TzK: Flow-Based Conditional Generative Model

  • 2019-03-14 16:00:49
  • Micha Livne, David Fleet
  • 0


We formulate a new class of conditional generative models based onprobability flows. Trained with maximum likelihood, it provides efficientinference and sampling from class-conditionals or the joint distribution, anddoes not require a priori knowledge of the number of classes or therelationships between classes. This allows one to train generative models frommultiple, heterogeneous datasets, while retaining strong prior models oversubsets of the data (e.g., from a single dataset, class label, or attribute).In this paper, in addition to end-to-end learning, we show how one can learn asingle model from multiple datasets with a relatively weak Glow architecture,and then extend it by conditioning on different knowledge types (e.g., a singledataset). This yields log likelihood comparable to state-of-the-art, compellingsamples from conditional priors.


Introduction (beta)



Conclusion (beta)