This repository contains PyTorch implementations of examples from the book GANs in Action by Jakub Langr and Vladimir Bok.
The goal of the examples is not to provide one-to-one translations of
the Keras source code, but to implement the examples in an idiomatic
way in PyTorch. Examples for that is the usage of Dataset and
DataLoader.
The provided Jupyter notebooks require Python 3.6 and were tested with PyTorch 1.3. The easiest way to use these examples is to set up a conda environment containing the follwoing packages.
$ conda install pytorch torchvision jupyter tqdm matplotlibThe package tqdm is used to have a nice progress bar during training
in the notebook. For plotting results and drawing images, the popular
matplotlib library is used.
Deep Convolutional GAN is a variant presented in the paper Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks by Alec Radford, Luke Metz, Soumith Chintala. This architecture introduced convolutional layers to GANs as well as some best practices like batch normalization.
The example in PyTorch is provided here and utilizes the MNIST dataset to train the generator. A following image shows some results.
Conditional GAN is a variant presented in the paper Conditional Generative Adversarial Nets by Mehdi Mirza and Simon Osindero. This architecture allows to direct the generator what samples to create.
The provided example in the notebook here is using the MNIST dataset to train a generator and allows to tell the generator what number should be generated. The following images shows some results.
