apaam/mlpack
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Buggy. Couldnt debug it

Browse Source
This commit is contained in:
gmravi2003
2008-02-09 21:58:58 +00:00
parent 47fd5fa9ad
commit bffbbdfcb3
5 changed files with 381 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastlib/u/gmravi/regression/main_regression_ll.cc
+61 -18
View File
@@ -11,6 +11,7 @@ int main(int argc, char *argv[]){
Matrix q_matrix;
Matrix r_matrix;
Vector rset_weights;
Vector true_regression_values;
//Reading parameters and loading data
@@ -18,7 +19,8 @@ int main(int argc, char *argv[]){
fx_submodule(NULL, "regression", "regression_module");
// The reference data file is a required parameter.
const char* reference_file_name = fx_param_str_req(regression_module, "data");
const char* reference_file_name =
fx_param_str_req(regression_module, "data");
// The query data file defaults to the references.
const char* query_file_name =
@@ -49,8 +51,8 @@ int main(int argc, char *argv[]){
//Get the bandwidth for kernel calculations and the tolerance limit.
//Both default to 0.2
double bandwidth=fx_param_double(regression_module,"bandwidth",0.15);
double tau=fx_param_double(regression_module,"tau",0.15);
double bandwidth=fx_param_double(regression_module,"bandwidth",0.125);
double tau=fx_param_double(regression_module,"tau",0.0025);
printf("tau is %f\n",tau);
//Get the weights for the reference set
@@ -65,7 +67,7 @@ int main(int argc, char *argv[]){
Dataset ref_weights;
ref_weights.InitFromFile (rwfname);
rset_weights.Copy (ref_weights.matrix ().GetColumnPtr (0), ref_weights.matrix ().n_rows ()); //Note rset_weights_ is a vector of weights
rset_weights.Copy (ref_weights.matrix ().GetColumnPtr (0), ref_weights.matrix ().n_rows ()); //Note rset_weights is a vector of weights
}
@@ -76,6 +78,30 @@ int main(int argc, char *argv[]){
}
//Also get the true regression values of all the reference points
const char *are=NULL;
if (fx_param_exists (NULL, "true_reg")){
//are is the filename having the regression estimates of the
are = fx_param_str (NULL, "true_reg", NULL);
}
if (are != NULL){
Dataset true_reg;
true_reg.InitFromFile (are);
true_regression_values.Copy (true_reg.matrix ().GetColumnPtr (0), true_reg.matrix ().n_rows ());
}
else{
true_regression_values.Init (r_matrix.n_cols ());
true_regression_values.SetAll (1);
}//Hence the true regression values have been stored
//Get the length of the leaf. Defaulted to 2
index_t leaf_length=fx_param_int(regression_module,"leaf_length",3);
@@ -88,23 +114,31 @@ int main(int argc, char *argv[]){
criteria=(char*)malloc(40*sizeof(char));
strcpy(criteria,fx_param_str(regression_module,"criteria","fnorm"));
fast_regression.Init(q_matrix, r_matrix, bandwidth, tau, leaf_length, rset_weights, criteria);
fast_regression.Init(q_matrix, r_matrix, bandwidth, tau,
leaf_length, rset_weights, criteria,true_regression_values);
fast_regression.Compute();
printf("FAST CALCULATIONS ALL DONE");
//These are the regression estimates provided by the fast
//algorithm
// Vector fast_regression_estimate;
//fast_regression_estimate.Copy(fast_regression.get_regression_estimate());
//Lets do naive calculations too............
//Lets first declare an object of the naive type
printf("FAST CALCULATIONS ALL DONE");
ArrayList<index_t> old_from_new_r;
old_from_new_r.Copy(fast_regression.get_old_from_new_r());
//Get the fast regression estimates.We shall use it to compare the
//accuracy
Vector fast_regression_estimate;
fast_regression_estimate.Copy(fast_regression.get_regression_estimate());
ArrayList<Matrix> fast_b_twy_estimates;
ArrayList<Matrix> fast_b_twb_estimates;
@@ -113,23 +147,32 @@ int main(int argc, char *argv[]){
fast_b_twb_estimates.Init(q_matrix.n_cols());
for(index_t q=0;q<q_matrix.n_cols();q++){
fast_b_twb_estimates[q].Init(r_matrix.n_rows()+1,r_matrix.n_rows()+1);
fast_b_twy_estimates[q].Init(r_matrix.n_rows()+1,1);
fast_b_twy_estimates[q].Alias(fast_regression.get_b_twy_estimates(q));
fast_b_twb_estimates[q].Alias(fast_regression.get_b_twb_estimates(q));
}
for(index_t q=0;q<q_matrix.n_cols();q++){
fast_b_twy_estimates[q].CopyValues(fast_regression.get_b_twy_estimates(q));
fast_b_twb_estimates[q].CopyValues(fast_regression.get_b_twb_estimates(q));
}
NaiveCalculation<GaussianKernel> naive;
naive.Init(q_matrix,r_matrix,old_from_new_r,bandwidth,rset_weights);
naive.Init(q_matrix,r_matrix,
old_from_new_r,bandwidth,rset_weights);
printf("started naive computations..\n");
naive.Compute();
printf("completed naive calculations..\n");
naive.ComputeMaximumRelativeError
(fast_b_twy_estimates,fast_b_twb_estimates,criteria);
naive.CompareFastWithNaive(fast_regression_estimate);
(fast_b_twy_estimates,fast_b_twb_estimates,criteria);
//naive.CompareFastWithNaive(fast_regression_estimate);
//Will need to verify if this is fine to do...
