src/hak.cpp

   1 /*  Copyright (C) 2006-2008  Joris Mooij  [j dot mooij at science dot ru dot nl]
   2     Radboud University Nijmegen, The Netherlands
   3
   4     This file is part of libDAI.
   5
   6     libDAI is free software; you can redistribute it and/or modify
   7     it under the terms of the GNU General Public License as published by
   8     the Free Software Foundation; either version 2 of the License, or
   9     (at your option) any later version.
  10
  11     libDAI is distributed in the hope that it will be useful,
  12     but WITHOUT ANY WARRANTY; without even the implied warranty of
  13     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14     GNU General Public License for more details.
  15
  16     You should have received a copy of the GNU General Public License
  17     along with libDAI; if not, write to the Free Software
  18     Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  19 */
  20
  21
  22 #include <map>
  23 #include <dai/hak.h>
  24 #include <dai/util.h>
  25 #include <dai/diffs.h>
  26 #include <dai/exceptions.h>
  27
  28
  29 namespace dai {
  30
  31
  32 using namespace std;
  33
  34
  35 const char *HAK::Name = "HAK";
  36
  37
  38 void HAK::setProperties( const PropertySet &opts ) {
  39     assert( opts.hasKey("tol") );
  40     assert( opts.hasKey("maxiter") );
  41     assert( opts.hasKey("verbose") );
  42     assert( opts.hasKey("doubleloop") );
  43     assert( opts.hasKey("clusters") );
  44
  45     props.tol = opts.getStringAs<double>("tol");
  46     props.maxiter = opts.getStringAs<size_t>("maxiter");
  47     props.verbose = opts.getStringAs<size_t>("verbose");
  48     props.doubleloop = opts.getStringAs<bool>("doubleloop");
  49     props.clusters = opts.getStringAs<Properties::ClustersType>("clusters");
  50
  51     if( opts.hasKey("loopdepth") )
  52         props.loopdepth = opts.getStringAs<size_t>("loopdepth");
  53     else
  54         assert( props.clusters != Properties::ClustersType::LOOP );
  55 }
  56
  57
  58 PropertySet HAK::getProperties() const {
  59     PropertySet opts;
  60     opts.Set( "tol", props.tol );
  61     opts.Set( "maxiter", props.maxiter );
  62     opts.Set( "verbose", props.verbose );
  63     opts.Set( "doubleloop", props.doubleloop );
  64     opts.Set( "clusters", props.clusters );
  65     opts.Set( "loopdepth", props.loopdepth );
  66     return opts;
  67 }
  68
  69
  70 void HAK::constructMessages() {
  71     // Create outer beliefs
  72     _Qa.clear();
  73     _Qa.reserve(nrORs());
  74     for( size_t alpha = 0; alpha < nrORs(); alpha++ )
  75         _Qa.push_back( Factor( OR(alpha).vars() ) );
  76
  77     // Create inner beliefs
  78     _Qb.clear();
  79     _Qb.reserve(nrIRs());
  80     for( size_t beta = 0; beta < nrIRs(); beta++ )
  81         _Qb.push_back( Factor( IR(beta) ) );
  82
  83     // Create messages
  84     _muab.clear();
  85     _muab.reserve( nrORs() );
  86     _muba.clear();
  87     _muba.reserve( nrORs() );
  88     for( size_t alpha = 0; alpha < nrORs(); alpha++ ) {
  89         _muab.push_back( vector<Factor>() );
  90         _muba.push_back( vector<Factor>() );
  91         _muab[alpha].reserve( nbOR(alpha).size() );
  92         _muba[alpha].reserve( nbOR(alpha).size() );
  93         foreach( const Neighbor &beta, nbOR(alpha) ) {
  94             _muab[alpha].push_back( Factor( IR(beta) ) );
  95             _muba[alpha].push_back( Factor( IR(beta) ) );
  96         }
  97     }
  98 }
  99
 100
 101 HAK::HAK(const RegionGraph & rg, const PropertySet &opts ) : DAIAlgRG(rg) {
 102     setProperties( opts );
 103
 104     constructMessages();
 105 }
 106
 107
 108 void HAK::findLoopClusters( const FactorGraph & fg, std::set<VarSet> &allcl, VarSet newcl, const Var & root, size_t length, VarSet vars ) {
 109     for( VarSet::const_iterator in = vars.begin(); in != vars.end(); in++ ) {
 110         VarSet ind = fg.delta( fg.findVar( *in ) );
 111         if( (newcl.size()) >= 2 && (ind >> root) ) {
 112             allcl.insert( newcl | *in );
 113         }
 114         else if( length > 1 )
 115             findLoopClusters( fg, allcl, newcl | *in, root, length - 1, ind / newcl );
 116     }
 117 }
 118
 119
 120 HAK::HAK(const FactorGraph & fg, const PropertySet &opts) : DAIAlgRG(), props(), maxdiff(0.0) {
 121     setProperties( opts );
 122
 123     vector<VarSet> cl;
 124     if( props.clusters == Properties::ClustersType::MIN ) {
 125         cl = fg.Cliques();
 126     } else if( props.clusters == Properties::ClustersType::DELTA ) {
