vnl_matrix <double> Normalize_Feature_Matrix(vnl_matrix<double> feats)
{
mbl_stats_nd stats;
for(int i = 0; i<feats.rows() ; ++i)
{
vnl_vector<double> temp_row = feats.get_row(i);
stats.obs(temp_row);
}
vnl_vector<double> std_vec = stats.sd();
vnl_vector<double> mean_vec = stats.mean();
for(int i = 0; i<feats.columns()-3 ; ++i)
{
vnl_vector<double> temp_col = feats.get_column(i);
if(std_vec(i) > 0)
{
for(int j =0; j<temp_col.size() ; ++j)
temp_col[j] = (temp_col[j] - mean_vec(i))/std_vec(i) ;
}
feats.set_column(i,temp_col);
}
return feats;
}
bool rgrsn_ldp::local_dynamic_programming(const vnl_matrix<double> & probMap, int nNeighborBin,
vcl_vector<int> & optimalBins)
{
const int N = probMap.rows();
const int nBin = probMap.cols();
// dynamic programming
vnl_matrix<double> accumulatedProbMap = vnl_matrix<double>(N, nBin);
accumulatedProbMap.fill(0.0);
vnl_matrix<int> lookbackTable = vnl_matrix<int>(N, nBin);
lookbackTable.fill(0);
// copy first row
for (int c = 0; c<probMap.cols(); c++) {
accumulatedProbMap[0][c] = probMap[0][c];
lookbackTable[0][c] = c;
}
for (int r = 1; r <N; r++) {
for (int c = 0; c<probMap.cols(); c++) {
// lookup all possible place in the window
double max_val = -1;
int max_index = -1;
for (int w = -nNeighborBin; w <= nNeighborBin; w++) {
if (c + w <0 || c + w >= probMap.cols()) {
continue;
}
double val = probMap[r][c] + accumulatedProbMap[r-1][c+w];
if (val > max_val) {
max_val = val;
max_index = c + w; // most probable path from the [r-1] row, in column c + w
}
}
assert(max_index != -1);
accumulatedProbMap[r][c] = max_val;
lookbackTable[r][c] = max_index;
}
}
// lookback the table
double max_prob = -1.0;
int max_prob_index = -1;
for (int c = 0; c<accumulatedProbMap.cols(); c++) {
if (accumulatedProbMap[N-1][c] > max_prob) {
max_prob = accumulatedProbMap[N-1][c];
max_prob_index = c;
}
}
// back track
optimalBins.push_back(max_prob_index);
for (int r = N-1; r > 0; r--) {
int bin = lookbackTable[r][optimalBins.back()];
optimalBins.push_back(bin);
}
assert(optimalBins.size() == N);
// vcl_reverse(optimalBins.begin(), optimalBins.end());
return true;
}
void denormalize(vnl_matrix<double>& x, const vnl_vector<double>& centroid, const double scale) {
int n = x.rows();
if (n==0) return;
int d = x.cols();
for (int i = 0; i < n; ++i) {
x.set_row(i, x.get_row(i) * scale + centroid);
}
}
//std::vector<double> Vectorize(const vnl_matrix<double> &M)
vnl_vector<double> Vectorize(const vnl_matrix<double> &M)
{
//std::vector<double> V;
vnl_vector<double> V(M.rows() * M.columns());
for (unsigned j = 0; j < M.rows(); j++)
{
for (unsigned i = 0; i < M.columns(); i++)
{
//V.push_back(M(i,j));
V[M.columns() * j + i] = M(i,j);
}
}
return V;
}
void compute_P(const vnl_matrix<double>& x,const vnl_matrix<double>& y, vnl_matrix<double>& P, double &E, double sigma, int outliers) {
double k;
k = -2*sigma*sigma;
//P.set_size(m,n); P.fill(0);
//vnl_vector<double> v_ij;
vnl_vector<double> column_sum;
int m = x.rows();
int s = y.rows();
int d = x.cols();
column_sum.set_size(s);
column_sum.fill(0);
double outlier_term = outliers*pow((2*sigma*sigma*3.1415926),0.5*d);
for (int i = 0; i < m; ++i) {
for (int j = 0; j < s; ++j) {
double r = 0;
for (int t = 0; t < d; ++t) {
