Menu
▾
▴
[Personalrobots-commit] SF.net SVN: personalrobots:[23343] pkg/trunk/sandbox/dorylus
|
From: <ale...@us...> - 2009-08-30 03:55:18
|
Revision: 23343
http://personalrobots.svn.sourceforge.net/personalrobots/?rev=23343&view=rev
Author: alexteichman
Date: 2009-08-30 03:55:07 +0000 (Sun, 30 Aug 2009)
Log Message:
-----------
Adding a pruning utility to the classifier.
Modified Paths:
--------------
pkg/trunk/sandbox/dorylus/dorylus.cpp
pkg/trunk/sandbox/dorylus/dtools.cpp
pkg/trunk/sandbox/dorylus/include/dorylus/dorylus.h
Modified: pkg/trunk/sandbox/dorylus/dorylus.cpp
===================================================================
--- pkg/trunk/sandbox/dorylus/dorylus.cpp 2009-08-30 02:59:32 UTC (rev 23342)
+++ pkg/trunk/sandbox/dorylus/dorylus.cpp 2009-08-30 03:55:07 UTC (rev 23343)
@@ -624,14 +624,15 @@
g_int = true;
}
-void Dorylus::relearnResponses(DorylusDataset& dd) {
+//! Relearn the responses for all weak classifiers using a new dataset. Weak classifier utilities will be
+//! recomputed for this dataset, and those with utility less than min_util will be thrown out.
+void Dorylus::relearnResponses(DorylusDataset& dd, double min_util, int max_wcs) {
assert(dd.nClasses_ == nClasses_);
useDataset(&dd);
vector<object*>& objs = dd_->objs_;
cout << "Objective before response relearning: " << classify(dd) << endl;
- save("tmp.d");
for(size_t t=0; t<pwcs_.size(); ++t) {
weak_classifier& wc = *pwcs_[t];
@@ -640,15 +641,11 @@
VectorXd numerators = VectorXd::Zero(nClasses_);
VectorXd denominators = VectorXd::Zero(nClasses_);
-// for(size_t c=0; c<nClasses_; ++c) {
-// wc.numerators(c) = 0;
-// wc.denominators(c) = 0;
-// wc.vals(c) = 0;
-// }
-
// -- Find which training examples fall in the hypersphere and increment numerators, denominators, responses, and utility.
vector<size_t> inside;
inside.reserve(objs.size());
+ VectorXd sum_weights_pos = VectorXd::Zero(nClasses_);
+ VectorXd sum_weights_neg = VectorXd::Zero(nClasses_);
for(size_t m=0; m<objs.size(); ++m) {
// -- If no descriptor of this type, ignore.
@@ -667,13 +664,6 @@
double weight = exp(log_weights_(c,m));
numerators(c) += weight * dd_->ymc_(c,m);
denominators(c) += weight;
-
-
- // Total weight of training examples in hypersphere (from *all* datasets, not just this one) which have ymc={1, -1}
-// double sum_weights_pos = (wc.denominators(c) + wc.numerators(c)) / 2;
-// double sum_weights_neg = (wc.denominators(c) - wc.numerators(c)) / 2;
-
-// wc.utility += ((1-exp(-wc.vals(c))) * sum_weights_pos + (1-exp(wc.vals(c))) * sum_weights_neg) / (nClasses_ * objs.size());
}
}
@@ -690,9 +680,103 @@
log_weights_(c,idx) += -dd_->ymc_(c,idx) * wc.vals(c);
}
}
+
+ // -- Compute the utility.
+ wc.utility = 0;
+ for(size_t c=0; c<nClasses_; ++c) {
+ for(size_t m=0; m<inside.size(); ++m) {
+ size_t idx = inside[m];
+ if(dd_->ymc_(c,idx) == 1)
+ sum_weights_pos(c) += exp(log_weights_(c,idx));
+ else if(dd_->ymc_(c,idx) == -1)
+ sum_weights_neg(c) += exp(log_weights_(c,idx));
+ else {
+ cout << "ymc must be in -1, +1" << endl;
+ assert(0);
+ }
+ }
+
+ wc.utility += (exp(wc.vals(c)) - 1) * sum_weights_pos(c);
+ wc.utility += (exp(-wc.vals(c)) - 1) * sum_weights_neg(c);
+ }
+
}
- cout << "Objective after response updating: " << computeObjective() << endl;
- cout << "Objective after response updating: " << classify(dd) << endl;
+
+ double objective = classify(dd);
+ cout << "Objective after response relearning: " << objective << endl;
+// cout << "Objective after response updating: " << computeObjective() << endl;
+
+ // -- Pruning.
