Random but consistent C stack error on Windows 10

I continue to get the C stack error that crashes R on Windows 10. I have run Stan with nothing in makevars.win and with CXX14FLAGS=-O3 -mmmx -msse -msse2 -msse3 -mssse3 -msse4.1 -msse4.2 but I continue to get this error.

Models compile fine and I can run the model but after running about 3 - 5 Stan models I’ll get the C stack error that crashes R. Normally this happens at the end of sampling but sometimes it just seems to happen randomly. This seems to be a problem with Rtools and not with Stan per se but is compounded by the fact that I need C++ to run rstan.

Can you try running them a few times and running gc() in between?

Still getting it

This successfully crashes R with the C stack error every time

library(rstan)
options(mc.cores = parallel::detectCores())
N <- 1000
mu <- c(-0.75, 0.75);
sigma <- c(1, 1);
lambda <- 0.4
z <- rbinom(N, 1, lambda) + 1;
y <- rnorm(N, mu[z], sigma[z]);


stan_rdump(c("N", "y"), file="mix.data.R")

input_data <- read_rdump("mix.data.R")
singular_fit <- stan(file='..\\mix_betancourt_fun.stan', data=input_data[1:2],
                     chains = 3, iter = 600, warmup = 200, 
                     control = list(adapt_delta = 0.9),
                     seed = 483892929)
gc()
singular_fit

singular_fit2 <- stan(file='..\\mix_betancourt.stan', data=input_data[1:2],
                     chains = 3, iter = 600, warmup = 200, 
                     control = list(adapt_delta = 0.9),
                     seed = 483892929)
gc()
singular_fit2 

singular_fit <- stan(file='..\\mix_betancourt_case.stan', data=input_data[1:2],
                     chains = 3, iter = 600, warmup = 200, 
                     control = list(adapt_delta = 0.9),
                     seed = 483892929)
gc()

mix_betancourt_case.stan (389 Bytes) mix_betancourt_fun.stan (677 Bytes) mix_betancourt.stan (545 Bytes)

This looks to be the same as an issue currently open on the RStan github: https://github.com/stan-dev/rstan/issues/844

1 Like

To be clear, I get this crash without calling gc() too.

So the following crashes for you?

library(rstan)
options(mc.cores = parallel::detectCores())
N <- 1000
mu <- c(-0.75, 0.75);
sigma <- c(1, 1);
lambda <- 0.4
z <- rbinom(N, 1, lambda) + 1;
y <- rnorm(N, mu[z], sigma[z]);


stan_rdump(c("N", "y"), file="mix.data.R")

input_data <- read_rdump("mix.data.R")
singular_fit <- stan(file='mix_betancourt_fun.stan', data=input_data[1:2],
                     chains = 3, iter = 600, warmup = 200, 
                     control = list(adapt_delta = 0.9),
                     seed = 483892929)

singular_fit

singular_fit2 <- stan(file='mix_betancourt.stan', data=input_data[1:2],
                      chains = 3, iter = 600, warmup = 200, 
                      control = list(adapt_delta = 0.9),
                      seed = 483892929)

singular_fit2 

singular_fit <- stan(file='mix_betancourt_case.stan', data=input_data[1:2],
                     chains = 3, iter = 600, warmup = 200, 
                     control = list(adapt_delta = 0.9),
                     seed = 483892929)

This successfully crashes R. I just need to trigger recompilation by saving a new empty line in the stan file.

library(rstan)
options(mc.cores = parallel::detectCores())
N <- 1000
mu <- c(-0.75, 0.75);
sigma <- c(1, 1);
lambda <- 0.4
z <- rbinom(N, 1, lambda) + 1;
y <- rnorm(N, mu[z], sigma[z]);


stan_rdump(c("N", "y"), file="mix.data.R")
input_data <- read_rdump("mix.data.R")

file_name <- '..\\mix_betancourt_fun.stan'
for (i in 1:10){
  print(i)
  sink(file_name, append = T)
  writeLines("\n")
  sink()
  singular_fit <- stan(file = file_name, data=input_data[1:2],
                       chains = 3, iter = 200, warmup = 100,
                       control = list(adapt_delta = 0.9),
                       seed = 483892929)
  # print(singular_fit)
}

Great, that narrows it down. So the crash only occurs when the model is re-compiled (or when a new model is compiled). The last thing to check is whether the crash occurs when the result of each model is assigned to a different output (rather than overwriting the singular_fit object every time).

