Files
mlpack/fastlib/u/nvasil/kernel_pca/kernel_pca_impl.h
T

284 lines
9.3 KiB
C++

/*
* =====================================================================================
*
* Filename: kernel_pca_impl.h
*
* Description:
*
* Version: 1.0
* Created: 11/30/2007 09:03:12 PM EST
* Revision: none
* Compiler: gcc
*
* Author: Nikolaos Vasiloglou (NV), nvasil@ieee.org
* Company: Georgia Tech Fastlab-ESP Lab
*
* =====================================================================================
*/
void KernelPCA::Init(std::string data_file, index_t knns,
index_t leaf_size) {
data::Load(data_file.c_str(), &data_);
dimension_ = data_.n_rows();
knns_ = knns;
allknn_.Init(data_, data_, leaf_size, knns);
}
void KernelPCA::Destruct() {
}
void KernelPCA::ComputeNeighborhoods() {
NOTIFY("Building tree...\n");
fflush(stdout);
ArrayList<index_t> resulting_neighbors;
ArrayList<double> distances;
NOTIFY("Computing Neighborhoods");
allknn_.ComputeNeighbors(&resulting_neighbors,
&distances);
FILE *fp=fopen("allnn.txt", "w");
if (fp==NULL) {
FATAL("Unable to open allnn for exporting the results, error %s\n",
strerror(errno));
}
for(index_t i=0; i<resulting_neighbors.size(); i++) {
fprintf(fp, "%lli %lli %lg\n", i / knns_, resulting_neighbors[i],
distances[i]);
}
fclose(fp);
}
template<typename DISTANCEKERNEL>
void KernelPCA::ComputeGeneralKernelPCA(DISTANCEKERNEL kernel,
index_t num_of_eigenvalues,
Matrix *eigen_vectors,
Vector *eigen_values){
kernel_matrix_.Copy(affinity_matrix_);
kernel_matrix_.ApplyFunction(kernel);
Vector temp;
temp.Init(kernel_matrix_.dimension());
temp.SetAll(1.0);
kernel_matrix_.SetDiagonal(temp);
kernel_matrix_.EndLoading();
NOTIFY("Computing eigen values...\n");
kernel_matrix_.Eig(num_of_eigenvalues,
"LM",
eigen_vectors,
eigen_values, NULL);
}
void KernelPCA::LoadAffinityMatrix() {
affinity_matrix_.Init("allnn.txt");
affinity_matrix_.MakeSymmetric();
}
void KernelPCA::SaveToTextFile(std::string file,
Matrix &eigen_vectors,
Vector &eigen_values) {
std::string vec_file(file);
vec_file.append(".vectors");
std::string lam_file(file);
lam_file.append(".lambdas");
FILE *fp = fopen(vec_file.c_str(), "w");
if (unlikely(fp==NULL)) {
FATAL("Unable to open file %s, error: %s", vec_file.c_str(),
strerror(errno));
}
for(index_t i=0; i<eigen_vectors.n_rows(); i++) {
for(index_t j=0; j<eigen_vectors.n_cols(); j++) {
fprintf(fp, "%lg\t", eigen_vectors.get(i, j));
}
fprintf(fp, "\n");
}
fclose(fp);
fp = fopen(lam_file.c_str(), "w");
if (unlikely(fp==NULL)) {
FATAL("Unable to open file %s, error: %s", lam_file.c_str(),
strerror(errno));
}
for(index_t i=0; i<(index_t)eigen_values.length(); i++) {
fprintf(fp, "%lg\n", eigen_values[i]);
}
fclose(fp);
}
void KernelPCA::EstimateBandwidth(double *bandwidth) {
FILE *fp=fopen("allnn.txt", "r");
if unlikely(fp==NULL) {
FATAL("Unable to open allnn.txt, error %s\n", strerror(errno));
}
uint64 p1, p2;
double dist;
double mean=0;
uint64 count=0;
while (!feof(fp)) {
fscanf(fp, "%llu %llu %lg", &p1, &p2, &dist);
mean+=dist;
count++;
}
*bandwidth=mean/count;
}
// It is not always know if the k nearest neighbors includes the same point with
// 0 distance. It is highly likely in cases where the query tree and the reference tree
// come from different structures that refer to the same dataset. We take care about
// that in this function.