 127         for( size_t i = 0; i < fg.nrVars(); i++ )
 128             cl.push_back(fg.Delta(i));
 129     } else if( props.clusters == Properties::ClustersType::LOOP ) {
 130         cl = fg.Cliques();
 131         set<VarSet> scl;
 132         for( size_t i0 = 0; i0 < fg.nrVars(); i0++ ) {
 133             VarSet i0d = fg.delta(i0);
 134             if( props.loopdepth > 1 )
 135                 findLoopClusters( fg, scl, fg.var(i0), fg.var(i0), props.loopdepth - 1, fg.delta(i0) );
 136         }
 137         for( set<VarSet>::const_iterator c = scl.begin(); c != scl.end(); c++ )
 138             cl.push_back(*c);
 139         if( props.verbose >= 3 ) {
 140             cout << "HAK uses the following clusters: " << endl;
 141             for( vector<VarSet>::const_iterator cli = cl.begin(); cli != cl.end(); cli++ )
 142                 cout << *cli << endl;
 143         }
 144     } else
 145         DAI_THROW(INTERNAL_ERROR);
 146
 147     RegionGraph rg(fg,cl);
 148     RegionGraph::operator=(rg);
 149     constructMessages();
 150
 151     if( props.verbose >= 3 )
 152         cout << "HAK regiongraph: " << *this << endl;
 153 }
 154
 155
 156 string HAK::identify() const {
 157     stringstream result (stringstream::out);
 158     result << Name << getProperties();
 159     return result.str();
 160 }
 161
 162
 163 void HAK::init( const VarSet &ns ) {
 164     for( vector<Factor>::iterator alpha = _Qa.begin(); alpha != _Qa.end(); alpha++ )
 165         if( alpha->vars().intersects( ns ) )
 166             alpha->fill( 1.0 / alpha->states() );
 167
 168     for( size_t beta = 0; beta < nrIRs(); beta++ )
 169         if( IR(beta).intersects( ns ) ) {
 170             _Qb[beta].fill( 1.0 );
 171             foreach( const Neighbor &alpha, nbIR(beta) ) {
 172                 size_t _beta = alpha.dual;
 173                 muab( alpha, _beta ).fill( 1.0 / IR(beta).states() );
 174                 muba( alpha, _beta ).fill( 1.0 / IR(beta).states() );
 175             }
 176         }
 177 }
 178
 179
 180 void HAK::init() {
 181     for( vector<Factor>::iterator alpha = _Qa.begin(); alpha != _Qa.end(); alpha++ )
 182         alpha->fill( 1.0 / alpha->states() );
 183
 184     for( vector<Factor>::iterator beta = _Qb.begin(); beta != _Qb.end(); beta++ )
 185         beta->fill( 1.0 / beta->states() );
 186
 187     for( size_t alpha = 0; alpha < nrORs(); alpha++ )
 188         foreach( const Neighbor &beta, nbOR(alpha) ) {
 189             size_t _beta = beta.iter;
 190             muab( alpha, _beta ).fill( 1.0 / muab( alpha, _beta ).states() );
 191             muba( alpha, _beta ).fill( 1.0 / muab( alpha, _beta ).states() );
 192         }
 193 }
 194
 195
 196 double HAK::doGBP() {
 197     if( props.verbose >= 1 )
 198         cout << "Starting " << identify() << "...";
 199     if( props.verbose >= 3)
 200         cout << endl;
 201
 202     double tic = toc();
 203
 204     // Check whether counting numbers won't lead to problems
 205     for( size_t beta = 0; beta < nrIRs(); beta++ )
 206         assert( nbIR(beta).size() + IR(beta).c() != 0.0 );
 207
 208     // Keep old beliefs to check convergence
 209     vector<Factor> old_beliefs;
 210     old_beliefs.reserve( nrVars() );
 211     for( size_t i = 0; i < nrVars(); i++ )
 212         old_beliefs.push_back( belief( var(i) ) );
 213
 214     // Differences in single node beliefs
 215     Diffs diffs(nrVars(), 1.0);
 216
 217     size_t iter = 0;
 218     // do several passes over the network until maximum number of iterations has
 219     // been reached or until the maximum belief difference is smaller than tolerance
 220     for( iter = 0; iter < props.maxiter && diffs.maxDiff() > props.tol; iter++ ) {
 221         for( size_t beta = 0; beta < nrIRs(); beta++ ) {
 222             foreach( const Neighbor &alpha, nbIR(beta) ) {
 223                 size_t _beta = alpha.dual;
 224                 muab( alpha, _beta ) = _Qa[alpha].marginal(IR(beta)).divided_by( muba(alpha,_beta) );
 225             }
 226
 227             Factor Qb_new;
 228             foreach( const Neighbor &alpha, nbIR(beta) ) {
 229                 size_t _beta = alpha.dual;
 230                 Qb_new *= muab(alpha,_beta) ^ (1 / (nbIR(beta).size() + IR(beta).c()));
 231             }
 232
 233             Qb_new.normalize( Prob::NORMPROB );
 234             if( Qb_new.hasNaNs() ) {
 235                 cout << "HAK::doGBP:  Qb_new has NaNs!" << endl;
 236                 return 1.0;
 237             }
 238 //          _Qb[beta] = Qb_new.makeZero(1e-100);    // damping?