fx_done();
}
}
fastlib/u/gmravi/regression/regression_ll1.h
+62 -37
View File
@@ -1,7 +1,7 @@
#ifndef REGRESSION_LL1_H
#define REGRESSION_LL1_H
#include "regression_ll2.h"
#include "pseudo_inverse.h"
//#include "pseudo_inverse.h"
/** In this code we shalll evaluate (B^TWB)^-1 by first calculating B^TWB
@@ -421,8 +421,8 @@ void FastRegression<TKernel>::UpdateBounds_(Tree *qnode,
//Add dl_b_twy and du_b_twy to mass_l and mass_u of B^TWY
la::AddTo(dl_b_twy, &qnode->stat().b_twy_mass_l);
la::AddTo(du_b_twy, &qnode->stat().b_twy_mass_u);
la::AddTo(dl_b_twy, &(qnode->stat().b_twy_mass_l));
la::AddTo(du_b_twy, &(qnode->stat().b_twy_mass_u));
if(!qnode->is_leaf()){
@@ -469,8 +469,8 @@ void FastRegression<TKernel>::UpdateBounds_(Tree *qnode,
//Add dl_b_twb and du_b_twb to mass_l and mass_u of B^TWB
la::AddTo(dl_b_twb, &qnode->stat().b_twb_mass_l);
la::AddTo(du_b_twb, &qnode->stat().b_twb_mass_u);
la::AddTo(dl_b_twb, &(qnode->stat().b_twb_mass_l));
la::AddTo(du_b_twb, &(qnode->stat().b_twb_mass_u));
if(!qnode->is_leaf()){
@@ -493,30 +493,30 @@ void FastRegression<TKernel>::UpdateBounds_(Tree *qnode,
else{
if(flag==CHECK_FOR_PRUNE_B_TWY){
//the flag is CHECK_FOR_PRUNE_B_TWY
//the flag is CHECK_FOR_PRUNE_B_TWY
//Add dl_b_twy and du_b_twy to mass_l and mass_u of B^TWY
//Add dl_b_twy and du_b_twy to mass_l and mass_u of B^TWY
la::AddTo(dl_b_twy, &qnode->stat().b_twy_mass_l);
la::AddTo(du_b_twy, &qnode->stat().b_twy_mass_u);
la::AddTo(dl_b_twy, &(qnode->stat().b_twy_mass_l));
la::AddTo(du_b_twy, &(qnode->stat().b_twy_mass_u));
if(!qnode->is_leaf()){
if(!qnode->is_leaf()){
// transmit these values to the children node
// transmit these values to the children node
la::AddTo(dl_b_twy,&(qnode->left()->stat().b_twy_owed_l));
la::AddTo(du_b_twy,&(qnode->left()->stat().b_twy_owed_u));
la::AddTo(dl_b_twy,&(qnode->left()->stat().b_twy_owed_l));
la::AddTo(du_b_twy,&(qnode->left()->stat().b_twy_owed_u));
la::AddTo(dl_b_twy,&(qnode->right()->stat().b_twy_owed_l));
la::AddTo(du_b_twy,&(qnode->right()->stat().b_twy_owed_u));
}
la::AddTo(dl_b_twy,&(qnode->right()->stat().b_twy_owed_l));
la::AddTo(du_b_twy,&(qnode->right()->stat().b_twy_owed_u));
}
else{
/* in case of leaf nodes add these values to more_l and more_u */
la::AddTo(dl_b_twy,&(qnode->stat().b_twy_more_l));
la::AddTo(du_b_twy,&(qnode->stat().b_twy_more_u));
}
else{
/* in case of leaf nodes add these values to more_l and more_u */
la::AddTo(dl_b_twy,&(qnode->stat().b_twy_more_l));
la::AddTo(du_b_twy,&(qnode->stat().b_twy_more_u));
}
}
}
@@ -565,9 +565,10 @@ FastRegression<TKernel>::Prunable_(Tree *qnode, Tree *rnode, Matrix &dl_b_twb,
/** this means both the quantities are prunable
*/
//For a moment lets makc a change
return PRUNE_BOTH;
//FOR DEBUGGING
//return PRUNE_NOT;
return PRUNE_BOTH;
}
else{
if(flag1==1){
@@ -587,6 +588,7 @@ FastRegression<TKernel>::Prunable_(Tree *qnode, Tree *rnode, Matrix &dl_b_twb,
else{
/** None of the quantities is prunable
*/
printf("Couldnt prune anything..\n");
return PRUNE_NOT;
}
@@ -615,7 +617,8 @@ FastRegression<TKernel>::Prunable_(Tree *qnode, Tree *rnode, Matrix &dl_b_twb,
}
//flag==CHECK_FOR_PRUNE_B_TWY
index_t flag2=FastRegression<GaussianKernel>::PrunableB_TWY_(qnode,rnode,dl_b_twy,du_b_twy);
index_t flag2=FastRegression<GaussianKernel>::
PrunableB_TWY_(qnode,rnode,dl_b_twy,du_b_twy);
if(flag2==1){
//this means B_TWY is prunable
@@ -624,6 +627,7 @@ FastRegression<TKernel>::Prunable_(Tree *qnode, Tree *rnode, Matrix &dl_b_twb,
}
else{
//it is not prunable.
