include/dai/bbp.h

   1 /*  This file is part of libDAI - http://www.libdai.org/
   2  *
   3  *  libDAI is licensed under the terms of the GNU General Public License version
   4  *  2, or (at your option) any later version. libDAI is distributed without any
   5  *  warranty. See the file COPYING for more details.
   6  *
   7  *  Copyright (C) 2009  Frederik Eaton [frederik at ofb dot net]
   8  */
   9
  10
  11 /// \file
  12 /// \brief Defines class BBP [\ref EaG09]
  13 /// \todo Fit more closely into libDAI framework
  14 /// \todo Improve documentation
  15
  16
  17 #ifndef ___defined_libdai_bbp_h
  18 #define ___defined_libdai_bbp_h
  19
  20
  21 #include <vector>
  22 #include <utility>
  23
  24 #include <dai/prob.h>
  25 #include <dai/daialg.h>
  26 #include <dai/factorgraph.h>
  27 #include <dai/enum.h>
  28 #include <dai/bp_dual.h>
  29
  30
  31 namespace dai {
  32
  33
  34 /// Computes the adjoint of the unnormed probability vector from the normalizer and the adjoint of the normalized probability vector @see eqn. (13) in [\ref EaG09]
  35 Prob unnormAdjoint( const Prob &w, Real Z_w, const Prob &adj_w );
  36
  37 /// Runs Gibbs sampling for \a iters iterations on ia.fg(), and returns state
  38 std::vector<size_t> getGibbsState( const InfAlg &ia, size_t iters );
  39
  40
  41 /// Implements BBP (Back-Belief-Propagation) [\ref EaG09]
  42 class BBP {
  43     protected:
  44     /// \name Inputs
  45     //@{
  46         BP_dual _bp_dual;
  47         const FactorGraph *_fg;
  48         const InfAlg *_ia;
  49     //@}
  50
  51         /// Number of iterations done
  52         size_t _iters;
  53
  54     /// \name Outputs
  55     //@{
  56         /// Variable factor adjoints
  57         std::vector<Prob> _adj_psi_V;
  58         /// Factor adjoints
  59         std::vector<Prob> _adj_psi_F;
  60         /// Variable->factor message adjoints (indexed [i][_I])
  61         std::vector<std::vector<Prob> > _adj_n;
  62         /// Factor->variable message adjoints (indexed [i][_I])
  63         std::vector<std::vector<Prob> > _adj_m;
  64         /// Normalized variable belief adjoints
  65         std::vector<Prob> _adj_b_V;
  66         /// Normalized factor belief adjoints
  67         std::vector<Prob> _adj_b_F;
  68     //@}
  69
  70     /// \name Helper quantities computed from the BP messages
  71     //@{
  72         /// _T[i][_I] (see eqn. (41) in [\ref EaG09])
  73         std::vector<std::vector<Prob > > _T;
  74         /// _U[I][_i] (see eqn. (42) in [\ref EaG09])
  75         std::vector<std::vector<Prob > > _U;
  76         /// _S[i][_I][_j] (see eqn. (43) in [\ref EaG09])
  77         std::vector<std::vector<std::vector<Prob > > > _S;
  78         /// _R[I][_i][_J] (see eqn. (44) in [\ref EaG09])
  79         std::vector<std::vector<std::vector<Prob > > > _R;
  80     //@}
  81
  82         /// Unnormalized variable belief adjoints
  83         std::vector<Prob> _adj_b_V_unnorm;
  84         /// Unnormalized factor belief adjoints
  85         std::vector<Prob> _adj_b_F_unnorm;
  86
  87         /// Initial variable factor adjoints
  88         std::vector<Prob> _init_adj_psi_V;
  89         /// Initial factor adjoints
  90         std::vector<Prob> _init_adj_psi_F;
  91
  92         /// Unnormalized variable->factor message adjoint (indexed [i][_I])
  93         std::vector<std::vector<Prob> > _adj_n_unnorm;
  94         /// Unnormalized factor->variable message adjoint (indexed [i][_I])
  95         std::vector<std::vector<Prob> > _adj_m_unnorm;
  96         /// Updated normalized variable->factor message adjoint (indexed [i][_I])
  97         std::vector<std::vector<Prob> > _new_adj_n;
  98         /// Updated normalized factor->variable message adjoint (indexed [i][_I])
