Alexey Radul

Add madeSpLogDensityOfDataBound methods for the categorical family of distributions

3

[Separated from #455]. (The "categorical family" is defined for the present purposes in #454). All those SPs that track sufficient statistics have easily computable non-trivial upper bounds on the probability...

Do we want make_gamma_crp?

5

According to http://web.cse.ohio-state.edu/~kulis/teaching/788_sp12/DP.learnalpha.pdf, the gamma distribution is a conjugate (*) prior for the alpha parameter of a Chinese Restaurant Process. Do we want to capture this observation in a collapsed...

calling itself the asymmetric Dirichlet categorical. This is due to excessive code de-duplication in https://github.com/probcomp/Venturecxx/commit/38a6c32f4f9f6ecd186aa67125a7e93bf4840f32, and absence of a standard system for SPs to know their names.

Language design push

2

I reviewed the ticket tracker in May 2016, and collected all the nits I could find that are reasonable to classify as "language design". The main characteristic these things share...

Puma's collapsed Dirichlet categorical distribution is asymptotically slower to simulate than Lite's

1

because it does a linear scan of the aux+weight vector instead of Lite's tree-based sampling algorithm. Implementing The Plan (#453) and porting it to Puma should fix this.

Increase politeness of C++ asserts

2

When a C++ assert fails, the user sees the location and message (if any), and "Aborted (core dumped)". No stack trace; no ability for an enclosing Python wrapper to recover....

Differences between Lite and Puma scaffold construction: - re-traversing an aaa node updates the index assignment in Lite but not Puma - re-traversing a drg node updates the index assignment...

Do we want to get Eigen out of the backend/new_cxx source tree?

2

Into external/? Into an actual build-time dependency?

Port implementer-defined records to Puma

1

The Lite implementation is in backend/lite/records.py Puma would need - a similar representation for record objects - the ability to translate records to and from stack dicts - matching definitions...

This may already be so, but would need to be tested. This might be nice to get by the 0.5 release, but maybe we don't care.