mirror of
https://github.com/aportelli/LatAnalyze.git
synced 2024-09-20 05:25:37 +01:00
significant optimisation of covariance routines + checks
This commit is contained in:
parent
c796187d1e
commit
57c6004797
@ -20,7 +20,8 @@ noinst_PROGRAMS = \
|
||||
exPValue \
|
||||
exRand \
|
||||
exRootFinder \
|
||||
exThreadPool
|
||||
exThreadPool \
|
||||
exVarBenchmark
|
||||
|
||||
exCompiledDoubleFunction_SOURCES = exCompiledDoubleFunction.cpp
|
||||
exCompiledDoubleFunction_CXXFLAGS = $(COM_CXXFLAGS)
|
||||
@ -78,4 +79,8 @@ exThreadPool_SOURCES = exThreadPool.cpp
|
||||
exThreadPool_CXXFLAGS = $(COM_CXXFLAGS)
|
||||
exThreadPool_LDFLAGS = -L../lib/.libs -lLatAnalyze
|
||||
|
||||
exVarBenchmark_SOURCES = exVarBenchmark.cpp
|
||||
exVarBenchmark_CXXFLAGS = $(COM_CXXFLAGS)
|
||||
exVarBenchmark_LDFLAGS = -L../lib/.libs -lLatAnalyze
|
||||
|
||||
ACLOCAL_AMFLAGS = -I .buildutils/m4
|
||||
|
||||
119
examples/exVarBenchmark.cpp
Normal file
119
examples/exVarBenchmark.cpp
Normal file
@ -0,0 +1,119 @@
|
||||
#include <LatAnalyze/Io/Io.hpp>
|
||||
#include <LatAnalyze/Functional/CompiledFunction.hpp>
|
||||
#include <LatAnalyze/Core/Plot.hpp>
|
||||
#include <LatAnalyze/Statistics/Random.hpp>
|
||||
#include <LatAnalyze/Statistics/MatSample.hpp>
|
||||
|
||||
using namespace std;
|
||||
using namespace Latan;
|
||||
|
||||
constexpr Index size = 1000;
|
||||
constexpr Index nSample = 2000;
|
||||
|
||||
int main(void)
|
||||
{
|
||||
random_device rd;
|
||||
DMat var(size, size);
|
||||
DVec mean(size);
|
||||
DMatSample sample(nSample, size, 1), sample2(nSample, size, 1);
|
||||
|
||||
cout << "-- generating " << nSample << " Gaussian random vectors..." << endl;
|
||||
var = DMat::Random(size, size);
|
||||
var *= var.adjoint();
|
||||
mean = DVec::Random(size);
|
||||
RandomNormal mgauss(mean, var, rd());
|
||||
sample[central] = mgauss();
|
||||
FOR_STAT_ARRAY(sample, s)
|
||||
{
|
||||
sample[s] = mgauss();
|
||||
}
|
||||
sample2[central] = mgauss();
|
||||
FOR_STAT_ARRAY(sample, s)
|
||||
{
|
||||
sample2[s] = mgauss();
|
||||
}
|
||||
|
||||
cout << "-- check new routines" << endl;
|
||||
|
||||
DMat v, vo;
|
||||
|
||||
cout << "var" << endl;
|
||||
auto start = chrono::high_resolution_clock::now();
|
||||
vo = sample.varianceOld();
|
||||
auto end = chrono::high_resolution_clock::now();
|
||||
chrono::duration<double> diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
v = sample.variance();
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
cout << "diff " << (v - vo).norm() << endl;
|
||||
|
||||
cout << "cov" << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
vo = sample.covarianceOld(sample2);
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
v = sample.covariance(sample2);
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
cout << "diff " << (v - vo).norm() << endl;
|
||||
|
||||
cout << "mean" << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
vo = sample.meanOld(3, 5);
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
v = sample.mean(3, 5);
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
cout << "diff " << (v - vo).norm() << endl;
|
||||
|
||||
cout << "varmat" << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
vo = sample.varianceMatrixOld();
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
v = sample.varianceMatrix();
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
cout << "diff " << (v - vo).norm() << endl;