r += (x(i,t) - y(j,t))*(x(i,t) - y(j,t));
}
P(i,j) = exp(r/k);
column_sum[j]+=P(i,j);
}
}
if (outliers!=0) {
for (int i = 0; i < s; ++i)
column_sum[i] += outlier_term;
}
if (column_sum.min_value()>(1e-12)) {
E = 0;
for (int i = 0; i < s; ++i) {
for (int j = 0; j < m; ++j){
P(j,i) = P(j,i)/column_sum[i];
}
E -= log(column_sum[i]);
}
//vcl_cerr< < s;
//vcl_cerr<<P.get_column(10);
}
else {
P.empty();
}
}
void normalize_same(vnl_matrix<double>& x,
vnl_vector<double>& centroid, double& scale) {
int n = x.rows();
if (n==0) return;
int d = x.cols();
for (int i = 0; i < n; ++i) {
x.set_row(i, (x.get_row(i) - centroid) / scale);
}
}
void vnl_flann::search(const vnl_matrix<double> & query_data,
vcl_vector<vcl_vector<int> > & indices,
vcl_vector<vcl_vector<double> > & dists, int knn) const
{
const int dim = (int)query_data.cols();
assert(dim == dim_);
Matrix<double> query_data_wrap((double *)query_data.data_block(), (int)query_data.rows(), dim);
index_.knnSearch(query_data_wrap, indices, dists, knn, flann::SearchParams(128));
}
bool rgrsn_ldp::dynamic_programming(const vnl_matrix<double> & data, double v_min, double v_max,
unsigned int nBin, int nJumpBin, unsigned int windowSize,
vnl_vector<double> & optimalSignal)
{
assert(v_min < v_max);
// raw data to probability map
const int N = data.rows();
vnl_matrix<double> probMap = vnl_matrix<double>(N, nBin);
double interval = (v_max - v_min)/nBin;
for (int r = 0; r<N; r++) {
for (int c = 0; c<data.cols(); c++) {
int num = value_to_bin_number(v_min, interval, data[r][c], nBin);
probMap[r][num] += 1.0;
}
}
probMap /= data.cols(); // normalize
// save prob
{
// vnl_matlab_filewrite awriter("prob.mat");
// awriter.write(probMap, "prob");
// printf("save to prob.mat.\n");
}
vcl_vector<double> optimalValues(N, 0);
vcl_vector<int> numValues(N, 0); // multiple values from local dynamic programming
for (int i = 0; i<=N - windowSize; i++) {
// get a local probMap;
vnl_matrix<double> localProbMap = probMap.extract(windowSize, probMap.cols(), i, 0);
vcl_vector<int> localOptimalBins;
rgrsn_ldp::local_dynamic_programming(localProbMap, nJumpBin, localOptimalBins);
assert(localOptimalBins.size() == windowSize);
for (int j = 0; j < localOptimalBins.size(); j++) {
numValues[j + i] += 1;
optimalValues[j + i] += bin_number_to_value(v_min, interval, localOptimalBins[j]);
}
// test
if (0 && i == 0)
{
printf("test first window output\n");
for (int j = 0; j<optimalValues.size() && j<windowSize; j++) {
printf("%f ", optimalValues[j]);
}
printf("\n");
}
}
//
for (int i = 0; i<optimalValues.size(); i++) {
optimalValues[i] /= numValues[i];
}
optimalSignal = vnl_vector<double>(&optimalValues[0], (int)optimalValues.size());
return true;
}
void LocalGeometryRef::Initialize(vnl_matrix<double> data)
{
this->num_row = data.rows();
this->num_col = data.columns();
this->data_matrix.set_size(this->num_row, this->num_col);
for(int row = 0; row < this->num_row; row++)
{
for(int col = 0; col < this->num_col; col++)
{
this->data_matrix(row, col) = data(row, col);
}
}
}
/*
Matlab code in cpd_G.m:
k=-2*beta^2;
[n, d]=size(x); [m, d]=size(y);
G=repmat(x,[1 1 m])-permute(repmat(y,[1 1 n]),[3 2 1]);
G=squeeze(sum(G.^2,2));
G=G/k;
G=exp(G);
*/
void ComputeGaussianKernel(const vnl_matrix<double>& model,
