Simple-GAN-MNIST

Pytorch implementation of a simple Generative Adversarial Network on the MNIST dataset.

Requirements

• Python 3.8.5
• pytorch 1.6.0
• torchvision 0.7.0
• tensorboard 2.3.0
• tqdm 4.50.2

Make sure to download the MNIST dataset.

1) Hyperparameters

Specify hyperparameters inside a .json file, e.g.:

{
    "dim_latent": 100,
    "batch_size": 16,
    "learning_rate": 0.0002,
    "beta_1": 0.5,
    "beta_2": 0.999,
    "num_epochs": 5,
    "num_fake_samples": 30
}

2) Training

Train the model:

python3 train_gan.py --datadir <path_to_mnist>

Alternatively you may want to specify

• path to your custom set of hyperparameters --params
• log directory --logdir
• model directory --modeldir
• model output name -o
• training device --device [specify None for cpu]
• number of fake sample outputs num_fake_samples_eval

3) Inspecting Results

Learning curves for the discriminator and generator can be inspected using tensorboard. Example fake numbers from training with the hyperparameters specifed above.