nan exception with variational inference

[wcbeard]

For the following model

data {
    int N;
    vector[N] p;
    int Ngrps;
    int<lower=1, upper=Ngrps> grp_index[N];
}
parameters {
    vector<lower=0.0001, upper=100>[Ngrps] sigmaGrp;
    vector<lower=-100, upper=1000>[Ngrps] muGrp;
}

model {
    int grpi;
    for (i in 1:N){
        grpi <- grp_index[i];
        p[i] ~ logistic(muGrp[grpi], sigmaGrp[grpi]);
    };
}

I'm trying to estimate group level parameters from the following data (I generated with python and saved data file test.R with pystan):

p = 63.1, 108.3, 1.0, 46.0, 22.9, 14.8, 28.8, 52.5, 60.1, 81.3
grp_index = [1, 1, 1, 1, 1, 1, 2, 3, 4, 5]

test_dat = dict(
    N=len(p),
    p=p,
    grp_index=grp_index,
    Ngrps=len(set(grp_index)),
)    

When I compile and sample with advi using

$ modeldir/model/test variational data file=modeldir/dat/test.R random seed=3021828106

I get the exception:

Exception: stan::variational::normal_meanfield::set_mu: Input vector[7] is nan, but must not be nan!
Diagnostic information: 
Dynamic exception type: std::domain_error
std::exception::what: stan::variational::normal_meanfield::set_mu: Input vector[7] is nan, but must not be nan!

This is a small dataset, and the error can be avoided if I run with most other random seeds, but I originally ran into this every time with a larger data set, that presumably has some groups only represented once in the grp_index array. I could get rid of those groups in the dataset, but wondered if this is a bug with the variational algorithm.

[bob-carpenter]

Thanks for filing the issue. This is a known bug with the way
errors are currently being trapped, which is why it hasn't
been ported to RStan yet --- this would've crashed RStan if
we'd included it.

We should have this fixed (and a version in RStan) by the
next release.

• Bob

On Jul 20, 2015, at 3:53 PM, Chris notifications@github.com wrote:

For the following model

data {
int N;
vector[N] p;
int Ngrps;
int<lower=1, upper=Ngrps> grp_index[N];
}
parameters {
vector<lower=0.0001, upper=100>[Ngrps] sigmaGrp;
vector<lower=-100, upper=1000>[Ngrps] muGrp;
}

model {
int grpi;
for (i in 1:N){
grpi <- grp_index[i];
p[i] ~ logistic(muGrp[grpi], sigmaGrp[grpi]);
};
}

I'm trying to estimate group level parameters from the following data (I generated with python and saved data file test.R with pystan):

p = 63.1, 108.3, 1.0, 46.0, 22.9, 14.8, 28.8, 52.5, 60.1, 81.3
grp_index = [1, 1, 1, 1, 1, 1, 2, 3, 4, 5]

test_dat = dict(
N=len(p),
p=p,
grp_index=grp_index,
Ngrps=len(set(grp_index)),
)

When I compile and sample with advi using

$ modeldir/model/test variational data file=modeldir/dat/test.R random seed=3021828106

I get the exception:

Exception: stan::variational::normal_meanfield::set_mu: Input vector[7] is nan, but must not be nan!
Diagnostic information:
Dynamic exception type: std::domain_error
std::exception::what: stan::variational::normal_meanfield::set_mu: Input vector[7] is nan, but must not be nan!

This is a small dataset, and the error can be avoided if I run with most other random seeds, but I originally ran into this every time with a larger data set, that presumably has some groups only represented once in the grp_index array. I could get rid of those groups in the dataset, but wondered if this is a bug with the variational algorithm.

—
Reply to this email directly or view it on GitHub.

[bgoodri]

Specifically it is #1544 and there is a pull request (that needs some work) for it. So, I think we can close this issue.

[dustinvtran]

Yup, hopefully when I have time later this week, I can repush some changes to the current pull request as it isn't working at the moment.