void KernelPCA::ComputeLLE(index_t num_of_eigenvalues,
Matrix *eigen_vectors,
Vector *eigen_values) {
FILE *fp=fopen("allnn.txt", "r");
if unlikely(fp==NULL) {
FATAL("Unable to open allnn.txt, error %s\n", strerror(errno));
}
uint64 p1, p2;
double dist;
uint64 last_point=numeric_limits<uint64>::max();
Vector point;
point.Init(dimension_);
Matrix neighbor_vals;
// We initialize everything with knns_ although it is highly likely that
// if the k nearest neighbors include the same point then we will need
// a smaller vector knns_-1
neighbor_vals.Init(dimension_, knns_-1);
Matrix covariance(knns_-1, knns_-1);
Vector ones;
ones.Init(knns_-1);
ones.SetAll(1);
Vector weights;
index_t neighbors[knns_];
index_t i=0;
kernel_matrix_.Init(data_.n_cols(),
data_.n_cols(),
knns_);
last_point=0;
point.CopyValues(data_.GetColumnPtr(0));
// Create a unitary matrix
Matrix covariance_regularizer;
covariance_regularizer.Init(knns_-1, knns_-1);
covariance_regularizer.SetZero();
for(index_t j=0; j<knns_-1; j++) {
covariance_regularizer.set(j, j, 1);
}
while (!feof(fp)) {
fscanf(fp, "%llu %llu %lg\n", &p1, &p2, &dist);
if (dist==0) {
continue;
}
if (p1!=last_point) {
point.CopyValues(data_.GetColumnPtr(p1));
last_point=p1;
la::MulTransAOverwrite(neighbor_vals, neighbor_vals, &covariance);
// calculate the covariance matrix trace
double trace=0;
for (index_t k=0; k<i; k++) {
trace+=covariance.get(k,k);
}
la::AddExpert(1e-3*trace, covariance_regularizer, &covariance);
la::SolveInit(covariance, ones, &weights);
double sum_weights=0;
for(index_t k=0; k<weights.length(); k++) {
sum_weights+=weights[k];
}
for(index_t k=0; k<weights.length(); k++) {
weights[k]/=sum_weights;
}
kernel_matrix_.LoadRow(p1, i, neighbors, weights.ptr());
i=0;
weights.Destruct();
}
memcpy(neighbor_vals.GetColumnPtr(i), data_.GetColumnPtr(p2),
sizeof(double)*dimension_);
neighbors[i]=p2;
la::SubFrom(dimension_, point.ptr(), neighbor_vals.GetColumnPtr(i));
i++;
}
kernel_matrix_.Negate();
kernel_matrix_.SetDiagonal(1.0);
NONFATAL("Computing eigen values...\n");
SparseMatrix kernel_matrix1;
Sparsem::MultiplyT(kernel_matrix_, &kernel_matrix1);
kernel_matrix1.ToFile("lle_mat.txt");
kernel_matrix1.EndLoading();
kernel_matrix1.Eig(num_of_eigenvalues,
"SM",
eigen_vectors,
eigen_values, NULL);
}
template<typename DISTANCEKERNEL>
void KernelPCA::ComputeSpectralRegression(DISTANCEKERNEL kernel,
std::map<index_t, index_t> &data_label,
Matrix *embedded_coordinates,
Vector *eigenvalues) {
// labels has the label of every point, it is not necessary
// for every point to have a label, that's why we are using a map and not
// a vector
// This map has the classes and the points
std::map<index_t, std::vector<index_t> > classes;
std::vector<index_t> default_bin;
// find how many classes we have
index_t num_of_classes=0;
std::map<index_t, index_t>::iterator it;
for(it=data_label.begin(); it!=data_label.end(); it++) {
if (classes.find(it->second)!=classes.end()) {
classes[it->second].push_back(it->first);
} else {
classes.insert(make_pair(it->second, default_bin));
classes[it->second].push_back(it->first);
num_of_classes++;
}
}
kernel_matrix_.Copy(affinity_matrix_);
kernel_matrix_.ApplyFunction(kernel);
// In the paper it is also called W^{SR}
SparseMatrix labeled_graph;
labeled_graph.Init(data_.n_cols(), data_.n_cols(), 2*knns_);
// In the paper this is D^{SR}
SparseMatrix d_sr_mat;
d_sr_mat.Init(data_.n_cols(), data_.n_cols(), 1);
Vector d_sr_mat_diag;
d_sr_mat_diag.Init(data_.n_cols());
d_sr_mat_diag.SetAll(0);
// Now put the label information
std::map<index_t, std::vector<index_t> >::iterator it1;
for(it1=classes.begin(); it1!=classes.end(); it1++) {
for(index_t i=0; i<(index_t)it1->second.size(); i++) {
for(index_t j=i+1; j<(index_t)it1->second.size(); j++) {
d_sr_mat_diag[it1->second[i]]+=1.0/(it1->first+1);
d_sr_mat_diag[it1->second[j]]+=1.0/(it1->first+1);
kernel_matrix_.set(it1->second[i], it1->second[j], 1.0);
labeled_graph.set(it1->second[i], it1->second[j], 1.0/(it1->first+1));
labeled_graph.set(it1->second[j], it1->second[i], 1.0/(it1->first+1));
}
}
}
labeled_graph.set_symmetric(true);
kernel_matrix_.MakeSymmetric();
kernel_matrix_.EndLoading();
d_sr_mat.SetDiagonal(d_sr_mat_diag);
// This is the matrix D that has the sum of the rows or columns
SparseMatrix d_mat;
d_mat.Init(data_.n_cols(), data_.n_cols(), 1);
Vector d_diagonal;
kernel_matrix_.RowSums(&d_diagonal);
d_mat.SetDiagonal(d_diagonal);
SparseMatrix laplacian_mat;
Sparsem::Subtract(d_mat, kernel_matrix_, &laplacian_mat);
SparseMatrix d_sr_plus_laplacian_mat;
Sparsem::Add(d_sr_mat, laplacian_mat, &d_sr_plus_laplacian_mat);
Matrix eigenvectors;
labeled_graph.EndLoading();
d_sr_plus_laplacian_mat.EndLoading();
labeled_graph.Eig(d_sr_plus_laplacian_mat, num_of_classes, "LM",
&eigenvectors, eigenvalues, NULL);
// this is the embedding therms alpha as shown in
// the paper
Matrix alpha_factors;
la::LeastSquareFitTrans(eigenvectors, data_, &alpha_factors);
la::MulTransAInit(alpha_factors, data_,embedded_coordinates);
}