libshogun-dev 3.1.1-1 / usr / include / shogun / machine / gp / InferenceMethod.h

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 2013 Roman Votyakov
 * Written (W) 2013 Heiko Strathmann
 * Copyright (C) 2012 Jacob Walker
 * Copyright (C) 2013 Roman Votyakov
 */

#ifndef CINFERENCEMETHOD_H_
#define CINFERENCEMETHOD_H_

#include <shogun/lib/config.h>

#ifdef HAVE_EIGEN3

#include <shogun/base/SGObject.h>
#include <shogun/kernel/Kernel.h>
#include <shogun/features/Features.h>
#include <shogun/labels/Labels.h>
#include <shogun/machine/gp/LikelihoodModel.h>
#include <shogun/machine/gp/MeanFunction.h>
#include <shogun/evaluation/DifferentiableFunction.h>

namespace shogun
{

/** inference type */
enum EInferenceType
{
	INF_NONE=0,
	INF_EXACT=10,
	INF_FITC=20,
	INF_LAPLACIAN=30,
	INF_EP=40
};

/** @brief The Inference Method base class.
 *
 * The Inference Method computes (approximately) the posterior distribution for
 * a given Gaussian Process.
 *
 * It is possible to sample the (true) log-marginal likelihood on the base of
 * any implemented approximation. See log_ml_estimate.
 */
class CInferenceMethod : public CDifferentiableFunction
{
public:
	/** default constructor */
	CInferenceMethod();

	/** constructor
	 *
	 * @param kernel covariance function
	 * @param features features to use in inference
	 * @param mean mean function
	 * @param labels labels of the features
	 * @param model likelihood model to use
	 */
	CInferenceMethod(CKernel* kernel, CFeatures* features,
			CMeanFunction* mean, CLabels* labels, CLikelihoodModel* model);

	virtual ~CInferenceMethod();

	/** return what type of inference we are, e.g. exact, FITC, Laplacian, etc.
	 *
	 * @return inference type
	 */
	virtual EInferenceType get_inference_type() const { return INF_NONE; }

	/** get negative log marginal likelihood
	 *
	 * @return the negative log of the marginal likelihood function:
	 *
	 * \f[
	 * -log(p(y|X, \theta))
	 * \f]
	 *
	 * where \f$y\f$ are the labels, \f$X\f$ are the features, and \f$\theta\f$
	 * represent hyperparameters.
	 */
	virtual float64_t get_negative_log_marginal_likelihood()=0;

	/** Computes an unbiased estimate of the log-marginal-likelihood,
	 *
	 * \f[
	 * log(p(y|X,\theta)),
	 * \f]
	 * where \f$y\f$ are the labels, \f$X\f$ are the features (omitted from in
	 * the following expressions), and \f$\theta\f$ represent hyperparameters.
	 *
	 * This is done via an approximation to the posterior
	 * \f$q(f|y, \theta)\approx p(f|y, \theta)\f$, which is computed by the
	 * underlying CInferenceMethod instance (if implemented, otherwise error),
	 * and then using an importance sample estimator
	 *
	 * \f[
	 * p(y|\theta)=\int p(y|f)p(f|\theta)df
	 * =\int p(y|f)\frac{p(f|\theta)}{q(f|y, \theta)}q(f|y, \theta)df
	 * \approx\frac{1}{n}\sum_{i=1}^n p(y|f^{(i)})\frac{p(f^{(i)}|\theta)}
	 * {q(f^{(i)}|y, \theta)},
	 * \f]
	 *
	 * where \f$ f^{(i)} \f$ are samples from the posterior approximation
	 * \f$ q(f|y, \theta) \f$. The resulting estimator has a low variance if
	 * \f$ q(f|y, \theta) \f$ is a good approximation. It has large variance
	 * otherwise (while still being consistent).
	 *
	 * @param num_importance_samples the number of importance samples \f$n\f$
	 * from \f$ q(f|y, \theta) \f$.
	 * @param ridge_size scalar that is added to the diagonal of the involved
	 * Gaussian distribution's covariance of GP prior and posterior
	 * approximation to stabilise things. Increase if Cholesky factorization
	 * fails.
	 *
	 * @return unbiased estimate of the log of the marginal likelihood function
	 * \f$ log(p(y|\theta)) \f$
	 */
	float64_t get_marginal_likelihood_estimate(int32_t num_importance_samples=1,
			float64_t ridge_size=1e-15);

