Bayesian Recurrent Neural Networks

This is a replication of the paper 'Bayesian Recurrent Neural Networks' by Meire Fortunato, Charles Blundell, Oriol Vinyals.

Paper: https://arxiv.org/abs/1704.02798

Status: Basic model replicated.

Requirements

• Python 3.5
• TensorFlow 1.3.0

Usage

$ sh download_ptb.sh
$ python bayesian_rnn.py -model medium -log_sigma1 -1.0 -log_sigma2 -7.0 -prior_pi 0.25

To-do:

• Implement posterior sharpening.
• Implement image captioning experiment.

Acknowledgements

Thanks to Meire Fortunato for providing the Bayes by Backprop/cell code and @alexkrk for an initial implementation.