-/* Copyright (C) 2006-2008 Joris Mooij [j dot mooij at science dot ru dot nl]
- Radboud University Nijmegen, The Netherlands
-
+/* Copyright (C) 2006-2008 Joris Mooij [joris dot mooij at tuebingen dot mpg dot de]
+ Radboud University Nijmegen, The Netherlands /
+ Max Planck Institute for Biological Cybernetics, Germany
+
This file is part of libDAI.
libDAI is free software; you can redistribute it and/or modify
}
-JTree::JTree( const FactorGraph &fg, const PropertySet &opts, bool automatic ) : DAIAlgRG(fg), _RTree(), _Qa(), _Qb(), _mes(), _logZ(), props() {
+JTree::JTree( const FactorGraph &fg, const PropertySet &opts, bool automatic ) : DAIAlgRG(fg), _mes(), _logZ(), RTree(), Qa(), Qb(), props() {
setProperties( opts );
+ if( !isConnected() )
+ DAI_THROW(FACTORGRAPH_NOT_CONNECTED);
+
if( automatic ) {
- // Copy VarSets of factors
+ // Create ClusterGraph which contains factors as clusters
vector<VarSet> cl;
cl.reserve( fg.nrFactors() );
for( size_t I = 0; I < nrFactors(); I++ )
for( size_t i = 0; i < Cliques.size(); i++ )
for( size_t j = i+1; j < Cliques.size(); j++ ) {
size_t w = (Cliques[i] & Cliques[j]).size();
- JuncGraph[UEdge(i,j)] = w;
+ if( w )
+ JuncGraph[UEdge(i,j)] = w;
}
// Construct maximal spanning tree using Prim's algorithm
- _RTree = MaxSpanningTreePrims( JuncGraph );
+ RTree = MaxSpanningTreePrims( JuncGraph );
// Construct corresponding region graph
size_t alpha;
for( alpha = 0; alpha < nrORs(); alpha++ )
if( OR(alpha).vars() >> factor(I).vars() ) {
-// OR(alpha) *= factor(I);
fac2OR.push_back( alpha );
break;
}
RecomputeORs();
// Create inner regions and edges
- IRs.reserve( _RTree.size() );
+ IRs.reserve( RTree.size() );
vector<Edge> edges;
- edges.reserve( 2 * _RTree.size() );
- for( size_t i = 0; i < _RTree.size(); i++ ) {
- edges.push_back( Edge( _RTree[i].n1, nrIRs() ) );
- edges.push_back( Edge( _RTree[i].n2, nrIRs() ) );
+ edges.reserve( 2 * RTree.size() );
+ for( size_t i = 0; i < RTree.size(); i++ ) {
+ edges.push_back( Edge( RTree[i].n1, nrIRs() ) );
+ edges.push_back( Edge( RTree[i].n2, nrIRs() ) );
// inner clusters have counting number -1
- IRs.push_back( Region( Cliques[_RTree[i].n1] & Cliques[_RTree[i].n2], -1.0 ) );
+ IRs.push_back( Region( Cliques[RTree[i].n1] & Cliques[RTree[i].n2], -1.0 ) );
}
// create bipartite graph
- G.create( nrORs(), nrIRs(), edges.begin(), edges.end() );
+ G.construct( nrORs(), nrIRs(), edges.begin(), edges.end() );
// Create messages and beliefs
- _Qa.clear();
- _Qa.reserve( nrORs() );
+ Qa.clear();
+ Qa.reserve( nrORs() );
for( size_t alpha = 0; alpha < nrORs(); alpha++ )
- _Qa.push_back( OR(alpha) );
+ Qa.push_back( OR(alpha) );
- _Qb.clear();
- _Qb.reserve( nrIRs() );
+ Qb.clear();
+ Qb.reserve( nrIRs() );
for( size_t beta = 0; beta < nrIRs(); beta++ )
- _Qb.push_back( Factor( IR(beta), 1.0 ) );
+ Qb.push_back( Factor( IR(beta), 1.0 ) );
_mes.clear();
_mes.reserve( nrORs() );
Factor JTree::belief( const VarSet &ns ) const {
vector<Factor>::const_iterator beta;