printf("Could not prune..\n");
return PRUNE_NOT;
}
}
@@ -713,9 +717,11 @@ void FastRegression<TKernel>::FRegression_(Tree *qnode, Tree *rnode,
if(what_is_prunable==PRUNE_B_TWB){
printf("BTWB pruned..\n");
UpdateBoundsForPruningB_TWB_(qnode, dl_b_twb, du_b_twb);
MergeChildBounds_(qnode,flag);
// MergeChildBounds_(qnode,what_is_prunable);
if((index_t)what_is_prunable==(index_t)flag){
//this means the job is done and we may return now
return;
@@ -731,9 +737,10 @@ void FastRegression<TKernel>::FRegression_(Tree *qnode, Tree *rnode,
else{
if(what_is_prunable==PRUNE_B_TWY){
printf("BTWY pruned..\n");
UpdateBoundsForPruningB_TWY_(qnode, dl_b_twy, du_b_twy);
MergeChildBounds_(qnode,flag);
//MergeChildBounds_(qnode,what_is_prunable);
if((index_t)what_is_prunable==(index_t)flag){
//this means the job is done and we may return now
@@ -748,10 +755,10 @@ void FastRegression<TKernel>::FRegression_(Tree *qnode, Tree *rnode,
}
else{
//both are prunable
printf("Both pruned...\n");
UpdateBoundsForPruningB_TWB_(qnode, dl_b_twb, du_b_twb);
UpdateBoundsForPruningB_TWY_(qnode, dl_b_twy, du_b_twy);
MergeChildBounds_(qnode,flag);
MergeChildBounds_(qnode,CHECK_FOR_PRUNE_BOTH);
return;
}
@@ -832,9 +839,13 @@ void FastRegression<TKernel>::Compute(){
check_for_prune_t flag=CHECK_FOR_PRUNE_BOTH;
fx_timer_start(NULL,"fast_timer");
printf("entered recursive function..\n");
FRegression_(qroot_,rroot_,flag);
printf("Completed recursion..\n");
PostProcess_(qroot_);
ObtainRegressionEstimate_();
printf("Will postprocess now..\n");
//ObtainRegressionEstimate_();
//CompareWithTrueValues_();
fx_timer_stop(NULL,"fast_timer");
//This will print the matrices B^TWB and B^TWY to an output file
@@ -916,15 +927,22 @@ void FastRegression<TKernel>::SetUpperBounds_(Tree *node){
if(node->is_leaf()){
node->stat().b_twy_mass_u.Copy(rroot_->stat().b_ty);
node->stat().b_twy_mass_u.Init(rset_.n_rows()+1,1);
node->stat().b_twy_mass_u.CopyValues(rroot_->stat().b_ty);
node->stat().b_twb_mass_u.Copy(rroot_->stat().b_tb);
node->stat().b_twb_mass_u.Init(rset_.n_rows()+1,rset_.n_rows()+1);
node->stat().b_twb_mass_u.CopyValues(rroot_->stat().b_tb);
}
else
{
node->stat().b_twy_mass_u .Copy(rroot_->stat().b_ty);
node->stat().b_twb_mass_u.Copy(rroot_->stat().b_tb);
node->stat().b_twy_mass_u.Init(rset_.n_rows()+1,1);
node->stat().b_twb_mass_u.Init(rset_.n_rows()+1,rset_.n_rows()+1);
node->stat().b_twy_mass_u .CopyValues(rroot_->stat().b_ty);
node->stat().b_twb_mass_u.CopyValues(rroot_->stat().b_tb);
SetUpperBounds_(node->left());
SetUpperBounds_(node->right());
}
@@ -936,18 +954,25 @@ void FastRegression<TKernel>::Init(Matrix &q_matrix, Matrix &r_matrix,
double bandwidth,
double tau,
index_t leaf_length,
Vector &rset_weights, char *criteria){
Vector &rset_weights,
char *criteria,
Vector &true_regression_values){
//Set up value of tau,qset_,rset_
//Set up value of tau,qset_,rset_, true_regression_values
qset_.Alias(q_matrix);
rset_.Alias(r_matrix);
true_regression_values_.Alias(true_regression_values);
tau_=tau;
index_t leaflen=leaf_length;
rset_weights_.Alias(rset_weights);
/* Construct Query and Reference trees */
fx_timer_start(NULL,"tree_create");
rroot_ = tree::MakeKdTreeMidpoint < Tree >
(rset_, leaflen, &old_from_new_r_, &new_from_old_r_);
fastlib/u/gmravi/regression/regression_ll2.h
+199 -111
View File
@@ -8,9 +8,9 @@ double FastRegression<TKernel>::Compute1NormLike_(Matrix &a){
//This function computes treats the natrix as a vector and computes
//it's 1-norm
double value=0;
for(index_t col=0;col<a.n_cols();col++){
for(index_t row=0;row<a.n_rows();row++){
double value=0.0;
for(index_t row=0;row<a.n_rows();row++){
for(index_t col=0;row<a.n_cols();row++){
value+=fabs(a.get(row,col));
}
}
@@ -21,12 +21,12 @@ template <typename TKernel>
double FastRegression<TKernel>::SquaredFrobeniusNorm_(Matrix &a){
double sqd_frobenius_norm=0.0;
for(index_t col=0;col<a.n_cols();col++){
for(index_t row=0;row<a.n_rows();row++){
//Along each column
for(index_t row=0;row<a.n_rows();row++){
for(index_t col=0;col<a.n_cols();col++){
//Alon each row
sqd_frobenius_norm+=a.get(row,col)*a.get(row,col);
sqd_frobenius_norm+=(a.get(row,col)*a.get(row,col));
}
}
return sqd_frobenius_norm;
@@ -128,19 +128,28 @@ index_t FastRegression<TKernel>::PrunableB_TWY_(Tree *qnode, Tree *rnode,
//THIS Means we are interested in the frobenius norm pruning................
//Maximum allowed error is the sqaured frobenius norm of the matrix max_error
//Maximum error is the sqaured frobenius norm of the matrix max_error
double squared_frobenius_norm_of_max_error=SquaredFrobeniusNorm_(max_error);
double ratio_of_1norms=
Compute1NormLike_(rnode->stat().b_ty)/Compute1NormLike_(rroot_->stat().b_ty);
(double) Compute1NormLike_(rnode->stat().b_ty)/Compute1NormLike_(rroot_->stat().b_ty);
//Now lets calculate the squared frobenius norm of new_mass_l
double sqd_frobenius_norm_of_new_mass_l=SquaredFrobeniusNorm_(new_mass_l);
double allowed_error= ratio_of_1norms*tau_*sqd_frobenius_norm_of_new_mass_l;
if(squared_frobenius_norm_of_max_error<=allowed_error)
if(squared_frobenius_norm_of_max_error<allowed_error)
{
//prune
printf("kernel value lo is %f\n",kernel_value_range.lo);
printf("kernel value hi is %f\n",kernel_value_range.hi);
printf("squared_frobenius_norm_of_max_error was %f\n",squared_frobenius_norm_of_max_error);
printf("allowed error is %f\n",allowed_error);
printf("\n\n");
return 1;
}
else{
@@ -195,7 +204,7 @@ index_t FastRegression<TKernel>:: PrunableB_TWB_(Tree *qnode, Tree *rnode, Matri
//Now lets calculate the allowed error.
