Andy Dienes
Andy Dienes
a challenge here is that the "best" choice can be any of 4-5 different algorithms (with significant performance implications) depending not only on `n` and `k` but also on the...
cleaned up the test failures, so this is ready for review. just as a reminder of the motivation, performance is improved across the board for all parameter choices, and for...
@devmotion I'm sorry to pester, I know your attention is probably stretched thin being basically one of only a small number of maintainers for a package as popular as this...
> the abstract super type and its "constructor" seems to be an approach that is not used by other samplers in Distributions. fair point, happy to address this only question...
thanks for the comments -- let me know if the most recent commit addresses them appropriately
friendly bump
I hope I have addressed all comments --- the benchmarks still look good on my end so nothing regressed (as expected)
took advantage of latest changes to `AliasTable` the approach ``` rand!(rng, s.scratch_alias_rng) for r in s.scratch_alias_rng x[AliasTables.sample(r, s.alias.at)] += 1 end ``` to batch sample the seed is slightly faster...
another friendly bump :)
I'm sorry to be a nag, but bumping again.