Classes
struct	MomentumHandler

struct	MomentumHandler< diag_var >

struct	MomentumHandler< unit_var >

struct	StepAdapter

struct	TreeInput

struct	TreeOutput

struct	VarAdapter

struct	VarAdapter< diag_var >

struct	VarAdapter< unit_var >

Functions
template<class ExprType , class ConfigType , class Sampler >
MCMCResult	base_mcmc (const ExprType &expr, const ConfigType &config, Sampler f)

double	hamiltonian (double potential, double kinetic)

template<class ADExprType , class MatType >
double	reset_autodiff (ADExprType &ad_expr, Eigen::MatrixBase< MatType > &adjoints, Eigen::MatrixBase< MatType > &tp_adjoints)

template<class ADExprType , class MatType , class MomentumHandlerType >
double	leapfrog (ADExprType &ad_expr, Eigen::MatrixBase< MatType > &theta, Eigen::MatrixBase< MatType > &theta_adj, Eigen::MatrixBase< MatType > &tp_adj, Eigen::MatrixBase< MatType > &r, const MomentumHandlerType &m_handler, double epsilon, bool reuse_adj)

template<class MatType1 , class MatType2 , class MatType3 >
bool	check_entropy (const MatType1 &rho, const MatType2 &p_beg_scaled, const MatType3 &p_end_scaled)

template<class InputType , class UniformDistType , class GenType , class MomentumHandlerType >
TreeOutput	build_tree (size_t n_params, InputType &input, uint8_t depth, UniformDistType &unif_sampler, GenType &gen, MomentumHandlerType &momentum_handler, double *tree_cache)

template<class ADExprType , class MatType , class GenType , class MomentumHandlerType >
double	find_reasonable_epsilon (double eps, ADExprType &ad_expr, MatType &theta, MatType &theta_adj, MatType &tp_adj, GenType &gen, MomentumHandlerType &momentum_handler)

template<class ProgramType , class OffsetPackType , class MCMCResultType , class NUTSConfigType = NUTSConfig<>>
void	nuts_ (ProgramType &program, const NUTSConfigType &config, const OffsetPackType &pack, MCMCResultType &res)

template<class ProgramType , class OffsetPackType , class MCMCResultType >
void	mh_ (const ProgramType &program, const MHConfig &config, const OffsetPackType &pack, MCMCResultType &res)

size_t	random_seed ()

template<class UniformDistType , class GenType >
bool	accept_or_reject (double p, UniformDistType &&unif_sampler, GenType &&gen)

Function Documentation

◆ accept_or_reject()

template<class UniformDistType , class GenType >

bool ppl::mcmc::accept_or_reject	(	double	p,
		UniformDistType &&	unif_sampler,
		GenType &&	gen
	)

inline

Accepts or rejects with given probability using UniformDistType object that works with GenType. The uniform sampler must sample from [0,1].

◆ base_mcmc()

template<class ExprType , class ConfigType , class Sampler >

MCMCResult ppl::mcmc::base_mcmc	(	const ExprType &	expr,
		const ConfigType &	config,
		Sampler	f
	)

inline

◆ build_tree()

template<class InputType , class UniformDistType , class GenType , class MomentumHandlerType >

TreeOutput ppl::mcmc::build_tree	(	size_t	n_params,
		InputType &	input,
		uint8_t	depth,
		UniformDistType &	unif_sampler,
		GenType &	gen,
		MomentumHandlerType &	momentum_handler,
		double *	tree_cache
	)

Building binary tree for sampling candidates. Helper function to obtain the forward/backward-most position and momentum. Accept/reject policy is based on UniformDistType parameter and GenType

Note that the caller, i.e. nuts(), MUST have theta_adj already pre-computed (theta_adj is a member of input and input will be an instance of TreeInput).

Parameters

n_params	number of (continuous) parameters
input	TreeInput-like input object
depth	current depth of building tree
unif_sampler	an object like std::uniform_distribution(0,1) used for metropolis acceptance
gen	rng device
momentum_handler	MomentumHandler-like object to compute kinetic energy and momentum
tree_cache	pointer to cache memory that will be used by build_tree. The array of doubles must be of size n_params * 7 * max_depth.

◆ check_entropy()

template<class MatType1 , class MatType2 , class MatType3 >

bool ppl::mcmc::check_entropy	(	const MatType1 &	rho,
		const MatType2 &	p_beg_scaled,
		const MatType3 &	p_end_scaled
	)

Checks if state is entroping based on integrated momentum vector across the path and the scaled momentum at the ends of the path.

Parameters

rho	integrated momentum vector
p_beg_scaled	scaled momentum beginning of current path (given the direction)
p_end_scaled	scaled momentum end of current path (given the direction).

