Page MenuHomeHEPForge
Files F11221278

test_LocalLogisticRegression.cc
No OneTemporary
Actions

test_LocalLogisticRegression.cc
View Options

#include "UnitTest++.h"
#include "test_utils.hh"
#include "npstat/stat/LocalLogisticRegression.hh"
#include "npstat/nm/ArrayND.hh"
#include "npstat/stat/DistributionsND.hh"
#include "geners/CPP11_array.hh"
using namespace npstat;
using namespace std;
#define KDDIM 2
#define NPOINTS 1000
namespace {
typedef CPP11_array<float,(KDDIM+1)> Point;
struct PassFunctor
{
inline bool operator()(const Point& pt) const
{
return pt[KDDIM] > 0.5;
}
};
TEST(LogisticRegressionOnKDTree)
{
const unsigned idim = KDDIM;
// Generate a random box
BoxND<double> box(idim);
for (unsigned i=0; i<idim; ++i)
box[i].setBounds(test_rng(), 1.0+test_rng());
// Generate a random array shape
ArrayShape shape(idim);
for (unsigned i=0; i<idim; ++i)
shape[i] = static_cast<unsigned>(3+5*test_rng());
// Generate a random weight array
ArrayND<double> w(shape);
const unsigned long arrlen = w.length();
for (unsigned long i=0; i<arrlen; ++i)
w.linearValue(i) = test_rng();
// Generate a density out of this data
std::vector<double> location(idim);
std::vector<double> scale(idim);
for (unsigned i=0; i<idim; ++i)
{
location[i] = box[i].min();
scale[i] = box[i].length();
}
BinnedDensityND dens(&location[0], &scale[0], idim, w, 0);
// Build the quadratic orthogonal polynomials out of it
QuadraticOrthoPolyND poly(dens, box,
&shape[0], shape.size());
// Generate a vector of points
std::vector<Point> data;
data.reserve(NPOINTS);
for (unsigned i=0; i<NPOINTS; ++i)
{
Point p;
for (unsigned i=0; i<=KDDIM; ++i)
p[i] = test_rng();
data.push_back(p);
}
// Generate the tree
unsigned dimUse[KDDIM] = {1U, 2U};
KDTree<Point,double> tree(data, dimUse, KDDIM);
// Generate the logistic regression object
PassFunctor f;
LogisticRegressionOnKDTree<Point,double,PassFunctor> reg(
tree, f, poly, true);
// Check the gradient calculations by this object
const unsigned nterms = poly.nTerms();
std::vector<double> coeffs(nterms);
std::vector<double> grad(nterms);
for (unsigned i=0; i<nterms; ++i)
coeffs[i] = test_rng();
reg.setRegressionBox(BoxND<double>::unitBox(KDDIM));
reg.calculateLogLikelihood(&coeffs[0], nterms);
reg.getGradient(&grad[0], grad.size());
double gEps = 1.e-7;
for (unsigned i=0; i<nterms; ++i)
{
const double center = coeffs[i];
volatile const double plus = center + gEps;
volatile const double minus = center - gEps;
coeffs[i] = plus;
const double vp = reg.calculateLogLikelihood(&coeffs[0], nterms);
coeffs[i] = minus;
const double vm = reg.calculateLogLikelihood(&coeffs[0], nterms);
coeffs[i] = center;
const double numgrad = (vp - vm)/(plus - minus);
CHECK_CLOSE(numgrad, grad[i], 1.e-4);
}
// Check the gradient calculations for some crazy
// values of the coefficients
gEps = 1.e13;
for (int j=-1; j<2; j+=2)
{
for (unsigned i=0; i<nterms; ++i)
{
const double oldval = coeffs[i];
coeffs[i] = j*1.e20;
reg.calculateLogLikelihood(&coeffs[0], nterms);
reg.getGradient(&grad[0], grad.size());
volatile const double plus = coeffs[i] + gEps;
volatile const double minus = coeffs[i] - gEps;
coeffs[i] = plus;
const double vp = reg.calculateLogLikelihood(&coeffs[0], nterms);
coeffs[i] = minus;
const double vm = reg.calculateLogLikelihood(&coeffs[0], nterms);
