Fast cmdstanr summary function

Modeling
1

Hi everyone. I wrote a small R library for quickly computing posterior summaries of draws_df objects. It’s available at the link below. I’ve seen that there’s been some discussion of speeding up summaries before, but this goes beyond speeding up the mean/median/sd and also improves the speed of computing the convergence metrics. My function is not as generic / deeply-tested as posterior::summarise_draws(), but it gives me a good performance boost for my purposes. Maybe someone will find it useful, just thought I’d share.

Blog post with a bit more detail: Fast cmdstanr summary function – andrewghazi.github.io
Github repo: GitHub - andrewGhazi/fsummary: Fast cmdstanr summaries

unnamed-chunk-1-1
unnamed-chunk-1-11344×960 12.4 KB

6 Likes
2

Cool, thanks for sharing this!

3

Very cool! I had the same issue and solved it with a wrapper of the stansummary command-line utility.

2 Likes
4

Ah interesting, I didn’t know about this. I’ll have to check how it compares.

5

The stansummary utility in CmdStan is using an older version of Rhat, not the rank-normalized computations introduced in this [1903.08008] Rank-normalization, folding, and localization: An improved $\widehat{R}$ for assessing convergence of MCMC paper. I’m working on a refactor that will correct this; also improve performance. Like all refactors, easier said than done.

1 Like
6

I wondered about this! Thanks for confirming. Compared to posterior, stansummary does not provide tail ESS either. Is this something we can hope for too?

7

yes. we were lucky to pick up a new Stan dev who implemented the rank-normalization needed to do the bulk, tail Rhat, it’s just a matter of using that for ESS as well. as I said, WIP.

2 Likes