const vnl_matrix<double>& ctrl_pts,
vnl_matrix<double>& G, vnl_matrix<double>& K,
double lambda) {
int m,n,d;
m = model.rows();
n = ctrl_pts.rows();
d = ctrl_pts.cols();
//asssert(model.cols()==d);
//assert(lambda>0);
G.set_size(m,n);
GaussianAffinityMatrix(model.data_block(), ctrl_pts.data_block(),
m, n, d, lambda, G.data_block());
if (model == ctrl_pts) {
K = G;
} else {
K.set_size(n,n);
GaussianAffinityMatrix(ctrl_pts.data_block(), ctrl_pts.data_block(),
n, n, d, lambda, K.data_block());
}
}
void ComputeTPSKernelU(const vnl_matrix<double>& model,
const vnl_matrix<double>& ctrl_pts,
vnl_matrix<double>& U) {
int m = model.rows();
int n = ctrl_pts.rows();
int d = ctrl_pts.cols();
//asssert(model.cols()==d==(2|3));
//K.set_size(n, n);
//K.fill(0);
U.set_size(m, n);
U.fill(0);
double eps = 1e-006;
vnl_vector<double> v_ij;
for (int i = 0; i < m; ++i) {
for (int j = 0; j < n; ++j)
{
v_ij = model.get_row(i) - ctrl_pts.get_row(j);
double r = v_ij.two_norm();
U(i, j) = -r;
}
}
}
//Reshape the matrix : columns first ; Similar to MATLAB
vnl_matrix<double> MCLR_SM::Reshape_Matrix(vnl_matrix<double>mat,int r,int c )
{
if(mat.rows()*mat.cols() != r*c)
{
cout<< "Number of elements in the matrix/vector should be equal to the total number of elements in the reshaped matrix";
getchar();
exit(1);
}
vnl_matrix<double>reshaped_matrix;
reshaped_matrix.set_size(r,c);
int count = 0;
for(int j=0;j<c;++j)
{
for(int i=0;i<r;++i)
{
reshaped_matrix(i,j) = mat(count%mat.rows(),floor(static_cast<double>(count/mat.rows())));
count++;
}
}
return reshaped_matrix;
}
void WriteMatrixImage(const vnl_matrix<double> &M, const std::string &Filename)
{
vil_image_view<vxl_byte> Image(M.rows(), M.columns(), 1, 1); //(ni, nj, n_planes, n_interleaved_planes)
for (unsigned j = 0; j < Image.nj(); j++)
{
for (unsigned i = 0; i < Image.ni(); i++)
{
Image(i,j) = static_cast<vxl_byte>(255 * M(i,j));
}
}
vil_save(Image, Filename.c_str());
}
bool CloseEnough(const vnl_matrix<double> &M1, const vnl_matrix<double> &M2, const double eps)
{
unsigned int NumRows = M1.rows();
unsigned int NumCols = M1.columns();
if((M2.rows() != NumRows) || (M2.columns() != NumCols))
{
std::cout << "Dimensions do not match!" << std::endl;
return false;
}
for(unsigned int r = 0; r < NumRows; r++)
{
for(unsigned int c = 0; c < NumCols; c++)
{
if(fabs(M1(r,c) - M2(r,c)) > eps)
{
std::cout << "Failed comparison: " << "M1: " << M1(r,c) << " M2: " << M2(r,c) << " diff: " << fabs(M1(r,c) - M2(r,c)) << std::endl;
return false;
}
}
}
return true;
}
bool rgrsn_ldp::local_viterbi(const vnl_matrix<double> & data,
double resolution,
const vnl_vector<double> & transition,
unsigned int window_size,
vnl_vector<double> & optimal_signal)
{
assert(resolution > 0.0);
assert(transition.size()%2 == 1);
const double min_v = data.min_value();
const double max_v = data.max_value();
const int nBin = (max_v - min_v)/resolution;
// raw data to probability map
// quantilization
const int N = data.rows();
vnl_matrix<double> probMap = vnl_matrix<double>(N, nBin);
for (int r = 0; r<N; r++) {
for (int c = 0; c<data.cols(); c++) {
int num = value_to_bin_number(min_v, resolution, data[r][c], nBin);
probMap[r][num] += 1.0;
}
}
probMap /= data.cols(); // normalization
