0.3.5

v0.3.5

version 0.3.5

0.3.2

This release adds support for diagonal and dense mass matrices, including adaptive tuning, and expands the sampler config API to include these.

It also adds hooks for displaying progress during sampling, and a basic progress display for Jupyter notebooks.

0.3.0

This is a complete overhaul of the API and of some parts of the implementation.

Most notably:

• RandomVariable has been removed
• Predictor has been replaced by a much more general Vec
• the various fit methods on those two have been replaced by Model.observe
• constants in the compute graph can now be vectors as well as scalars

See https://rainier.fit for more complete documentation of the new API.

0.2.3

Release 0.2.3 (#398)

* Setting version to 0.2.3

* update for 0.23 release

0.2.2

• Fixes bugs in auto-differentiation (thanks to @arthur-bit-monnot )
• introduces trig functions on Real (ditto)
• adds basic https://almond.sh/ jupyter kernel integration to rainier-plot (see Example.ipynb)

0.2.1

This is a bug fix release for 0.2.0, but since I didn't write up release notes for that one I'll include them here.

This includes:

• deterministic evaluation, to simplify unit testing and reproducing bugs
• unit tests use simulation-based calibration for validating correctness
• better discrete distribution support, such as DiscreteConstant and DiscreteMixture (thanks to @DavidRdgz)
• detailed logging for samplers using google's Flogger
• graphviz-based visualization of compute and IR graphs
• a Lookup node for more compact compute graphs
• a significant overhaul of the compute layer to allow batched evaluation of large observation vectors
• an "empirical HMC" sampler courtesy of @jonnylaw that provides a lightweight NUTS alternative

Mixture params and performance improvements

Merge pull request #175 from stripe/minmax

Rectifier

0.1.1

This release includes:

• significant scalability improvements to HMC; CPU time should scale linearly with the number of iterations, rather than hitting a garbage collection wall; some quadratic performance bugs in deriving gradients have also been fixed.
• much improved diagnostics, including: effectiveSampleSize computation, simulation-based calibration support, and fancy trace plots and corner plots using EvilPlot
• assorted bug fixes
• richer examples

Special thanks to @darrenjw for all his contributions and stress-testing in this release cycle.