syclik
commented
9 years ago From @ksvanhorn on March 19, 2015 18:9
Here are the stan models I'm using:
data {
int<lower=0> K; // # of product covariates; no intercept
int<lower=0> G; // # of respondent covariates
int<lower=0> S; // # of scenarios per respondent
int<lower=0> C; // # of alternatives (choices) per scenario
matrix[C, K] X[S]; // X[s] is covariate matrix of scenario s for respondent.
vector[G] Z; // vector of covariates describing respondent r
int<lower=1,upper=C> Y1[S]; // forced-choice responses
real<lower=0,upper=1> Y2[S]; // whether respondent would choose any of the options
int<lower=0> n_shared;
int<lower=0, upper=1> use_quad_form;
matrix[n_shared, n_shared] S_mat;
vector[n_shared] mu;
}
transformed data {
int n_preTheta;
int<lower=0, upper=0> n_shared_check;
n_preTheta <- (K + 1) * G;
n_shared_check <- n_shared - (1 + (K + 1) + n_preTheta);
}
parameters {
real logit_lambda;
vector[K+1] log_sigma;
matrix[K+1, G] preTheta;
vector[K+1] Epsilon;
}
transformed parameters {
}
model {
real lambda;
vector[K+1] sigma;
lambda <- inv_logit(logit_lambda);
for (i in 1:(K+1))
sigma[i] <- exp(log_sigma[i]);
Epsilon ~ normal(0, 1);
{
vector[S] ll;
vector[K+1] Beta;
Beta <- 10.0* preTheta * Z + (sigma .* Epsilon);
{
vector[K] b;
real alpha2;
b <- segment(Beta, 2, K);
alpha2 <- Beta[1];
for (s in 1:S) {
vector[C] u;
real u_buy;
u <- X[s] * b;
u_buy <- alpha2 + lambda * log_sum_exp(u);
ll[s] <-
categorical_logit_log(Y1[s], u) +
Y2[s] * binomial_logit_log(1, 1, u_buy) +
(1 - Y2[s]) * binomial_logit_log(0, 1, u_buy);
}
}
increment_log_prob(ll);
}
if (use_quad_form) {
vector[n_preTheta] vec_preTheta;
vector[n_shared] shparms;
int idx;
vec_preTheta <- to_vector(preTheta);
idx <- 1;
shparms[idx] <- logit_lambda; idx <- idx + 1;
for (i in 1:(K+1)) {
shparms[idx] <- log_sigma[i]; idx <- idx + 1;
}
for (i in 1:n_preTheta) {
shparms[idx] <- vec_preTheta[i]; idx <- idx + 1;
}
increment_log_prob(-0.5 * quad_form(S_mat, shparms - mu));
}
}
generated quantities {
}and this:
data {
int<lower=0> K; // # of product covariates; no intercept
int<lower=0> G; // # of respondent covariates
int<lower=0> n_shared;
int<lower=0, upper=1> use_quad_form;
matrix[n_shared, n_shared] S_mat;
vector[n_shared] mu;
}
transformed data {
int n_preTheta;
int<lower=0, upper=0> n_shared_check;
n_preTheta <- (K + 1) * G;
n_shared_check <- n_shared - (1 + (K + 1) + n_preTheta);
}
parameters {
real logit_lambda;
vector[K+1] log_sigma;
matrix[K+1, G] preTheta;
}
transformed parameters {
}
model {
real lambda;
vector[K+1] sigma;
lambda <- inv_logit(logit_lambda);
//lambda ~ uniform(0, 1);
increment_log_prob(-2 * log1p_exp(-logit_lambda) - logit_lambda);
for (i in 1:(K+1))
sigma[i] <- exp(log_sigma[i]);
sigma ~ normal(0, 5.0);
increment_log_prob(log_sigma);
{
vector[n_preTheta] vec_preTheta;
vec_preTheta <- to_vector(preTheta);
vec_preTheta ~ normal(0, 1);
if (use_quad_form) {
vector[n_shared] shparms;
int idx;
idx <- 1;
shparms[idx] <- logit_lambda; idx <- idx + 1;
for (i in 1:(K+1)) {
shparms[idx] <- log_sigma[i]; idx <- idx + 1;
}
for (i in 1:n_preTheta) {
shparms[idx] <- vec_preTheta[i]; idx <- idx + 1;
}
increment_log_prob(-0.5 * quad_form(S_mat, shparms - mu));
}
}
}
generated quantities {
}[Arrg! I hate that github won't let you attach text files to an issue!]
I'm using the models in two ways: 1) doing optimization, with use_quad_form = TRUE, and 2) computing a Hessian at a particular point by calling sampling() with use_quad_form = FALSE in the data, and test_grad=TRUE.
bgoodri
rBatt
From @ksvanhorn on March 19, 2015 18:0
I have been seeing some strange behavior from rstan that I have not been able to boil down to a simple case, as it's a "Heisenbug" -- problems come and go with seemingly inconsequential changes to my program, including adding print statements or running the debugger. Nor do the problems seem to be entirely repeatable.
Here is one error message I got from stan_model (the traceback indicates that it was calling stanc):
Another common symptom is to call a function
where the.function() includes calls to various rstan functions (stan_model and sampling), whereupon the call to the.function() fails at a point that appears to be a call to an rstan function, but no error message is ever output. The only reason I know the function fails is that nothing gets assigned to lpres, and expected console output is never produced.
On another occasion an rstan call caused R itself to outright crash.
I've encountered these problems with both rstan 2.5 and rstan 2.6, and on two different machines.
I will add more information as I continue investigating this.
Copied from original issue: stan-dev/stan#1400