vcl_vector<double> optimalValues(N, 0);
vcl_vector<int> numValues(N, 0); // multiple values from local dynamic programming
for (int i = 0; i <= N - window_size; i++) {
// get a local probMap;
vnl_matrix<double> localProbMap = probMap.extract(window_size, probMap.cols(), i, 0);
vcl_vector<int> localOptimalBins;
rgrsn_ldp::viterbi(localProbMap, transition, localOptimalBins);
assert(localOptimalBins.size() == window_size);
for (int j = 0; j < localOptimalBins.size(); j++) {
double value = bin_number_to_value(min_v, resolution, localOptimalBins[j]);
numValues[j + i] += 1;
optimalValues[j + i] += value;
}
}
// average all optimal path as final result
for (int i = 0; i<optimalValues.size(); i++) {
optimalValues[i] /= numValues[i];
}
optimal_signal = vnl_vector<double>(&optimalValues[0], (int)optimalValues.size());
return true;
}
void itk::BiExpFitFunctor::operator()(vnl_matrix<double> & newSignal,const vnl_matrix<double> & SignalMatrix, const double & S0)
{
vnl_vector<double> initalGuess(3);
// initialize Least Squres Function
// SignalMatrix.cols() defines the number of shells points
lestSquaresFunction model(SignalMatrix.cols());
model.set_bvalues(m_BValueList);// set BValue Vector e.g.: [1000, 2000, 3000] <- shell b Values
// initialize Levenberg Marquardt
vnl_levenberg_marquardt minimizer(model);
minimizer.set_max_function_evals(1000); // Iterations
minimizer.set_f_tolerance(1e-10); // Function tolerance
// for each Direction calculate LSF Coeffs ADC & AKC
for(unsigned int i = 0 ; i < SignalMatrix.rows(); i++)
{
model.set_measurements(SignalMatrix.get_row(i));
model.set_reference_measurement(S0);
initalGuess.put(0, 0.f); // ADC_slow
initalGuess.put(1, 0.009f); // ADC_fast
initalGuess.put(2, 0.7f); // lambda
// start Levenberg-Marquardt
minimizer.minimize_without_gradient(initalGuess);
const double & ADC_slow = initalGuess.get(0);
const double & ADC_fast = initalGuess.get(1);
const double & lambda = initalGuess(2);
newSignal.put(i, 0, S0 * (lambda * std::exp(-m_TargetBvalue * ADC_slow) + (1-lambda)* std::exp(-m_TargetBvalue * ADC_fast)));
newSignal.put(i, 1, minimizer.get_end_error()); // RMS Error
//OUTPUT FOR EVALUATION
/*std::cout << std::scientific << std::setprecision(5)
<< ADC_slow << "," // lambda
<< ADC_fast << "," // alpha
<< lambda << "," // lambda
<< S0 << "," // S0 value
<< minimizer.get_end_error() << ","; // End error
for(unsigned int j = 0; j < SignalMatrix.get_row(i).size(); j++ ){
std::cout << std::scientific << std::setprecision(5) << SignalMatrix.get_row(i)[j]; // S_n Values corresponding to shell 1 to shell n
if(j != SignalMatrix.get_row(i).size()-1) std::cout << ",";
}
std::cout << std::endl;*/
}
}
int LoadMatrixFromTxt(const char* filename, vnl_matrix<double>& matrix) {
std::ifstream infile(filename, std::ios_base::in);
if (infile.is_open()) {
if (matrix.read_ascii(infile)) {
return matrix.rows();
} else {
std::cerr << "unable to parse input file " << filename
<< " as a matrix." << std::endl;
return -1;
}
} else {
std::cerr << "unable to open model file " << filename << std::endl;
return -1;
}
}
bool rgrsn_ldp::compact_transition_matrix(const vcl_vector<int> & fns,
const vcl_vector<double> & values,
vnl_matrix<double> & transition,
const double resolution)
