Distribution Augmentation for Generative Modeling

Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, Ilya Sutskever
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:5006-5019, 2020.

Abstract

We present distribution augmentation (DistAug), a simple and powerful method of regularizing generative models. Our approach applies augmentation functions to data and, importantly, conditions the generative model on the specific function used. Unlike typical data augmentation, DistAug 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 152M 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 and problem domains.

Cite this Paper


BibTeX
@InProceedings{pmlr-v119-jun20a, title = {Distribution Augmentation for Generative Modeling}, author = {Jun, Heewoo and Child, Rewon and Chen, Mark and Schulman, John and Ramesh, Aditya and Radford, Alec and Sutskever, Ilya}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {5006--5019}, year = {2020}, editor = {III, Hal Daumé and Singh, Aarti}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/jun20a/jun20a.pdf}, url = {http://proceedings.mlr.press/v119/jun20a.html}, abstract = {We present distribution augmentation (DistAug), a simple and powerful method of regularizing generative models. Our approach applies augmentation functions to data and, importantly, conditions the generative model on the specific function used. Unlike typical data augmentation, DistAug 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 152M 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 and problem domains.} }
Endnote
%0 Conference Paper %T Distribution Augmentation for Generative Modeling %A Heewoo Jun %A Rewon Child %A Mark Chen %A John Schulman %A Aditya Ramesh %A Alec Radford %A Ilya Sutskever %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-jun20a %I PMLR %P 5006--5019 %U http://proceedings.mlr.press/v119/jun20a.html %V 119 %X We present distribution augmentation (DistAug), a simple and powerful method of regularizing generative models. Our approach applies augmentation functions to data and, importantly, conditions the generative model on the specific function used. Unlike typical data augmentation, DistAug 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 152M 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 and problem domains.
APA
Jun, H., Child, R., Chen, M., Schulman, J., Ramesh, A., Radford, A. & Sutskever, I.. (2020). Distribution Augmentation for Generative Modeling. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:5006-5019 Available from http://proceedings.mlr.press/v119/jun20a.html.

Related Material