|
||||
|
||||
cout << "covarmat" << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
vo = sample.covarianceMatrixOld(sample2);
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
v = sample.covarianceMatrix(sample2);
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
cout << "diff " << (v - vo).norm() << endl;
|
||||
|
||||
cout << "corrmat" << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
vo = sample.correlationMatrixOld();
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
start = chrono::high_resolution_clock::now();
|
||||
v = sample.correlationMatrix();
|
||||
end = chrono::high_resolution_clock::now();
|
||||
diff = end - start;
|
||||
cout << "time " << diff.count() << endl;
|
||||
cout << "diff " << (v - vo).norm() << endl;
|
||||
|
||||
return EXIT_SUCCESS;
|
||||
}
|
||||
@ -103,6 +103,10 @@ public:
|
||||
const Index nCol);
|
||||
// resize all matrices
|
||||
void resizeMat(const Index nRow, const Index nCol);
|
||||
// covariance matrix
|
||||
Mat<T> covarianceMatrix(const MatSample<T> &sample) const;
|
||||
Mat<T> varianceMatrix(void) const;
|
||||
Mat<T> correlationMatrix(void) const;
|
||||
};
|
||||
|
||||
// non-member operators
|
||||
@ -379,6 +383,78 @@ void MatSample<T>::resizeMat(const Index nRow, const Index nCol)
|
||||
}
|
||||
}
|
||||
|
||||
// covariance matrix ///////////////////////////////////////////////////////////
|
||||
template <typename T>
|
||||
Mat<T> MatSample<T>::covarianceMatrix(const MatSample<T> &sample) const
|
||||
{
|
||||
if (((*this)[central].cols() != 1) or (sample[central].cols() != 1))
|
||||
{
|
||||
LATAN_ERROR(Size, "samples have more than one column");
|
||||
}
|
||||
|
||||
Index n1 = (*this)[central].rows(), n2 = sample[central].rows();
|
||||
Index nSample = this->size();
|
||||
Mat<T> tmp1(n1, nSample), tmp2(n2, nSample), res(n1, n2);
|
||||
Mat<T> s1(n1, 1), s2(n2, 1), one(nSample, 1);
|
||||
|
||||
one.fill(1.);
|
||||
s1.fill(0.);
|
||||
s2.fill(0.);
|
||||
for (unsigned int s = 0; s < nSample; ++s)
|
||||
{
|
||||
s1 += (*this)[s];
|
||||
tmp1.col(s) = (*this)[s];
|
||||
}
|
||||
tmp1 -= s1*one.transpose()/static_cast<double>(nSample);
|
||||
for (unsigned int s = 0; s < nSample; ++s)
|
||||
{
|
||||
s2 += sample[s];
|
||||
tmp2.col(s) = sample[s];
|
||||
}
|
||||
tmp2 -= s2*one.transpose()/static_cast<double>(nSample);
|
||||
res = tmp1*tmp2.transpose()/static_cast<double>(nSample - 1);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
Mat<T> MatSample<T>::varianceMatrix(void) const
|
||||
{
|
||||
if ((*this)[central].cols() != 1)
|
||||
{
|
||||
LATAN_ERROR(Size, "samples have more than one column");
|
||||
}
|
||||
|
||||
Index n1 = (*this)[central].rows();
|
||||
Index nSample = this->size();
|
||||
Mat<T> tmp1(n1, nSample), res(n1, n1);
|
||||
Mat<T> s1(n1, 1), one(nSample, 1);
|
||||
|
||||
one.fill(1.);
|
||||
s1.fill(0.);
|
||||
for (unsigned int s = 0; s < nSample; ++s)
|
||||
{
|
||||
s1 += (*this)[s];
|
||||
tmp1.col(s) = (*this)[s];
|
||||
}
|
||||
tmp1 -= s1*one.transpose()/static_cast<double>(nSample);
|
||||
res = tmp1*tmp1.transpose()/static_cast<double>(nSample - 1);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
Mat<T> MatSample<T>::correlationMatrix(void) const
|
||||
{
|
||||
Mat<T> res = varianceMatrix();
|
||||
Mat<T> invDiag(res.rows(), 1);
|
||||
|
||||
invDiag = res.diagonal();
|
||||
invDiag = invDiag.cwiseInverse().cwiseSqrt();
