Dimension mismatch when using Pathfinder object as initial values

[mhollanders]

Hi,

The following Stan program declares mu as a matrix with 1 column.

data {
  int N;
  vector[N] y;
}

parameters {
  matrix[N, 1] mu;
  real<lower=0> sigma;
}

model {
  target += normal_lupdf(y | mu[:, 1] , sigma);
}

After running Pathfinder and using this object as initial values for HMC, the following error is reported:

mod <- cmdstan_model("test.stan")
data <- list(N = 100, y = rnorm(100))
pf <- mod$pathfinder(data = data)
fit <- mod$sample(data = data, init = pf)


Chain 1 Unrecoverable error evaluating the log probability at the initial value.
Chain 1 Exception: mismatch in number dimensions declared and found in context; processing stage=parameter initialization; variable name=mu; dims declared=(100,1); dims found=(100)

So it seems like Stan doesn't recognise that mu should be a matrix.

Thanks,

Matt

[avehtari]

Probably the usual problem due to difference between R and Stan data types, and R tending to prefer vector instead Nx1 matrix, but can be fixed