- for( beta = _Qb.begin(); beta != _Qb.end(); beta++ )
+ for( beta = Qb.begin(); beta != Qb.end(); beta++ )
if( beta->vars() >> ns )
break;
- if( beta != _Qb.end() )
+ if( beta != Qb.end() )
return( beta->marginal(ns) );
else {
vector<Factor>::const_iterator alpha;
- for( alpha = _Qa.begin(); alpha != _Qa.end(); alpha++ )
+ for( alpha = Qa.begin(); alpha != Qa.end(); alpha++ )
if( alpha->vars() >> ns )
break;
- assert( alpha != _Qa.end() );
+ assert( alpha != Qa.end() );
return( alpha->marginal(ns) );
}
}
vector<Factor> JTree::beliefs() const {
vector<Factor> result;
for( size_t beta = 0; beta < nrIRs(); beta++ )
- result.push_back( _Qb[beta] );
+ result.push_back( Qb[beta] );
for( size_t alpha = 0; alpha < nrORs(); alpha++ )
- result.push_back( _Qa[alpha] );
+ result.push_back( Qa[alpha] );
return result;
}
// Needs no init
void JTree::runHUGIN() {
for( size_t alpha = 0; alpha < nrORs(); alpha++ )
- _Qa[alpha] = OR(alpha);
+ Qa[alpha] = OR(alpha);
for( size_t beta = 0; beta < nrIRs(); beta++ )
- _Qb[beta].fill( 1.0 );
+ Qb[beta].fill( 1.0 );
// CollectEvidence
_logZ = 0.0;
- for( size_t i = _RTree.size(); (i--) != 0; ) {
-// Make outer region _RTree[i].n1 consistent with outer region _RTree[i].n2
-// IR(i) = seperator OR(_RTree[i].n1) && OR(_RTree[i].n2)
- Factor new_Qb = _Qa[_RTree[i].n2].partSum( IR( i ) );
- _logZ += log(new_Qb.normalize( Prob::NORMPROB ));
- _Qa[_RTree[i].n1] *= new_Qb.divided_by( _Qb[i] );
- _Qb[i] = new_Qb;
+ for( size_t i = RTree.size(); (i--) != 0; ) {
+// Make outer region RTree[i].n1 consistent with outer region RTree[i].n2
+// IR(i) = seperator OR(RTree[i].n1) && OR(RTree[i].n2)
+ Factor new_Qb = Qa[RTree[i].n2].partSum( IR( i ) );
+ _logZ += log(new_Qb.normalize());
+ Qa[RTree[i].n1] *= new_Qb.divided_by( Qb[i] );
+ Qb[i] = new_Qb;
}
- if( _RTree.empty() )
- _logZ += log(_Qa[0].normalize( Prob::NORMPROB ) );
+ if( RTree.empty() )
+ _logZ += log(Qa[0].normalize() );
else
- _logZ += log(_Qa[_RTree[0].n1].normalize( Prob::NORMPROB ));
+ _logZ += log(Qa[RTree[0].n1].normalize());
// DistributeEvidence
- for( size_t i = 0; i < _RTree.size(); i++ ) {
-// Make outer region _RTree[i].n2 consistent with outer region _RTree[i].n1
-// IR(i) = seperator OR(_RTree[i].n1) && OR(_RTree[i].n2)
- Factor new_Qb = _Qa[_RTree[i].n1].marginal( IR( i ) );
- _Qa[_RTree[i].n2] *= new_Qb.divided_by( _Qb[i] );
- _Qb[i] = new_Qb;
+ for( size_t i = 0; i < RTree.size(); i++ ) {
+// Make outer region RTree[i].n2 consistent with outer region RTree[i].n1
+// IR(i) = seperator OR(RTree[i].n1) && OR(RTree[i].n2)
+ Factor new_Qb = Qa[RTree[i].n1].marginal( IR( i ) );
+ Qa[RTree[i].n2] *= new_Qb.divided_by( Qb[i] );
+ Qb[i] = new_Qb;
}
// Normalize
for( size_t alpha = 0; alpha < nrORs(); alpha++ )
- _Qa[alpha].normalize( Prob::NORMPROB );
+ Qa[alpha].normalize();
}
// First pass
_logZ = 0.0;
for( size_t e = nrIRs(); (e--) != 0; ) {
- // send a message from _RTree[e].n2 to _RTree[e].n1