{
assert(fns.size() == values.size());
double min_v = *vcl_min_element(values.begin(), values.end());
double max_v = *vcl_max_element(values.begin(), values.end());
unsigned num_bin = (max_v - min_v)/resolution;
unsigned max_bin_transition = 0; // biggest transition between frames quantized in bin
for (int i = 0; i<fns.size(); i++) {
if (fns[i] + 1 == fns[i+1]) {
double cur_v = values[i];
double next_v = values[i+1];
int cur_bin = value_to_bin_number(min_v, resolution, cur_v, num_bin);
int next_bin = value_to_bin_number(min_v, resolution, next_v, num_bin);
int dif = abs(next_bin - cur_bin);
if (dif > max_bin_transition) {
max_bin_transition = dif;
}
}
}
transition = vnl_matrix<double>(max_bin_transition * 2 + 1, num_bin, 0.0);
vnl_vector<double> column(num_bin, 0.0);
for (int i = 0; i<fns.size()-1; i++) {
if (fns[i] + 1 == fns[i+1]) {
double cur_v = values[i];
double next_v = values[i+1];
int cur_bin = value_to_bin_number(min_v, resolution, cur_v, num_bin);
int next_bin = value_to_bin_number(min_v, resolution, next_v, num_bin);
int row = max_bin_transition + (next_bin - cur_bin);
transition[row][cur_bin] += 1.0;
column[cur_bin] += 1.0;
}
}
// normalize each column
for (int r = 0; r < transition.rows(); r++) {
for (int c = 0; c < transition.cols(); c++) {
// transition[r][c] /= column[c];
}
}
return true;
}
int vnl_matrix2vtkPolyData(vtkPolyData* A, vnl_matrix<double>& matrix)
{
int dim = matrix.cols();;
int n = matrix.rows();
//int n = A->GetNumberOfPoints();
//matrix.set_size(n, dim);
double P[3];
for(int i = 0;i < n; i++)
{
for(int j = 0;j < dim; j++ )
{
P[j] = matrix(i,j);
}
A->GetPoints()->SetPoint(i,P);
}
return 1;
}
vnl_matrix<double> FiniteDiffOdfMaximaExtractionFilter< PixelType, ShOrder, NrOdfDirections>
::CalcShBasis(vnl_matrix<double>& sphCoords)
{
int M = sphCoords.rows();
int j, m; double mag, plm;
vnl_matrix<double> shBasis;
shBasis.set_size(M, m_NumCoeffs);
for (int p=0; p<M; p++)
{
j=0;
for (int l=0; l<=ShOrder; l=l+2)
for (m=-l; m<=l; m++)
{
switch (m_Toolkit)
{
case FSL:
plm = legendre_p<double>(l,abs(m),cos(sphCoords(p,0)));
mag = sqrt((double)(2*l+1)/(4.0*M_PI)*factorial<double>(l-abs(m))/factorial<double>(l+abs(m)))*plm;
if (m<0)
shBasis(p,j) = sqrt(2.0)*mag*cos(fabs((double)m)*sphCoords(p,1));
else if (m==0)
shBasis(p,j) = mag;
else
shBasis(p,j) = pow(-1.0, m)*sqrt(2.0)*mag*sin(m*sphCoords(p,1));
break;
case MRTRIX:
plm = legendre_p<double>(l,abs(m),-cos(sphCoords(p,0)));
mag = sqrt((double)(2*l+1)/(4.0*M_PI)*factorial<double>(l-abs(m))/factorial<double>(l+abs(m)))*plm;
if (m>0)
shBasis(p,j) = mag*cos(m*sphCoords(p,1));
else if (m==0)
shBasis(p,j) = mag;
else
shBasis(p,j) = mag*sin(-m*sphCoords(p,1));
break;
}
j++;
}
}
return shBasis;
}
void normalize(vnl_matrix<double>& x,
vnl_vector<double>& centroid, double& scale) {
int n = x.rows();
if (n==0) return;
int d = x.cols();
centroid.set_size(d);
vnl_vector<double> col;
for (int i = 0; i < d; ++i) {
col = x.get_column(i);
centroid(i) = col.mean();
}
for (int i = 0; i < n; ++i) {
x.set_row(i, x.get_row(i) - centroid);
}
scale = x.frobenius_norm() / sqrt(double(n));
x = x / scale;
}
void WriteMatrixImageScaled(const vnl_matrix<double> &M, const std::string &Filename)
{
