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

Browse Source
This commit is contained in:
houyang
2011-10-23 16:43:15 +00:00
parent 587f0efa29
commit 31c7b43554
6 changed files with 30 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastlib/trunk/contrib/houyang/pole/exp_nasa.txt
+1 -1
View File
@@ -52,7 +52,7 @@ Examples:
================================ Batch learning ================================
./pole_pt --batch 1 -d svmguide1 -t svmguide1 --random 1 -m nasa --type classification -l hinge --reg 2 --bias 1 --lambda 0.01 --threads_test 2 -e 600 --dbound 1.5
472.062 0.94788 2.45
461.375 0.950146 2.77
./pole_pt --batch 1 -d rcv1_train.binary -t rcv1_test.binary --random 1 -m nasa --type classification -l hinge --reg 2 --bias 1 --lambda 0.00001 --threads_test 4 -e 1 --dbound 16
83341.9 0.95506 2:26.56
fastlib/trunk/contrib/houyang/pole/exp_ogd.txt
+4
View File
@@ -51,6 +51,8 @@ Examples:
================================ Batch learning ================================
./pole_pt --batch 1 -d svmguide1 -t svmguide1 --random 1 -m ogd --type classification -l hinge --reg 2 --bias 1 --lambda 0.01 --threads_test 2 -e 600 --dbound 0.025
485.553 0.939786 1.72
./pole_pt --batch 1 -d svmguide1 -t svmguide1 --random 1 -m ogd_a --type classification -l hinge --reg 2 --bias 1 --lambda 0.01 --threads_test 2 -e 600 --dbound 5500 --strongness 0.01
487.144 0.942052 2.25
./pole_pt --batch 1 -d rcv1_train.binary -t rcv1_test.binary --random 1 -m ogd --type classification -l hinge --reg 2 --bias 1 --lambda 0.00001 --threads_test 4 -e 1 --dbound 17
92298.2 0.95138 2:19.310817
@@ -64,3 +66,5 @@ Examples:
53690 0.77541 2.50
./pole_pt --batch 1 -d zeta_train --td_ratio 0.2 --random 1 -m ogd --type classification -l hinge --reg 2 --bias 1 --lambda 0.00001 --threads_test 4 -e 1 --dbound 65
98849.5 0.50529 10.31
fastlib/trunk/contrib/houyang/pole/learner.cc
+1
View File
@@ -11,6 +11,7 @@ Learner::Learner() {
TE_ = NULL;
LF_ = NULL;
LOG_ = NULL;
batch_ = false; // default: online learning
}
///////////////
fastlib/trunk/contrib/houyang/pole/learner.h
+1
View File
@@ -24,6 +24,7 @@ class Learner {
Data *VA_; // validation set
Data *TE_; // testing set
bool batch_; // batch learning or not
bool random_data_;
T_IDX n_source_, port_;
bool read_port_; // read data from port or file
fastlib/trunk/contrib/houyang/pole/opt_ogd.cc
+22 -12
View File
@@ -61,14 +61,15 @@ void* OGD::LearnThread(void *in_par) {
// Make prediction and get loss
for (T_IDX b = 0; b<Lp->mb_size_; b++) {
/*
// calculate w_avg and make logs
Lp->w_avg_pool_[tid] *= (Lp->t_n_it_[tid] - 1.0);
Lp->w_avg_pool_[tid] += Lp->w_pool_[tid];
Lp->w_avg_pool_[tid] *= (1.0/Lp->t_n_it_[tid]);
*/
Lp->w_avg_pool_[tid] = Lp->w_pool_[tid];
double pred_val = Lp->LinearPredictBias(Lp->w_avg_pool_[tid],
