Stan produces a high amount of inf at init phase

Modeling
1

Due to a problem https://github.com/stan-dev/cmdstan/issues/665 with CMDSTAN 2.18.1 with Ubuntu 16.04 I checked out the development version.
And the initialization phase I am receiving a bunch of errors because of -nan, sometimes nan.
So I added a print statement to my model. I get the following output:

sigma_scale_unscaled=[inf,inf,0,0,inf,inf,0,inf,inf,inf,inf,0,0,inf,inf,0,0,inf,inf,0,0,inf,inf,0,0,inf,0,inf,0,inf,inf,0,inf,0,inf,inf,0,inf,inf,0,0,0,0,inf,0,inf,0,0,0]

My parameter and model section:

  vector<lower=0>[N] sigma_scale_unscaled;
  sigma_scale_unscaled ~ inv_chi_square(5);

The other parameters got a lot of 0’s two.

After some iterations the inf values vanish and become real numbers.
The output of the run looks good.

Is this a normal behavior?

1 Like