vil_image_view<vxl_byte> Image(M.rows(), M.columns(), 1, 1); //(ni, nj, n_planes, n_interleaved_planes)
//double Max = Tools::VectorMax(Vectorize(M));
double Max = M.max_value();
for (unsigned j = 0; j < Image.nj(); j++)
{
for (unsigned i = 0; i < Image.ni(); i++)
{
Image(i,j) = static_cast<vxl_byte>(255 * M(i,j)/Max);
//cout << "M: " << M(i,j) << endl;
//cout << "Image: " << Image(i,j) << endl;
}
}
vil_save(Image, Filename.c_str());
}
vnl_vector<double> mitk::GeneralizedLinearModel::ExpMu(const vnl_matrix<double> &x)
{
LogItLinking link;
vnl_vector<double> mu(x.rows());
int cols = m_B.size();
for (unsigned int r = 0 ; r < mu.size(); ++r)
{
mu(r) = 0;
for (int c = 0; c < cols; ++c)
{
if (!m_AddConstantColumn)
mu(r) += x(r,c)*m_B(c);
else if ( c == 0)
mu(r) += m_B(c);
else
mu(r) += x(r,c-1)*m_B(c);
}
mu(r) = exp(-mu(r));
}
return mu;
}
void mitk::GeneralizedLinearModel::EstimatePermutation(const vnl_matrix<double> &xData)
{
v3p_netlib_integer rows = xData.rows();
v3p_netlib_integer cols = xData.cols();
if (m_AddConstantColumn)
++cols;
v3p_netlib_doublereal *x = new v3p_netlib_doublereal[rows* cols];
_UpdateXMatrix(xData, m_AddConstantColumn, x);
v3p_netlib_doublereal *qraux = new v3p_netlib_doublereal[cols];
v3p_netlib_integer *jpvt = new v3p_netlib_integer[cols];
std::fill_n(jpvt,cols,0);
v3p_netlib_doublereal *work = new v3p_netlib_doublereal[cols];
std::fill_n(work,cols,0);
v3p_netlib_integer job = 16;
// Make a call to Lapack-DQRDC which does QR with permutation
// Permutation is saved in JPVT.
v3p_netlib_dqrdc_(x, &rows, &rows, &cols, qraux, jpvt, work, &job);
double limit = std::abs(x[0]) * std::max(cols, rows) * std::numeric_limits<double>::epsilon();
// Calculate the rank of the matrix
int m_Rank = 0;
for (int i = 0; i <cols; ++i)
{
m_Rank += (std::abs(x[i*rows + i]) > limit) ? 1 : 0;
}
// Create a permutation vector
m_Permutation.set_size(m_Rank);
for (int i = 0; i < m_Rank; ++i)
{
m_Permutation(i) = jpvt[i]-1;
}
delete x;
delete qraux;
delete jpvt;
delete work;
}
void pick_indices(const vnl_matrix<double>&dist,
std::vector<int>&row_index, std::vector<int>&col_index,
const double threshold) {
int m = dist.rows();
int n = dist.cols();
vnl_vector<int> row_flag, col_flag;
col_flag.set_size(n); col_flag.fill(0);
row_flag.set_size(n); row_flag.fill(0);
for (int i = 0; i < m; ++i) {
double min_dist = dist.get_row(i).min_value();
if (min_dist < threshold) {
for (int j = 0; j < n; ++j){
if (dist(i,j) == min_dist && col_flag[j] == 0){
row_index.push_back(i);
row_flag[i] = 1;
col_index.push_back(j);
col_flag[j] = 1;
}
}
}
}
}
// Fills the value of the xData-matrix into the x-matrix. Adds a constant
// column if required. Permutes the rows corresponding to the permutation vector.
static void _UpdatePermXMatrix(const vnl_matrix<double> &xData, bool addConstant, const vnl_vector<unsigned int> &permutation, vnl_matrix<double> &x)
{
int rows = xData.rows();
int cols = permutation.size();
x.set_size(rows, cols);
for (int r=0; r < rows; ++r)
{
for (int c=0; c<cols; ++c)
{
unsigned int newCol = permutation(c);
if (!addConstant)
{
x(r, c) = xData(r,newCol);
} else if (newCol == 0)
{
x(r, c) = 1.0;
} else
{
x(r, c) = xData(r, newCol-1);
}
}
}
}
// Copy a vnl-matrix to an c-array with row-wise representation.