const double numgrad = (vp - vm)/(plus - minus);
CHECK_CLOSE(numgrad, grad[i], 1.e-5);
coeffs[i] = oldval;
}
}
}
TEST(LogisticRegressionOnGrid)
{
// const double eps = 1.e-8;
const unsigned idim = KDDIM;
// Generate a random box
BoxND<double> box(idim);
for (unsigned i=0; i<idim; ++i)
box[i].setBounds(test_rng(), 1.0+test_rng());
// Generate a random array shape
ArrayShape shape(idim);
for (unsigned i=0; i<idim; ++i)
shape[i] = static_cast<unsigned>(3+5*test_rng());
// Generate a random weight array
ArrayND<double> w(shape);
const unsigned long arrlen = w.length();
for (unsigned long i=0; i<arrlen; ++i)
w.linearValue(i) = test_rng();
// Generate a density out of this data
std::vector<double> location(idim);
std::vector<double> scale(idim);
for (unsigned i=0; i<idim; ++i)
{
location[i] = box[i].min();
scale[i] = box[i].length();
}
BinnedDensityND dens(&location[0], &scale[0], idim, w, 0);
// Build the quadratic orthogonal polynomials out of it
QuadraticOrthoPolyND poly(dens, box,
&shape[0], shape.size());
// Generate the denominator array
ArrayND<int> denom(shape);
const unsigned len = denom.length();
for (unsigned i=0; i<len; ++i)
denom.linearValue(i) = static_cast<int>(3 + 15*test_rng());
// Generate the numerator array
ArrayND<int> num(shape);
for (unsigned i=0; i<len; ++i)
num.linearValue(i) = static_cast<int>(test_rng()*denom.linearValue(i));
// Generate the logistic regression object
LogisticRegressionOnGrid<int> reg(num, denom, poly, true);
// Check the gradient calculations by this object
const unsigned nterms = poly.nTerms();
std::vector<double> coeffs(nterms);
std::vector<double> grad(nterms);
for (unsigned i=0; i<nterms; ++i)
coeffs[i] = test_rng();
reg.setRegressionBox(BoxND<int>(shape));
reg.calculateLogLikelihood(&coeffs[0], nterms);
reg.getGradient(&grad[0], grad.size());
double gEps = 1.e-7;
for (unsigned i=0; i<nterms; ++i)
{
const double center = coeffs[i];
volatile const double plus = center + gEps;
volatile const double minus = center - gEps;
coeffs[i] = plus;
const double vp = reg.calculateLogLikelihood(&coeffs[0], nterms);
coeffs[i] = minus;
const double vm = reg.calculateLogLikelihood(&coeffs[0], nterms);
coeffs[i] = center;
const double numgrad = (vp - vm)/(plus - minus);
CHECK_CLOSE(0.0, (numgrad - grad[i])/
(fabs(numgrad) + fabs(grad[i])), 2.e-6);
}
// Check the gradient calculations for some crazy
// values of the coefficients
gEps = 1.e13;
for (int j=-1; j<2; j+=2)
{
for (unsigned i=0; i<nterms; ++i)
{
const double oldval = coeffs[i];
coeffs[i] = j*1.e20;
reg.calculateLogLikelihood(&coeffs[0], nterms);
reg.getGradient(&grad[0], grad.size());
volatile const double plus = coeffs[i] + gEps;
volatile const double minus = coeffs[i] - gEps;
coeffs[i] = plus;
const double vp = reg.calculateLogLikelihood(&coeffs[0], nterms);
coeffs[i] = minus;
const double vm = reg.calculateLogLikelihood(&coeffs[0], nterms);
const double numgrad = (vp - vm)/(plus - minus);
CHECK_CLOSE(0.0, (numgrad - grad[i])/
(fabs(numgrad) + fabs(grad[i])), 1.e-6);
coeffs[i] = oldval;
}
}
}
}

File Metadata

Mime Type
text/x-c
Expires
Wed, May 14, 10:10 AM (1 d, 13 h)
Storage Engine
blob
Storage Format
Raw Data
Storage Handle
5041499
Default Alt Text
test_LocalLogisticRegression.cc (7 KB)

Event Timeline