Getting more details on the error message

Lei_Hua · February 9, 2021, 7:04pm

I have encountered the ‘Error in sampler$call_sampler(args_list[[i]]) : Initialization failed’. But not sure how to figure out exactly what went wrong. How can I get more details on the error? Any help is appreciated!

data {
  int<lower=0> J;
  real<lower=0> t[J];
  
  int<lower=0> n;
  real adt[n];
  real evt[n];
  
  real alpha0;
  real beta0;
}
transformed data{
  real<lower=0> delta_t[J];
  delta_t[1] = t[1];
  for (j in 2:J){
    delta_t[j]=t[j]-t[j-1];
  }
}
parameters {
  vector[J] lambda;
}
model {
  for (j in 1:J) lambda[j] ~ gamma(alpha0, beta0); 
  for (i in 1:n){
    target += -lambda[1]*delta_t[1]*(adt[i] >= t[1]) +
          log1m_exp(-lambda[1]*delta_t[1])*(adt[i]<=t[1] && evt[i]==1);
  }
  for (j in 2:J){
    for (i in 1:n){
      target += -lambda[j]*delta_t[j]*(adt[i] >= t[j]) +
            log1m_exp(-lambda[j]*delta_t[j])*(t[j-1]<adt[i] && adt[i]<=t[j] && evt[i]==1);
      }
    }
}

R code:

mydata_mod ← list(
J = J,
t = myt,
n = n,
adt = adt,
evt = evt,
alpha0 = 1/1000,
beta0 = 48/log(2)/1000
)

fit ← sampling(sm,
data = mydata_mod, # named list of data
# seed=1234,
chains = 1, # number of Markov chains
warmup = 1000, # number of warmup iterations per chain
iter = 2500, # total number of iterations per chain
cores = 2, # number of cores (could use one per chain)
refresh = 0 # no progress shown
)

bbbales2 · February 11, 2021, 12:57pm

I was able to run this with some made up data:

library(cmdstanr)

mod = cmdstan_model("test5.stan")

fit = mod$sample(data = list(J = 2,
                             t = c(1,2),
                             n = 2,
                             adt = c(1, 2),
                             evt = c(1, 2),
                             alpha0 = 1,
                             beta0 = 2))

I suspect there is a problem with one of the variables J, myt, n, adt, or evt. They should match the types defined in the Stan program. J and n should be integers, and myt, adt, and evt should be vectors (lists might work too but I’m not sure).

Lei_Hua · February 11, 2021, 3:23pm

Thanks! You are right. The issue is the type of variables and parameters. I tried to update the lambda to real lambda[J], it seems solve the problem. I am always confused when to the difference between the two below. Any insight?

vector[J] x
real x[J]

bbbales2 · February 11, 2021, 3:41pm

As far as input goes, I don’t think there is. Are you seeing a difference?

Using them in the code, a vector should be thought of as one object (like the think you do vector math on and such). An array of reals is an array of little objects – it is not a mathematical vector. You can convert between the two with to_array_1d and to_vector.

Lei_Hua · February 11, 2021, 3:48pm

For my case here, I had to change vecotr[J] lambda to real lambda[J]. Using the former, I have the error messages, while after i change, the code run without problems.

bbbales2 · February 11, 2021, 7:26pm

Oh that’s surprising. This sounds like something that we should fix.

Do you know what type the input argument was?

Lei_Hua · February 11, 2021, 7:38pm

Here is the stan code

data {
  int<lower=0> J;
  real<lower=0> t[J];
  
  int<lower=0> n;
  real adt[n];
  real evt[n];
  
  real alpha0;
  real beta0;
}
transformed data{
  real<lower=0> delta_t[J];
  delta_t[1] = t[1];
  for (j in 2:J){
    delta_t[j]=t[j]-t[j-1];
  }
}
parameters {
  // for piecewise hazard;
  real<lower=0> lambda[J];
}
model {
  for (j in 1:J){
    lambda[j] ~ gamma(alpha0, beta0);
    if (j==1){
      for (i in 1:n){
        target += -lambda[1]*delta_t[1]*(adt[i] >= t[1] && !(adt[i]<=t[1] && evt[i]==1)) +
          log1m_exp(-lambda[1]*delta_t[1])*(adt[i]<=t[1] && evt[i]==1);
          }
    }else{
      // lambda[j] ~ gamma(alpha0, beta0);
      for (i in 1:n){
        target += -lambda[j]*delta_t[j]*(adt[i] >= t[j] && !(t[j-1]<adt[i] && adt[i]<=t[j] && evt[i]==1)) +
              log1m_exp(-lambda[j]*delta_t[j])*(t[j-1]<adt[i] && adt[i]<=t[j] && evt[i]==1);
        }
      }
  }
}

when I change the parameter to

vector[J] lambda;