// Adds a constant column if required.
static void _UpdateXMatrix(const vnl_matrix<double> &xData, bool addConstant, v3p_netlib_doublereal *x)
{
v3p_netlib_integer rows = xData.rows();
v3p_netlib_integer cols = xData.cols();
if (addConstant)
++cols;
for (int r=0; r < rows; ++r)
{
for (int c=0; c <cols; ++c)
{
if (!addConstant)
{
x[c*rows + r] = xData(r,c);
} else if (c == 0)
{
x[c*rows + r] = 1.0;
} else
{
x[c*rows + r] = xData(r, c-1);
}
}
}
}
bool rgrsn_ldp::dynamic_programming(const vnl_matrix<double> & data,
double v_min, double v_max,
unsigned int nBin,
int nJumpBin,
unsigned int windowSize,
vnl_vector<double> & optimalSignal,
vnl_vector<double> & signal_variance)
{
assert(v_min < v_max);
// raw data to probability map
// quantilization
const int N = data.rows();
vnl_matrix<double> probMap = vnl_matrix<double>(N, nBin);
double interval = (v_max - v_min)/nBin;
for (int r = 0; r<N; r++) {
for (int c = 0; c<data.cols(); c++) {
int num = value_to_bin_number(v_min, interval, data[r][c], nBin);
probMap[r][num] += 1.0;
}
}
probMap /= data.cols(); // normalization
vcl_vector<double> optimalValues(N, 0);
vcl_vector<int> numValues(N, 0); // multiple values from local dynamic programming
vcl_vector<vcl_vector<double> > all_values(N);
for (int i = 0; i<=N - windowSize; i++) {
// get a local probMap;
vnl_matrix<double> localProbMap = probMap.extract(windowSize, probMap.cols(), i, 0);
vcl_vector<int> localOptimalBins;
rgrsn_ldp::local_dynamic_programming(localProbMap, nJumpBin, localOptimalBins);
assert(localOptimalBins.size() == windowSize);
for (int j = 0; j < localOptimalBins.size(); j++) {
double value = bin_number_to_value(v_min, interval, localOptimalBins[j]);
assert(j + i < N);
all_values[j + i].push_back(value);
numValues[j + i] += 1;
optimalValues[j + i] += value;
}
}
// mean value
for (int i = 0; i<optimalValues.size(); i++) {
optimalValues[i] /= numValues[i];
}
// variance
signal_variance = vnl_vector<double>(N, 0);
for (int i = 0; i<optimalValues.size(); i++) {
assert(all_values[i].size() > 0);
if (all_values[i].size() == 1) {
signal_variance[i] = 0.0001;
}
else
{
double dump_mean = 0.0;
double sigma = 0.0;
VnlPlus::mean_std(&all_values[i][0], (int)all_values[i].size(), dump_mean, sigma);
signal_variance[i] = sigma + 0.0001; // avoid zero
}
}
optimalSignal = vnl_vector<double>(&optimalValues[0], (int)optimalValues.size());
// save variance with the size of window size, for test purpose
if(0)
{
vcl_vector<vnl_vector<double> > all_value_vecs;
for (int i = 0; i<all_values.size(); i++) {
if (all_values[i].size() == windowSize) {
all_value_vecs.push_back(VnlPlus::vector_2_vec(all_values[i]));
}
}
vcl_string save_file("ldp_all_prediction.mat");
vnl_matlab_filewrite awriter(save_file.c_str());
awriter.write(VnlPlus::vector_2_mat(all_value_vecs), "ldp_all_opt_path");
printf("save to %s\n", save_file.c_str());
}
return true;
}
bool rgrsn_ldp::local_dynamic_programming_log(const vnl_matrix<double> & probMap, int nNeighborBin,
vcl_vector<int> & optimalBins)
{
// find minimum path
const int N = probMap.rows();
const int nBin = probMap.cols();
const double epsilon = 0.01;
vnl_matrix<double> negLogProbMap(N, nBin);
for (int r = 0; r<N; r++) {
for (int c = 0; c <nBin; c++) {
negLogProbMap(r, c) = -log(probMap(r, c) + epsilon);
}
}
