Estimating Gradients for Discrete Random Variables by Sampling without Replacement

This repository contains implementations of the Unordered set estimator as derived in our paper Estimating Gradients for Discrete Random Variables by Sampling without Replacement.

• The Bernoulli gradient variance experiment is in PyTorch
• The Categorical VAE experiment is in TensorFlow

Paper

For more details, please see our paper Estimating Gradients for Discrete Random Variables by Sampling without Replacement which has been accepted at ICLR 2020. If this code is useful for your work, please cite our paper:

@inproceedings{
    Kool2020Estimating,
    title={Estimating Gradients for Discrete Random Variables by Sampling without Replacement},
    author={Wouter Kool and Herke van Hoof and Max Welling},
    booktitle={International Conference on Learning Representations},
    year={2020},
    url={https://openreview.net/forum?id=rklEj2EFvB}
}

Stochastic Beam Search

For sampling without replacement from a fully factorized distribution for the VAE experiment, this code contains an implementation of Stochastic Beam Search in TensorFlow in vae/lib/beam_search.py.