Hey all, I’ve been working to translate a hierarchical nonlinear model from JAGS to STAN, mostly to allow timely inference on a large-ish dataset of ~100,000 observations. I’ve got a version in STAN that works well for simulated data, but performance isn’t much better than the JAGS version. I susp…

Cauchy is advised against of late as I recall, but if you really want it, try the tan trick illustrated in the “reparameterizing the Cauchy” section of the manual.

[image] ibreckhe: can’t vectorize the exponentiation But you can vectorize everything inside the exp, storing as a temporary variable that you then loop over to exponentiate.

The for loops in the model block are unnecessary, you can do just: opt_group ~ normal(0,group_sd[1]); width_group ~ normal(0,group_sd[2]); height_group ~ normal(0,group_sd[3]); opt_spp ~ normal(0,spp_sd[1]); width_spp ~ normal(0,spp_sd[2]); height_spp ~ normal(0,spp_sd[3]);

[image] ibreckhe: group_sd[1] ~ cauchy(0,3); group_sd[2] ~ cauchy(0,3); group_sd[3] ~ cauchy(0,3); spp_sd[1] ~ cauchy(0,5); spp_sd[2] ~ cauchy(0,5); spp_sd[3] ~ cauchy(0,5); Can be just: group_sd ~ cauchy(0,3); spp_sd ~ cauchy(0,5);

If the number of unique rows in a table produced by cbind(group,species) is far fewer than N, then your computation of opt, width and height have some redundant computations that you could instead do for just those unique combinations, indexing into the result to continue with the test of the comput…

Thanks Mike! I didn’t realize the distribution assignment worked on vectors, and I’ll give some thought to the redundant computation problem. An example would be really useful! Cheers, Ian

Mike provided some good hints [image] ibreckhe: I’ve got a version in STAN that works well for simulated data, but performance isn’t much better than the JAGS version. but can you be more specific how do you measure performance? N_eff/s ?

Yeah…nothing exact. I’d have to go back and compute N_eff for the JAGS version. Mixing is relatively poor, though, even with the current version in STAN incorporating some of Mike’s suggestions. For the worst-behaving parameters, I’m getting about 0.75 N_eff / sec on a run with 200 simulated observa…

This: beta_width[1] ~ normal(0,5); //prior for the intercept beta_width[2] ~ normal(0,5); //prior for the slope beta_opt[1] ~ normal(0,5); //prior for the intercept beta_opt[2] ~ normal(0,5); //prior for the slope beta_height[1] ~ normal(0,5); //prior for the intercept beta_height[2] ~ normal(0,5);…

Speeding up a hierarchical nonlinear model

Modeling

mike-lawrence February 26, 2018, 9:02pm 16

Finally got around to doing this: Tip for speeding up models with repeated observations

Topic		Replies	Views
My first Stan model - hierarchical logistic regression Modeling cognitive-science	28	8863	June 6, 2017
Suggestions for improving model of bird monitoring data Modeling performance	11	1166	November 17, 2020
Slow Stan Code Modeling fitting-issues , performance	33	3500	June 27, 2018
NLME model in STAN for missing + unbalanced longitudinal data Modeling rstan , mixed-model , missing-data	20	752	October 21, 2024
Scaling up a hierarchical model Modeling bioinformatics	28	3480	June 6, 2019

Speeding up a hierarchical nonlinear model

Related topics