Brandon T. Willard

FYI: I've added a lot of relevant functionality in https://github.com/pymc-devs/symbolic-pymc/pull/113 and https://github.com/pymc-devs/symbolic-pymc/pull/114. Part of the functionality introduced there allows us to automatically transform Theano graphs with random variable outputs from...

Since the recent improvements and fixes to the shape handling in our custom `Distribution`s, we've been able to dramatically increase the speed of posterior predictive sampling by constructing the posterior...

> Will posterior [predictive] sampling speed still be an issue with PyMC3 version 4? No, not at all.

> It [appears](https://github.com/pytoolz/toolz/issues/188) that `toolz` does not use trampoline evaluation, and has no intention to do so. We may need to implement a version of `interleave` that uses it to...

> The error is raised in the `interleave` function from `toolz`. I'm not able to reproduce the error locally.

FYI: If I set `N` to something large, the issue appears. Anyway, yes, we should make our own version of `interleave` that doesn't use recursion.

> just throwing the idea here to just generate the types from running the tests. > > (/mypy/ typing? or just typing?) This issue is for Mypy (i.e. checked typing),...

This needs more/better tests before merging. It's still just a quick port of my old `logpy` PR.

That's a pretty useful numerical study. The work they do to break down negative-binomials into Poisson-like terms might help with the derivation of a negative-binomial parameterization based on the standard...

> Here is a [notebook report](https://gist.github.com/xjing76/3c99a09930faf57ca2562cf9b2b1ae7a) on the findings. This is very helpful! It's missing a plot of the simulated data, though, which would help illustrate some of the challenges...