Can you test whether the following crashes for you?

library(rstan)
options(mc.cores = parallel::detectCores())
N <- 1000
mu <- c(-0.75, 0.75);
sigma <- c(1, 1);
lambda <- 0.4
z <- rbinom(N, 1, lambda) + 1;
y <- rnorm(N, mu[z], sigma[z]);


stan_rdump(c("N", "y"), file="mix.data.R")
input_data <- read_rdump("mix.data.R")

file_name <- '..\\mix_betancourt_fun.stan'
res = list()
for (i in 1:10){
    print(i)
    sink(file_name, append = T)
    writeLines("\n")
    sink()
    res[[i]] <- stan(file = file_name, data=input_data[1:2],
                         chains = 3, iter = 200, warmup = 100,
                         control = list(adapt_delta = 0.9),
                         seed = 483892929)
    # print(singular_fit)
}
2 Likes

Great, that does not crash my R session.

2 Likes

Great! It looks like the crashing is the same as the previously-linked RStan github issue. So it looks like repeatedly assigning new RStan models to the same output/object causes a crash. Good to have confirmation!

3 Likes

Hi!

Have you managed to find a solution to this issue? I have similar issue with predicting fitted RStan model and using loo for the model and I am wondering if this has been solved in some version of RStan.

1 Like

Unfortunately not sorry, the only available workaround at the moment is to assign the results of each call to stan to a different output.

In other words, instead of:

fit <- stan(model1, data=data)

fit <- stan(model2, data=data)

Use:

fit1 <- stan(model1, data=data)

fit2 <- stan(model2, data=data)
2 Likes

Not yet in RStan like @andrjohns said but you could try out the new CmdStanR interface.

1 Like

C stack error occurs when using rstan in a for-loop or functions of purrr

I have the very same issue as the thread starter had. In my case, I’m using brms::brm() with rstan backend in a for-loop (actually purrr::walk2). Moreover, although @andrjohns suggests assigning a unique name for different models is helpful, my case may be incurable by that method.

MWE

What I intend to do

I computed a lot of models using brm() with cmdstanr backend with different prior settings for a parameter. Now, I want to compute the log marginal likelihood of each model. To obtain the likelihood, I need to compile brm() with rstan backend first, as discussed on here. But compiling brm() with rstan backend multiple times cause the C stack error, even if (1) rstan appears in a loop environment and (2) every rstan model has no name.

Attach packages

library(lme4) ## For the dataset called sleepstudy
library(tidyverse)
library(magrittr)
library(brms)
library(cmdstanr)
library(rstan)
## Parallelize the chains using all the cores:
options(mc.cores = parallel::detectCores())
## Save compiled models:
rstan_options(auto_write = TRUE)

Set priors

I provided here 9 different combinations of the prior mean and sd for a explanatory variable called Days.

priors <- data.frame(
  prior_mean = prior_mean <- c(-5, 0, 5) %>% 
    rep(., each = length(c(0.01, 0.1, 1))),
  prior_sd = prior_sd <- c(0.01, 0.1, 1) %>% 
    rep(., length(c(-5, 0, 5))),
  id = case_when(
    prior_mean == 0 ~ paste0(prior_mean, "_", prior_sd),
    prior_mean < 0 ~ paste0("m", abs(prior_mean), "_", prior_sd),
    prior_mean > 0 ~ paste0("p", prior_mean, "_", prior_sd)
  )
)

Sample the model with various parameter settings

Since I want to compute the models with cmdstanr using 9 sets of priors, I made an original function called computation, which takes the prior mean and sd and compose the brm(). Then I compute the 9 models in purrr::walk2, which is one of dplyr's loop function.

