This repository contains an implementation of a Deep Convolutional Generative Adversarial Network (DCGAN) for MNIST. It implements the suggested architectural constraints for stable learning from the paper Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks by Alec Radford, Luke Metz, Soumith Chintala.
The generator and discriminator are all convolutional networks without max pooling but instead uses strided (and fractionally strided) convolutions. Fruthermore, batch normalization is used and ReLu and Leaky ReLU activations.
Usage: dcgan-mnist.py [OPTIONS]
Options:
--root TEXT Root directory for MNIST dataset
--epochs INTEGER Number of epochs
--batch-size INTEGER Batch size
--latent-vector INTEGER Size of latent vector Z
--disable-cuda TEXT Disable CUDA acceleration
--help Show this message and exit.
dcgan-mnist is Copyright © 2019 Alexander Stante. It is free software, and may be redistributed under the terms specified in the LICENSE file.