src/bp.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 <iostream>
  23 #include <sstream>
  24 #include <map>
  25 #include <set>
  26 #include <algorithm>
  27 #include <dai/bp.h>
  28 #include <dai/diffs.h>
  29 #include <dai/util.h>
  30 #include <dai/properties.h>
  31
  32
  33 namespace dai {
  34
  35
  36 using namespace std;
  37
  38
  39 const char *BP::Name = "BP";
  40
  41
  42 bool BP::checkProperties() {
  43     if( !HasProperty("updates") )
  44         return false;
  45     if( !HasProperty("tol") )
  46         return false;
  47     if (!HasProperty("maxiter") )
  48         return false;
  49     if (!HasProperty("verbose") )
  50         return false;
  51     if (!HasProperty("logdomain") )
  52         return false;
  53
  54     ConvertPropertyTo<double>("tol");
  55     ConvertPropertyTo<size_t>("maxiter");
  56     ConvertPropertyTo<size_t>("verbose");
  57     ConvertPropertyTo<UpdateType>("updates");
  58     ConvertPropertyTo<bool>("logdomain");
  59     logDomain = GetPropertyAs<bool>("logdomain");
  60
  61     return true;
  62 }
  63
  64
  65 void BP::create() {
  66     // create edge properties
  67     edges.clear();
  68     edges.reserve( nrVars() );
  69     for( size_t i = 0; i < nrVars(); ++i ) {
  70         edges.push_back( vector<EdgeProp>() );
  71         edges[i].reserve( nbV(i).size() );
  72         foreach( const Neighbor &I, nbV(i) ) {
  73             EdgeProp newEP;
  74             newEP.message = Prob( var(i).states() );
  75             newEP.newMessage = Prob( var(i).states() );
  76
  77             newEP.index.reserve( factor(I).states() );
  78             for( IndexFor k( var(i), factor(I).vars() ); k >= 0; ++k )
  79                 newEP.index.push_back( k );
  80
  81             newEP.residual = 0.0;
  82             edges[i].push_back( newEP );
  83         }
  84     }
  85 }
  86
  87
  88 void BP::init() {
  89     assert( checkProperties() );
  90     for( size_t i = 0; i < nrVars(); ++i ) {
  91         foreach( const Neighbor &I, nbV(i) ) {
  92             if( logDomain ) {
  93                 message( i, I.iter ).fill( 0.0 );
  94                 newMessage( i, I.iter ).fill( 0.0 );
  95             } else {
  96                 message( i, I.iter ).fill( 1.0 );
  97                 newMessage( i, I.iter ).fill( 1.0 );
  98             }
  99         }
 100     }
 101 }
 102
 103
 104 void BP::findMaxResidual( size_t &i, size_t &_I ) {
 105     i = 0;
 106     _I = 0;
 107     double maxres = residual( i, _I );
 108     for( size_t j = 0; j < nrVars(); ++j )
 109         foreach( const Neighbor &I, nbV(j) )
 110             if( residual( j, I.iter ) > maxres ) {
 111                 i = j;
 112                 _I = I.iter;
 113                 maxres = residual( i, _I );
 114             }
 115 }
 116
 117
 118 void BP::calcNewMessage( size_t i, size_t _I ) {
 119     // calculate updated message I->i
 120     size_t I = nbV(i,_I);
 121
 122 /*  UNOPTIMIZED (SIMPLE TO READ, BUT SLOW) VERSION
 123
 124     Factor prod( factor( I ) );
 125     for( _nb_cit j = nb2(I).begin(); j != nb2(I).end(); j++ )
 126         if( *j != i ) {     // for all j in I \ i
 127             for( _nb_cit J = nb1(*j).begin(); J != nb1(*j).end(); J++ )
 128                 if( *J != I ) {     // for all J in nb(j) \ I
 129                     prod *= Factor( *j, message(*j,*J) );
 130     Factor marg = prod.marginal(var(i));
 131 */
 132
 133     Prob prod( factor(I).p() );
 134     if( logDomain )
 135         prod.takeLog();
 136
 137     // Calculate product of incoming messages and factor I
 138     foreach( const Neighbor &j, nbF(I) ) {
 139         if( j != i ) {     // for all j in I \ i
 140             size_t _I = j.dual;
 141             // ind is the precalculated IndexFor(j,I) i.e. to x_I == k corresponds x_j == ind[k]
 142             const ind_t & ind = index(j, _I);
 143
 144             // prod_j will be the product of messages coming into j
 145             Prob prod_j( var(j).states(), logDomain ? 0.0 : 1.0 );
 146             foreach( const Neighbor &J, nbV(j) )
 147                 if( J != I ) { // for all J in nb(j) \ I
 148                     if( logDomain )
 149                         prod_j += message( j, J.iter );
 150                     else
 151                         prod_j *= message( j, J.iter );
 152                 }
 153
 154             // multiply prod with prod_j
 155             for( size_t r = 0; r < prod.size(); ++r )
 156                 if( logDomain )
 157                     prod[r] += prod_j[ind[r]];
 158                 else
 159                     prod[r] *= prod_j[ind[r]];
 160         }
 161     }
 162     if( logDomain ) {
 163         prod -= prod.maxVal();
 164         prod.takeExp();
 165     }
 166
 167     // Marginalize onto i
 168     Prob marg( var(i).states(), 0.0 );
 169     // ind is the precalculated IndexFor(i,I) i.e. to x_I == k corresponds x_i == ind[k]
 170     const ind_t ind = index(i,_I);
 171     for( size_t r = 0; r < prod.size(); ++r )
 172         marg[ind[r]] += prod[r];
 173     marg.normalize( _normtype );
 174
 175     // Store result
 176     if( logDomain )
 177         newMessage(i,_I) = marg.log();
 178     else
 179         newMessage(i,_I) = marg;
 180 }
 181
 182
 183 // BP::run does not check for NANs for performance reasons
 184 // Somehow NaNs do not often occur in BP...