◆ find_reasonable_epsilon()

template<class ADExprType , class MatType , class GenType , class MomentumHandlerType >

double ppl::mcmc::find_reasonable_epsilon	(	double	eps,
		ADExprType &	ad_expr,
		MatType &	theta,
		MatType &	theta_adj,
		MatType &	tp_adj,
		GenType &	gen,
		MomentumHandlerType &	momentum_handler
	)

Finds a reasonable epsilon for NUTS algorithm.

Parameters

eps	initial epsilon (see Gelman's paper)
ad_expr	AD expression bound to theta and theta_adj
theta	vector of theta values
theta_adj	vector of theta adjoints
gen	rng device
momentum_handler	MomentumHandler-like object

◆ hamiltonian()

double ppl::mcmc::hamiltonian	(	double	potential,
		double	kinetic
	)

inline

Compute Hamiltonian given potential and kinetic energy. Returns simply the sum of the two.

◆ leapfrog()

template<class ADExprType , class MatType , class MomentumHandlerType >

double ppl::mcmc::leapfrog	(	ADExprType &	ad_expr,
		Eigen::MatrixBase< MatType > &	theta,
		Eigen::MatrixBase< MatType > &	theta_adj,
		Eigen::MatrixBase< MatType > &	tp_adj,
		Eigen::MatrixBase< MatType > &	r,
		const MomentumHandlerType &	m_handler,
		double	epsilon,
		bool	reuse_adj
	)

inline

Leapfrog algorithm. Expects theta, theta_adj, tp_adj, r to be Eigen column-like objects.

Updates theta, theta_adj, tp_adj, and r to contain the new leaped values and adjoints.

Parameters

ad_expr	AD expression representing L(theta) It must be built such that values are read from theta and adjoints are placed into theta_adj.
theta	theta at which we want to start leaping
theta_adj	adjoint for theta. If not reusing, resets adjoints first.
tp_adj	adjoints for transformed parameters. If not reusing, resets adjoints first.
r	momentum vector to start leaping
m_handler	momentum handler to compute correct dkinetic/dr
epsilon	step size
reuse_adj	flag to not compute gradient of L(theta) if user can guarantee that theta_adj currently has it.

Returns: new potential energy (L(theta'))

◆ mh_()

template<class ProgramType , class OffsetPackType , class MCMCResultType >

void ppl::mcmc::mh_	(	const ProgramType &	program,
		const MHConfig &	config,
		const OffsetPackType &	pack,
		MCMCResultType &	res
	)

inline

Metropolis-Hastings algorithm to sample from posterior distribution. Any variables that model references which are parameters are sampled.

Template Parameters

ProgramType	program expression type
OffsetPackType	offset pack type (likely util::OffsetPack)

Parameters

program	program expression
config	configuration object
pack	offset pack from activating program expression
res	sampling result object to populate

◆ nuts_()

template<class ProgramType , class OffsetPackType , class MCMCResultType , class NUTSConfigType = NUTSConfig<>>

void ppl::mcmc::nuts_	(	ProgramType &	program,
		const NUTSConfigType &	config,
		const OffsetPackType &	pack,
		MCMCResultType &	res
	)

No-U-Turn Sampler (NUTS)

User must ensure that the program does not have any discrete parameters. Discrete data is allowed.

Parameters

program	program expression used to determine log-pdf
config	NUTS configuration object
pack	offset pack result of activating program. It will likely be util::OffsetPack where each offset value is equivalent to the total number of values needed, i.e. if pack.uc_offset is 10, there is exactly 10 unconstrained values for the program.
res	result object of calling NUTS that will be populated with samples and other information.

◆ random_seed()

size_t ppl::mcmc::random_seed ( )

inline

Get current time in milliseconds for random seeding.

◆ reset_autodiff()

template<class ADExprType , class MatType >

double ppl::mcmc::reset_autodiff	(	ADExprType &	ad_expr,
		Eigen::MatrixBase< MatType > &	adjoints,
		Eigen::MatrixBase< MatType > &	tp_adjoints
	)

inline

Helper function for leapfrog algorithm. Resets adjoints and then differentiates AD expression.

Parameters

ad_expr	AD expression to differentiate
adjoints	Eigen generic matrix type that supports member fn "setZero"
tp_adjoints	Eigen generic matrix type that supports member fn "setZero"

Returns: result of calling ad::autodiff on ad_expr after resetting adjoints

Classes

Functions

Function Documentation

◆ accept_or_reject()

◆ base_mcmc()

◆ build_tree()

◆ check_entropy()

◆ find_reasonable_epsilon()

◆ hamiltonian()

◆ leapfrog()

◆ mh_()

◆ nuts_()

◆ random_seed()

◆ reset_autodiff()