 239             _Qb[beta] = Qb_new;
 240
 241             foreach( const Neighbor &alpha, nbIR(beta) ) {
 242                 size_t _beta = alpha.dual;
 243
 244                 muba(alpha,_beta) = _Qb[beta].divided_by( muab(alpha,_beta) );
 245
 246                 Factor Qa_new = OR(alpha);
 247                 foreach( const Neighbor &gamma, nbOR(alpha) )
 248                     Qa_new *= muba(alpha,gamma.iter);
 249                 Qa_new ^= (1.0 / OR(alpha).c());
 250                 Qa_new.normalize( Prob::NORMPROB );
 251                 if( Qa_new.hasNaNs() ) {
 252                     cout << "HAK::doGBP:  Qa_new has NaNs!" << endl;
 253                     return 1.0;
 254                 }
 255 //              _Qa[alpha] = Qa_new.makeZero(1e-100); // damping?
 256                 _Qa[alpha] = Qa_new;
 257             }
 258         }
 259
 260         // Calculate new single variable beliefs and compare with old ones
 261         for( size_t i = 0; i < nrVars(); i++ ) {
 262             Factor new_belief = belief( var( i ) );
 263             diffs.push( dist( new_belief, old_beliefs[i], Prob::DISTLINF ) );
 264             old_beliefs[i] = new_belief;
 265         }
 266
 267         if( props.verbose >= 3 )
 268             cout << "HAK::doGBP:  maxdiff " << diffs.maxDiff() << " after " << iter+1 << " passes" << endl;
 269     }
 270
 271     if( diffs.maxDiff() > maxdiff )
 272         maxdiff = diffs.maxDiff();
 273
 274     if( props.verbose >= 1 ) {
 275         if( diffs.maxDiff() > props.tol ) {
 276             if( props.verbose == 1 )
 277                 cout << endl;
 278             cout << "HAK::doGBP:  WARNING: not converged within " << props.maxiter << " passes (" << toc() - tic << " clocks)...final maxdiff:" << diffs.maxDiff() << endl;
 279         } else {
 280             if( props.verbose >= 2 )
 281                 cout << "HAK::doGBP:  ";
 282             cout << "converged in " << iter << " passes (" << toc() - tic << " clocks)." << endl;
 283         }
 284     }
 285
 286     return diffs.maxDiff();
 287 }
 288
 289
 290 double HAK::doDoubleLoop() {
 291     if( props.verbose >= 1 )
 292         cout << "Starting " << identify() << "...";
 293     if( props.verbose >= 3)
 294         cout << endl;
 295
 296     double tic = toc();
 297
 298     // Save original outer regions
 299     vector<FRegion> org_ORs = ORs;
 300
 301     // Save original inner counting numbers and set negative counting numbers to zero
 302     vector<double> org_IR_cs( nrIRs(), 0.0 );
 303     for( size_t beta = 0; beta < nrIRs(); beta++ ) {
 304         org_IR_cs[beta] = IR(beta).c();
 305         if( IR(beta).c() < 0.0 )
 306             IR(beta).c() = 0.0;
 307     }
 308
 309     // Keep old beliefs to check convergence
 310     vector<Factor> old_beliefs;
 311     old_beliefs.reserve( nrVars() );
 312     for( size_t i = 0; i < nrVars(); i++ )
 313         old_beliefs.push_back( belief( var(i) ) );
 314
 315     // Differences in single node beliefs
 316     Diffs diffs(nrVars(), 1.0);
 317
 318     size_t  outer_maxiter   = props.maxiter;
 319     double  outer_tol       = props.tol;
 320     size_t  outer_verbose   = props.verbose;
 321     double  org_maxdiff     = maxdiff;
 322
 323     // Set parameters for inner loop
 324     props.maxiter = 5;
 325     props.verbose = outer_verbose ? outer_verbose - 1 : 0;
 326
 327     size_t outer_iter = 0;
 328     for( outer_iter = 0; outer_iter < outer_maxiter && diffs.maxDiff() > outer_tol; outer_iter++ ) {
 329         // Calculate new outer regions
 330         for( size_t alpha = 0; alpha < nrORs(); alpha++ ) {