I have issues. Not sure if it matters, the data used to fit the model have J smaller than n.

bbbales2 · February 11, 2021, 7:39pm

Oh sorry, I meant the r-code that calls the Stan model. What type are the things you pass in to the data argument? I think this would be a problem with how IO is handled at the interface.

Lei_Hua · February 11, 2021, 7:53pm

I kept updating my model, but I think below is the one I used to call this:

mydata_mod <- list(
  J = J,
  t = myt,
  n = n,
  adt = adt,
  evt = evt
)

fit <- sampling(sm,
                data = mydata_mod,      # named list of data
                # seed=1234,
                chains = 4,             # number of Markov chains
                warmup = 1000,          # number of warmup iterations per chain
                iter = 3500,            # total number of iterations per chain
                cores = 2,              # number of cores (could use one per chain)
                refresh = 0             # no progress shown
)

bbbales2 · February 12, 2021, 1:55pm

Oh oh oh sorry, yes lambda is a parameter my apologies, I thought it was data.

Looking at this model more closely, I’m worried there is a segfault or something weird happening here.

I do not think there should be a difference in using a vector or an array in the code you posted.

Do you mind posting the data you used to run the model? That way I can look at this more closely with the two different things and see if I can reproduce the error on my computer.

You can dump your data to a Stan-friendly format using:

stan_rdump(names(mydata_mod), "mydata.dat", env = list2env(mydata_mod))

Lei_Hua · February 17, 2021, 10:06pm

Here is the data. Part of the reason could also be my J is the number of unique values in adt. Once I reduce it, whether I use vector[J] lambda or real lambda[J], doesn’t matter any more. Looking for more insights in this (480 Bytes)

betanalpha · February 17, 2021, 11:24pm

Try running again with cores=1 – that will turn off parallelization and should let you see the full error message from Stan.

bbbales2 · February 19, 2021, 1:56pm

The attachment isn’t working for me (it’s showing up like the link got corrupted somehow). Can you try again?

Lei_Hua · February 19, 2021, 3:09pm

Copied below directly

J <- 12
t <- 
c(3, 7, 8, 11, 17, 18, 25, 27, 32, 42, 56, 57)
n <- 30
b <- 1
cov <- 
structure(c(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1),
.Dim = c(30, 1))
adt <- 
c(57, 25, 8, 25, 57, 7, 18, 18, 57, 57, 11, 42, 57, 57, 8, 17, 27, 57, 3, 57, 57, 57, 57, 56, 57, 57, 57, 32, 57, 57)
evt <- 
c(0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 1, 0, 0)
alpha0 <- 0.01
beta0 <- 0.692493619626703
B0 <- 5

bbbales2 · February 19, 2021, 10:58pm

Thanks for the data! I just tried this out.

I think the vector is missing a lower bound.

So:

real<lower=0> lambda[J];

should be:

vector<lower=0>[J] lambda;

The problem is just that the constraints on the parameters should match the support of the distribution.

So with this line:

for (j in 1:J) lambda[j] ~ gamma(alpha0, beta0);

You need to make sure and constrain lambda to be positive to keep the sampler from moving it into an invalid range.

Topic		Replies	Views
Initialization failed when fitting a relatively simple RL model Modeling rstan , fitting-issues	2	455	May 13, 2022
Error: Chain 1: Rejecting initial value: mu[1] is nan, but must be greater than or equal to 0 Modeling	28	2394	July 1, 2020
Error evaluating the log probability at the initial value in Ordered Logistic Model Modeling fitting-issues	2	1029	July 31, 2019
Initialization error Modeling	10	3159	January 22, 2019
RuntimeError : Initialization failed Modeling	5	496	September 2, 2022

Getting more details on the error message

Related Topics