GAN You Do the GAN GAN? arXiv

This repository contains the Paper PDF, full source code, and sample models for the short paper "GAN You Do the GAN GAN?" in which we train a GAN over 100 snapshots from each of 35 trained GANs.

Quickstart

We assume a standard Anaconda setup with GPU Pytorch installed. This is a minimal project with standard libraries -- if you are missing anything, snag it with pip. This project was conducted on a single desktop with a 6 core i7 and a GTX 1080 Ti. If training from scratch, expect to generate ~20GB of data.

Recommended Setup:

#Train all the GANS (overnight this)
python train.py gan

#Train the GAN-GAN over GANs (lunch break this)
python train.py gangan

#Reproduce the paper figures (using prebuilt models)
python figures.py

Note: GANs are notoriously difficult to train. While the GAN training code is stable. the GAN-GAN code is not. Stability and convergence speed vary dramatically with random seed. While most runs yield something reasonable, you should not expect to reproduce our results exactly without a bit of patience and tweaking.

TL;DR

GAN You Do the GAN-GAN? Yes you GAN! Generative models can model other generative models in an interpretable and meaningful latent space.

Summary

This is an April 1 paper, but the results are real. This is what the samples from a small MNIST GAN look like over the course of training:

(Image in case the video does not load)

And this is what the images (horizontal) sampled from the GANs sampled from the 1D latent space (vertical) of a GAN-GAN look like. Not only are they somewhat better than the images sampled from the GAN, but the latent space of the GAN-GAN appears to have a structure that roughly corresponds to the different training stages of a GAN.