//allowed_error=tau*(rnode->stat().b_tb/rrot->stat().b_tb)(b_twb_mass_l+dl)
//allowed_error=tau*(rnode->stat().b_tb/rroot->stat().b_tb)(b_twb_mass_l+dl)
//Lets define a matrix new_mass_l as shown below.This matrix takes
//into account the addditional dl that will be added if pruning takes
//place
@@ -228,7 +237,7 @@ index_t FastRegression<TKernel>:: PrunableB_TWB_(Tree *qnode, Tree *rnode, Matri
double alpha=
tau_*(double)(rnode->stat().b_tb.get(row,col))/
(rroot_->stat().b_tb.get(row,col));
if(rroot_->stat().b_tb.get(row,col)==0){
if(rroot_->stat().b_tb.get(row,col)==0){
printf("btb.is not properly defined\n");
}
@@ -249,6 +258,7 @@ index_t FastRegression<TKernel>:: PrunableB_TWB_(Tree *qnode, Tree *rnode, Matri
}
}
//This means the matrix is compoenent wise prunable. hence return 1
return 1;
}
//THIS MEANS WE ARE INTERESTED IN PRUNING BY FROBENIUS NORM
@@ -256,14 +266,15 @@ index_t FastRegression<TKernel>:: PrunableB_TWB_(Tree *qnode, Tree *rnode, Matri
double squared_frobenius_norm_of_max_error=SquaredFrobeniusNorm_(max_error);
double ratio_of_1norms=
Compute1NormLike_(rnode->stat().b_tb)/Compute1NormLike_(rroot_->stat().b_tb);
(double)Compute1NormLike_(rnode->stat().b_tb)/Compute1NormLike_(rroot_->stat().b_tb);
//Now lets calculate the squared frobenius norm of new_mass_l
double squared_frobenius_norm_of_new_mass_l=SquaredFrobeniusNorm_(new_mass_l);
double allowed_error=tau_*ratio_of_1norms*squared_frobenius_norm_of_new_mass_l;
if(squared_frobenius_norm_of_max_error > allowed_error){
if(squared_frobenius_norm_of_max_error >= allowed_error){
//then this matrix is NOT runable
dl.SetAll(0);
@@ -271,12 +282,19 @@ index_t FastRegression<TKernel>:: PrunableB_TWB_(Tree *qnode, Tree *rnode, Matri
return 0;
}
//This means that the quantity is prunable
return 1;
else{
//This means that the quantity is prunable
printf("squared_frobenius_norm_of_max_error was %f\n",squared_frobenius_norm_of_max_error);
printf("allowed error is %f\n",allowed_error);
printf("kernel value lo is %f\n",kernel_value_range.lo);
printf("kernel value hi is %f\n",kernel_value_range.hi);
printf("\n\n");
return 1;
}
}
/* The Update Boundws function is independent of the pruning criteria */
/* The Update Bounds function is independent of the pruning criteria */
template <typename TKernel>
void FastRegression<TKernel>::
@@ -290,8 +308,8 @@ UpdateBoundsForPruningB_TWY_(Tree *qnode, Matrix &dl_b_twy, Matrix &du_b_twy){
//b_twy_mass_u <- b_twy_mass_u+du
la::AddTo(dl_b_twy, &qnode->stat().b_twy_mass_l);
la::AddTo(du_b_twy, &qnode->stat().b_twy_mass_u);
la::AddTo(dl_b_twy, &(qnode->stat().b_twy_mass_l));
la::AddTo(du_b_twy, &(qnode->stat().b_twy_mass_u));
// for a leaf node, incorporate the lower and upper bound changes into
@@ -396,31 +414,31 @@ MergeChildBoundsB_TWB_( FastRegression<TKernel>::
// parent
//So lets find the componentwise minimum and maximum
if(pruning_criteria==CRITERIA_FOR_PRUNE_COMPONENT){
Matrix max_children;
Matrix min_children;
max_children.Init(parent_stat.b_twb_mass_l.n_rows(),
parent_stat.b_twb_mass_l.n_cols());
min_children.Init(parent_stat.b_twb_mass_l.n_rows(),
parent_stat.b_twb_mass_l.n_cols());
Matrix max_children;
Matrix min_children;
max_children.Init(parent_stat.b_twb_mass_l.n_rows(),
parent_stat.b_twb_mass_l.n_cols());
min_children.Init(parent_stat.b_twb_mass_l.n_rows(),
parent_stat.b_twb_mass_l.n_cols());
if(pruning_criteria==CRITERIA_FOR_PRUNE_COMPONENT){
for(index_t col=0;col<parent_stat.b_twb_mass_l.n_cols();col++){
for(index_t row=0;row<parent_stat.b_twb_mass_l.n_rows();row++){
if(left_stat->b_twb_mass_l.get(row,col) <=
right_stat->b_twb_mass_l.get(row,col)){
//left child has lesser mass_l value
min_children.set(row,col,
left_stat->b_twb_mass_l.get(row,col));
max_children.set(row,col,
right_stat->b_twb_mass_l.get(row,col));
}
for(index_t col=0;col<parent_stat.b_twb_mass_l.n_cols();col++){
for(index_t row=0;row<parent_stat.b_twb_mass_l.n_rows();row++){
if(left_stat->b_twb_mass_l.get(row,col) <=
right_stat->b_twb_mass_l.get(row,col)){
//left child has lesser mass_l value
min_children.set(row,col,
left_stat->b_twb_mass_l.get(row,col));
max_children.set(row,col,
right_stat->b_twb_mass_l.get(row,col));
}
else{
//right child has lesser mass-l value
@@ -432,9 +450,9 @@ MergeChildBoundsB_TWB_( FastRegression<TKernel>::
}
}
}
}
}
//Now compare with parent...
for(index_t col=0;col<parent_stat.b_twb_mass_l.n_cols();col++){
@@ -453,7 +471,8 @@ MergeChildBoundsB_TWB_( FastRegression<TKernel>::
}
}
}
}
return;
}
//THe pruning criteria is Frobenius norm pruning...
