Files
mlpack/fastlib/trunk/contrib/nvasil/nmf_tree/nmf_tree_impl.h
T

260 lines
8.7 KiB
C++

/*
* =====================================================================================
*
* Filename: nmf_tree_impl.h
*
* Description:
*
* Version: 1.0
* Created: 08/18/2008 01:12:28 PM EDT
* Revision: none
* Compiler: gcc
*
* Author: Nikolaos Vasiloglou (NV), nvasil@ieee.org
* Company: Georgia Tech Fastlab-ESP Lab
*
* =====================================================================================
*/
#ifdef NMF_TREE_IMPL_H_
#define NMF_TREE_IMP_H_
// we use this macro to avoid including this file directly
#ifdef NMF_TREE_H_
template<typename TKdTree, typename T>
void NmfTreeConstructor<TKdTree, T>::Init(fx_module *module,
ArrayList<index_t> &rows,
ArrayList<index_t> &columns,
ArrayList<double> &values) {
module_=fx_submodule(module, "tree");;
fx_module *l_bfgs_module=fx_submodule(module_, "optimizer");
fx_module *relaxed_nmf_module=fx_submodule(module_, "l_bfgs");
rows_.Copy(rows);
columns_.Copy(columns);
values_.Copy(values);
index_t num_of_rows=*std::max_element(rows_.begin(), rows_.end())+1;
index_t num_of_columns=*std::max_element(columns_.begin(), columns_.end())+1;
new_dimension_ = fx_param_int(module, "new_dimension", 5);
data_matrix_.Init(new_dimension_, num_of_rows+num_of_columns);
w_matrix_.Alias(data_matrix_.ptr(), new_dimension_, num_of_rows_);
h_matrix_.Alias(data_matrix_.ptr(), new_dimension_, num_of_columns_);
lower_bound_.Init(new_dimension_, num_of_rows+num_of_columns);
upper_bound_.Init(new_dimension_, num_of_rows+num_of_columns);
lower_bound_.SetAll(fx_param_double(module_, "lower_bound", 1e-5));
upper_bound_.SetAll(fx_param_double(module_, "upper_bound", 1.0));
opt_fun_.Init(relaxed_nmf_module,
&rows_,
&columns_,
&values_,
&lower_bound,
& upper_bound);
l_bfgs_engine_.Init(&opt_fun, l_bfgs_module_);
leaf_size_=fx_param_int(module, "leaf_size", 20);
old_from_new_h_.Init(num_of_columns_);
for (index_t i = 0; i < num_of_rows_; i++) {
old_from_new_h_[i] = i;
}
old_from_new_w_.Init(num_of_rows_);
for (index_t i = 0; i < num_of_columns_; i++) {
old_from_new_w_[i] = i;
}
}
template<typename TKdTree, typename T>
void NmfTreeConstructor<TKdTree, T>::MakeNmfTree() {
parent_= new TKdTree();
parent_->Init(0, data_matrix.n_cols());
MakeNmfTreeMidpointSelective(parent_w_, parent_h_);
}
template<typename TKdTree, typename T>
void NmfTreeConstructor<TKdTree, T>::MakeNmfTreeMidpointSelective(
TKdTree *node1, TKdTree *node2) {
GenVector<T> split_dimensions;
split_dimensions.Init(data_matrix_.n_rows());
int i;
for (i = 0; i < data_matrix.n_rows(); i++){
split_dimensions[i] = i;
}
SelectSplitKdTreeMidpoint(node1, node2, split_dimensions);
}
template<typename TKdTree, typename T>
void NmfTreeConstructor<TKdTree, T>::SelectSplitKdTreeMidpoint(TKdTree *node1,
TKdTree *node2,
const GenVector<T>& split_dimensions) {
TKdTree *left = NULL;
TKdTree *right = NULL;
optimizer.Reset();
GenMatrix<T> init_matrix;
opt_fun.GiveInitMatrix(&init_matrix);
optimizer_.set(init_matrix);
optimizer_ComputeLocalOptimumBFGS();
tree_kdtree_private::SelectFindBoundFromMatrix(data_matrix_,
split_dimensions,
node->begin(),
node->count(),
&node->bound());
if (node1->count() > leaf_size_ && node2->count() > leaf_size_) {
SplitAndBound(node1, w_matrix_, split_dimensions);
SplitAndBound(node2, h_matrix_, split_dimensions);
SelectSplitKdTreeMidpoint(node1->left(), node2->left(), split_dimensions);
SelectSplitKdTreeMidpoint(node1->right(), node2->right(), split_dimensions);
} else {
if (node1->count() > leaf_size_ && node2->count() < leaf_size_) {
SplitAndBound(node1, w_matrix_, split_dimensions);
SelectSplitKdTreeMidpoint(node1->left(), node2->left(), split_dimensions);
SelectSplitKdTreeMidpoint(node1->right(), node2->right(), split_dimensions);
} else {
if (node1->count() < leaf_size_ && node2->count() > leaf_size_) {
SplitAndBound(node2, h_matrix_, split_dimensions);
SelectSplitKdTreeMidpoint(node1->left(), node2->left(), split_dimensions);
SelectSplitKdTreeMidpoint(node1->right(), node2->right(), split_dimensions);
}
}
}
}
template<typename TBound, typename T>
void NmfTreeConstructor<TKdTree, T>::SplitAndBound(TKdTree *node,
GenMatrix<T> *data_matrix,
GenVector<T> &split_dimensions) {
index_t split_dim = BIG_BAD_NUMBER;
double max_width = -1;
for (index_t d = 0; d < split_dimensions.length(); d++) {
double w = node->bound().get(d).width();
if (w > max_width) {
max_width = w;
split_dim = d;
}
}
double split_val = node->bound().get(split_dim).mid();
if (max_width == 0) {
// Okay, we can't do any splitting, because all these points are the
// same. We have to give up.