|
||||
res = (invDiag*invDiag.transpose()).cwiseProduct(res);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
END_LATAN_NAMESPACE
|
||||
|
||||
|
||||
@ -51,14 +51,17 @@ public:
|
||||
const T & operator[](const Index s) const;
|
||||
// statistics
|
||||
void bin(Index binSize);
|
||||
T sum(const Index pos = 0, const Index n = -1) const;
|
||||
T meanOld(const Index pos = 0, const Index n = -1) const;
|
||||
T mean(const Index pos = 0, const Index n = -1) const;
|
||||
T covariance(const StatArray<T, os> &array, const Index pos = 0,
|
||||
const Index n = -1) const;
|
||||
T covarianceMatrix(const StatArray<T, os> &array, const Index pos = 0,
|
||||
T covarianceOld(const StatArray<T, os> &array) const;
|
||||
T covariance(const StatArray<T, os> &array) const;
|
||||
T covarianceMatrixOld(const StatArray<T, os> &array, const Index pos = 0,
|
||||
const Index n = -1) const;
|
||||
T variance(const Index pos = 0, const Index n = -1) const;
|
||||
T varianceMatrix(const Index pos = 0, const Index n = -1) const;
|
||||
T correlationMatrix(const Index pos = 0, const Index n = -1) const;
|
||||
T varianceOld(void) const;
|
||||
T variance(void) const;
|
||||
T varianceMatrixOld(const Index pos = 0, const Index n = -1) const;
|
||||
T correlationMatrixOld(const Index pos = 0, const Index n = -1) const;
|
||||
// IO type
|
||||
virtual IoType getType(void) const;
|
||||
public:
|
||||
@ -66,7 +69,7 @@ public:
|
||||
};
|
||||
|
||||
// reduction operations
|
||||
namespace ReducOp
|
||||
namespace StatOp
|
||||
{
|
||||
// general templates
|
||||
template <typename T>
|
||||
@ -149,42 +152,82 @@ void StatArray<T, os>::bin(Index binSize)
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::mean(const Index pos, const Index n) const
|
||||
T StatArray<T, os>::meanOld(const Index pos, const Index n) const
|
||||
{
|
||||
T result = T();
|
||||
const Index m = (n >= 0) ? n : size();
|
||||
|
||||
if (m)
|
||||
{
|
||||
result = this->segment(pos+os, m).redux(&ReducOp::sum<T>);
|
||||
result = this->segment(pos+os, m).redux(&StatOp::sum<T>);
|
||||
}
|
||||
return result/static_cast<double>(m);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::covariance(const StatArray<T, os> &array, const Index pos,
|
||||
const Index n) const
|
||||
T StatArray<T, os>::sum(const Index pos, const Index n) const
|
||||
{
|
||||
T result;
|
||||
const Index m = (n >= 0) ? n : size();
|
||||
|
||||
result = (*this)[pos];
|
||||
for (Index i = pos + 1; i < pos + m; ++i)
|
||||
{
|
||||
result += (*this)[i];
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::mean(const Index pos, const Index n) const
|
||||
{
|
||||
const Index m = (n >= 0) ? n : size();
|
||||
|
||||
return sum(pos, n)/static_cast<double>(m);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::covarianceOld(const StatArray<T, os> &array) const
|
||||
{
|
||||
T s1, s2, prs, res = T();
|
||||
const Index m = (n >= 0) ? n : size();
|
||||
const Index m = size();
|
||||
|
||||
if (m)
|
||||
{
|
||||
auto arraySeg = array.segment(pos+os, m);
|
||||
auto thisSeg = this->segment(pos+os, m);
|
||||
auto arraySeg = array.segment(os, m);
|
||||
auto thisSeg = this->segment(os, m);
|
||||
|
||||
s1 = thisSeg.redux(&ReducOp::sum<T>);
|
||||
s2 = arraySeg.redux(&ReducOp::sum<T>);
|
||||
prs = thisSeg.binaryExpr(arraySeg, &ReducOp::prod<T>)
|
||||
.redux(&ReducOp::sum<T>);
|
||||