@@ -538,6 +557,7 @@ MergeChildBoundsB_TWY_( FastRegression<TKernel>::
if(pruning_criteria==CRITERIA_FOR_PRUNE_COMPONENT){
//This means we want to merge the bounds of b_twb
//b_twb_mass_l_parent=
//max(min(b_twy_mass_l,left_child,b_twy_mass_l_right_child),
@@ -601,7 +621,9 @@ MergeChildBoundsB_TWY_( FastRegression<TKernel>::
}
}
}
return;
}
//IF Pruning criteria is Frobenius Norm Pruning criteria.....
else{
@@ -609,17 +631,20 @@ MergeChildBoundsB_TWY_( FastRegression<TKernel>::
// lower_of_parent=max (parent, min (children))
// upper_of_parent=min(parent, max(children))
//lets first frind out the squared frobenius norms of the lower
//lets first find out the squared frobenius norms of the lower
//bound masses of both the children
double sqd_frobenius_norm_of_left_child=SquaredFrobeniusNorm_(left_stat->b_twy_mass_l);
double sqd_frobenius_norm_of_right_child=SquaredFrobeniusNorm_(right_stat->b_twy_mass_l);
double sqd_frobenius_norm_of_left_child=
SquaredFrobeniusNorm_(left_stat->b_twy_mass_l);
double sqd_frobenius_norm_of_right_child=
SquaredFrobeniusNorm_(right_stat->b_twy_mass_l);
//Now compare it with that of the parent
double sqd_frobenius_norm_of_parent=SquaredFrobeniusNorm_(parent_stat.b_twy_mass_l);
double sqd_frobenius_norm_of_parent=
SquaredFrobeniusNorm_(parent_stat.b_twy_mass_l);
if(sqd_frobenius_norm_of_left_child <sqd_frobenius_norm_of_right_child)
if(sqd_frobenius_norm_of_left_child < sqd_frobenius_norm_of_right_child)
{
//This means the left child has lower frobenius norm compared
//to the right child
@@ -647,10 +672,13 @@ MergeChildBoundsB_TWY_( FastRegression<TKernel>::
}
//A similar logic for mass_u values
sqd_frobenius_norm_of_left_child=SquaredFrobeniusNorm_(left_stat->b_twy_mass_u);
sqd_frobenius_norm_of_right_child=SquaredFrobeniusNorm_(right_stat->b_twy_mass_u);
sqd_frobenius_norm_of_left_child=
SquaredFrobeniusNorm_(left_stat->b_twy_mass_u);
//Now compute it with that of the parent
sqd_frobenius_norm_of_right_child=
SquaredFrobeniusNorm_(right_stat->b_twy_mass_u);
//Now compare it with that of the parent
sqd_frobenius_norm_of_parent=SquaredFrobeniusNorm_(parent_stat.b_twy_mass_u);
@@ -680,6 +708,43 @@ MergeChildBoundsB_TWY_( FastRegression<TKernel>::
* and the children
*/
template <typename TKernel>
void FastRegression<TKernel>::
MergeChildBounds_( Tree *qnode,prune_t flag){
//We will merge bounds depending on the value of the flag
//But firstly we will check if the qnode is a leaf node. if it is
//then we will simply return
if(qnode->is_leaf()){
return;
}
//The statistics of the parent and the children node
FastRegressionStat &parent_stat=qnode->stat();
FastRegressionStat *left_stat=&(qnode->left()->stat());
FastRegressionStat *right_stat=&(qnode->right()->stat());
if(flag==PRUNE_B_TWB){
MergeChildBoundsB_TWB_(left_stat,right_stat,parent_stat);
}
else{
if(flag==PRUNE_B_TWY){
MergeChildBoundsB_TWY_(left_stat,right_stat,parent_stat);
}
else{
if(flag==PRUNE_BOTH){
MergeChildBoundsB_TWB_(left_stat,right_stat,parent_stat);
MergeChildBoundsB_TWY_(left_stat,right_stat,parent_stat);
}
}
}
}
template <typename TKernel>
void FastRegression<TKernel>::
MergeChildBounds_( Tree *qnode,check_for_prune_t flag){
@@ -708,10 +773,10 @@ MergeChildBounds_( Tree *qnode,check_for_prune_t flag){
MergeChildBoundsB_TWY_(left_stat,right_stat,parent_stat);
}
else{
//printf("will merge both the bounds...\n");
MergeChildBoundsB_TWB_(left_stat,right_stat,parent_stat);
MergeChildBoundsB_TWY_(left_stat,right_stat,parent_stat);
if(flag==CHECK_FOR_PRUNE_BOTH){
MergeChildBoundsB_TWB_(left_stat,right_stat,parent_stat);
MergeChildBoundsB_TWY_(left_stat,right_stat,parent_stat);
}
}
}
}
@@ -725,13 +790,15 @@ template <typename TKernel>
void FastRegression<TKernel>::FRegressionBaseB_TWY_(Tree *qnode, Tree *rnode){
//printf("Hit the base case for BTWY..\n");
//subtract along each dimension as we are now doing exhaustive calculations
//more_u <- more_u - rnode->stat().b_ty
la::SubFrom (rnode->stat().b_ty, &qnode->stat().b_twy_more_u);
//Having subtracted calculate B^TWY exhaustively
//Having subtracted, calculate B^TWY exhaustively
//One can do this by using linear algebra routines available
//in LaPack. however we shall not use them because B can be a
//very large matrix
@@ -759,27 +826,27 @@ void FastRegression<TKernel>::FRegressionBaseB_TWY_(Tree *qnode, Tree *rnode){
//This is nothing but B^TWY being evaluated dimension wise
if (row != 0){
double val=b_twy_l_estimate_[q].get(row,col)+
double val1=b_twy_l_estimate_[q].get(row,col)+
ker_value * rset_weights_[old_from_new_r_[r]] * rset_.get (row- 1, r);
b_twy_l_estimate_[q].set(row,col,val);
b_twy_l_estimate_[q].set(row,col,val1);
val=b_twy_u_estimate_[q].get(row,col)+
double val2 =b_twy_u_estimate_[q].get(row,col)+
ker_value * rset_weights_[old_from_new_r_[r]] * rset_.get (row- 1, r);
b_twy_u_estimate_[q].set(row,col,val);
b_twy_u_estimate_[q].set(row,col,val2);
}
else{
double val= b_twy_l_estimate_[q].get(row,col) +
