {Goodfellow, Ian J. and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua},
  keywords = {Machine Learning (stat.ML), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Generative Adversarial Networks},
  publisher = {arXiv},
  year = {2014},
  copyright = {arXiv.org perpetual, non-exclusive license}
}