search_branch_and_bound_fast.hpp

Go to the documentation of this file.

#ifndef FSTSEARCHBRANCHANDBOUNDFAST_H
#define FSTSEARCHBRANCHANDBOUNDFAST_H
 /* 
=========================================================================
Copyright and license:
  * FST4 software (with the exception of any bundled or externally linked
    third-party components) is Copyright (C) 2000-present Institute of
    Information Theory and Automation of the Czech Academy of Sciences (UTIA).
  * FST4 is free software: you can redistribute it and/or modify it under
    the terms of the GNU Affero General Public License as published by the
    Free Software Foundation, version 3. The full text of the license is in
    the file LICENSE in the distribution root and at
    https://www.gnu.org/licenses/agpl-3.0.html.
  * FST4 is distributed in the hope that it will be useful, but WITHOUT ANY
    WARRANTY; without even the implied warranty of MERCHANTABILITY or
    FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public
    License for more details.
  * Closed-source use without the AGPL's copyleft obligations requires a
    commercial license from UTIA; in case of interest write to
    utia@utia.cas.cz. LICENSING.md in the distribution root explains both
    options, the contributor license agreement that every contribution
    requires, and the contacts (technical questions and pull requests:
    somol@utia.cas.cz, with "FST" in the subject line).
========================================================================== */
 
#include <memory>
#include <cassert>
#include <iostream>
#include "branch_and_bound_predictor.hpp"
#include "search_branch_and_bound.hpp"
#include "error.hpp"
#include "global.hpp"
#include "result_tracker.hpp"
 
/*============== Template parameter type naming conventions ==============
--------- Numeric types: -------------------------------------------------
DATATYPE - data sample values - usually real numbers (but may be integers
          in text processing etc.)
REALTYPE - must be real numbers - for representing intermediate results of 
          calculations like mean, covariance etc.
IDXTYPE - index values for enumeration of data samples - (nonnegative) integers, 
          extent depends on numbers of samples in data
DIMTYPE - index values for enumeration of features (dimensions), or classes (not 
          class sizes) - (nonnegative) integers, usually lower extent than IDXTYPE, 
          but be aware of expressions like _classes*_features*_features ! 
          in linearized representations of feature matrices for all classes
BINTYPE - feature selection marker type - represents ca. <10 different feature 
          states (selected, deselected, sel./desel. temporarily 1st nested loop, 2nd...)
RETURNTYPE - criterion value: real value, but may be extended in future to support 
          multiple values 
--------- Class types: ---------------------------------------------------
SUBSET       - class of class type Subset 
CLASSIFIER   - class implementing interface defined in abstract class Classifier 
EVALUATOR    - class implementing interface defined in abstract class Sequential_Step 
DISTANCE     - class implementing interface defined in abstract class Distance 
DATAACCESSOR - class implementing interface defined in abstract class Data_Accessor 
INTERVALCONTAINER - class of class type Data_Intervaller 
CONTAINER    - STL container of class type Data_Interval  
========================================================================== */
 
namespace FST {

 
template<class RETURNTYPE, typename DIMTYPE, class SUBSET, class CRITERION, class PREDICTOR>
class Search_Branch_And_Bound_Fast : public Search_Branch_And_Bound<RETURNTYPE,DIMTYPE,SUBSET,CRITERION>{
public:
        typedef Search_Branch_And_Bound<RETURNTYPE,DIMTYPE,SUBSET,CRITERION> parent; 
        typedef typename parent::PCriterion PCriterion; 
        typedef typename parent::PSubset PSubset; 
        typedef typename parent::PNode PNode; 
        typedef typename parent::Node Node; 
        typedef typename parent::NodeType NodeType; 
        typedef typename parent::PEvaluatorEngine PEvaluatorEngine; 
        explicit Search_Branch_And_Bound_Fast(PEvaluatorEngine engine=PEvaluatorEngine()):Search_Branch_And_Bound<RETURNTYPE,DIMTYPE,SUBSET,CRITERION>(engine) {notify("Search_Branch_And_Bound_Fast constructor.");} 
        virtual ~Search_Branch_And_Bound_Fast() {notify("Search_Branch_And_Bound_Fast destructor.");}
 
        // for search() implementation see parent class
 
        void set_gamma(const RETURNTYPE gamma=1.0) {_predictor.set_gamma(gamma);} 
        void set_delta(const unsigned int delta=1) {_predictor.set_delta(delta);} 

        virtual std::ostream& print(std::ostream& os) const override {
                if(parent::get_engine()->is_threaded()) os << "Fast Branch & Bound (FBB)  [Search_Branch_And_Bound_Fast(threads="<<parent::get_engine()->get_max_threads()<<") with " << _predictor << "]";
                else os << "Fast Branch & Bound (FBB)  [Search_Branch_And_Bound_Fast() with " << _predictor << "]";
                return os;
        };
protected:
        PREDICTOR _predictor; 
 
