src/fbp.cpp

   1 /*  This file is part of libDAI - http://www.libdai.org/
   2  *
   3  *  Copyright (c) 2006-2011, The libDAI authors. All rights reserved.
   4  *
   5  *  Use of this source code is governed by a BSD-style license that can be found in the LICENSE file.
   6  */
   7
   8
   9 #include <dai/fbp.h>
  10
  11
  12 #define DAI_FBP_FAST 1
  13
  14
  15 namespace dai {
  16
  17
  18 using namespace std;
  19
  20
  21 // This code has been copied from bp.cpp, except where comments indicate FBP-specific behaviour
  22 Real FBP::logZ() const {
  23     Real sum = 0.0;
  24     for( size_t I = 0; I < nrFactors(); I++ ) {
  25         sum += (beliefF(I) * factor(I).log(true)).sum();  // FBP
  26         sum += Weight(I) * beliefF(I).entropy();  // FBP
  27     }
  28     for( size_t i = 0; i < nrVars(); ++i ) {
  29         Real c_i = 0.0;
  30         bforeach( const Neighbor &I, nbV(i) )
  31             c_i += Weight(I);
  32         if( c_i != 1.0 )
  33             sum += (1.0 - c_i) * beliefV(i).entropy();  // FBP
  34     }
  35     return sum;
  36 }
  37
  38
  39 // This code has been copied from bp.cpp, except where comments indicate FBP-specific behaviour
  40 Prob FBP::calcIncomingMessageProduct( size_t I, bool without_i, size_t i ) const {
  41     Real c_I = Weight(I); // FBP: c_I
  42
  43     Factor Fprod( factor(I) );
  44     Prob &prod = Fprod.p();
  45
  46     if( props.logdomain ) {
  47         prod.takeLog();
  48         prod /= c_I; // FBP
  49     } else
  50         prod ^= (1.0 / c_I); // FBP
  51
  52     // Calculate product of incoming messages and factor I
  53     bforeach( const Neighbor &j, nbF(I) )
  54         if( !(without_i && (j == i)) ) {
  55             // prod_j will be the product of messages coming into j
  56             // FBP: corresponds to messages n_jI
  57             Prob prod_j( var(j).states(), props.logdomain ? 0.0 : 1.0 );
  58             bforeach( const Neighbor &J, nbV(j) )
  59                 if( J != I ) { // for all J in nb(j) \ I
  60                     if( props.logdomain )
  61                         prod_j += message( j, J.iter );
  62                     else
  63                         prod_j *= message( j, J.iter );
  64                 } else if( c_I != 1.0 ) {
  65                     // FBP: multiply by m_Ij^(1-1/c_I)
  66                     if( props.logdomain )
  67                         prod_j += newMessage( j, J.iter) * (1.0 - 1.0 / c_I);
  68                     else
  69                         prod_j *= newMessage( j, J.iter) ^ (1.0 - 1.0 / c_I);
  70                 }
  71
  72             // multiply prod with prod_j
  73             if( !DAI_FBP_FAST ) {
  74                 // UNOPTIMIZED (SIMPLE TO READ, BUT SLOW) VERSION
  75                 if( props.logdomain )
  76                     Fprod += Factor( var(j), prod_j );
  77                 else
  78                     Fprod *= Factor( var(j), prod_j );
  79             } else {
  80                 // OPTIMIZED VERSION
  81                 size_t _I = j.dual;
  82                 // ind is the precalculated IndexFor(j,I) i.e. to x_I == k corresponds x_j == ind[k]
  83                 const ind_t &ind = index(j, _I);
  84
  85                 for( size_t r = 0; r < prod.size(); ++r )
  86                     if( props.logdomain )
  87                         prod.set( r, prod[r] + prod_j[ind[r]] );
  88                     else
  89                         prod.set( r, prod[r] * prod_j[ind[r]] );
  90             }
  91     }
  92     return prod;
  93 }
  94
  95
  96 // This code has been copied from bp.cpp, except where comments indicate FBP-specific behaviour
  97 void FBP::calcNewMessage( size_t i, size_t _I ) {
  98     // calculate updated message I->i
  99     size_t I = nbV(i,_I);
 100
 101     Real c_I = Weight(I); // FBP: c_I
 102
 103     Factor Fprod( factor(I) );
 104     Prob &prod = Fprod.p();
 105     prod = calcIncomingMessageProduct( I, true, i );
 106
 107     if( props.logdomain ) {
 108         prod -= prod.max();
 109         prod.takeExp();
 110     }
 111
 112     // Marginalize onto i
 113     Prob marg;
 114     if( !DAI_FBP_FAST ) {
 115         // UNOPTIMIZED (SIMPLE TO READ, BUT SLOW) VERSION
 116         if( props.inference == Properties::InfType::SUMPROD )
 117             marg = Fprod.marginal( var(i) ).p();
 118         else
 119             marg = Fprod.maxMarginal( var(i) ).p();
 120     } else {
 121         // OPTIMIZED VERSION
 122         marg = Prob( var(i).states(), 0.0 );
 123         // ind is the precalculated IndexFor(i,I) i.e. to x_I == k corresponds x_i == ind[k]
 124         const ind_t ind = index(i,_I);
 125         if( props.inference == Properties::InfType::SUMPROD )
 126             for( size_t r = 0; r < prod.size(); ++r )
 127                 marg.set( ind[r], marg[ind[r]] + prod[r] );
 128         else
 129             for( size_t r = 0; r < prod.size(); ++r )
 130                 if( prod[r] > marg[ind[r]] )
 131                     marg.set( ind[r], prod[r] );
 132         marg.normalize();
 133     }
 134
 135     // FBP
 136     marg ^= c_I;
 137
 138     // Store result
 139     if( props.logdomain )
 140         newMessage(i,_I) = marg.log();
 141     else
 142         newMessage(i,_I) = marg;
 143
 144     // Update the residual if necessary
 145     if( props.updates == Properties::UpdateType::SEQMAX )
 146         updateResidual( i, _I , dist( newMessage( i, _I ), message( i, _I ), DISTLINF ) );
 147 }
 148
 149
 150 void FBP::construct() {
 151     BP::construct();
 152     _weight.resize( nrFactors(), 1.0 );
 153 }
 154
 155
 156 } // end of namespace dai