// for batch: calculate w_avg
if (Lp->batch_) {
Lp->w_avg_pool_[tid] *= (Lp->t_n_it_[tid] - 1.0);
Lp->w_avg_pool_[tid] += Lp->w_pool_[tid];
Lp->w_avg_pool_[tid] *= (1.0/Lp->t_n_it_[tid]);
//Lp->w_avg_pool_[tid] = Lp->w_pool_[tid];
}
// online prediction
double pred_val = Lp->LinearPredictBias(Lp->w_pool_[tid],
*exs[b], Lp->b_pool_[tid]);
Lp->MakeLearnLog(tid, exs[b], pred_val);
update = Lp->LF_->GetUpdate(pred_val, (double)exs[b]->y_);
@@ -83,6 +84,15 @@ void* OGD::LearnThread(void *in_par) {
else if (Lp->opt_name_ == "ogd") {
eta = Lp->dbound_ / sqrt(Lp->t_n_it_[tid]);
}
else if (Lp->opt_name_ == "ogd_str_a") { // Xu's Averaged SGD
// 1/( M(1+t\lambda_0/M)^2/3 )
eta = 1 / ( Lp->dbound_ * pow(1+Lp->strongness_*Lp->t_n_it_[tid]/Lp->dbound_, 2/3) );
}
// Assuming general convexity: ogd
else if (Lp->opt_name_ == "ogd_a") { // Xu's Averaged SGD
// 1/( M(1+t\lambda_0/M)^3/4 )
eta = 1 / ( Lp->dbound_ * pow(1+Lp->strongness_*Lp->t_n_it_[tid]/Lp->dbound_, 0.75) );
}
else {
cout << "ERROR! Unkown OGD method."<< endl;
exit(1);
@@ -170,7 +180,7 @@ void* OGD::TestThread(void *in_par) {
while (true) {
if ( Lp->GetTestExample(Lp->TE_, exs, tid) ) { // new test example read
// testing using Thread[0]'s w & b. TODO...
// testing using Thread[0]'s running-averaged w & b. TODO...
double pred_val = Lp->LinearPredictBias(Lp->w_avg_pool_[0],
*exs[0], Lp->b_pool_[0]);
Lp->MakeTestLog(tid, exs[0], pred_val);
@@ -203,11 +213,11 @@ void OGD::MakeLearnLog(T_IDX tid, Example *x, double pred_val) {
if (reg_type_ == 2 && reg_factor_ != 0) {
//L + \lambda/2 \|w\|^2 <=> CL + 1/2 \|w\|^2
t_loss_[tid] = t_loss_[tid] +
0.5 * reg_factor_ * w_avg_pool_[tid].SparseSqL2Norm();
0.5 * reg_factor_ * w_pool_[tid].SparseSqL2Norm();
}
// for classification only: calc # of misclassifications
if (type_ == "classification") {
T_LBL pred_lbl = LinearPredictBiasLabelBinary(w_avg_pool_[tid], *x, b_pool_[tid]);
T_LBL pred_lbl = LinearPredictBiasLabelBinary(w_pool_[tid], *x, b_pool_[tid]);
//cout << x->y_ << " : " << pred_lbl << endl;
if (pred_lbl != x->y_) {
t_err_[tid] = t_err_[tid] + 1;
@@ -292,7 +302,7 @@ void OGD::SaveLearnLog() {
void OGD::MakeTestLog(T_IDX tid, Example *x, double pred_val) {
//cout << "pred: " << pred_val <<", y: " << (double)x->y_ << endl;
t_test_loss_[tid] = t_test_loss_[tid] + LF_->GetLoss(pred_val, (double)x->y_);
// testing using Thread[0]'s w. TODO...
// testing using Thread[0]'s running-averaged w. TODO...
if (reg_type_ == 2 && reg_factor_ != 0) {
//L + \lambda/2 \|w\|^2 <=> CL + 1/2 \|w\|^2
t_loss_[tid] = t_loss_[tid] +
fastlib/trunk/contrib/houyang/pole/pole.cc
+1
View File
@@ -305,6 +305,7 @@ void Pole::Run() {
L_->OnlineLearn();
}
else { // Batch Learning
L_->batch_ = true;
L_->BatchLearn();
}
}