- // or, actually, from the seperator IR(e) to _RTree[e].n1
+ // send a message from RTree[e].n2 to RTree[e].n1
+ // or, actually, from the seperator IR(e) to RTree[e].n1
- size_t i = nbIR(e)[1].node; // = _RTree[e].n2
- size_t j = nbIR(e)[0].node; // = _RTree[e].n1
+ size_t i = nbIR(e)[1].node; // = RTree[e].n2
+ size_t j = nbIR(e)[0].node; // = RTree[e].n1
size_t _e = nbIR(e)[0].dual;
Factor piet = OR(i);
if( k != e )
piet *= message( i, k.iter );
message( j, _e ) = piet.partSum( IR(e) );
- _logZ += log( message(j,_e).normalize( Prob::NORMPROB ) );
+ _logZ += log( message(j,_e).normalize() );
}
// Second pass
for( size_t e = 0; e < nrIRs(); e++ ) {
- size_t i = nbIR(e)[0].node; // = _RTree[e].n1
- size_t j = nbIR(e)[1].node; // = _RTree[e].n2
+ size_t i = nbIR(e)[0].node; // = RTree[e].n1
+ size_t j = nbIR(e)[1].node; // = RTree[e].n2
size_t _e = nbIR(e)[1].dual;
Factor piet = OR(i);
foreach( const Neighbor &k, nbOR(i) )
- if( k != e )
+ if( k != e )
piet *= message( i, k.iter );
message( j, _e ) = piet.marginal( IR(e) );
}
foreach( const Neighbor &k, nbOR(alpha) )
piet *= message( alpha, k.iter );
if( nrIRs() == 0 ) {
- _logZ += log( piet.normalize( Prob::NORMPROB ) );
- _Qa[alpha] = piet;
- } else if( alpha == nbIR(0)[0].node /*_RTree[0].n1*/ ) {
- _logZ += log( piet.normalize( Prob::NORMPROB ) );
- _Qa[alpha] = piet;
+ _logZ += log( piet.normalize() );
+ Qa[alpha] = piet;
+ } else if( alpha == nbIR(0)[0].node /*RTree[0].n1*/ ) {
+ _logZ += log( piet.normalize() );
+ Qa[alpha] = piet;
} else
- _Qa[alpha] = piet.normalized( Prob::NORMPROB );
+ Qa[alpha] = piet.normalized();
}
// Only for logZ (and for belief)...
for( size_t beta = 0; beta < nrIRs(); beta++ )
- _Qb[beta] = _Qa[nbIR(beta)[0].node].marginal( IR(beta) );
+ Qb[beta] = Qa[nbIR(beta)[0].node].marginal( IR(beta) );
}
Real JTree::logZ() const {
Real sum = 0.0;
for( size_t beta = 0; beta < nrIRs(); beta++ )
- sum += IR(beta).c() * _Qb[beta].entropy();
+ sum += IR(beta).c() * Qb[beta].entropy();
for( size_t alpha = 0; alpha < nrORs(); alpha++ ) {
- sum += OR(alpha).c() * _Qa[alpha].entropy();
- sum += (OR(alpha).log0() * _Qa[alpha]).totalSum();
+ sum += OR(alpha).c() * Qa[alpha].entropy();
+ sum += (OR(alpha).log0() * Qa[alpha]).totalSum();
}
return sum;
}
// find new root clique (the one with maximal statespace overlap with ns)
size_t maxval = 0, maxalpha = 0;
for( size_t alpha = 0; alpha < nrORs(); alpha++ ) {
- size_t val = (ns & OR(alpha).vars()).states();
+ size_t val = VarSet(ns & OR(alpha).vars()).nrStates();
if( val > maxval ) {
maxval = val;
maxalpha = alpha;
}
}
-// for( size_t e = 0; e < _RTree.size(); e++ )
-// cout << OR(_RTree[e].n1).vars() << "->" << OR(_RTree[e].n2).vars() << ", ";
-// cout << endl;
// grow new tree
Graph oldTree;
- for( DEdgeVec::const_iterator e = _RTree.begin(); e != _RTree.end(); e++ )
+ for( DEdgeVec::const_iterator e = RTree.begin(); e != RTree.end(); e++ )
oldTree.insert( UEdge(e->n1, e->n2) );
DEdgeVec newTree = GrowRootedTree( oldTree, maxalpha );
-// cout << ns << ": ";
-// for( size_t e = 0; e < newTree.size(); e++ )
-// cout << OR(newTree[e].n1).vars() << "->" << OR(newTree[e].n2).vars() << ", ";
-// cout << endl;
// identify subtree that contains variables of ns which are not in the new root
VarSet nsrem = ns / OR(maxalpha).vars();
-// cout << "nsrem:" << nsrem << endl;
set<DEdge> subTree;
// for each variable in ns that is not in the root clique
for( VarSet::const_iterator n = nsrem.begin(); n != nsrem.end(); n++ ) {
pos = newTree[e-1].n1;
}
}
-// cout << "subTree: " << endl;
-// for( set<DEdge>::const_iterator sTi = subTree.begin(); sTi != subTree.end(); sTi++ )
-// cout << OR(sTi->n1).vars() << "->" << OR(sTi->n2).vars() << ", ";
-// cout << endl;
// Resulting Tree is a reordered copy of newTree
// First add edges in subTree to Tree
for( DEdgeVec::const_iterator e = newTree.begin(); e != newTree.end(); e++ )
if( subTree.count( *e ) ) {
Tree.push_back( *e );
-// cout << OR(e->n1).vars() << "->" << OR(e->n2).vars() << ", ";
}
-// cout << endl;
// Then add edges pointing away from nsrem
// FIXME
/* for( DEdgeVec::const_iterator e = newTree.begin(); e != newTree.end(); e++ )
if( e->n1 == sTi->n1 || e->n1 == sTi->n2 ||
e->n2 == sTi->n1 || e->n2 == sTi->n2 ) {
Tree.push_back( *e );
-// cout << OR(e->n1).vars() << "->" << OR(e->n2).vars() << ", ";
}
}*/
// FIXME
}
if( found ) {
Tree.push_back( *e );
- cout << OR(e->n1).vars() << "->" << OR(e->n2).vars() << ", ";
}
- }
- cout << endl;*/
+ }*/
size_t subTreeSize = Tree.size();
// Then add remaining edges
for( DEdgeVec::const_iterator e = newTree.begin(); e != newTree.end(); e++ )
// assumes that run() has been called already
Factor JTree::calcMarginal( const VarSet& ns ) {
vector<Factor>::const_iterator beta;
- for( beta = _Qb.begin(); beta != _Qb.end(); beta++ )
+ for( beta = Qb.begin(); beta != Qb.end(); beta++ )
if( beta->vars() >> ns )
break;
- if( beta != _Qb.end() )
+ if( beta != Qb.end() )
return( beta->marginal(ns) );
else {
vector<Factor>::const_iterator alpha;
- for( alpha = _Qa.begin(); alpha != _Qa.end(); alpha++ )
+ for( alpha = Qa.begin(); alpha != Qa.end(); alpha++ )
if( alpha->vars() >> ns )
break;
- if( alpha != _Qa.end() )
+ if( alpha != Qa.end() )
return( alpha->marginal(ns) );
else {
// Find subtree to do efficient inference
VarSet nsrem = ns / OR(T.front().n1).vars();
Factor Pns (ns, 0.0);
- // Save _Qa and _Qb on the subtree
- map<size_t,Factor> _Qa_old;
- map<size_t,Factor> _Qb_old;
+ // Save Qa and Qb on the subtree
+ map<size_t,Factor> Qa_old;
+ map<size_t,Factor> Qb_old;
vector<size_t> b(Tsize, 0);
for( size_t i = Tsize; (i--) != 0; ) {
size_t alpha1 = T[i].n1;
size_t alpha2 = T[i].n2;
size_t beta;
for( beta = 0; beta < nrIRs(); beta++ )
- if( UEdge( _RTree[beta].n1, _RTree[beta].n2 ) == UEdge( alpha1, alpha2 ) )
+ if( UEdge( RTree[beta].n1, RTree[beta].n2 ) == UEdge( alpha1, alpha2 ) )
break;
assert( beta != nrIRs() );
b[i] = beta;
- if( !_Qa_old.count( alpha1 ) )
- _Qa_old[alpha1] = _Qa[alpha1];
- if( !_Qa_old.count( alpha2 ) )
- _Qa_old[alpha2] = _Qa[alpha2];
- if( !_Qb_old.count( beta ) )
- _Qb_old[beta] = _Qb[beta];
+ if( !Qa_old.count( alpha1 ) )
+ Qa_old[alpha1] = Qa[alpha1];
+ if( !Qa_old.count( alpha2 ) )
+ Qa_old[alpha2] = Qa[alpha2];
+ if( !Qb_old.count( beta ) )
+ Qb_old[beta] = Qb[beta];
}
// For all states of nsrem
for( State s(nsrem); s.valid(); s++ ) {
-
// CollectEvidence
double logZ = 0.0;
for( size_t i = Tsize; (i--) != 0; ) {
// IR(i) = seperator OR(T[i].n1) && OR(T[i].n2)
for( VarSet::const_iterator n = nsrem.begin(); n != nsrem.end(); n++ )
- if( _Qa[T[i].n2].vars() >> *n ) {
+ if( Qa[T[i].n2].vars() >> *n ) {
Factor piet( *n, 0.0 );
piet[s(*n)] = 1.0;
- _Qa[T[i].n2] *= piet;
+ Qa[T[i].n2] *= piet;
}
- Factor new_Qb = _Qa[T[i].n2].partSum( IR( b[i] ) );
- logZ += log(new_Qb.normalize( Prob::NORMPROB ));
- _Qa[T[i].n1] *= new_Qb.divided_by( _Qb[b[i]] );
- _Qb[b[i]] = new_Qb;
+ Factor new_Qb = Qa[T[i].n2].partSum( IR( b[i] ) );
+ logZ += log(new_Qb.normalize());
+ Qa[T[i].n1] *= new_Qb.divided_by( Qb[b[i]] );
+ Qb[b[i]] = new_Qb;
}
- logZ += log(_Qa[T[0].n1].normalize( Prob::NORMPROB ));
+ logZ += log(Qa[T[0].n1].normalize());
Factor piet( nsrem, 0.0 );
piet[s] = exp(logZ);
- Pns += piet * _Qa[T[0].n1].partSum( ns / nsrem ); // OPTIMIZE ME
+ Pns += piet * Qa[T[0].n1].partSum( ns / nsrem ); // OPTIMIZE ME
// Restore clamped beliefs
- for( map<size_t,Factor>::const_iterator alpha = _Qa_old.begin(); alpha != _Qa_old.end(); alpha++ )
- _Qa[alpha->first] = alpha->second;
- for( map<size_t,Factor>::const_iterator beta = _Qb_old.begin(); beta != _Qb_old.end(); beta++ )
- _Qb[beta->first] = beta->second;
+ for( map<size_t,Factor>::const_iterator alpha = Qa_old.begin(); alpha != Qa_old.end(); alpha++ )
+ Qa[alpha->first] = alpha->second;
+ for( map<size_t,Factor>::const_iterator beta = Qb_old.begin(); beta != Qb_old.end(); beta++ )
+ Qb[beta->first] = beta->second;
}
- return( Pns.normalized(Prob::NORMPROB) );
+ return( Pns.normalized() );
}
}
}
+/// Calculates upper bound to the treewidth of a FactorGraph
+/** \relates JTree
+ * \return a pair (number of variables in largest clique, number of states in largest clique)
+ */
+std::pair<size_t,size_t> treewidth( const FactorGraph & fg ) {
+ ClusterGraph _cg;
+
+ // Copy factors
+ for( size_t I = 0; I < fg.nrFactors(); I++ )
+ _cg.insert( fg.factor(I).vars() );
+
+ // Retain only maximal clusters
+ _cg.eraseNonMaximal();
+
+ // Obtain elimination sequence
+ vector<VarSet> ElimVec = _cg.VarElim_MinFill().eraseNonMaximal().toVector();
+
+ // Calculate treewidth
+ size_t treewidth = 0;
+ size_t nrstates = 0;
+ for( size_t i = 0; i < ElimVec.size(); i++ ) {
+ if( ElimVec[i].size() > treewidth )
+ treewidth = ElimVec[i].size();
+ size_t s = ElimVec[i].nrStates();
+ if( s > nrstates )
+ nrstates = s;
+ }
+
+ return pair<size_t,size_t>(treewidth, nrstates);
+}
+
+
} // end of namespace dai
projects / libdai.git / blobdiff