} else {
left = new TKdTree();
node->left->bound().Init(split_dimensions.length());
node->right = new TKdTree();
node->right->bound().Init(split_dimensions.length());
index_t split_col = tree_kdtree_private::SelectMatrixPartition(data_matrix,
split_dimensions,
(int)split_dimensions[split_dim], split_val,
node->begin(), node->count(),
&node->left->bound(), &node->right->bound(),
old_from_new_);
VERBOSE_MSG(3.0,"split (%d,[%d],%d) dim %d on %f (between %f, %f)",
node->begin(), split_col,
node->begin() + node->count(), (int)split_dimensions[split_dim],
split_val,
node->bound().get(split_dim).lo,
node->bound().get(split_dim).hi);
node->left->Init(node->begin(), split_col - node->begin());
node->right->Init(split_col, node->begin() + node->count() - split_col);
// This should never happen if max_width > 0
DEBUG_ASSERT(node->left->count() != 0 && node->right->count() != 0);
}
template<typename TBound, typename T>
index_t NmfTreeConstructor<TKdTree, T>::SelectMatrixPartition(GenMatrix<T>& matrix,
const Vector& split_dimensions, index_t dim, double splitvalue,
index_t first, index_t count, TBound* left_bound, TBound* right_bound) {
index_t left = first;
index_t right = first + count - 1;
/* At any point:
*
* everything < left is correct
* everything > right is correct
*/
for (;;) {
while (matrix.get(dim, left) < splitvalue && likely(left <= right)) {
GenVector<T> left_vector;
matrix.MakeColumnVector(left, &left_vector);
if (split_dimensions.length() == matrix.n_rows()) {
*left_bound |= left_vector;
} else {
GenVector<T> sub_left_vector;
sub_left_vector.Init(split_dimensions.length());
MakeBoundVector(left_vector, split_dimensions, &sub_left_vector);
*left_bound |= sub_left_vector;
}
left++;
}
while (matrix.get(dim, right) >= splitvalue && likely(left <= right)) {
GenVector<T> right_vector;
matrix.MakeColumnVector(right, &right_vector);
if (split_dimensions.length() == matrix.n_rows()) {
*right_bound |= right_vector;
} else {
GenVector<T> sub_right_vector;
sub_right_vector.Init(split_dimensions.length());
MakeBoundVector(right_vector, split_dimensions, &sub_right_vector);
*right_bound |= sub_right_vector;
}
right--;
}
if (unlikely(left > right)) {
/* left == right + 1 */
break;
}
GenVector<T> left_vector;
GenVector<T> right_vector;
matrix.MakeColumnVector(left, &left_vector);
matrix.MakeColumnVector(right, &right_vector);
left_vector.SwapValues(&right_vector);
if (split_dimensions.length() == matrix.n_rows()) {
*left_bound |= left_vector;
} else {
GenVector<T> sub_left_vector;
sub_left_vector.Init(split_dimensions.length());
MakeBoundVector(left_vector, split_dimensions, &sub_left_vector);
*left_bound |= sub_left_vector;
}
if (split_dimensions.length() == matrix.n_rows()){
*right_bound |= right_vector;
} else {
GenVector<T> sub_right_vector;
sub_right_vector.Init(split_dimensions.length());
MakeBoundVector(right_vector, split_dimensions, &sub_right_vector);
*right_bound |= sub_right_vector;
}
index_t t = old_from_new[left];
old_from_new[left] = old_from_new[right];
old_from_new[right] = t;
DEBUG_ASSERT(left <= right);
right--;
// this conditional is always true, I belueve
//if (likely(left <= right)) {
// right--;
//}
}
DEBUG_ASSERT(left == right + 1);
return left;
}
#endif // NMF_TREE_H_
#endif // NMF_TREE_IMPL_H_