        // the following may be overriden in descendant Branch and Bound specialization classes
        virtual void initialize(const DIMTYPE d, const DIMTYPE n, const PCriterion crit) override {_predictor.initialize(n);} 
        virtual void leaf_result(const DIMTYPE feature, const RETURNTYPE value) override; 
        virtual void pre_evaluate_availables() override; 
        virtual void post_process_tree_level() override {} 
        virtual bool cut_possible() override; 
};
 
template<class RETURNTYPE, typename DIMTYPE, class SUBSET, class CRITERION, class PREDICTOR>
void Search_Branch_And_Bound_Fast<RETURNTYPE,DIMTYPE,SUBSET,CRITERION,PREDICTOR>::leaf_result(const DIMTYPE feature, const RETURNTYPE value)
{
        // update the prediction info - but only from a parent whose value was actually
        // computed: the drop learned here is parent value - leaf value, and against a
        // merely PREDICTED parent value the predictor would be learning from its own
        // guesses. The base class fires this after each leaf's bound update, in leaf order
        Node &parentnode=parent::get_parent_node();
        if(parentnode.type==parent::COMPUTED) _predictor.learn(feature,parentnode.value-value);
}
 
template<class RETURNTYPE, typename DIMTYPE, class SUBSET, class CRITERION, class PREDICTOR>
void Search_Branch_And_Bound_Fast<RETURNTYPE,DIMTYPE,SUBSET,CRITERION,PREDICTOR>::pre_evaluate_availables()
{
        assert(parent::get_criterion());
        assert(parent::get_engine());
        const PSubset &currentset=parent::get_currentset(); assert(currentset);
        // no assert on the parent node's type here: FBB legitimately descends into
        // PREDICTED nodes whenever cut_possible() declines to cut, so the parent node
        // need not be COMPUTED
        Node &parentnode=parent::get_parent_node();
 
        // pass 1 (master thread): decide predict-vs-compute for every available feature.
        // Deciding the whole level upfront gives the same decisions as deciding one
        // feature at a time would, because predictor state is per-feature (the PREDICTOR
        // contract): predict(f) reads only feature f's own statistics, each available is
        // visited exactly once per level, and learn(f) fires only at f's own turn - so no
        // decision can observe another feature's this-level learn
        PNode avail;
        RETURNTYPE val;
        DIMTYPE to_compute=0;
        for(bool got=this->getFirstAvailable(avail);got;got=this->getNextAvailable(avail))
        {
                if(_predictor.predict(avail->feature,val)) { //value could be predicted
                        avail->value=parentnode.value-val;
                        avail->type=parent::PREDICTED;
                } else { //value could not be predicted
                        avail->type=parent::UNKNOWN; // computed by the engine batch below
                        ++to_compute;
                }
        }
        if(to_compute==0) return;
 
        // pass 2: batch-evaluate the non-predictable features through the engine; the
        // per-result callback fires in add-order under both engine implementations, so
        // values are assigned and prediction contrasts learned in one fixed order -
        // master-thread decisions plus in-order result replay make a threaded run
        // reproduce the sequential one exactly
        typename parent::EVALUATORENGINE::ResultCallback on_result=[this,&parentnode](const RETURNTYPE value, const PSubset&, const DIMTYPE feat) {
                PNode node;
                if(this->getAvailable(feat,node)) {
                        node->value=value;
                        node->type=parent::COMPUTED;
                        // conditionally update prediction info
                        if(parentnode.type==parent::COMPUTED) _predictor.learn(feat,parentnode.value-value);
                }
        };
        parent::get_engine()->initialize(parent::get_criterion(),parent::get_n(),on_result);
        for(bool got=this->getFirstAvailable(avail);got;got=this->getNextAvailable(avail))
        {
                if(avail->type!=parent::PREDICTED) {
                        assert(currentset->selected_raw(avail->feature));
                        currentset->deselect_raw(avail->feature);
                        if(!parent::get_engine()->add_candidate(currentset,avail->feature)) throw FST::fst_error("Criterion evaluation failure.");
                        currentset->select_raw(avail->feature);
                }
        }
        // a criterion exception thrown on a worker is rethrown here; a failed evaluation
        // marks the batch failed with no callbacks fired - the nodes would stay UNKNOWN
        // and poison the level sort-in, so the search must fail instead (same rationale
        // as in the IBB body)
        if(!parent::get_engine()->collect()) throw FST::fst_error("Criterion evaluation failure.");
}
 
template<class RETURNTYPE, typename DIMTYPE, class SUBSET, class CRITERION, class PREDICTOR>
bool Search_Branch_And_Bound_Fast<RETURNTYPE,DIMTYPE,SUBSET,CRITERION,PREDICTOR>::cut_possible()
{       // NOTE: the lazy compute-before-cut below deliberately stays on the master thread -
        // it is a single evaluation on the traversal's critical path (nothing to batch)
        if(parent::is_bound_valid()) {
                Node &currentnode=parent::get_current_node();
                // in FBB the PREDICTED value is enough to evaluate cut as not possible
                if(currentnode.value>parent::get_bound_value()) return false;
                // but decision to cut must be always based on COMPUTED value...
                if(currentnode.type!=parent::COMPUTED) {
                        const PSubset &currentset=parent::get_currentset();     assert(currentset);
                        assert(parent::get_criterion());
                        assert(currentnode.feature<parent::get_n());
                        assert(currentset->selected_raw(currentnode.feature));
                        currentset->deselect_raw(currentnode.feature);
                        if(!parent::get_criterion()->evaluate(currentnode.value,currentset)) throw FST::fst_error("Criterion evaluation failure."); 
                        currentset->select_raw(currentnode.feature);
                        currentnode.type=parent::COMPUTED;
                        // conditionally update prediction info
                        // (no assert on the parent's type - the if below already handles the legitimate PREDICTED-parent case)
                        Node &parentnode=parent::get_parent_node();
                        if(parentnode.type==parent::COMPUTED) _predictor.learn(currentnode.feature,parentnode.value-currentnode.value);
                }
                if(currentnode.value<=parent::get_bound_value()) {
                        return true;
                }
        }
        return false;
}
 
} // namespace
#endif // FSTSEARCHBRANCHANDBOUNDFAST_H ///:~