Emiel Hoogeboom,
Rianne Van Den Berg,
Max Welling
;
Proceedings of the 36th International Conference on Machine Learning, PMLR 97:2771-2780, 2019.
Abstract
Generative flows are attractive because they admit exact likelihood optimization and efficient image synthesis. Recently, Kingma & Dhariwal (2018) demonstrated with Glow that generative flows are capable of generating high quality images. We generalize the 1 {\texttimes} 1 convolutions proposed in Glow to invertible d {\texttimes} d convolutions, which are more flexible since they operate on both channel and spatial axes. We propose two methods to produce invertible convolutions, that have receptive fields identical to standard convolutions: Emerging convolutions are obtained by chaining specific autoregressive convolutions, and periodic convolutions are decoupled in the frequency domain. Our experiments show that the flexibility of d {\texttimes} d convolutions significantly improves the performance of generative flow models on galaxy images, CIFAR10 and ImageNet.
@InProceedings{pmlr-v97-hoogeboom19a,
title = {Emerging Convolutions for Generative Normalizing Flows},
author = {Hoogeboom, Emiel and Van Den Berg, Rianne and Welling, Max},
booktitle = {Proceedings of the 36th International Conference on Machine Learning},
pages = {2771--2780},
year = {2019},
editor = {Chaudhuri, Kamalika and Salakhutdinov, Ruslan},
volume = {97},
series = {Proceedings of Machine Learning Research},
address = {Long Beach, California, USA},
month = {09--15 Jun},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v97/hoogeboom19a/hoogeboom19a.pdf},
url = {http://proceedings.mlr.press/v97/hoogeboom19a.html},
abstract = {Generative flows are attractive because they admit exact likelihood optimization and efficient image synthesis. Recently, Kingma & Dhariwal (2018) demonstrated with Glow that generative flows are capable of generating high quality images. We generalize the 1 {\texttimes} 1 convolutions proposed in Glow to invertible d {\texttimes} d convolutions, which are more flexible since they operate on both channel and spatial axes. We propose two methods to produce invertible convolutions, that have receptive fields identical to standard convolutions: Emerging convolutions are obtained by chaining specific autoregressive convolutions, and periodic convolutions are decoupled in the frequency domain. Our experiments show that the flexibility of d {\texttimes} d convolutions significantly improves the performance of generative flow models on galaxy images, CIFAR10 and ImageNet.}
}
%0 Conference Paper
%T Emerging Convolutions for Generative Normalizing Flows
%A Emiel Hoogeboom
%A Rianne Van Den Berg
%A Max Welling
%B Proceedings of the 36th International Conference on Machine Learning
%C Proceedings of Machine Learning Research
%D 2019
%E Kamalika Chaudhuri
%E Ruslan Salakhutdinov
%F pmlr-v97-hoogeboom19a
%I PMLR
%J Proceedings of Machine Learning Research
%P 2771--2780
%U http://proceedings.mlr.press
%V 97
%W PMLR
%X Generative flows are attractive because they admit exact likelihood optimization and efficient image synthesis. Recently, Kingma & Dhariwal (2018) demonstrated with Glow that generative flows are capable of generating high quality images. We generalize the 1 {\texttimes} 1 convolutions proposed in Glow to invertible d {\texttimes} d convolutions, which are more flexible since they operate on both channel and spatial axes. We propose two methods to produce invertible convolutions, that have receptive fields identical to standard convolutions: Emerging convolutions are obtained by chaining specific autoregressive convolutions, and periodic convolutions are decoupled in the frequency domain. Our experiments show that the flexibility of d {\texttimes} d convolutions significantly improves the performance of generative flow models on galaxy images, CIFAR10 and ImageNet.
Hoogeboom, E., Van Den Berg, R. & Welling, M.. (2019). Emerging Convolutions for Generative Normalizing Flows. Proceedings of the 36th International Conference on Machine Learning, in PMLR 97:2771-2780
This site last compiled Mon, 16 Sep 2019 16:05:04 +0000