res = prs - ReducOp::prod(s1, s2)/static_cast<double>(m);
|
||||
s1 = thisSeg.redux(&StatOp::sum<T>);
|
||||
s2 = arraySeg.redux(&StatOp::sum<T>);
|
||||
prs = thisSeg.binaryExpr(arraySeg, &StatOp::prod<T>)
|
||||
.redux(&StatOp::sum<T>);
|
||||
res = prs - StatOp::prod(s1, s2)/static_cast<double>(m);
|
||||
}
|
||||
|
||||
return res/static_cast<double>(m - 1);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::covarianceMatrix(const StatArray<T, os> &array,
|
||||
T StatArray<T, os>::covariance(const StatArray<T, os> &array) const
|
||||
{
|
||||
T s1, s2, res;
|
||||
|
||||
s1 = array.sum();
|
||||
s2 = this->sum();
|
||||
res = StatOp::prod<T>(array[0], (*this)[0]);
|
||||
for (Index i = 1; i < size(); ++i)
|
||||
{
|
||||
res += StatOp::prod<T>(array[i], (*this)[i]);
|
||||
}
|
||||
res -= StatOp::prod<T>(s1, s2)/static_cast<double>(size());
|
||||
res /= static_cast<double>(size() - 1);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::covarianceMatrixOld(const StatArray<T, os> &array,
|
||||
const Index pos, const Index n) const
|
||||
{
|
||||
T s1, s2, prs, res = T();
|
||||
@ -195,32 +238,38 @@ T StatArray<T, os>::covarianceMatrix(const StatArray<T, os> &array,
|
||||
auto arraySeg = array.segment(pos+os, m);
|
||||
auto thisSeg = this->segment(pos+os, m);
|
||||
|
||||
s1 = thisSeg.redux(&ReducOp::sum<T>);
|
||||
s2 = arraySeg.redux(&ReducOp::sum<T>);
|
||||
prs = thisSeg.binaryExpr(arraySeg, &ReducOp::tensProd<T>)
|
||||
.redux(&ReducOp::sum<T>);
|
||||
res = prs - ReducOp::tensProd(s1, s2)/static_cast<double>(m);
|
||||
s1 = thisSeg.redux(&StatOp::sum<T>);
|
||||
s2 = arraySeg.redux(&StatOp::sum<T>);
|
||||
prs = thisSeg.binaryExpr(arraySeg, &StatOp::tensProd<T>)
|
||||
.redux(&StatOp::sum<T>);
|
||||
res = prs - StatOp::tensProd(s1, s2)/static_cast<double>(m);
|
||||
}
|
||||
|
||||
return res/static_cast<double>(m - 1);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::variance(const Index pos, const Index n) const
|
||||
T StatArray<T, os>::variance(void) const
|
||||
{
|
||||
return covariance(*this, pos, n);
|
||||
return covariance(*this);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::varianceMatrix(const Index pos, const Index n) const
|
||||
T StatArray<T, os>::varianceOld(void) const
|
||||
{
|
||||
return covarianceMatrix(*this, pos, n);
|
||||
return covarianceOld(*this);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::correlationMatrix(const Index pos, const Index n) const
|
||||
T StatArray<T, os>::varianceMatrixOld(const Index pos, const Index n) const
|
||||
{
|
||||
T res = varianceMatrix(pos, n);
|
||||
return covarianceMatrixOld(*this, pos, n);
|
||||
}
|
||||
|
||||
template <typename T, Index os>
|
||||
T StatArray<T, os>::correlationMatrixOld(const Index pos, const Index n) const
|
||||
{
|
||||
T res = varianceMatrixOld(pos, n);
|
||||
T invDiag(res.rows(), 1);
|
||||
|
||||
invDiag = res.diagonal();
|
||||
@ -231,8 +280,14 @@ T StatArray<T, os>::correlationMatrix(const Index pos, const Index n) const
|
||||
}
|
||||
|
||||
// reduction operations ////////////////////////////////////////////////////////
|
||||
namespace ReducOp
|
||||
namespace StatOp
|
||||
{
|
||||
template <typename T>
|
||||
inline void zero(T &a)
|
||||
{
|
||||
a = 0.;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
inline T sum(const T &a, const T &b)
|
||||
{
|
||||
|
||||
Loading…
Reference in New Issue
Block a user