Use certain internal brms functions in non-linear formulas #1635
Comments
|
you need to define the step function in an environment that can be found by
the environment of your model formula. unfortunately the brms name space is
no longer in scope during nonlinear formula evaluation and I didn't find a
safe way to make it in scope without breaking other things. does that help?
Diego Barneche ***@***.***> schrieb am Di., 26. März 2024,
07:13:
… Hello,
Since the latest CRAN release, I've been having issues adding LOO via
add_criterion for a non-linear model.
This has prompted CRAN to ask us for a fix in one of our packages (
bayesnec) that depends on brms.
Please see reprex below.
Thanks for your help!
library(brms)
rstan::rstan_options(auto_write = TRUE)
options(mc.cores = parallel::detectCores())
fct <- function(beta, bot, nec, top, x) {
bot + (top - bot) * exp(-exp(beta) * (x - nec) *
ifelse(x - nec < 0, 0, 1))
}
set.seed(10)
x <- abs(rnorm(100, 1))
y <- fct(beta = 1.2, bot = 0.01, nec = 1.5, top = 0.95, x)
data <- data.frame(x, y)
# plot(y ~ x, data)
# linear
test_l <- brm(
y ~ x, data = data, iter = 200, warmup = 150, chains = 2, seed = 10
)
add_criterion(test_l, "loo")
nl_mod <- brms::bf(
y ~ bot + (top - bot) * exp(-exp(beta) * (x - nec) * step(x - nec)),
bot + top + beta + nec ~ 1, nl = TRUE
)
test_nl <- brm(
nl_mod, data = data, iter = 200, warmup = 150, chains = 2, seed = 10
)
add_criterion(test_nl, "loo") # returns an error
# Error: Error in x$terms : $ operator is invalid for atomic vectors
Session
> sessionInfo()
R version 4.3.0 (2023-04-21)
Platform: aarch64-apple-darwin20 (64-bit)
Running under: macOS 14.2.1
Matrix products: default
BLAS: /Library/Frameworks/R.framework/Versions/4.3-arm64/Resources/lib/libRblas.0.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/4.3-arm64/Resources/lib/libRlapack.dylib; LAPACK version 3.11.0
locale:
[1] en_AU.UTF-8/en_AU.UTF-8/en_AU.UTF-8/C/en_AU.UTF-8/en_AU.UTF-8
time zone: Australia/Perth
tzcode source: internal
attached base packages:
[1] stats graphics grDevices utils datasets methods base
other attached packages:
[1] brms_2.21.0 Rcpp_1.0.12
loaded via a namespace (and not attached):
[1] gtable_0.3.4 tensorA_0.36.2.1 ggplot2_3.5.0 QuickJSR_1.1.3 processx_3.8.4 inline_0.3.19 lattice_0.21-8
[8] callr_3.7.5 vctrs_0.6.5 tools_4.3.0 ps_1.7.6 generics_0.1.3 stats4_4.3.0 curl_5.0.2
[15] parallel_4.3.0 sandwich_3.0-2 tibble_3.2.1 fansi_1.0.6 pkgconfig_2.0.3 Matrix_1.6-1 checkmate_2.3.1
[22] distributional_0.4.0 RcppParallel_5.1.7 lifecycle_1.0.4 compiler_4.3.0 stringr_1.5.1 Brobdingnag_1.2-9 munsell_0.5.0
[29] codetools_0.2-19 bayesplot_1.11.1 pillar_1.9.0 MASS_7.3-60 StanHeaders_2.32.6 bridgesampling_1.1-2 abind_1.4-5
[36] multcomp_1.4-25 nlme_3.1-163 posterior_1.5.0 rstan_2.32.6 tidyselect_1.2.1 digest_0.6.35 mvtnorm_1.2-4
[43] stringi_1.8.3 dplyr_1.1.4 reshape2_1.4.4 splines_4.3.0 grid_4.3.0 colorspace_2.1-0 cli_3.6.2
[50] magrittr_2.0.3 loo_2.7.0 survival_3.5-7 pkgbuild_1.4.4 utf8_1.2.4 TH.data_1.1-2 scales_1.3.0
[57] backports_1.4.1 estimability_1.4.1 matrixStats_1.2.0 emmeans_1.8.7 gridExtra_2.3 zoo_1.8-12 coda_0.19-4.1
[64] V8_4.3.3 rstantools_2.4.0 rlang_1.1.3 xtable_1.8-4 glue_1.7.0 jsonlite_1.8.8 R6_2.5.1
[71] plyr_1.8.9
—
Reply to this email directly, view it on GitHub
<#1635>, or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ADCW2AESPLYBTILFWZKHD33Y2EGX7AVCNFSM6AAAAABFIJFAFOVHI2DSMVQWIX3LMV43ASLTON2WKOZSGIYDOMZUHEYTAOI>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
|
Thanks very much Paul, that certainly clarified things. 1 - was there any reason why you decided not to export 2 - are there any other functions like |
|
1 - because I don't want to clutter the global env with a lot of these brms functions, in particular if they mask other functions as is the case for step. TBH I implemented a lot of the stan functions in brms for precisely the reason you are using them now (in non-linear formulas). However, I have yet to find a safe way to expose just exactly these functions to the formula environment without exposing all of brms internal functions to it which may lead to all sort of hard to debug problems for users. 2 - all internal brms functions one might want to use in non-linear formulas. If you have a suggestion for how to fix 1 from the brms side, I would be happy to hear! |
paul-buerkner
changed the title
|
Thanks Paul, that makes sense. I don't have any solutions for problem 1, unfortunately, but I'll keep my eyes peeled in case I come across any potential ideas. For now, we'll stick with exporting |
|
Could one define inside the brms namespace an environment which has the Stan internal functions and then one places this environment into the search path in some magic way? E.g. could we say that the formula has as parent environment a copy of this Stan functions namespace which itself has as parent the environement of the user which was used to define the brms formula? I am not a guru on these R search path stuff, but maybe this is a solution to follow on? |
|
Yeah, I like the idea. I think this is doable. |
|
I was going to suggest something very similar. A quick test with just the step function as an example show it works: Should be pretty safe, since this internal stan functions environment won't be exposed to any other functions, just to things that work with the formula |
|
for new formula envs it is safe. for existing formula envs like global-env it may not be super safe without making sure that the env changes are reverted after evaluation. |
|
But how about this pattern: cur_env <- environment(test_nl$formula$formula) ## not sure if this copies...
env <- new.env(parent = cur=env)
env$step <- brms:::step
environment(test_nl$formula$formula) <- new.env(parent = env)
add_criterion(test_nl, "loo")Essentially one should be able to "sandwich" a pre-populated Stan functions environment into the child-parent structure of environments. |
|
yeah, I think this could work. Are there any dangers of side effects here? |
|
I'd hope not, but I am not confident. I read through the R documentation of environments and how they are searched - but I have not messed myself with environments at this level. A ramification is obviously that the functions in the Stan internal functions environment get precedence over what is found in the environment of the formula - but that would be intended, I guess. It get's a bit tricky with the vectorisation if the functions, I guess (so the Stan internal functions as implemented in R need to be clever in this regard). |
|
This should now be working. Thank you all for the discussion about this issue. |
Hello,
Since the latest CRAN release, I've been having issues adding LOO via
add_criterionfor a non-linear model.This has prompted CRAN to ask us for a fix in one of our packages (
bayesnec) that depends onbrms.Please see reprex below.
Thanks for your help!
Session
The text was updated successfully, but these errors were encountered: