 |
Stan
1.0
probability, sampling & optimization
|
|
Stan
1.0
probability, sampling & optimization
|
No-U-Turn Sampler (NUTS) with varying step sizes. More...
#include <nuts_diag.hpp>
Public Member Functions | |
| nuts_diag (stan::model::prob_grad &model, int maxdepth=10, double epsilon=-1, double epsilon_pm=0.0, bool epsilon_adapt=true, double delta=0.6, double gamma=0.05, BaseRNG base_rng=BaseRNG(std::time(0)), const std::vector< double > *params_r=0, const std::vector< int > *params_i=0) | |
| Construct a No-U-Turn Sampler (NUTS) for the specified model, using the specified step size and number of leapfrog steps, with the specified random seed for randomization. More... | |
| ~nuts_diag () | |
| Destroy this sampler. More... | |
| virtual sample | next_impl () |
| Return the next sample. More... | |
| virtual void | write_sampler_param_names (std::ostream &o) |
| Write out any sampler-specific parameter names for output. More... | |
| virtual void | write_sampler_params (std::ostream &o) |
| Write out any sampler-specific parameters for output. More... | |
| virtual void | write_adaptation_params (std::ostream &o) |
| Use this method to write the adaptation parameters into the output. More... | |
| virtual void | get_sampler_param_names (std::vector< std::string > &names) |
| Get any sampler-specific parameter namess. More... | |
| virtual void | get_sampler_params (std::vector< double > &values) |
| Get any sampler-specific parameters. More... | |
| void | build_tree (const std::vector< double > &x, const std::vector< double > &m, const std::vector< double > &grad, double u, int direction, int depth, double H0, std::vector< double > &xminus, std::vector< double > &mminus, std::vector< double > &gradminus, std::vector< double > &xplus, std::vector< double > &mplus, std::vector< double > &gradplus, std::vector< double > &newx, std::vector< double > &newgrad, double &newlogp, int &nvalid, bool &criterion, double &prob_sum, int &n_considered) |
| The core recursion in NUTS. More... | |
 Public Member Functions inherited from stan::mcmc::hmc_base< boost::mt19937 > | |
| hmc_base (stan::model::prob_grad &model, double epsilon=-1, double epsilon_pm=0.0, bool epsilon_adapt=true, double delta=0.651, double gamma=0.05, boost::mt19937 rand_int=boost::mt19937(std::time(0)), const std::vector< double > *params_r=0, const std::vector< int > *params_i=0) | |
| Construct a base HMC sampler. More... | |
| virtual | ~hmc_base () |
| virtual void | set_params (const std::vector< double > &x, const std::vector< int > &z) |
| Sets the model real and integer parameters to the specified values. More... | |
| void | set_params_r (const std::vector< double > &x) |
| Sets the model real parameters to the specified values and update gradients and log probability to match. More... | |
| void | set_params_i (const std::vector< int > &z) |
| Sets the model integer parameters to the specified values and update gradients and log probability to match. More... | |
| bool | varying_epsilon () |
| virtual void | adapt_off () |
| Turn off parameter adaptation. More... | |
| virtual void | get_parameters (std::vector< double > ¶ms) |
| Write the step size into position 1 of the specified vector. More... | |
| Public Member Functions inherited from stan::mcmc::adaptive_sampler | |
| adaptive_sampler (bool adapt, std::ostream *error_msgs=0, std::ostream *output_msgs=0) | |
| Constructs an adaptive sampler with specified adaptation status. More... | |
| virtual | ~adaptive_sampler () |
| Destructor. More... | |
| void | set_error_stream (std::ostream &error_msgs) |
| Set the stream into which errors will be written as the sampler runs. More... | |
| void | unset_error_stream () |
| Unset the stream into which errors are written to 0 so that error messages are ignored. More... | |
| void | set_output_stream (std::ostream &output_msgs) |
| Set the stream into which output will be written as the sampler runs. More... | |
| void | unset_output_stream () |
| Unset the stream into which errors are written to 0 so that output messages are ignored. More... | |
| sample | next () |
| Returns the next sample from this sampler. More... | |
| double | mean_stat () |
| Returns the value of the statistic we are trying to coerce. More... | |
| void | set_mean_stat (double v) |
| Sets the mean statistic to the specified value. More... | |
| void | update_mean_stat (double avg_eta, double adapt_stat) |
| Updates the mean statistic given the specified adaptation statistic and weighting. More... | |
| unsigned int | nfevals () |
| Returns the number of times that the (possibly unnormalized) log probability function has been evaluated by this sampler. More... | |
| void | nfevals_plus_eq (int n) |
| Add the specified number of evaluations to the number of function evaluations. More... | |
| int | n_steps () |
| Return the number of iterations for this sampler. More... | |
| int | n_adapt_steps () |
| Return how many iterations parameter adaptation has happened for. More... | |
| virtual void | adapt_on () |
| Turn on parameter adaptation. More... | |
| bool | adapting () |
| Return whether or not parameter adaptation is on. More... | |
Additional Inherited Members | |
| Protected Member Functions inherited from stan::mcmc::hmc_base< boost::mt19937 > | |
| virtual void | find_reasonable_parameters () |
| Search for a roughly reasonable (within a factor of 2) setting of the step size epsilon. More... | |
| void | adaptation_init (double epsilon_scale) |
| Protected Attributes inherited from stan::mcmc::hmc_base< boost::mt19937 > | |
| stan::model::prob_grad & | _model |
| double | _epsilon |
| double | _epsilon_pm |
| double | _epsilon_last |
| bool | _epsilon_adapt |
| const double | _delta |
| const double | _gamma |
| DualAverage | _da |
| boost::mt19937 | _rand_int |
| boost::variate_generator< boost::mt19937 &, boost::normal_distribution<> > | _rand_unit_norm |
| boost::uniform_01< boost::mt19937 & > | _rand_uniform_01 |
| std::vector< double > | _x |
| std::vector< int > | _z |
| std::vector< double > | _g |
| double | _logp |
| Protected Attributes inherited from stan::mcmc::adaptive_sampler | |
| bool | _adapt |
| unsigned int | _n_steps |
| int | _n_adapt_steps |
| unsigned int | _nfevals |
| double | _mean_stat |
| std::ostream * | _error_msgs |
| std::ostream * | _output_msgs |
No-U-Turn Sampler (NUTS) with varying step sizes.
