Conditional Augmentation for Generative Modeling
Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, Ilya Sutskever
Unverified — Be the first to reproduce this paper.
ReproduceAbstract
We present conditional augmentation (CondAugment), a simple and powerful method of regularizing generative models. Core to our approach is applying augmentation functions to data and then conditioning the generative model on the specific function used. Unlike typical data augmentation, CondAugment allows usage of functions which modify the target density, enabling aggressive augmentations more commonly seen in supervised and self-supervised learning. We demonstrate this is a more effective regularizer than standard methods, and use it to train a 150M parameter autoregressive model on CIFAR-10 to 2.56 bits per dim (relative to the state-of-the-art 2.80). Samples from this model attain FID 12.75 and IS 8.40, outperforming the majority of GANs. We further demonstrate the technique is broadly applicable across model architectures, objectives, and problem domains.