// the find
goodfeli/adversarial
Code and hyperparameters for the paper "Generative Adversarial Networks"
The original GAN implementation from Goodfellow et al.'s 2014 paper — the code that started the generative modeling gold rush. It runs on Pylearn2 and Theano, both of which have been dead for years. This is a historical artifact, not a usable library.
The YAML-based experiment configs are a clean way to reproduce exact paper hyperparameters. The Parzen window log-likelihood estimation script (parzen_ll.py) is a nice reference for evaluating generative models the way the original paper did. The flat structure means you can read the entire codebase in an afternoon — there's no abstraction hiding what GANs actually do.
Depends on Theano and Pylearn2, both abandoned circa 2017 — getting this to run requires pinning to decade-old commits of dead libraries. No PyTorch or TensorFlow port, so you can't drop this into any modern stack. The README explicitly warns about hardware-dependent rounding differences affecting reproduced numbers, which makes exact replication nearly impossible without a GTX 580. Last touched in 2020 and only to update a license file — effectively unmaintained since 2014.