double val1= b_twy_l_estimate_[q].get(row,col) +
ker_value * rset_weights_[old_from_new_r_[r]];
b_twy_l_estimate_[q].set(row,col,val);
b_twy_l_estimate_[q].set(row,col,val1);
val= b_twy_u_estimate_[q].get(row,col) +
double val2= b_twy_u_estimate_[q].get(row,col) +
ker_value * rset_weights_[old_from_new_r_[r]];
b_twy_u_estimate_[q].set(row,col,val);
b_twy_u_estimate_[q].set(row,col,val2);
}
@@ -833,13 +900,15 @@ void FastRegression<TKernel>::FRegressionBaseB_TWY_(Tree *qnode, Tree *rnode){
//having looped over each point
qnode->stat().b_twy_mass_u.CopyValues(max_u);
qnode->stat().b_twy_mass_l.CopyValues(min_l);
qnode->stat().b_twy_mass_l.CopyValues(min_l);
return;
}
else{
//We are doing frobenius norm pruning
/** get a tighter lower and upper boiounf by looping over each query
/** get a tighter lower and upper bound by looping over each query
* point to find that particular query point that has the least sqd
* frobenius norm of the matrix b_twy_l_estimate_[q]+qnode->stat().b_twy_more_l
*/
@@ -850,8 +919,8 @@ void FastRegression<TKernel>::FRegressionBaseB_TWY_(Tree *qnode, Tree *rnode){
//The min_pointer and the max_pointer store the index number of the
//query point which possibly has the least and the highest squared frobenius norm
index_t min_pointer;
index_t max_pointer;
index_t min_pointer=qnode->begin();
index_t max_pointer=qnode->begin();
temp.Init(rset_.n_rows()+1,1);
for(index_t i=qnode->begin();i<qnode->end();i++){
@@ -879,9 +948,13 @@ void FastRegression<TKernel>::FRegressionBaseB_TWY_(Tree *qnode, Tree *rnode){
//Once done with the looping process set up the mass_l and mass_u values
la::AddOverwrite(b_twy_l_estimate_[min_pointer],qnode->stat().b_twy_more_l,&qnode->stat().b_twy_mass_l);
la::AddOverwrite(b_twy_u_estimate_[max_pointer],qnode->stat().b_twy_more_u,&qnode->stat().b_twy_mass_u);
la::AddOverwrite(b_twy_l_estimate_[min_pointer],
qnode->stat().b_twy_more_l,&qnode->stat().b_twy_mass_l);
la::AddOverwrite(b_twy_u_estimate_[max_pointer],
qnode->stat().b_twy_more_u,&qnode->stat().b_twy_mass_u);
}
}
@@ -890,6 +963,7 @@ void FastRegression<TKernel>::FRegressionBaseB_TWY_(Tree *qnode, Tree *rnode){
template <typename TKernel>
void FastRegression<TKernel>::FRegressionBaseB_TWB_(Tree *qnode,Tree *rnode){
printf("Hit the base case for BTWB\n");
la::SubFrom (rnode->stat().b_tb, &qnode->stat().b_twb_more_u);
@@ -973,7 +1047,7 @@ void FastRegression<TKernel>::FRegressionBaseB_TWB_(Tree *qnode,Tree *rnode){
}
}
}
}
}
if(pruning_criteria==CRITERIA_FOR_PRUNE_COMPONENT){
//Loop over each point and set the max and min
@@ -1022,35 +1096,37 @@ void FastRegression<TKernel>::FRegressionBaseB_TWB_(Tree *qnode,Tree *rnode){
qnode->stat().b_twb_mass_u.CopyValues(max_u);
qnode->stat().b_twb_mass_l.CopyValues(min_l);
return;
}
//We are interested in frobenius norm pruning
else{
//We are doing frobenius norm pruning
/** get a tighter lower and upper boiounf by looping over each query
* point to find that particular query point that has the least sqd
* frobenius norm of the matrix b_twy_l_estimate_[q]+qnode->stat().b_twy_more_l
* frobenius norm of the matrix b_twb_l_estimate_[q]+qnode->stat().b_twb_more_l
*/
double min_norm=DBL_MAX;
double max_norm=DBL_MIN;
Matrix temp;
Matrix temp1;
Matrix temp2;
//The min_pointer and the max_pointer store the index number of the
//query point which possibly has the least and the highest squared frobenius norm
index_t min_pointer;
index_t max_pointer;
index_t min_pointer=qnode->begin();
index_t max_pointer=qnode->begin();
temp.Init(rset_.n_rows()+1,rset_.n_rows()+1);
temp1.Init(rset_.n_rows()+1,rset_.n_rows()+1);
temp2.Init(rset_.n_rows()+1,rset_.n_rows()+1);
for(index_t i=qnode->begin();i<qnode->end();i++){
//look for the lower bound
la::AddOverwrite(b_twb_l_estimate_[i],qnode->stat().b_twb_more_l,&temp);
double var=SquaredFrobeniusNorm_(temp);
la::AddOverwrite(b_twb_l_estimate_[i],qnode->stat().b_twb_more_l,&temp1);
double var=SquaredFrobeniusNorm_(temp1);
if(var< min_norm){
min_pointer=i;
@@ -1058,20 +1134,22 @@ void FastRegression<TKernel>::FRegressionBaseB_TWB_(Tree *qnode,Tree *rnode){
}
//Look for the upper bound
la::AddOverwrite(b_twb_u_estimate_[i],qnode->stat().b_twb_more_u,&temp);
var=SquaredFrobeniusNorm_(temp);
if(var>max_norm){
la::AddOverwrite(b_twb_u_estimate_[i],qnode->stat().b_twb_more_u,&temp2);
var=SquaredFrobeniusNorm_(temp2);
if(var > max_norm){
max_pointer=i;
max_norm=var;
}
}
//Once done with the looping process set up the mass_l and mass_u values
la::AddOverwrite(b_twb_l_estimate_[min_pointer],qnode->stat().b_twb_more_l,&qnode->stat().b_twb_mass_l);
la::AddOverwrite(b_twb_u_estimate_[max_pointer],qnode->stat().b_twb_more_u,&qnode->stat().b_twb_mass_u);
la::AddOverwrite(b_twb_l_estimate_[min_pointer],
qnode->stat().b_twb_more_l,&qnode->stat().b_twb_mass_l);
la::AddOverwrite(b_twb_u_estimate_[max_pointer],
qnode->stat().b_twb_more_u,&qnode->stat().b_twb_mass_u);
}
}
