Shogun Gradient Model Selection example bugs

gistfile1.cpp

/*
 * 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 Heiko Strathmann
 */

#include <shogun/lib/config.h>

#ifdef HAVE_EIGEN3
#include <shogun/labels/RegressionLabels.h>
#include <shogun/features/DenseFeatures.h>
#include <shogun/kernel/GaussianKernel.h>
#include <shogun/regression/GaussianProcessRegression.h>
#include <shogun/machine/gp/ExactInferenceMethod.h>
#include <shogun/machine/gp/ZeroMean.h>
#include <shogun/machine/gp/GaussianLikelihood.h>
#include <shogun/base/init.h>
#include <shogun/evaluation/GradientEvaluation.h>
#include <shogun/modelselection/GradientModelSelection.h>
#include <shogun/modelselection/ModelSelectionParameters.h>
#include <shogun/modelselection/ParameterCombination.h>
#include <shogun/evaluation/GradientCriterion.h>

using namespace shogun;

CModelSelectionParameters* build_modelselection_tree(CInferenceMethod* inf,
		CLikelihoodModel* lik, CKernel* kernel)
{
	CModelSelectionParameters* root=new CModelSelectionParameters();

	CModelSelectionParameters* c1=new CModelSelectionParameters(
			"inference_method", inf);
	root->append_child(c1);

	CModelSelectionParameters* c2=new CModelSelectionParameters(
			"likelihood_model", lik);
	c1->append_child(c2);

	CModelSelectionParameters* c3=new CModelSelectionParameters("sigma");
	c2->append_child(c3);
	c3->build_values(1.0, 4.0, R_LINEAR);

	CModelSelectionParameters* c4=new CModelSelectionParameters("scale");
	c1->append_child(c4);
	c4->build_values(1.0, 1.0, R_LINEAR);

	CModelSelectionParameters* c5=new CModelSelectionParameters("kernel",
			kernel);
	c1->append_child(c5);

	CModelSelectionParameters* c6=new CModelSelectionParameters("width");
	c5->append_child(c6);
	c6->build_values(1.0, 4.0, R_LINEAR);

	return root;
}

void test()
{
	/* create some easy regression data: 1d noisy sine wave */
	index_t n=100;
	float64_t x_range=6;

	SGMatrix<float64_t> X(1, n);
	SGMatrix<float64_t> X_test(1, n);
	SGVector<float64_t> Y(n);

	for (index_t i=0; i<n; ++i)
	{
		X[i]=CMath::random(0.0, x_range);
		X_test[i]=(float64_t)i/n*x_range;
		Y[i]=CMath::sin(X[i]);
	}

	/* shogun representation */
	CDenseFeatures<float64_t>* feat_train=new CDenseFeatures<float64_t>(X);
	CDenseFeatures<float64_t>* feat_test=new CDenseFeatures<float64_t>(X_test);
	CRegressionLabels* label_train=new CRegressionLabels(Y);

	// width: 0.000002; scale: 1.400435; sigma: 95.395590

	/* specity GPR with exact inference */
	float64_t width=0.000002;
	float64_t scale=1.400435;
	float64_t sigma=95.395590;

	CGaussianKernel* kernel=new CGaussianKernel(10, width);
	CZeroMean* mean=new CZeroMean();
	CGaussianLikelihood* lik=new CGaussianLikelihood(sigma);
	CExactInferenceMethod* inf=new CExactInferenceMethod(kernel, feat_train,
			mean, label_train, lik);
	inf->set_scale(scale);
	CGaussianProcessRegression* gp=new CGaussianProcessRegression(inf);

	CModelSelectionParameters* tree=build_modelselection_tree(inf, lik, kernel);
	tree->print_tree();

	CGradientCriterion* crit=new CGradientCriterion();
	CGradientEvaluation* grad=new CGradientEvaluation(gp, feat_train,
			label_train, crit);
	grad->set_function(inf);
	CGradientModelSelection* grad_search=new CGradientModelSelection(tree,
			grad);

	grad->set_autolock(false);
	CParameterCombination* best_combination=grad_search->select_model(true);

	SG_UNREF(best_combination);
	SG_UNREF(grad_search);
	SG_UNREF(tree);
	SG_UNREF(feat_test);
	SG_UNREF(gp);
}

int main(int argc, char** argv)
{
	init_shogun_with_defaults();

	test();

	exit_shogun();
	return 0;
}
#else
int main(int argc, char **argv)
{
	return 0;
}
#endif