	/** get log marginal likelihood gradient
	 *
	 * @return vector of the marginal likelihood function gradient with respect
	 * to hyperparameters (under the current approximation to the posterior
	 * \f$q(f|y)\approx p(f|y)\f$:
	 *
	 * \f[
	 * -\frac{\partial log(p(y|X, \theta))}{\partial \theta}
	 * \f]
	 *
	 * where \f$y\f$ are the labels, \f$X\f$ are the features, and \f$\theta\f$
	 * represent hyperparameters.
	 */
	virtual CMap<TParameter*, SGVector<float64_t> >* get_negative_log_marginal_likelihood_derivatives(
			CMap<TParameter*, CSGObject*>* parameters);

	/** get alpha vector
	 *
	 * @return vector to compute posterior mean of Gaussian Process:
	 *
	 * \f[
	 * \mu = K\alpha
	 * \f]
	 *
	 * where \f$\mu\f$ is the mean and \f$K\f$ is the prior covariance matrix.
	 */
	virtual SGVector<float64_t> get_alpha()=0;

	/** get Cholesky decomposition matrix
	 *
	 * @return Cholesky decomposition of matrix:
	 *
	 * \f[
	 * L = cholesky(sW*K*sW+I)
	 * \f]
	 *
	 * where \f$K\f$ is the prior covariance matrix, \f$sW\f$ is the vector
	 * returned by get_diagonal_vector(), and \f$I\f$ is the identity matrix.
	 */
	virtual SGMatrix<float64_t> get_cholesky()=0;

	/** get diagonal vector
	 *
	 * @return diagonal of matrix used to calculate posterior covariance matrix:
	 *
	 * \f[
	 * Cov = (K^{-1}+sW^{2})^{-1}
	 * \f]
	 *
	 * where \f$Cov\f$ is the posterior covariance matrix, \f$K\f$ is the prior
	 * covariance matrix, and \f$sW\f$ is the diagonal vector.
	 */
	virtual SGVector<float64_t> get_diagonal_vector()=0;

	/** returns mean vector \f$\mu\f$ of the Gaussian distribution
	 * \f$\mathcal{N}(\mu,\Sigma)\f$, which is an approximation to the
	 * posterior:
	 *
	 * \f[
	 * p(f|y) \approx q(f|y) = \mathcal{N}(\mu,\Sigma)
	 * \f]
	 *
	 * in case if particular inference method doesn't compute posterior
	 * \f$p(f|y)\f$ exactly, and it returns covariance matrix \f$\Sigma\f$ of
	 * the posterior Gaussian distribution \f$\mathcal{N}(\mu,\Sigma)\f$
	 * otherwise.
	 *
	 * @return mean vector
	 */
	virtual SGVector<float64_t> get_posterior_mean()=0;

	/** returns covariance matrix \f$\Sigma\f$ of the Gaussian distribution
	 * \f$\mathcal{N}(\mu,\Sigma)\f$, which is an approximation to the
	 * posterior:
	 *
	 * \f[
	 * p(f|y) \approx q(f|y) = \mathcal{N}(\mu,\Sigma)
	 * \f]
	 *
	 * in case if particular inference method doesn't compute posterior
	 * \f$p(f|y)\f$ exactly, and it returns covariance matrix \f$\Sigma\f$ of
	 * the posterior Gaussian distribution \f$\mathcal{N}(\mu,\Sigma)\f$
	 * otherwise.
	 *
	 * @return covariance matrix
	 */
	virtual SGMatrix<float64_t> get_posterior_covariance()=0;

	/** get the gradient
	 *
	 * @param parameters parameter's dictionary
	 *
	 * @return map of gradient. Keys are names of parameters, values are values
	 * of derivative with respect to that parameter.
	 */
	virtual CMap<TParameter*, SGVector<float64_t> >* get_gradient(
			CMap<TParameter*, CSGObject*>* parameters)
	{
        return get_negative_log_marginal_likelihood_derivatives(parameters);
	}

	/** get the function value
	 *
	 * @return vector that represents the function value
	 */
	virtual SGVector<float64_t> get_value()
	{
		SGVector<float64_t> result(1);
		result[0]=get_negative_log_marginal_likelihood();
		return result;
	}

	/** get features
	*
	* @return features
	*/
	virtual CFeatures* get_features() { SG_REF(m_features); return m_features; }

	/** set features
	*
	* @param feat features to set
	*/
	virtual void set_features(CFeatures* feat)
	{
		SG_REF(feat);
		SG_UNREF(m_features);
		m_features=feat;
	}

	/** get kernel
	 *
	 * @return kernel
	 */
	virtual CKernel* get_kernel() { SG_REF(m_kernel); return m_kernel; }

	/** set kernel
	 *
	 * @param kern kernel to set
	 */
	virtual void set_kernel(CKernel* kern)
	{
		SG_REF(kern);
		SG_UNREF(m_kernel);
		m_kernel=kern;
	}