@@ -1118,31 +1196,28 @@ void FastRegression<TKernel>::PostProcess_(Tree *qnode){
//mass_u <- owed_u+mass_u
UpdateBounds_(qnode, qnode->stat().b_twb_owed_l,
qnode->stat().b_twb_owed_u,
qnode->stat().b_twy_owed_l,
qnode->stat().b_twy_owed_u, CHECK_FOR_PRUNE_BOTH);
if(!qnode->is_leaf()){
//Having completed all the tree calculations we now update bounds for both
//the qunatities. this can be done by calling the UpdateBounds_ function
//with the flag set to CHECK_FOR_PRUNE_BOTH
UpdateBounds_(qnode, qnode->stat().b_twb_owed_l,
qnode->stat().b_twb_owed_u,
qnode->stat().b_twy_owed_l,
qnode->stat().b_twy_owed_u, CHECK_FOR_PRUNE_BOTH);
PostProcess_(qnode->left());
PostProcess_(qnode->right());
}
else{
UpdateBounds_(qnode, qnode->stat().b_twb_owed_l,
qnode->stat().b_twb_owed_u,
qnode->stat().b_twy_owed_l,
qnode->stat().b_twy_owed_u, CHECK_FOR_PRUNE_BOTH);
// b_twb_e_estimate=
//0.5*(b_twy_l_estimate+b_twy_u_estimate+b_twy_more_l+b_twy_more_u)
//0.5*(b_twb_l_estimate+b_twb_u_estimate+b_twb_more_l+b_twb_more_u)
for(index_t q=qnode->begin();q<qnode->end();q++){
FastRegressionStat qstat=qnode->stat();
@@ -1150,16 +1225,16 @@ void FastRegression<TKernel>::PostProcess_(Tree *qnode){
la::AddInit(b_twb_l_estimate_[q],
b_twb_u_estimate_[q],&estimate_mean1);
la::Scale(0.50,&estimate_mean1);
// la::Scale(0.50,&estimate_mean1);
Matrix more_mean1;
la::AddInit(qstat.b_twb_more_l,qstat.b_twb_more_u,&more_mean1);
la::Scale(0.50,&more_mean1);
//la::Scale(0.50,&more_mean1);
//Final estimate is just the sum of estimate_mean and more_mean
la::AddOverwrite(estimate_mean1,more_mean1,&b_twb_e_estimate_[q]);
la::Scale(0.50,&b_twb_e_estimate_[q]);
// b_twy_e_estimate=
//0.5*(b_twy_l_estimate+b_twy_u_estimate+b_twy_more_l+b_twy_more_u)
@@ -1168,15 +1243,17 @@ void FastRegression<TKernel>::PostProcess_(Tree *qnode){
la::AddInit(b_twy_l_estimate_[q],
b_twy_u_estimate_[q],&estimate_mean2);
la::Scale(0.50,&estimate_mean2);
// la::Scale(0.50,&estimate_mean2);
Matrix more_mean2;
la::AddInit(qstat.b_twy_more_l,qstat.b_twy_more_u,&more_mean2);
la::Scale(0.50,&more_mean2);
// la::Scale(0.50,&more_mean2);
//Final estimate is just the sum of estimate_mean and more_mean
la::AddOverwrite(estimate_mean2,more_mean2,&b_twy_e_estimate_[q]);
la::Scale(0.50,&b_twy_e_estimate_[q]);
}
}
}
@@ -1298,6 +1375,20 @@ void FastRegression<TKernel>::PrintRegressionEstimate_(){
fclose(gp);
}
template<typename TKernel>
void FastRegression<TKernel>::CompareWithTrueValues_(){
//We have regression_estimate_ and true_regression_values_.Lets get
//mean squared error
double error=0.0;
for(index_t q=0;q<qset_.n_cols();q++){
double diff=regression_estimate_[q]-true_regression_values_[q];
error+=diff*diff;
}
double mean_squared_error=error/qset_.n_cols();
printf("Mean squared error when compared with true regression values is %f\n",mean_squared_error);
}
#endif
@@ -1308,6 +1399,3 @@ void FastRegression<TKernel>::PrintRegressionEstimate_(){
fastlib/u/gmravi/regression/regression_ll_naive.h
+51 -16
View File
@@ -1,7 +1,7 @@
#ifndef REGRESSION_LL_NAIVE_H
#define REGRESSION_LL_NAIVE_H
#include "fastlib/fastlib_int.h"
#include "pseudo_inverse.h"
//#include "pseudo_inverse.h"
template <typename TKernel>
double NaiveCalculation<TKernel>::SquaredFrobeniusNorm_(Matrix &a){
@@ -9,9 +9,9 @@ double NaiveCalculation<TKernel>::SquaredFrobeniusNorm_(Matrix &a){
//This function computes treats the natrix as a vector and computes
//it's 1-norm
double value=0;
for(index_t col=0;col<a.n_cols();col++){
for(index_t row=0;row<a.n_rows();row++){
value+=a.get(row,col)*a.get(row,col);
for(index_t row=0;row<a.n_rows();row++){
for(index_t col=0;col<a.n_cols();col++){
value+=(a.get(row,col)*a.get(row,col));
}
}
return value;
@@ -62,7 +62,7 @@ void NaiveCalculation<TKernel>::Compute (){
//Fill B^TWY naive
double val=b_twy_naive_estimate_[q].get(row,col)+
ker_value * rset_weights_[old_from_new_r_[r]] * rset_.get (row- 1, r);
ker_value * rset_weights_[old_from_new_r_[r]] * rset_.get (row-1, r);
b_twy_naive_estimate_[q].set(row,col,val);
@@ -75,6 +75,7 @@ void NaiveCalculation<TKernel>::Compute (){
}
else{
//this is row =0
//Fill B^TWY naive
@@ -120,7 +121,7 @@ void NaiveCalculation<TKernel>::Compute (){
Print_();
//Having done this get regression estimates by calling the function
//ObtainRegressionEstimates
ObtainRegressionEstimate_();
//ObtainRegressionEstimate_();
//PrintRegressionEstimate_(fast_regression_estimate);
}
@@ -181,10 +182,16 @@ void NaiveCalculation<TKernel>::ObtainRegressionEstimate_(){
template <typename TKernel>
void NaiveCalculation<TKernel>:: ComputeMaximumRelativeError(ArrayList<Matrix> &fast_b_twy_estimate, ArrayList<Matrix> & fast_b_twb_estimate, char *pruning_criteria){
void NaiveCalculation<TKernel>::
ComputeMaximumRelativeError(ArrayList<Matrix> &fast_b_twy_estimate,
ArrayList<Matrix> & fast_b_twb_estimate,
char *pruning_criteria){
if(!strcmp(pruning_criteria,"fnorm")){
FILE *fp;
fp=fopen("errors.txt","w+");
