Trouble adding a function of a vector with known gradients

Modeling
1

I’m trying to add a function demo to Stan as a learning exercise. I’m struggling with adding a scalar function of a vector. I figure I’m missing something pretty fundamental, so any advice would be appreciated.

New function:

namespace foo {
    double foo(std::vector<double>& x) {
        return x[0] + x[1];
    }

    std::vector<double> grad_foo(std::vector<double>& x) {
        std::vector<double> v;
        v[0] = 1;
        v[1] = 1;
        return v;
    }
}

namespace stan {
    namespace math {

        double demo(std::vector<double>& x) {
            return foo::foo(x);
        }

        var demo(std::vector<var>& x) {
            std::vector<double> a = value_of(x); // x.val();
            double fa = foo::foo(a);
            std::vector<double> grad_fa = foo::grad_foo(a);
            return precomputed_gradients(fa, x, grad_fa);
        }

    }
}

Signature:

add("demo", bare_expr_type(double_type()), bare_expr_type(vector_type()));

Stan file:

data {
  int<lower=0> N;
  real alpha[N];
}
parameters {
  vector[2] alpha_est;
}
transformed parameters {
  real transformed_alpha;
  transformed_alpha = demo(alpha_est);
}
model {
  alpha_est[0] ~ normal(0,10);
  alpha_est[1] ~ normal(0,10);
  for (n in 1:N)
    alpha[n] ~ normal(transformed_alpha, 1);
}

I get the error:

./examples/demo/demo.hpp:289:51: error: no matching function for call to ‘demo’ stan::math::assign(transformed_alpha, demo(alpha_est));

Followed by several more lines of related error messaging. Thanks for any help!

2

Probably trivial but I’ve always seen additional namespaces such as you’re using foo included within stan and math. What is var wouldn’t you typically use return type of auto?

3

I had a misunderstanding of the relationship between std::vector in C++ and the Stan type vector. The Stan type vector maps on to the “Eigen” vector type. std::vector maps onto the Stan real[] type.

The fix was:

Signature:

add("demo", bare_expr_type(double_type()), bare_expr_type(bare_array_type(double_type(), 1)));

Stan file:

data {
  int<lower=0> N;
  real alpha[N];
}
parameters {
  real alpha_est[2];
}
transformed parameters {
  real transformed_alpha;
  transformed_alpha = demo(alpha_est);
}
model {
  alpha_est[1] ~ normal(0,10);
  alpha_est[2] ~ normal(0,10);
  for (n in 1:N)
    alpha[n] ~ normal(transformed_alpha, .1);
}
1 Like