TEST(McmcNuts, instantiaton_test) {
rng_t base_rng(4839294);
std::stringstream output;
stan::callbacks::stream_writer writer(output);
std::stringstream error_stream;
stan::callbacks::stream_writer error_writer(error_stream);
std::fstream empty_stream("", std::fstream::in);
stan::io::dump data_var_context(empty_stream);
gauss3D_model_namespace::gauss3D_model model(data_var_context);
stan::mcmc::unit_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
unit_e_sampler(model, base_rng);
stan::mcmc::diag_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
diag_e_sampler(model, base_rng);
stan::mcmc::dense_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
dense_e_sampler(model, base_rng);
stan::mcmc::adapt_unit_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
adapt_unit_e_sampler(model, base_rng);
stan::mcmc::adapt_diag_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
adapt_diag_e_sampler(model, base_rng);
stan::mcmc::adapt_dense_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
adapt_dense_e_sampler(model, base_rng);
}
TEST(McmcNutsBaseNuts, transition) {
rng_t base_rng(0);
int model_size = 1;
double init_momentum = 1.5;
stan::mcmc::ps_point z_init(model_size);
z_init.q(0) = 0;
z_init.p(0) = init_momentum;
stan::mcmc::mock_model model(model_size);
stan::mcmc::mock_nuts sampler(model, base_rng);
sampler.set_nominal_stepsize(1);
sampler.set_stepsize_jitter(0);
sampler.sample_stepsize();
sampler.z() = z_init;
std::stringstream output_stream;
stan::interface_callbacks::writer::stream_writer writer(output_stream);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
stan::mcmc::sample init_sample(z_init.q, 0, 0);
stan::mcmc::sample s = sampler.transition(init_sample, writer, error_writer);
EXPECT_EQ(31.5, s.cont_params()(0));
EXPECT_EQ(0, s.log_prob());
EXPECT_EQ(1, s.accept_stat());
EXPECT_EQ("", output_stream.str());
EXPECT_EQ("", error_stream.str());
}
TEST(McmcNutsBaseNuts, build_tree) {
rng_t base_rng(0);
int model_size = 1;
double init_momentum = 1.5;
stan::mcmc::ps_point z_init(model_size);
z_init.q(0) = 0;
z_init.p(0) = init_momentum;
stan::mcmc::ps_point z_propose(model_size);
Eigen::VectorXd rho = z_init.p;
double log_sum_weight = -std::numeric_limits<double>::infinity();
double H0 = -0.1;
int n_leapfrog = 0;
double sum_metro_prob = 0;
stan::mcmc::mock_model model(model_size);
stan::mcmc::mock_nuts sampler(model, base_rng);
sampler.set_nominal_stepsize(1);
sampler.set_stepsize_jitter(0);
sampler.sample_stepsize();
sampler.z() = z_init;
std::stringstream output;
stan::interface_callbacks::writer::stream_writer writer(output);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
bool valid_subtree = sampler.build_tree(3, rho, z_propose,
H0, 1, n_leapfrog, log_sum_weight,
sum_metro_prob, writer, error_writer);
EXPECT_TRUE(valid_subtree);
EXPECT_EQ(init_momentum * (n_leapfrog + 1), rho(0));
EXPECT_EQ(8 * init_momentum, sampler.z().q(0));
EXPECT_EQ(init_momentum, sampler.z().p(0));
EXPECT_EQ(8, n_leapfrog);
EXPECT_FLOAT_EQ(H0 + std::log(n_leapfrog), log_sum_weight);
EXPECT_FLOAT_EQ(std::exp(H0) * n_leapfrog, sum_metro_prob);
EXPECT_EQ("", output.str());
EXPECT_EQ("", error_stream.str());
}
TEST(McmcHmcIntegratorsImplLeapfrog, unit_e_energy_conservation) {
rng_t base_rng(0);
std::fstream data_stream(std::string("").c_str(), std::fstream::in);
stan::io::dump data_var_context(data_stream);
data_stream.close();
std::stringstream model_output;
std::stringstream metric_output;