  99         std::vector<std::vector<Prob> > _new_adj_m;
 100
 101     /// \name Optimized indexing (for performance)
 102     //@{
 103         /// Calculates _indices, which is a cache of IndexFor @see bp.cpp
 104         void RegenerateInds();
 105
 106         /// Index type
 107         typedef std::vector<size_t>  _ind_t;
 108         /// Cached indices (indexed [i][_I])
 109         std::vector<std::vector<_ind_t> >  _indices;
 110         /// Returns an index from the cache
 111         const _ind_t& _index(size_t i, size_t _I) const { return _indices[i][_I]; }
 112     //@}
 113
 114     /// \name Initialization
 115     //@{
 116         /// Calculate T values; see eqn. (41) in [\ref EaG09]
 117         void RegenerateT();
 118         /// Calculate U values; see eqn. (42) in [\ref EaG09]
 119         void RegenerateU();
 120         /// Calculate S values; see eqn. (43) in [\ref EaG09]
 121         void RegenerateS();
 122         /// Calculate R values; see eqn. (44) in [\ref EaG09]
 123         void RegenerateR();
 124         /// Calculate _adj_b_V_unnorm and _adj_b_F_unnorm from _adj_b_V and _adj_b_F
 125         void RegenerateInputs();
 126         /// Initialise members for factor adjoints (call after RegenerateInputs)
 127         void RegeneratePsiAdjoints();
 128         /// Initialise members for message adjoints (call after RegenerateInputs) for parallel algorithm
 129         void RegenerateParMessageAdjoints();
 130         /// Initialise members for message adjoints (call after RegenerateInputs) for sequential algorithm
 131         /** Same as RegenerateMessageAdjoints, but calls sendSeqMsgN rather
 132          *  than updating _adj_n (and friends) which are unused in the sequential algorithm.
 133          */
 134         void RegenerateSeqMessageAdjoints();
 135     //@}
 136
 137         /// Returns reference to T value; see eqn. (41) in [\ref EaG09]
 138         Prob & T(size_t i, size_t _I) { return _T[i][_I]; }
 139         /// Returns constant reference to T value; see eqn. (41) in [\ref EaG09]
 140         const Prob & T(size_t i, size_t _I) const { return _T[i][_I]; }
 141         /// Returns reference to U value; see eqn. (42) in [\ref EaG09]
 142         Prob & U(size_t I, size_t _i) { return _U[I][_i]; }
 143         /// Returns constant reference to U value; see eqn. (42) in [\ref EaG09]
 144         const Prob & U(size_t I, size_t _i) const { return _U[I][_i]; }
 145         /// Returns reference to S value; see eqn. (43) in [\ref EaG09]
 146         Prob & S(size_t i, size_t _I, size_t _j) { return _S[i][_I][_j]; }
 147         /// Returns constant reference to S value; see eqn. (43) in [\ref EaG09]
 148         const Prob & S(size_t i, size_t _I, size_t _j) const { return _S[i][_I][_j]; }
 149         /// Returns reference to R value; see eqn. (44) in [\ref EaG09]
 150         Prob & R(size_t I, size_t _i, size_t _J) { return _R[I][_i][_J]; }
 151         /// Returns constant reference to R value; see eqn. (44) in [\ref EaG09]
 152         const Prob & R(size_t I, size_t _i, size_t _J) const { return _R[I][_i][_J]; }
 153
 154     /// \name Parallel algorithm
 155     //@{
 156         /// Calculates new variable->factor message adjoint
 157         /** Increases variable factor adjoint according to eqn. (27) in [\ref EaG09] and
 158          *  calculates the new variable->factor message adjoint according to eqn. (29) in [\ref EaG09].
 159          */
 160         void calcNewN( size_t i, size_t _I );
 161         /// Calculates new factor->variable message adjoint
 162         /** Increases factor adjoint according to eqn. (28) in [\ref EaG09] and
 163          *  calculates the new factor->variable message adjoint according to the r.h.s. of eqn. (30) in [\ref EaG09].
