Adversarial training of partially invertible variational autoencoders

Abstract : Adversarial generative image models yield outstanding sample quality, but suffer from two drawbacks: (i) they mode-drop, \ie, do not cover the full support of the target distribution, and (ii) they do not allow for likelihood evaluations on held-out data. Conversely, maximum likelihood estimation encourages models to cover the full support of the training data, but yields poor samples. To address these mutual shortcomings, we propose a generative model that can be jointly trained with both procedures. In our approach, the conditional independence assumption typically made in variational autoencoders is relaxed by leveraging invertible models. This leads to improved sample quality, as well as improved likelihood on held-out data. Our model significantly improves on existing hybrid models, yielding GAN-like samples, and IS and FID scores that are competitive with fully adversarial models, while offering likelihoods measures on held-out data comparable to recent likelihood-based methods.
Document type :
Preprints, Working Papers, ...
Liste complète des métadonnées

https://hal.archives-ouvertes.fr/hal-01886285
Contributor : Thomas Lucas <>
Submitted on : Friday, March 1, 2019 - 4:36:40 PM
Last modification on : Wednesday, March 20, 2019 - 1:29:04 AM

File

ms.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01886285, version 2

Collections

Citation

Thomas Lucas, Konstantin Shmelkov, Karteek Alahari, Cordelia Schmid, Jakob Verbeek. Adversarial training of partially invertible variational autoencoders. 2019. ⟨hal-01886285v2⟩

Share

Metrics

Record views

166

Files downloads

273