Hello Stan Aficionados! The gist of my request is as follows: how the heck can I speed-up a mixed gaussian copula so that it does not take 2 weeks to fit on an I9 with 14 cores and 32 gb of ram. Details and Specification: I am incredibly indebted to @spinkney 's code here as well as discussions w…

Glad it’s helping! Some of the speed may be less the code and more numerical struggles. One thing to try is calculating on the log scale using ‘inv_Phi_log’ function. Replace all the ‘normal(0, 1)’ with ‘std_normal()’. One thing I’m not sure about is how the compiler handles all those different 1:N…

Thanks Sean! I’ll get to work vectorizing the j-loop - that has been something I’ve been meaning to do. Perhaps a dumb question - is it possible in Stan to do elementwise addition and subtraction? I know we can do products, quotients, and powers. Trying to figure out the best way to vectorize. Rel…

[image] spinkney: One thing I’m not sure about is how the compiler handles all those different 1:N loops. @WardBrian does the compiler merge those together? Stanc3 does not. It’s possible clang/gcc would, but it requires some sophisticated analysis to merge loops together. Whether you will ge…

The irony! I ran a test combining that loop in the Transformed Parameters block to assign all of the means and sd’s with one pass of N instead of re-running for every data type (i.e., a loop for dentition, a loop for ef, etc.) and it did in fact slow things down. Not by a huge margin - ~ an hour or…

You can try this as the inverse is just done on a J x J instead of J x N. real gaussian_copula_cholesky_lpdf( matrix U, matrix L ) { int N = rows(U); int J = cols(U); matrix[N,J] Q = inv_Phi(U); matrix[J,J] Gammainv = crossprod(mdiv…

What about this: real multi_normal_cholesky_copula_lpdf(matrix U, matrix L) { int N = rows(U); // total number of observations int J = cols(U); // total number of responses (J+K) matrix[J, J] Gammainv = chol2inv(L); return -N * sum(log(diagonal(L))) - 0.5 * sum(add_diag(Gammainv, -1…

Yes, just depends on the input. I have like 10 versions of the multi normal copula, lol.

Mind if I ask the difference numerically? I believe when I started this long ago I used the matrix[J, J] Gammainv = chol2inv(L); and then a suggestion was to try matrix[J, N] G = mdivide_left_tri_low(L, U');. I believe there is a slight slowdown when using mdivide_left_tri_low as compared to chol2…

Just an update for you: Running a 6-variable model (2 continuous, 4 ordinal) with the exact code you have on gihub for the multi_normal_cholesky_function works great. This is using the chol2inv function I described in my last post with a J x J matrix. Alternatively, if I run the same model with th…

I have had a lot of difficulty with fitting multivariate normal models in Stan with a factor structure (or really in any case where there is a large amount of correlation between some of the variables), particularly as the number of variables increases. As far as I can tell, the problem is associat…

Speed-up for Mixed Discrete-Continuous Gaussian Copula - Reduce Sum? - #12 by spinkney - Modeling

Topic		Replies	Views	Activity
Feature request: copulas for multivariate responses of mixed types brms	50	6775	June 17, 2024
Slow run time when fitting multi_normal_cholesky likelihood in bivariate model Modeling rstan , fitting-issues , specification , performance	3	356	May 8, 2024
Vine copula examples in Stan Modeling techniques , specification	12	2332	April 14, 2023
Optimizing stan code runtime for matrix factorization Modeling	6	392	May 18, 2024
Vectorization and improving speed for a complex mixture type latent trait model for response times Modeling	11	865	January 16, 2025