Roadmap

The items below are considered high priority for future development, and are at various stages of planning and implementation.

Data-parallel gradient evaluation

Taking advantage of multiple CPU cores to parallelize sampling for larger datasets.

Multivariate Normal

MVNormal is a very commonly used distribution that is currently not supported by Rainier.

Discrete latent variables

Support latent for Discrete distributions, at least in some cases, with automatic Rao-Blackwellization.

Automatic Reparameterization

Rainier currently only supports non-centered parameterizations, which is a good default, but automatic reparameterization as in Gorinova et al would be an improvement in some cases.

Mass Matrix adaptation

Currently Rainier's HMC always uses the identity mass matrix. Mass matrix adaptation would improve performance on correlated parameters.

Feedback

Feel free to file issues at GitHub if something important to you is missing from this list.

You may also send email with any feedback (good or bad!) to avi@avibryant.com.