stan::interface_callbacks::writer::stream_writer writer(metric_output);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
gauss_model_namespace::gauss_model model(data_var_context, &model_output);
stan::mcmc::impl_leapfrog<
stan::mcmc::unit_e_metric<gauss_model_namespace::gauss_model, rng_t> >
integrator;
stan::mcmc::unit_e_metric<gauss_model_namespace::gauss_model, rng_t> metric(model);
stan::mcmc::unit_e_point z(1);
z.q(0) = 1;
z.p(0) = 1;
metric.init(z, writer, error_writer);
double H0 = metric.H(z);
double aveDeltaH = 0;
double epsilon = 1e-3;
double tau = 6.28318530717959;
size_t L = tau / epsilon;
for (size_t n = 0; n < L; ++n) {
integrator.evolve(z, metric, epsilon, writer, error_writer);
double deltaH = metric.H(z) - H0;
aveDeltaH += (deltaH - aveDeltaH) / double(n + 1);
}
// Average error in Hamiltonian should be O(epsilon^{2})
// in general, smaller for the gauss_model in this case due to cancellations
EXPECT_NEAR(aveDeltaH, 0, epsilon * epsilon);
EXPECT_EQ("", model_output.str());
EXPECT_EQ("", metric_output.str());
}
TEST(McmcUnitENuts, transition_test) {
rng_t base_rng(4839294);
stan::mcmc::unit_e_point z_init(3);
z_init.q(0) = 1;
z_init.q(1) = -1;
z_init.q(2) = 1;
z_init.p(0) = -1;
z_init.p(1) = 1;
z_init.p(2) = -1;
std::stringstream output_stream;
stan::interface_callbacks::writer::stream_writer writer(output_stream);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
std::fstream empty_stream("", std::fstream::in);
stan::io::dump data_var_context(empty_stream);
gauss3D_model_namespace::gauss3D_model model(data_var_context);
stan::mcmc::unit_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
sampler(model, base_rng);
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.set_nominal_stepsize(0.1);
sampler.set_stepsize_jitter(0);
sampler.sample_stepsize();
stan::mcmc::sample init_sample(z_init.q, 0, 0);
stan::mcmc::sample s = sampler.transition(init_sample, writer, error_writer);
EXPECT_EQ(4, sampler.depth_);
EXPECT_EQ((2 << 3) - 1, sampler.n_leapfrog_);
EXPECT_FALSE(sampler.divergent_);
EXPECT_FLOAT_EQ(1.8718261, s.cont_params()(0));
EXPECT_FLOAT_EQ(-0.74208695, s.cont_params()(1));
EXPECT_FLOAT_EQ( 1.5202962, s.cont_params()(2));
EXPECT_FLOAT_EQ(-3.1828632, s.log_prob());
EXPECT_FLOAT_EQ(0.99629009, s.accept_stat());
EXPECT_EQ("", output_stream.str());
EXPECT_EQ("", error_stream.str());
}
TEST(McmcUnitENuts, tree_boundary_test) {
rng_t base_rng(4839294);
stan::mcmc::unit_e_point z_init(3);
z_init.q(0) = 1;
z_init.q(1) = -1;
z_init.q(2) = 1;
z_init.p(0) = -1;
z_init.p(1) = 1;
z_init.p(2) = -1;
std::stringstream output;
stan::interface_callbacks::writer::stream_writer writer(output);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
std::fstream empty_stream("", std::fstream::in);
stan::io::dump data_var_context(empty_stream);
typedef gauss3D_model_namespace::gauss3D_model model_t;
model_t model(data_var_context);
// Compute expected tree boundaries
typedef stan::mcmc::unit_e_metric<model_t, rng_t> metric_t;
metric_t metric(model);
stan::mcmc::expl_leapfrog<metric_t> unit_e_integrator;
double epsilon = 0.1;
stan::mcmc::unit_e_point z_test = z_init;
metric.init(z_test, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_forward_1 = metric.dtau_dp(z_test);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_forward_2 = metric.dtau_dp(z_test);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_forward_3 = metric.dtau_dp(z_test);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_forward_4 = metric.dtau_dp(z_test);
z_test = z_init;
