Test_grad: No Output

General
1

I’m trying to compare autodiff and finite differencing results to make sure everything is working okay, but when I run sampling in rstan with test_grad = TRUE and file ‘reparam.stan’, all it says is:

“TESTING GRADIENT FOR MODEL ‘reparam’ NOW (CHAIN 1).”, with nothing after it. I’m a fairly new Stan user, is this a common problem? I’m new to the forum as well, but did my best to search around and couldn’t find anything. I’ve tried running on a Linux cluster and on Windows, so I don’t think it’s an OS issue. I also made sure I have the latest version of rstan. If it’s helpful, here’s my model block:

model {
// Likelihood
Z ~ normal(theta , t);

// Priors
x_tilde ~ normal(0,1);
y_tilde ~ normal(0,1);
mu_tilde ~ normal(0,1);
tau ~ inv_gamma(0.1, 0.1);

for (n in 1:J){
w_tilde[n] ~ normal(0,1);
}
lat ~ gamma(nu/2, nu/2);

// Hyperpriors
v ~ lognormal(-2, 3);
nu ~ lognormal(-2, 1);
sigma ~ inv_gamma(0.1, 0.1);
}

Thanks,
Matt

2

That seems to be all it prints in rstan (just checked for my own model).

Cmdstan gives more verbose output. What was it you were looking for?

3

Oh you already told me I guess. Hmm, I guess if you want to check gradients make sense in different places, you can do grad_log_prob and check against finite differences of log_prob.

I guess that’s what you were hoping test grad would do for you? I agree, seems a little cludgy to do it yourself but usually we just assume the autodiff is correct unless we have a reason to believe otherwise. Which I guess sounds dangerous. Is there something particularly finicky about your model that isn’t shown there? All the stuff there I think it’s fair to assume it works (knock on wood).

4

Yeah, based on the documentation I thought I’d get a table printout with autodiff and finite difference calculations and the difference beteween them. They probably are the same, but I thought it would be a quick check to see. My tree depth needs to be set to 14 and adapt_delta to 0.99, and while I think it’s probably because of other issues in the model, I wanted to check this first.

Thanks for the grad_log_prob/log_prob and the cmdstan suggestions. I’ll check those out.