wiseodd/generative-models

This repository is a collection of Python implementations of generative models, which are a category of machine learning models that learn to produce new data that resembles their training data. The most common application is generating images. The implementations use two popular machine learning libraries: PyTorch and TensorFlow. The bulk of the collection focuses on Generative Adversarial Networks, or GANs. A GAN works by pitting two neural networks against each other: one tries to create convincing fake samples, and the other tries to tell the fakes from real ones. The repository includes twenty different GAN variants, from the original design up through later approaches with names like Wasserstein GAN, InfoGAN, and Coupled GAN, each of which addresses different weaknesses in earlier methods. The papers that proposed each variant are linked directly in the README. The collection also includes five variants of Variational Autoencoders, a different approach to generation that compresses data into a compact representation and then reconstructs it. Additional models cover Restricted Boltzmann Machines and the Helmholtz Machine, which are older probabilistic approaches that predate the GAN era. When you run any of these models, the training process saves generated image samples to an output folder so you can watch the quality improve over time. Setup requires Miniconda, a Python package manager, along with separate installations of TensorFlow and PyTorch. The README is sparse on usage detail beyond the setup steps, but each model is a self-contained script that follows the same folder structure.