	/** get mean
	 *
	 * @return mean
	 */
	virtual CMeanFunction* get_mean() { SG_REF(m_mean); return m_mean; }

	/** set mean
	 *
	 * @param m mean function to set
	 */
	virtual void set_mean(CMeanFunction* m)
	{
		SG_REF(m);
		SG_UNREF(m_mean);
		m_mean=m;
	}

	/** get labels
	 *
	 * @return labels
	 */
	virtual CLabels* get_labels() { SG_REF(m_labels); return m_labels; }

	/** set labels
	 *
	 * @param lab label to set
	 */
	virtual void set_labels(CLabels* lab)
	{
		SG_REF(lab);
		SG_UNREF(m_labels);
		m_labels=lab;
	}

	/** get likelihood model
	 *
	 * @return likelihood
	 */
	CLikelihoodModel* get_model() { SG_REF(m_model); return m_model; }

	/** set likelihood model
	 *
	 * @param mod model to set
	 */
	virtual void set_model(CLikelihoodModel* mod)
	{
		SG_REF(mod);
		SG_UNREF(m_model);
		m_model=mod;
	}

	/** get kernel scale
	 *
	 * @return kernel scale
	 */
	virtual float64_t get_scale() const { return m_scale; }

	/** set kernel scale
	 *
	 * @param scale scale to be set
	 */
	virtual void set_scale(float64_t scale) { m_scale=scale; }

	/** whether combination of inference method and given likelihood function
	 * supports regression
	 *
	 * @return false
	 */
	virtual bool supports_regression() const { return false; }

	/** whether combination of inference method and given likelihood function
	 * supports binary classification
	 *
	 * @return false
	 */
	virtual bool supports_binary() const { return false; }

	/** whether combination of inference method and given likelihood function
	 * supports multiclass classification
	 *
	 * @return false
	 */
	virtual bool supports_multiclass() const { return false; }

	/** update all matrices */
	virtual void update();

protected:
	/** check if members of object are valid for inference */
	virtual void check_members() const;

	/** update alpha vector */
	virtual void update_alpha()=0;

	/** update cholesky matrix */
	virtual void update_chol()=0;

	/** update matrices which are required to compute negative log marginal
	 * likelihood derivatives wrt hyperparameter
	 */
	virtual void update_deriv()=0;

	/** update train kernel matrix */
	virtual void update_train_kernel();

	/** returns derivative of negative log marginal likelihood wrt parameter of
	 * CInferenceMethod class
	 *
	 * @param param parameter of CInferenceMethod class
	 *
	 * @return derivative of negative log marginal likelihood
	 */
	virtual SGVector<float64_t> get_derivative_wrt_inference_method(
			const TParameter* param)=0;

	/** returns derivative of negative log marginal likelihood wrt parameter of
	 * likelihood model
	 *
	 * @param param parameter of given likelihood model
	 *
	 * @return derivative of negative log marginal likelihood
	 */
	virtual SGVector<float64_t> get_derivative_wrt_likelihood_model(
			const TParameter* param)=0;

	/** returns derivative of negative log marginal likelihood wrt kernel's
	 * parameter
	 *
	 * @param param parameter of given kernel
	 *
	 * @return derivative of negative log marginal likelihood
	 */
	virtual SGVector<float64_t> get_derivative_wrt_kernel(
			const TParameter* param)=0;

	/** returns derivative of negative log marginal likelihood wrt mean
	 * function's parameter
	 *
	 * @param param parameter of given mean function
	 *
	 * @return derivative of negative log marginal likelihood
	 */
	virtual SGVector<float64_t> get_derivative_wrt_mean(
			const TParameter* param)=0;

	/** pthread helper method to compute negative log marginal likelihood
	 * derivatives wrt hyperparameter
	 */
	static void* get_derivative_helper(void* p);

private:
	void init();

protected:
	/** covariance function */
	CKernel* m_kernel;

	/** mean function */
	CMeanFunction* m_mean;

	/** likelihood function to use */
	CLikelihoodModel* m_model;

	/** features to use */
	CFeatures* m_features;

	/** labels of features */
	CLabels* m_labels;

	/** alpha vector used in process mean calculation */
	SGVector<float64_t> m_alpha;

	/** upper triangular factor of Cholesky decomposition */
	SGMatrix<float64_t> m_L;

	/** kernel scale */
	float64_t m_scale;

	/** kernel matrix from features (non-scalled by inference scalling) */
	SGMatrix<float64_t> m_ktrtr;
};
}
#endif /* HAVE_EIGEN3 */
#endif /* CINFERENCEMETHOD_H_ */