metric.init(z_test, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_backward_1 = metric.dtau_dp(z_test);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_backward_2 = metric.dtau_dp(z_test);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_backward_3 = metric.dtau_dp(z_test);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
unit_e_integrator.evolve(z_test, metric, -epsilon, writer, error_writer);
Eigen::VectorXd p_sharp_backward_4 = metric.dtau_dp(z_test);
// Check expected tree boundaries to those dynamically geneated by NUTS
stan::mcmc::unit_e_nuts<model_t, rng_t> sampler(model, base_rng);
sampler.set_nominal_stepsize(epsilon);
sampler.set_stepsize_jitter(0);
sampler.sample_stepsize();
stan::mcmc::ps_point z_propose = z_init;
Eigen::VectorXd p_sharp_left = Eigen::VectorXd::Zero(z_init.p.size());
Eigen::VectorXd p_sharp_right = Eigen::VectorXd::Zero(z_init.p.size());
Eigen::VectorXd rho = z_init.p;
double log_sum_weight = -std::numeric_limits<double>::infinity();
double H0 = -0.1;
int n_leapfrog = 0;
double sum_metro_prob = 0;
// Depth 0 forward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(0, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, 1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_1(n), p_sharp_right(n));
// Depth 1 forward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(1, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, 1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_2(n), p_sharp_right(n));
// Depth 2 forward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(2, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, 1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_3(n), p_sharp_right(n));
// Depth 3 forward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(3, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, 1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_forward_4(n), p_sharp_right(n));
// Depth 0 backward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(0, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, -1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_1(n), p_sharp_right(n));
// Depth 1 backward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(1, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, -1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_2(n), p_sharp_right(n));
// Depth 2 backward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(2, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, -1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_3(n), p_sharp_right(n));
// Depth 3 backward
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.build_tree(3, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, -1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_1(n), p_sharp_left(n));
for (int n = 0; n < rho.size(); ++n)
EXPECT_FLOAT_EQ(p_sharp_backward_4(n), p_sharp_right(n));
}
TEST(McmcUnitENuts, build_tree_test) {
rng_t base_rng(4839294);
stan::mcmc::unit_e_point z_init(3);
z_init.q(0) = 1;
z_init.q(1) = -1;
z_init.q(2) = 1;
z_init.p(0) = -1;
z_init.p(1) = 1;
z_init.p(2) = -1;
std::stringstream output;
stan::interface_callbacks::writer::stream_writer writer(output);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
std::fstream empty_stream("", std::fstream::in);
stan::io::dump data_var_context(empty_stream);
gauss3D_model_namespace::gauss3D_model model(data_var_context);
stan::mcmc::unit_e_nuts<gauss3D_model_namespace::gauss3D_model, rng_t>
sampler(model, base_rng);
sampler.z() = z_init;
