Improve documentation and implementation of custom distributions #1431
Conversation
Codecov Report
@@ Coverage Diff @@
## master #1431 +/- ##
==========================================
- Coverage 67.03% 65.56% -1.48%
==========================================
Files 25 25
Lines 1617 1661 +44
==========================================
+ Hits 1084 1089 +5
- Misses 533 572 +39
Continue to review full report at Codecov.
|
| struct LogPoisson{T<:Real,S} <: DiscreteUnivariateDistribution | ||
| logλ::T | ||
| λ::S |
There was a problem hiding this comment.
Is it worth adding an extra field here to remove the exp call later?
There was a problem hiding this comment.
I don't have a strong opinion on this one. It will be (slightly) more efficient if you evaluate the logpdf at least twice or sample twice, and it won't be less efficient if you evaluate or sample at least once. I noticed that it is possible to cache it after having updated the BinomialLogit distribution (in this case the cached computations are definitely more expensive though).
There was a problem hiding this comment.
Alright, then I'm happy. Let's stick with this for now.
2 participants
This PR adds proper docstrings to the included distributions, improves the implementation by implementing
samplerand fixing some type instability issues, and prepares the code for Distributions 0.24 (which fixes thelogpdf/pdfweirdness for univariate distributions).I noticed that we currently can't support Distributions 0.24 since it requires TuringLang/DynamicPPL.jl#150 (both AdvancedMH and DynamicPPL block the update, and the latest versions of AdvancedMH are only compatible with AbstractMCMC 2, so it has to be supported by DynamicPPL and Turing as well).
Probably similarly the random measures could be improved.