 185 double BP::run() {
 186     if( Verbose() >= 1 )
 187         cout << "Starting " << identify() << "...";
 188     if( Verbose() >= 3)
 189        cout << endl;
 190
 191     double tic = toc();
 192     Diffs diffs(nrVars(), 1.0);
 193
 194     vector<Edge> update_seq;
 195
 196     vector<Factor> old_beliefs;
 197     old_beliefs.reserve( nrVars() );
 198     for( size_t i = 0; i < nrVars(); ++i )
 199         old_beliefs.push_back( beliefV(i) );
 200
 201     size_t iter = 0;
 202     size_t nredges = nrEdges();
 203
 204     if( Updates() == UpdateType::SEQMAX ) {
 205         // do the first pass
 206         for( size_t i = 0; i < nrVars(); ++i )
 207             foreach( const Neighbor &I, nbV(i) ) {
 208                 calcNewMessage( i, I.iter );
 209                 // calculate initial residuals
 210                 residual( i, I.iter ) = dist( newMessage( i, I.iter ), message( i, I.iter ), Prob::DISTLINF );
 211             }
 212     } else {
 213         update_seq.reserve( nredges );
 214         for( size_t i = 0; i < nrVars(); ++i )
 215             foreach( const Neighbor &I, nbV(i) )
 216                 update_seq.push_back( Edge( i, I.iter ) );
 217     }
 218
 219     // do several passes over the network until maximum number of iterations has
 220     // been reached or until the maximum belief difference is smaller than tolerance
 221     for( iter=0; iter < MaxIter() && diffs.maxDiff() > Tol(); ++iter ) {
 222         if( Updates() == UpdateType::SEQMAX ) {
 223             // Residuals-BP by Koller et al.
 224             for( size_t t = 0; t < nredges; ++t ) {
 225                 // update the message with the largest residual
 226
 227                 size_t i, _I;
 228                 findMaxResidual( i, _I );
 229                 message( i, _I ) = newMessage( i, _I );
 230                 residual( i, _I ) = 0.0;
 231
 232                 // I->i has been updated, which means that residuals for all
 233                 // J->j with J in nb[i]\I and j in nb[J]\i have to be updated
 234                 foreach( const Neighbor &J, nbV(i) ) {
 235                     if( J.iter != _I ) {
 236                         foreach( const Neighbor &j, nbF(J) ) {
 237                             size_t _J = j.dual;
 238                             if( j != i ) {
 239                                 calcNewMessage( j, _J );
 240                                 residual( j, _J ) = dist( newMessage( j, _J ), message( j, _J ), Prob::DISTLINF );
 241                             }
 242                         }
 243                     }
 244                 }
 245             }
 246         } else if( Updates() == UpdateType::PARALL ) {
 247             // Parallel updates
 248             for( size_t i = 0; i < nrVars(); ++i )
 249                 foreach( const Neighbor &I, nbV(i) )
 250                     calcNewMessage( i, I.iter );
 251
 252             for( size_t i = 0; i < nrVars(); ++i )
 253                 foreach( const Neighbor &I, nbV(i) )
 254                     message( i, I.iter ) = newMessage( i, I.iter );
 255         } else {
 256             // Sequential updates
 257             if( Updates() == UpdateType::SEQRND )
 258                 random_shuffle( update_seq.begin(), update_seq.end() );
 259
 260             foreach( const Edge &e, update_seq ) {
 261                 calcNewMessage( e.first, e.second );
 262                 message( e.first, e.second ) = newMessage( e.first, e.second );
 263             }
 264         }
 265
 266         // calculate new beliefs and compare with old ones
 267         for( size_t i = 0; i < nrVars(); ++i ) {
 268             Factor nb( beliefV(i) );
 269             diffs.push( dist( nb, old_beliefs[i], Prob::DISTLINF ) );
 270             old_beliefs[i] = nb;
 271         }
 272
 273         if( Verbose() >= 3 )
 274             cout << "BP::run:  maxdiff " << diffs.maxDiff() << " after " << iter+1 << " passes" << endl;
 275     }
 276
 277     updateMaxDiff( diffs.maxDiff() );
 278
 279     if( Verbose() >= 1 ) {
 280         if( diffs.maxDiff() > Tol() ) {
 281             if( Verbose() == 1 )
 282                 cout << endl;
 283                 cout << "BP::run:  WARNING: not converged within " << MaxIter() << " passes (" << toc() - tic << " clocks)...final maxdiff:" << diffs.maxDiff() << endl;