sampler.init_hamiltonian(writer, error_writer);
sampler.set_nominal_stepsize(0.1);
sampler.set_stepsize_jitter(0);
sampler.sample_stepsize();
stan::mcmc::ps_point z_propose = z_init;
Eigen::VectorXd p_sharp_left = Eigen::VectorXd::Zero(z_init.p.size());
Eigen::VectorXd p_sharp_right = Eigen::VectorXd::Zero(z_init.p.size());
Eigen::VectorXd rho = z_init.p;
double log_sum_weight = -std::numeric_limits<double>::infinity();
double H0 = -0.1;
int n_leapfrog = 0;
double sum_metro_prob = 0;
bool valid_subtree = sampler.build_tree(3, z_propose,
p_sharp_left, p_sharp_right, rho,
H0, 1, n_leapfrog, log_sum_weight,
sum_metro_prob,
writer, error_writer);
EXPECT_EQ(0.1, sampler.get_nominal_stepsize());
EXPECT_TRUE(valid_subtree);
EXPECT_FLOAT_EQ(-11.401228, rho(0));
EXPECT_FLOAT_EQ(11.401228, rho(1));
EXPECT_FLOAT_EQ(-11.401228, rho(2));
EXPECT_FLOAT_EQ(-0.022019938, sampler.z().q(0));
EXPECT_FLOAT_EQ(0.022019938, sampler.z().q(1));
EXPECT_FLOAT_EQ(-0.022019938, sampler.z().q(2));
EXPECT_FLOAT_EQ(-1.4131583, sampler.z().p(0));
EXPECT_FLOAT_EQ(1.4131583, sampler.z().p(1));
EXPECT_FLOAT_EQ(-1.4131583, sampler.z().p(2));
EXPECT_EQ(8, n_leapfrog);
EXPECT_FLOAT_EQ(std::log(0.36134657), log_sum_weight);
EXPECT_FLOAT_EQ(0.36134657, sum_metro_prob);
EXPECT_EQ("", output.str());
EXPECT_EQ("", error_stream.str());
}
TEST(McmcHmcIntegratorsImplLeapfrog, unit_e_symplecticness) {
rng_t base_rng(0);
std::fstream data_stream(std::string("").c_str(), std::fstream::in);
stan::io::dump data_var_context(data_stream);
data_stream.close();
std::stringstream model_output;
std::stringstream metric_output;
stan::interface_callbacks::writer::stream_writer writer(metric_output);
std::stringstream error_stream;
stan::interface_callbacks::writer::stream_writer error_writer(error_stream);
gauss_model_namespace::gauss_model model(data_var_context, &model_output);
stan::mcmc::impl_leapfrog<
stan::mcmc::unit_e_metric<gauss_model_namespace::gauss_model, rng_t> >
integrator;
stan::mcmc::unit_e_metric<gauss_model_namespace::gauss_model, rng_t> metric(model);
// Create a circle of points
const int n_points = 1000;
double pi = 3.141592653589793;
double r = 1.5;
double q0 = 1;
double p0 = 0;
std::vector<stan::mcmc::unit_e_point> z;
for (int i = 0; i < n_points; ++i) {
z.push_back(stan::mcmc::unit_e_point(1));
double theta = 2 * pi * (double)i / (double)n_points;
z.back().q(0) = r * cos(theta) + q0;
z.back().p(0) = r * sin(theta) + p0;
}
// Evolve circle
double epsilon = 1e-3;
size_t L = pi / epsilon;
for (int i = 0; i < n_points; ++i)
metric.init(z.at(i), writer, error_writer);
for (size_t n = 0; n < L; ++n)
for (int i = 0; i < n_points; ++i)
integrator.evolve(z.at(i), metric, epsilon, writer, error_writer);
// Compute area of evolved shape using divergence theorem in 2D
double area = 0;
for (int i = 0; i < n_points; ++i) {
double x1 = z[i].q(0);
double y1 = z[i].p(0);
double x2 = z[(i + 1) % n_points].q(0);
double y2 = z[(i + 1) % n_points].p(0);
double x_bary = 0.5 * (x1 + x2);
double y_bary = 0.5 * (y1 + y2);
double x_delta = x2 - x1;
double y_delta = y2 - y1;
double a = sqrt( x_delta * x_delta + y_delta * y_delta);
double x_norm = 1;
double y_norm = - x_delta / y_delta;
double norm = sqrt( x_norm * x_norm + y_norm * y_norm );
a *= (x_bary * x_norm + y_bary * y_norm) / norm;
a = a < 0 ? -a : a;
area += a;
}
area *= 0.5;
// Symplectic integrators preserve volume (area in 2D)
EXPECT_NEAR(area, pi * r * r, 1e-2);
EXPECT_EQ("", model_output.str());
EXPECT_EQ("", metric_output.str());
}
