Generative Adversarial Network Implementation - MNIST

1.0 About

This project is a simple implementation of the paper Generative Adversarial Net. As stated in the paper, Generative Adversarial Network (GAN) consists of 2 models. The Generator model and the Discriminator model.

Using the MNIST Dataset, the GAN is aimed to generate acurate image representation of digits from 0 - 9. The following screenshot is an example generated result on a model that was trained with epoch=50.

2.0 Getting Started

2.1 Installation

Install the required packages

pip3 install -r requirements.txt

2.2 Training

Training the model is pretty straight forward. Once you have created an instance of MNISTGAN, you can immediately use its train method to start training.

mnist_gan = MNISTGAN()
mnist_gan.train(epochs=100, batch_size=64, lr=0.001, is_save_images=True)

2.3 Testing

Similar to testing, you can use the test method to generate images using your trained model.

mnist_gan.test(model_path='mymodel.pt')