computation = function(prior_mean, prior_sd) {
  
  file_name = case_when(
    prior_mean == 0 ~ paste0("D:/test/c/c_", prior_mean, "_", prior_sd, "_fit.Rds"),
    prior_mean < 0 ~  paste0("D:/test/c/c_m", abs(prior_mean), "_", prior_sd, "_fit.Rds"),
    prior_mean > 0 ~  paste0("D:/test/c/c_p", prior_mean, "_", prior_sd, "_fit.Rds")
  )
  # pass formula, prior, data, options, etc.
  brms_object = brm(
        Reaction ~ 0 + Intercept + Days + (0 + Intercept + Days|Subject), 
        data = sleepstudy,
        family = "normal",
        prior =
          c(
            prior(normal(0, 1), class = b, coef = Intercept),
            set_prior(
              paste0(
                "normal(",
                prior_mean,
                ", ",
                prior_sd,
                ")"
              ),
              class = "b",
              coef = "Days"
            ),
            prior(normal(0, 1), class = sd),
            prior(lkj(2),       class = cor)    
          ),
        warmup = 2000,
        iter = 10000,
        chains = 4,
        control = list(adapt_delta = 0.95),
        save_pars = save_pars(all = TRUE),
        backend = "cmdstanr"
  )$fit
  
  saveRDS(brms_object, file_name) 
}

walk2(
  priors$prior_mean,
  priors$prior_sd,
  computation  
)

So far, I had no problem to sample the models with cmdstanr backend and to save the results as .Rds file.

Compute log marginal likelihood with various prior settings

By creating a original function ml, I set the brms models with rstan backend as iter = 0 (and chain = 0) and pass them bridgesampling::bridge_sampler(..., stanfit_model = .) to calculate log marginal likelihood in different prior parameter settings. However, this flow hangs in a loop environment, i.e. walk2(), at third or fourth iteration. Note that I do not assign any object name for the stanfits.

ml = function(prior_mean, prior_sd) {
  
  brm(
    Reaction ~ 0 + Intercept + Days + (0 + Intercept + Days|Subject), 
    data = sleepstudy,
    family = "normal",
    prior =
      c(
        prior(normal(0, 1), class = b, coef = Intercept),
        set_prior(
          paste0(
            "normal(",
            prior_mean,
            ", ",
            prior_sd,
            ")"
          ),
          class = "b",
          coef = "Days"
        ),
        prior(normal(0, 1), class = sd),
        prior(lkj(2),       class = cor)    
      ),
    iter = 0,
    chains = 0,
    control = list(adapt_delta = 0.95),
    save_pars = save_pars(all = TRUE),
    backend = "rstan"
  )$fit %>% 
    bridgesampling::bridge_sampler(
      samples = case_when(
        prior_mean == 0 ~ paste0("D:/test/c/c_", prior_mean, "_", prior_sd, "_fit.Rds"),
        prior_mean < 0 ~  paste0("D:/test/c/c_m", abs(prior_mean), "_", prior_sd, "_fit.Rds"),
        prior_mean > 0 ~  paste0("D:/test/c/c_p", prior_mean, "_", prior_sd, "_fit.Rds")
      ) %>% readRDS(), 
      stanfit_model = .  
    ) %>% 
    saveRDS(
      .,
      case_when(
        prior_mean == 0 ~ paste0("D:/test/ml/ml_r1_", prior_mean, "_", prior_sd, ".Rds"),
        prior_mean < 0 ~  paste0("D:/test/ml/ml_r1_m", abs(prior_mean), "_", prior_sd, ".Rds"),
        prior_mean > 0 ~  paste0("D:/test/ml/ml_r1_p", prior_mean, "_", prior_sd, ".Rds")
      )
    )
}

purrr::walk2(
  priors$prior_mean, 
  priors$prior_sd, 
  ml
) ##### The loop hangs at fourth iteration

Session Info

R version 4.0.2 (2020-06-22)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 19042)
rstan_2.21.2
StanHeaders_2.21.0-7
cmdstanr_0.3.0.9000
brms_2.14.4

I struggle with the very same issue as yours. I wanna if you have solved it. Appreciated a lot.