Reduce_sum for multi-level model

sonicking · June 26, 2021, 4:45pm

Hello. I am using CmdStanR.

Is there any example of a multi-level model using reduce_sum, slicing over the group-level parameter vector? I’d believe slicing over the data is rather similar for multi-level models and the minimal example. (Please confirm?) Much appreciated. Thank you very much.

wds15 · June 26, 2021, 5:52pm

github.com/paul-buerkner/brms

improvements for threading with mixed models

opened 02:37PM - 02 Oct 20 UTC

wds15

efficiency feature

As already discussed, it would be better for the performance of threading models… with random effects to slice over the random effect of a group level. Below is some example model and code which demonstrates how that could be coded up. The code is written in a hopefully generic enough way that it can be fit into the brms code generation machinery. Essentially the code resorts the entire data set along one grouping level and then the `reduce_sum` is done over levels of the grouping factor instead of data rows. The resorting seems to be beneficial for single core performance as well (at least in this examples, but that's not huge though still there is a speedup). ```R > ## base model brms, 1 core > stime1b user system elapsed 180.172 0.614 180.434 > ## optimized model, 1 core > stime1c user system elapsed 167.554 0.473 168.505 > ## base model brms, 3 cores > ptime1b user system elapsed 324.204 2.862 109.954 > ## optimized model, 3 core > ptime1c user system elapsed 282.691 1.989 95.492 > ## base model brms speedup 1 core vs 3 cores > stime1b / ptime1b user system elapsed 2.142594 1.731050 1.640995 > ## optimized model speedup 1 core vs 3 cores > stime1c / ptime1c user system elapsed 2.235256 1.905097 1.764598 > ## optimized model 3 core vs 1 core base model > stime1b / ptime1c user system elapsed 2.475910 1.991306 1.889520 > ``` [horseshoe-scale_R.txt](https://github.com/paul-buerkner/brms/files/5318580/horseshoe-scale_R.txt)

sonicking · June 29, 2021, 3:06am

Thank you. When you have time, are you able to provide a bit more commentary on what is going on? I am not fully able to follow it. I will also try again when my mind is fresher. Thanks.

wds15 · June 30, 2021, 11:00am

I have a varying intercept by group “g”. It makes thus sense to do the partial sums over groups formed by “g”. To do that I am sorting the data-set first by group membership. The code exemplifies everything.

Anything more specific you have questions on?

sonicking · July 3, 2021, 2:54am

What are N_1 in line 127, J_1 in line 128, and N in the next few lines?

I believe J_1 is an array or vector, but I don’t know what it looks like.

Thanks.

wds15 · July 3, 2021, 7:09am

That’s standard brms way of coding. You should look at the Stan data if in doubt.

sonicking · July 3, 2021, 11:13am

Ah. I use RStan and now CmdStanR. I never use brms. Sorry about that. I will look into it. Thanks again for your help.?

sonicking · July 4, 2021, 1:34am

Sorry to bother you again. Can you explain r_1_1? I see that it is not even initialized anywhere…thanks.

Topic		Replies	Views
Reduce_sum() no time saving for multilevel model Modeling techniques	1	447	December 24, 2020
Parallelization via reduce_sum for hierarchical model Modeling performance , hierarchical-model , paralellization	2	636	January 26, 2022
Help with reduce_sum Modeling	32	1392	August 4, 2020
Model with reduce_sum takes too long Modeling	28	1455	December 28, 2020
Reduce_sum parallelisation issue Modeling cmdstanr , multivariate-normal	12	1029	February 24, 2022

Reduce_sum for multi-level model

Related topics