+ if(min_util > 0) {
+ cout << "WARNING: Minimum utility pruning is not yet implemented." << endl;
+ }
+ if(max_wcs > 0) {
+ vector< pair<double, int> > util_idx(pwcs_.size());
+ for(size_t t=0; t<pwcs_.size(); ++t) {
+ util_idx[t].first = pwcs_[t]->utility;
+ util_idx[t].second = t;
+ }
+ greater< pair<double, int> > emacs = greater< pair<double, int> >();
+ sort(util_idx.begin(), util_idx.end(), emacs); //Descending.
+
+ cout << "Pruning " << pwcs_.size() - max_wcs << " weak classifiers." << endl;
+// double min_util_in_classifier = util_idx.back().first;
+
+// cout << status() << endl;
+
+// map<string, vector<weak_classifier*> >::iterator it;
+// for(it=battery_.begin(); it != battery_.end(); it++) {
+// //cout << it->first << endl;
+// vector<weak_classifier*> wcs = it->second;
+// for(size_t j=0; j<wcs.size(); ++j)
+// cout << displayWeakClassifier(*wcs[j]) << endl;
+// }
+
+
+ // -- Make new pwcs_ and battery_.
+ vector<weak_classifier*> pwcs_new;
+ map<string, vector<weak_classifier*> > battery_new;
+ pwcs_new.reserve(max_wcs);
+ for(size_t t=0; t<pwcs_.size(); ++t) {
+ if((int)t < max_wcs) {
+ //cout << "keeping wc " << util_idx[t].second << " with util " << util_idx[t].first << endl;
+ weak_classifier* pwc = pwcs_[util_idx[t].second];
+ pwc->id = t;
+
+ pwcs_new.push_back(pwc);
+ battery_new[pwc->descriptor].push_back(pwc);
+ }
+ else {
+ //cout << "deleting wc " << util_idx[t].second << " with util " << util_idx[t].first << endl;
+ delete pwcs_[util_idx[t].second];
+ }
+ }
+
+
+ pwcs_ = pwcs_new;
+ battery_ = battery_new;
+
+
+
+ double objective_after_pruning = classify(dd);
+// cout << "same: " << objective_after_pruning - objective << " " << min_util_in_classifier << endl;
+ cout << "Objective after pruning: " << objective_after_pruning << endl;
+ cout << status() << endl;
+
+
+// for(it=battery_.begin(); it != battery_.end(); it++) {
+// //cout << it->first << endl;
+// vector<weak_classifier*> wcs = it->second;
+// for(size_t j=0; j<wcs.size(); ++j)
+// cout << displayWeakClassifier(*wcs[j]) << endl;
+// }
+
+
+ //cout << "Relearning responses." << endl;
+ // relearnResponses(dd); // TODO: Should instead be able to decrement log_weights_ for those pruned, then recompute responses without redoing euc distances.
+ }
+
+
}
void Dorylus::resumeTraining(int num_candidates, int max_secs, int max_wcs, double min_util, void (*debugHook)(weak_classifier)) {
Modified: pkg/trunk/sandbox/dorylus/dtools.cpp
===================================================================
--- pkg/trunk/sandbox/dorylus/dtools.cpp 2009-08-30 02:59:32 UTC (rev 23342)
+++ pkg/trunk/sandbox/dorylus/dtools.cpp 2009-08-30 03:55:07 UTC (rev 23343)
@@ -66,7 +66,7 @@
return 1;
}
- d.relearnResponses(dd);
+ d.relearnResponses(dd, 0, max_wcs);
d.save(argv[4]);
}
@@ -145,6 +145,7 @@
cout << endl;
cout << argv[0] << " --relearnResponses DATASET OLD_CLASSIFIER NEW_CLASSIFIER_SAVENAME" << endl;
+ cout << " MAX_WCS=x is the max number of weak classifiers to allow in the final classifier: the rest will be pruned by utility. Default infinite." << endl;
cout << endl;
cout << argv[0] << " --resumeTraining DATASET OLD_CLASSIFIER NEW_CLASSIFIER_SAVENAME" << endl;
Modified: pkg/trunk/sandbox/dorylus/include/dorylus/dorylus.h
===================================================================
--- pkg/trunk/sandbox/dorylus/include/dorylus/dorylus.h 2009-08-30 02:59:32 UTC (rev 23342)
+++ pkg/trunk/sandbox/dorylus/include/dorylus/dorylus.h 2009-08-30 03:55:07 UTC (rev 23343)
@@ -186,7 +186,7 @@
//! Continues training a classifier, possibly on a new dataset.
void resumeTraining(int num_candidates, int max_secs, int max_wcs, double min_util, void (*debugHook)(weak_classifier)=NULL);
//! Relearns the weak classifier responses on a new labeled dataset
- void relearnResponses(DorylusDataset& dd);
+ void relearnResponses(DorylusDataset& dd, double min_util=0, int max_wcs=0);
//! Sets log_weights_, classes_, nClasses_, and dd_.
void useDataset(DorylusDataset *dd);
bool save(std::string filename, std::string *user_data_str=NULL);
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