double max_frobenius_error_b_twb=0;
double max_frobenius_error_b_twy=0;
@@ -195,27 +202,49 @@ void NaiveCalculation<TKernel>:: ComputeMaximumRelativeError(ArrayList<Matrix> &
la::SubInit (b_twy_naive_estimate_[q],fast_b_twy_estimate[q] , &temp1);
double f_norm=SquaredFrobeniusNorm_(temp1);
double f_norm_naive=SquaredFrobeniusNorm_(b_twy_naive_estimate_[q]);
double rel_error_b_twy=fabs(f_norm)/f_norm_naive;
//printf("relative frobenius norm error for BTWY is %f\n",rel_error_b_twy);
double rel_error_b_twy;
if(f_norm==0.0 && f_norm_naive==0.0){
rel_error_b_twy=0;
}
else{
rel_error_b_twy=fabs(f_norm)/f_norm_naive;
}
fprintf(fp,"\n");
fprintf(fp,"diff norm is %f\n",f_norm);
fprintf(fp,"naive norm is %f\n",f_norm_naive);
fprintf(fp,"relative frobenius norm error for BTWY is %f\n",rel_error_b_twy);
if(max_frobenius_error_b_twy<rel_error_b_twy){
max_frobenius_error_b_twy=rel_error_b_twy;
}
//Comparing BTWB estimates with the naive estimate
Matrix temp2;
la::SubInit (b_twb_naive_estimate_[q],fast_b_twb_estimate[q] , &temp2);
f_norm=SquaredFrobeniusNorm_(temp2);
f_norm_naive=SquaredFrobeniusNorm_(b_twb_naive_estimate_[q]);
double rel_error_b_twb=fabs(f_norm)/f_norm_naive;
// printf("relative frobenius norm error for BTWB is %f\n",rel_error_b_twb);
double rel_error_b_twb=0.0;
if(f_norm==0.0 && f_norm_naive==0.0){
rel_error_b_twb=0;
}
else{
rel_error_b_twb=fabs(f_norm)/f_norm_naive;
}
fprintf(fp,"\n");
fprintf(fp,"diff norm is %f\n",f_norm);
fprintf(fp,"naive norm is %f\n",f_norm_naive);
fprintf(fp,"relative frobenius norm error for BTWB is %f\n",rel_error_b_twb);
if(max_frobenius_error_b_twb < rel_error_b_twb){
max_frobenius_error_b_twb= rel_error_b_twb;
}
}
fprintf(fp,"The max frobenius error for BTWY is %f\n",max_frobenius_error_b_twy);
fprintf(fp,"The max frobenius error for BTWB is %f\n",max_frobenius_error_b_twb);
printf("The max frobenius error for BTWY is %f\n",max_frobenius_error_b_twy);
printf("The max frobenius error for BTWB is %f\n",max_frobenius_error_b_twb);
}
@@ -237,6 +266,9 @@ void NaiveCalculation<TKernel>:: ComputeMaximumRelativeError(ArrayList<Matrix> &
fabs(fast_b_twb_estimate[q].get(row,col)-b_twb_naive_estimate_[q].get(row,col));
double error=diff/b_twb_naive_estimate_[q].get(row,col);
printf("diff =%f\n",diff);
printf("Being compared against %f\n",b_twb_naive_estimate_[q].get(row,col));
if(error>max_error_for_this_point){
max_error_for_this_point=error;
@@ -245,8 +277,8 @@ void NaiveCalculation<TKernel>:: ComputeMaximumRelativeError(ArrayList<Matrix> &
}
}//Calculations complete for this point
printf("Maximum error for this point is %f\n",max_error_for_this_point);
if(max_error_for_this_point<max_error_on_the_whole){
printf("Maximum error for this point BTWB is %f\n",max_error_for_this_point);
if(max_error_for_this_point>max_error_on_the_whole){
max_error_on_the_whole=max_error_for_this_point;
}
@@ -266,7 +298,10 @@ void NaiveCalculation<TKernel>:: ComputeMaximumRelativeError(ArrayList<Matrix> &
double diff=
fabs(fast_b_twy_estimate[q].get(row,col)-b_twy_naive_estimate_[q].get(row,col));
double error=diff/b_twb_naive_estimate_[q].get(row,col);
double error=diff/b_twy_naive_estimate_[q].get(row,col);
printf("diff =%f\n",diff);
printf("Being compared against %f\n",b_twy_naive_estimate_[q].get(row,col));
if(error>max_error_for_this_point){
max_error_for_this_point=error;
@@ -274,9 +309,9 @@ void NaiveCalculation<TKernel>:: ComputeMaximumRelativeError(ArrayList<Matrix> &
}
} //Calculations complete for this point
printf("Maximum error for this point is %f\n",max_error_for_this_point);
printf("Maximum error for this point BTWY is %f\n",max_error_for_this_point);
if(max_error_for_this_point<max_error_on_the_whole){
if(max_error_for_this_point>max_error_on_the_whole){
max_error_on_the_whole=max_error_for_this_point;
}
}
fastlib/u/gmravi/regression/regression_parent.h
+8 -2
View File
@@ -207,6 +207,9 @@ public:
/** Regression estimates of the query points */
Vector regression_estimate_;
/* True regression values of the query points */
Vector true_regression_values_;
/** Mappings from old dataset to new dataset
* Remember that when we build the query tree and the
* reference tree out of the query and reference
@@ -246,8 +249,9 @@ public:
void FRegressionBase_(Tree *qnode, Tree *rnode, check_for_prune_t flag);
void MergeChildBounds_(Tree *,check_for_prune_t flag);
void MergeChildBounds_(Tree *,prune_t flag);
void MergeChildBounds_(Tree *,check_for_prune_t flag);
void CallRecursively_(Tree *qnode, Tree *rnode,check_for_prune_t flag);
@@ -283,6 +287,8 @@ public:
void PrintRegressionEstimate_();
double SquaredFrobeniusNorm_(Matrix &);
double Compute1NormLike_(Matrix &);
void CompareWithTrueValues_();
public:
@@ -305,7 +311,7 @@ public:
void Compute();
void Init(Matrix &q_matrix, Matrix &r_matrix, double bandwidth,
double tau,index_t leaf_length,Vector &rset_weights, char *);
double tau,index_t leaf_length,Vector &rset_weights, char *, Vector &);
};