GAN implementations w/ PyTorch.
-
Unzip mnist_csv.zip directly into repo.
-
Pip install requirements.txt
python3 -m pip install -r requirements.txt- Run GAN
python3 gan/gan.pyGANs, Generative Adversarial Networks, are adversarial models meaning multiple models compete with eachother in a game. In this scenario, there are two players, both represented as neural networks, the image generator and the image discriminator. The discriminator tries to determine which images are real and which are generated, encouraging both models to become more effective at their tasks.
More specifically, the generator defines a distribution that maps input noise to an image. The discriminator maps images to the probability that they are real images.
Of course, the discriminator is trained to maximize the probability of correctly labelling the images as fake or real, while the generator is trained to minimize the probablity of its images being labelled as fake. Thus models are playing a minimax game with the nash equilibra where the generator is able to produce images that are indistinguishable from those in the dataset.
Training is done by alternating k steps optimizing the discriminator and one step for the generator. It is important that both players improve at roughly the same rate, else will get issues like mode collapse and poor generated images.
(More in relevant files)