The NUTS sampler requires a probability model with the ability to compute gradients, characterized as an instance of prob_grad.
Definition at line 33 of file nuts_diag.hpp.
|
inline |
Construct a No-U-Turn Sampler (NUTS) for the specified model, using the specified step size and number of leapfrog steps, with the specified random seed for randomization.
If the same seed is used twice, the series of samples should be the same. This property is most helpful for testing. If no random seed is specified, the std::time(0) function is called from the ctime library.
| model | Probability model with gradients. |
| maxdepth | |
| epsilon | Optional (initial) Hamiltonian dynamics simulation step size. If not specified or set < 0, find_reasonable_parameters() will be called to initialize epsilon. |
| epsilon_pm | Plus/minus range for uniformly sampling epsilon around its value. |
| epsilon_adapt | True if epsilon is adapted during warmup. |
| delta | Optional target value between 0 and 1 used to tune epsilon. Lower delta => higher epsilon => more efficiency, unless epsilon gets too big in which case efficiency suffers. If not specified, defaults to the usually reasonable value of 0.6. |
| gamma | Gamma tuning parameter for dual averaging adaptation. |
| base_rng | Optional Seed for random number generator; if not specified, generate new seed based on system time. |
Definition at line 99 of file nuts_diag.hpp.
|
inline |
Destroy this sampler.
The implementation for this class is a no-op.
Definition at line 138 of file nuts_diag.hpp.
|
inline |
The core recursion in NUTS.
| x | The position value to start from. |
| m | The momentum value to start from. |
| grad | The gradient at the initial position. |
| u | The slice variable. |
| direction | Simulate backwards if -1, forwards if 1. |
| depth | The depth of the tree to build—we'll run 2^depth leapfrog steps. |
| H0 | The joint probability of the position-momentum we started from initially—used to compute statistic to adapt epsilon. |
| xminus | Returns the position of the backwardmost leaf of this subtree. |
| mminus | Returns the momentum of the backwardmost leaf of this subtree. |
| gradminus | Returns the gradient at xminus. |
| xplus | Returns the position of the forwardmost leaf of this subtree. |
| mplus | Returns the momentum of the forwardmost leaf of this subtree. |
| gradplus | Returns the gradient at xplus. |
| newx | Returns the new position sample selected from this subtree. |
| newgrad | Returns the gradient at the new sample selected from this subtree. |
| newlogp | Returns the log-probability of the new sample selected from this subtree. |
| nvalid | Returns the number of usable points in the subtree. |
| criterion | Returns true if the subtree is usable, false if not. |
| prob_sum | Returns the sum of the HMC acceptance probabilities at each point in the subtree. |
| n_considered | Returns the number of states in the subtree. |
Definition at line 327 of file nuts_diag.hpp.
|
inlinevirtual |
Get any sampler-specific parameter namess.
| [out] | names | Output vector to which param names are written. |
Reimplemented from stan::mcmc::adaptive_sampler.
Definition at line 279 of file nuts_diag.hpp.
|
inlinevirtual |
Get any sampler-specific parameters.
| [out] | values | Output vector to which params are written. All values are casted to type double. This function should match get_sampler_param_names. |
Reimplemented from stan::mcmc::adaptive_sampler.
Definition at line 286 of file nuts_diag.hpp.
|
inlinevirtual |
Return the next sample.
Implements stan::mcmc::adaptive_sampler.
Definition at line 145 of file nuts_diag.hpp.
|
inlinevirtual |
Use this method to write the adaptation parameters into the output.
These should be written as one or more comment lines, each starting with a pound (#) character.
| o | Output stream to which adaptation information is written. |
Reimplemented from stan::mcmc::adaptive_sampler.
Definition at line 267 of file nuts_diag.hpp.
|
inlinevirtual |
Write out any sampler-specific parameter names for output.
This method must match write_sampler_params() in terms of number of parameters written.
Params should be writte starting with a comma, then the first parameter, then a comma, then the second parameter, ending on the final parameter.
The base class implementation is a no-op.
| [out] | o | Output stream to which param names are written. |
Reimplemented from stan::mcmc::adaptive_sampler.
Definition at line 255 of file nuts_diag.hpp.
|
inlinevirtual |
Write out any sampler-specific parameters for output.
This method must match write_sampler_param_names() in terms of number of parameters written.
Params should be writte starting with a comma, then the first parameter, then a comma, then the second parameter, ending on the final parameter.
The base class implementation is a no-op.
| [out] | o | Output stream to which params are written. |
Reimplemented from stan::mcmc::adaptive_sampler.
Definition at line 261 of file nuts_diag.hpp.