// dynamic programming
vnl_matrix<double> accumulatedMap = vnl_matrix<double>(N, nBin);
accumulatedMap.fill(0.0);
vnl_matrix<int> lookbackTable = vnl_matrix<int>(N, nBin);
lookbackTable.fill(0);
// copy first row
for (int c = 0; c<negLogProbMap.cols(); c++) {
accumulatedMap[0][c] = negLogProbMap[0][c];
}
for (int r = 1; r <N; r++) {
for (int c = 0; c<negLogProbMap.cols(); c++) {
// lookup all possible place in the window
double min_val = INT_MAX;
int index = -1;
for (int w = -nNeighborBin; w <= nNeighborBin; w++) {
if (c + w <0 || c + w >= negLogProbMap.cols()) {
continue;
}
double val = negLogProbMap[r][c] + accumulatedMap[r-1][c+w];
if (val < min_val) {
min_val = val;
index = c + w;
}
}
assert(index != -1);
accumulatedMap[r][c] = min_val;
lookbackTable[r][c] = index;
}
}
// lookback the table
double min_val = INT_MAX;
int initIndex = -1;
for (int c = 0; c<accumulatedMap.cols(); c++) {
if (accumulatedMap[N-1][c] < min_val) {
min_val = accumulatedMap[N-1][c];
initIndex = c;
}
}
// back track
optimalBins.push_back(initIndex);
for (int r = N-1; r > 0; r--) {
int bin = lookbackTable[r][optimalBins.back()];
optimalBins.push_back(bin);
}
assert(optimalBins.size() == N);
vcl_reverse(optimalBins.begin(), optimalBins.end());
return true;
}
bool rgrsn_ldp::viterbi(const vnl_matrix<double> & prob_map, const vnl_vector<double> & transition,
vcl_vector<int> & optimal_bins)
{
const int N = prob_map.rows();
const int nBin = prob_map.cols();
const int nNeighborBin = transition.size()/2;
const double epsilon = 0.01;
// dynamic programming
vnl_matrix<double> log_accumulatedProbMap = vnl_matrix<double>(N, nBin);
log_accumulatedProbMap.fill(0.0);
vnl_matrix<int> lookbackTable = vnl_matrix<int>(N, nBin);
lookbackTable.fill(0);
// copy first row
for (int c = 0; c<prob_map.cols(); c++) {
log_accumulatedProbMap[0][c] = log(prob_map[0][c] + epsilon);
lookbackTable[0][c] = c;
}
vnl_vector<double> log_transition = vnl_vector<double>(transition.size(), 0);
for (int i = 0; i<transition.size(); i++) {
log_transition[i] = log(transition[i] + epsilon);
}
for (int r = 1; r <N; r++) {
for (int c = 0; c<prob_map.cols(); c++) {
// lookup all possible place in the window
double max_val = vcl_numeric_limits<int>::min();
int max_index = -1;
for (int w = -nNeighborBin; w <= nNeighborBin; w++) {
if (c + w < 0 || c + w >= prob_map.cols()) {
continue;
}
assert(w + nNeighborBin >= 0 && w + nNeighborBin < transition.size());
double val = log_accumulatedProbMap[r-1][c+w] + log_transition[w + nNeighborBin];
if (val > max_val) {
max_val = val;
max_index = c + w; // most probable path from the [r-1] row, in column c + w
}
}
assert(max_index != -1);
log_accumulatedProbMap[r][c] = max_val + log(prob_map[r][c] + epsilon);
lookbackTable[r][c] = max_index;
}
}
// lookback the table
double max_prob = vcl_numeric_limits<int>::min();
int max_prob_index = -1;
for (int c = 0; c<log_accumulatedProbMap.cols(); c++) {
if (log_accumulatedProbMap[N-1][c] > max_prob) {
max_prob = log_accumulatedProbMap[N-1][c];
max_prob_index = c;
}
}
// back track
optimal_bins.push_back(max_prob_index);
for (int r = N-1; r > 0; r--) {
int bin = lookbackTable[r][optimal_bins.back()];
optimal_bins.push_back(bin);
}
assert(optimal_bins.size() == N);
vcl_reverse(optimal_bins.begin(), optimal_bins.end());
return true;
}