 331             OR(alpha) = org_ORs[alpha];
 332             foreach( const Neighbor &beta, nbOR(alpha) )
 333                 OR(alpha) *= _Qb[beta] ^ ((IR(beta).c() - org_IR_cs[beta]) / nbIR(beta).size());
 334         }
 335
 336         // Inner loop
 337         if( isnan( doGBP() ) )
 338             return 1.0;
 339
 340         // Calculate new single variable beliefs and compare with old ones
 341         for( size_t i = 0; i < nrVars(); ++i ) {
 342             Factor new_belief = belief( var( i ) );
 343             diffs.push( dist( new_belief, old_beliefs[i], Prob::DISTLINF ) );
 344             old_beliefs[i] = new_belief;
 345         }
 346
 347         if( props.verbose >= 3 )
 348             cout << "HAK::doDoubleLoop:  maxdiff " << diffs.maxDiff() << " after " << outer_iter+1 << " passes" << endl;
 349     }
 350
 351     // restore _maxiter, _verbose and _maxdiff
 352     props.maxiter = outer_maxiter;
 353     props.verbose = outer_verbose;
 354     maxdiff = org_maxdiff;
 355
 356     if( diffs.maxDiff() > maxdiff )
 357         maxdiff = diffs.maxDiff();
 358
 359     // Restore original outer regions
 360     ORs = org_ORs;
 361
 362     // Restore original inner counting numbers
 363     for( size_t beta = 0; beta < nrIRs(); ++beta )
 364         IR(beta).c() = org_IR_cs[beta];
 365
 366     if( props.verbose >= 1 ) {
 367         if( diffs.maxDiff() > props.tol ) {
 368             if( props.verbose == 1 )
 369                 cout << endl;
 370                 cout << "HAK::doDoubleLoop:  WARNING: not converged within " << outer_maxiter << " passes (" << toc() - tic << " clocks)...final maxdiff:" << diffs.maxDiff() << endl;
 371             } else {
 372                 if( props.verbose >= 3 )
 373                     cout << "HAK::doDoubleLoop:  ";
 374                 cout << "converged in " << outer_iter << " passes (" << toc() - tic << " clocks)." << endl;
 375             }
 376         }
 377
 378     return diffs.maxDiff();
 379 }
 380
 381
 382 double HAK::run() {
 383     if( props.doubleloop )
 384         return doDoubleLoop();
 385     else
 386         return doGBP();
 387 }
 388
 389
 390 Factor HAK::belief( const VarSet &ns ) const {
 391     vector<Factor>::const_iterator beta;
 392     for( beta = _Qb.begin(); beta != _Qb.end(); beta++ )
 393         if( beta->vars() >> ns )
 394             break;
 395     if( beta != _Qb.end() )
 396         return( beta->marginal(ns) );
 397     else {
 398         vector<Factor>::const_iterator alpha;
 399         for( alpha = _Qa.begin(); alpha != _Qa.end(); alpha++ )
 400             if( alpha->vars() >> ns )
 401                 break;
 402         assert( alpha != _Qa.end() );
 403         return( alpha->marginal(ns) );
 404     }
 405 }
 406
 407
 408 Factor HAK::belief( const Var &n ) const {
 409     return belief( (VarSet)n );
 410 }
 411
 412
 413 vector<Factor> HAK::beliefs() const {
 414     vector<Factor> result;
 415     for( size_t beta = 0; beta < nrIRs(); beta++ )
 416         result.push_back( Qb(beta) );
 417     for( size_t alpha = 0; alpha < nrORs(); alpha++ )
 418         result.push_back( Qa(alpha) );
 419     return result;
 420 }
 421
 422
 423 Real HAK::logZ() const {
 424     Real sum = 0.0;
 425     for( size_t beta = 0; beta < nrIRs(); beta++ )
 426         sum += IR(beta).c() * Qb(beta).entropy();
 427     for( size_t alpha = 0; alpha < nrORs(); alpha++ ) {
 428         sum += OR(alpha).c() * Qa(alpha).entropy();
 429         sum += (OR(alpha).log0() * Qa(alpha)).totalSum();
 430     }
 431     return sum;
 432 }
 433
 434
 435 } // end of namespace dai