Generative Adversarial Networks for MNIST Generation

Overview

Implemented a vanilla GAN to generate MNIST images.

Generator: four fully connected layers with ReLU activation functions and a Tanh output Layer, mapping a 100 dimensional noise vector to a 28x28 image.

Discriminator: four fully connected layers with leakyReLU activation and a sigmoid output layer to classify if generated output was fake or real.

Trained 25 epochs resulting in Discriminator loss to ~0.777, and Generator loss to ~1.62.

Results:

Next Steps:

Building a cGAN (conditional GAN) to take an integer as input and ouput an artifically generated MNIST image.