Setting initial values causing segmentation fault

Interfaces CmdStan
1

I am running a HMC sampling model to sample over a number of parameters based on a normal distribution with a predicted mean. The main prediction is done using the disp_eq_root_par function (which works fine and has been tested with known values to return correct results).

The model runs fine without any initial values, however, as soon as I set some I immediately return a segmentation fault, after the gradient evaluation has taken place.

The initial values are sensible values for the regression model given the data points as they are material properties and so are estimated to be of the correct magnitude.

The code is written in such a way as to be able to parallelise using map_rect.

data {
  int<lower=0> N;
  real<lower=0> x[N];
  real<lower=0> y[N];
}
transformed data{
    // non-parallel
    int xi[0];
    vector[0] theta;
}
parameters {
  real<lower=0> log_C[4];
  real<lower=0> log_rho;
  real<lower=0> log_sigma;
}
transformed parameters {
   real<lower=0> C[4];
   real<lower=0> rho;
   real<lower=0> sigma;
   for (i in 1:4) { C[i] = exp(log_C[i]); } 
   rho = exp(log_rho);
   sigma = exp(log_sigma);
}
model {

    // Set parameters
    real h = 1e-3;

    // MCMC model

    vector[N] ypred;

    print(C);
    print(rho);
    print(sigma);

    real a0_C = 4.9;
    real b0_C = 0.07;
    C[1] ~ gamma(a0_C, b0_C);
    C[2] ~ gamma(a0_C, b0_C);
    C[3] ~ gamma(a0_C, b0_C);
    C[4] ~ gamma(12.5,2.5);

    rho ~ normal(1600.0, 300.0);
    sigma ~ gamma(2.0, 2.5e-5);
    
    vector[6] phi = [C[1],C[2],C[3],C[4],rho,h]';
    ypred = disp_eq_root_par(phi, theta, x, xi);
    y ~ normal(ypred, sigma);

}

From the cmd folder, the code is compiled with:
“”"
make ~/dispersion_stan/dispersion USER_HEADER=~/dispersion_stan/dispersion_eigs.hpp
“”"

And then from the ~dispersion_stan/ folder, the compiled binary is ran with:
“”"
./dispersion sample num_samples=50 num_warmup=20 init=init_vals_log.json data file=dispersion.data_0deg.json output file=dispersion_output.csv
“”"
Where `init=init_vals_log.json’ is or isn’t included.

Here is the output from the terminal when initial values are included:
“”"
method = sample (Default)
sample
num_samples = 50
num_warmup = 20
save_warmup = 0 (Default)
thin = 1 (Default)
adapt
engaged = 1 (Default)
gamma = 0.050000000000000003 (Default)
delta = 0.80000000000000004 (Default)
kappa = 0.75 (Default)
t0 = 10 (Default)
init_buffer = 75 (Default)
term_buffer = 50 (Default)
window = 25 (Default)
algorithm = hmc (Default)
hmc
engine = nuts (Default)
nuts
max_depth = 10 (Default)
metric = diag_e (Default)
metric_file = (Default)
stepsize = 1 (Default)
stepsize_jitter = 0 (Default)
id = 0 (Default)
data
file = dispersion.data_0deg.json
init = init_vals_log.json
random
seed = 598831642 (Default)
output
file = dispersion_output.csv
diagnostic_file = (Default)
refresh = 100 (Default)
sig_figs = -1 (Default)
profile_file = profile.csv (Default)

Gradient evaluation took 0.088989 seconds
1000 transitions using 10 leapfrog steps per transition would take 889.89 seconds.
Adjust your expectations accordingly!

WARNING: There aren’t enough warmup iterations to fit the
three stages of adaptation as currently configured.
Reducing each adaptation stage to 15%/75%/10% of
the given number of warmup iterations:
init_buffer = 3
adapt_window = 15
term_buffer = 2

Segmentation fault
“”"

If possible, add also code to simulate data or attach a (subset of) the dataset you work with.

Any help on this would be much appreciated!

  • Operating System: WSL Ubuntu 20.04
  • CmdStan Version: 2.27
  • Compiler/Toolkit: g++ and gnu make
2

Can you post the contents of your json containing the inits?

3

The contents of the init_vals_log.json are:

{
    "log_C" : [3.91202, 3.91202, 3.91202, 1.60944],
    "log_rho" : 7.09008,
    "log_sigma" : 9.21034
}

Some additional info: I thought the segmentation fault might be a RAM issue but the Ubuntu WSL has 10GB available and this is only for 50 data points for predicting. So I would think it wouldn’t be a RAM issue but could be wrong.

1 Like
4

Not necessarily related to the issues you’re having, but I just wanted to ask: what is going on here? I personally haven’t seen 0-dimensional structures like this before and am curious as to what they are for.

1 Like
5

I’ve written the function `disp_eq_root_par’ in order to conform with the signature that is required for using map_rect for parallelization purposes. For this the signature of the function must follow:

vector f(vector phi, vector theta, data real[] x_r, data int[] x_i);

where x_r is the real inputs for all data points and x_i is the integer inputs for each data points, phi is constant for each data point, and theta is a vector of constants the same length as x_r/x_i. However, as in my function any inputs are constant across the data and there are no integer inputs, I just create some empty vectors as they must be passed into the mapped function anyhow. Hope that makes sense! I followed the documentation and they did a similar thing for the examples in there.

1 Like
6

Today I learned – thanks!