 284         } else {
 285             if( Verbose() >= 3 )
 286                 cout << "BP::run:  ";
 287                 cout << "converged in " << iter << " passes (" << toc() - tic << " clocks)." << endl;
 288         }
 289     }
 290
 291     return diffs.maxDiff();
 292 }
 293
 294
 295 Factor BP::beliefV( size_t i ) const {
 296     Prob prod( var(i).states(), logDomain ? 0.0 : 1.0 );
 297     foreach( const Neighbor &I, nbV(i) )
 298         if( logDomain )
 299             prod += newMessage( i, I.iter );
 300         else
 301             prod *= newMessage( i, I.iter );
 302     if( logDomain ) {
 303         prod -= prod.maxVal();
 304         prod.takeExp();
 305     }
 306
 307     prod.normalize( Prob::NORMPROB );
 308     return( Factor( var(i), prod ) );
 309 }
 310
 311
 312 Factor BP::belief (const Var &n) const {
 313     return( beliefV( findVar( n ) ) );
 314 }
 315
 316
 317 vector<Factor> BP::beliefs() const {
 318     vector<Factor> result;
 319     for( size_t i = 0; i < nrVars(); ++i )
 320         result.push_back( beliefV(i) );
 321     for( size_t I = 0; I < nrFactors(); ++I )
 322         result.push_back( beliefF(I) );
 323     return result;
 324 }
 325
 326
 327 Factor BP::belief( const VarSet &ns ) const {
 328     if( ns.size() == 1 )
 329         return belief( *(ns.begin()) );
 330     else {
 331         size_t I;
 332         for( I = 0; I < nrFactors(); I++ )
 333             if( factor(I).vars() >> ns )
 334                 break;
 335         assert( I != nrFactors() );
 336         return beliefF(I).marginal(ns);
 337     }
 338 }
 339
 340
 341 Factor BP::beliefF (size_t I) const {
 342     Prob prod( factor(I).p() );
 343     if( logDomain )
 344         prod.takeLog();
 345
 346     foreach( const Neighbor &j, nbF(I) ) {
 347         size_t _I = j.dual;
 348         // ind is the precalculated IndexFor(j,I) i.e. to x_I == k corresponds x_j == ind[k]
 349         const ind_t & ind = index(j, _I);
 350
 351         // prod_j will be the product of messages coming into j
 352         Prob prod_j( var(j).states(), logDomain ? 0.0 : 1.0 );
 353         foreach( const Neighbor &J, nbV(j) ) {
 354             if( J != I ) { // for all J in nb(j) \ I
 355                 if( logDomain )
 356                     prod_j += newMessage( j, J.iter );
 357                 else
 358                     prod_j *= newMessage( j, J.iter );
 359             }
 360         }
 361
 362         // multiply prod with prod_j
 363         for( size_t r = 0; r < prod.size(); ++r ) {
 364             if( logDomain )
 365                 prod[r] += prod_j[ind[r]];
 366             else
 367                 prod[r] *= prod_j[ind[r]];
 368         }
 369     }
 370
 371     if( logDomain ) {
 372         prod -= prod.maxVal();
 373         prod.takeExp();
 374     }
 375
 376     Factor result( factor(I).vars(), prod );
 377     result.normalize( Prob::NORMPROB );
 378
 379     return( result );
 380
 381 /*  UNOPTIMIZED VERSION
 382
 383     Factor prod( factor(I) );
 384     for( _nb_cit i = nb2(I).begin(); i != nb2(I).end(); i++ ) {
 385         for( _nb_cit J = nb1(*i).begin(); J != nb1(*i).end(); J++ )
 386             if( *J != I )
 387                 prod *= Factor( var(*i), newMessage(*i,*J)) );
 388     }
 389     return prod.normalize( Prob::NORMPROB );*/
 390 }
 391
 392
 393 Complex BP::logZ() const {
 394     Complex sum = 0.0;
 395     for(size_t i = 0; i < nrVars(); ++i )
 396         sum += Complex(1.0 - nbV(i).size()) * beliefV(i).entropy();
 397     for( size_t I = 0; I < nrFactors(); ++I )
 398         sum -= KL_dist( beliefF(I), factor(I) );
 399     return sum;
 400 }
 401
 402
 403 string BP::identify() const {
 404     stringstream result (stringstream::out);
 405     result << Name << GetProperties();
 406     return result.str();
 407 }
 408
 409
 410 void BP::init( const VarSet &ns ) {
 411     for( VarSet::const_iterator n = ns.begin(); n != ns.end(); ++n ) {
 412         size_t ni = findVar( *n );
 413         foreach( const Neighbor &I, nbV( ni ) )
 414             message( ni, I.iter ).fill( logDomain ? 0.0 : 1.0 );
 415     }
 416 }
 417
 418
 419 } // end of namespace dai