 164          */
 165         void calcNewM( size_t i, size_t _I );
 166         /// Calculates unnormalized variable->factor message adjoint from the normalized one
 167         void calcUnnormMsgN( size_t i, size_t _I );
 168         /// Calculates unnormalized factor->variable message adjoint from the normalized one
 169         void calcUnnormMsgM( size_t i, size_t _I );
 170         /// Updates (un)normalized variable->factor message adjoints
 171         void upMsgN( size_t i, size_t _I );
 172         /// Updates (un)normalized factor->variable message adjoints
 173         void upMsgM( size_t i, size_t _I );
 174         /// Do one parallel update of all message adjoints
 175         void doParUpdate();
 176     //@}
 177
 178     /// \name Sequential algorithm
 179     //@{
 180         /// Helper function for sendSeqMsgM: increases factor->variable message adjoint by p and calculates the corresponding unnormalized adjoint
 181         void incrSeqMsgM( size_t i, size_t _I, const Prob& p );
 182         //  DISABLED BECAUSE IT IS BUGGY:
 183         //  void updateSeqMsgM( size_t i, size_t _I );
 184         /// Sets normalized factor->variable message adjoint and calculates the corresponding unnormalized adjoint
 185         void setSeqMsgM( size_t i, size_t _I, const Prob &p );
 186         /// Implements routine Send-n in Figure 5 in [\ref EaG09]
 187         void sendSeqMsgN( size_t i, size_t _I, const Prob &f );
 188         /// Implements routine Send-m in Figure 5 in [\ref EaG09]
 189         void sendSeqMsgM( size_t i, size_t _I );
 190     //@}
 191
 192         /// Calculates averaged L-1 norm of unnormalized message adjoints
 193         Real getUnMsgMag();
 194         /// Calculates averaged L-1 norms of current and new normalized message adjoints
 195         void getMsgMags( Real &s, Real &new_s );
 196
 197         /// Sets all vectors _adj_b_F to zero
 198         void zero_adj_b_F() {
 199             _adj_b_F.clear();
 200             _adj_b_F.reserve( _fg->nrFactors() );
 201             for( size_t I = 0; I < _fg->nrFactors(); I++ )
 202                 _adj_b_F.push_back( Prob( _fg->factor(I).states(), Real( 0.0 ) ) );
 203         }
 204
 205         /// Returns indices and magnitude of the largest normalized factor->variable message adjoint
 206         void getArgmaxMsgM( size_t &i, size_t &_I, Real &mag );
 207         /// Returns magnitude of the largest (in L1-norm) normalized factor->variable message adjoint
 208         Real getMaxMsgM();
 209         /// Calculates sum of L1 norms of all normalized factor->variable message adjoints
 210         Real getTotalMsgM();
 211         /// Calculates sum of L1 norms of all updated normalized factor->variable message adjoints
 212         Real getTotalNewMsgM();
 213         /// Calculates sum of L1 norms of all normalized variable->factor message adjoints
 214         Real getTotalMsgN();
 215
 216     public:
 217         /// Called by \a init, recalculates intermediate values
 218         void Regenerate();
 219
 220         /// Constructor
 221         BBP( const InfAlg *ia, const PropertySet &opts ) : _bp_dual(ia), _fg(&(ia->fg())), _ia(ia) {
 222             props.set(opts);
 223         }
 224
 225         /// Returns a vector of Probs (filled with zeroes) with state spaces corresponding to the factors in the factor graph fg
 226         std::vector<Prob> getZeroAdjF( const FactorGraph &fg );
 227         /// Returns a vector of Probs (filled with zeroes) with state spaces corresponding to the variables in the factor graph fg
 228         std::vector<Prob> getZeroAdjV( const FactorGraph &fg );
 229
 230         /// Initializes belief adjoints and initial factor adjoints and regenerates
 231         void init( const std::vector<Prob> &adj_b_V, const std::vector<Prob> &adj_b_F, const std::vector<Prob> &adj_psi_V, const std::vector<Prob> &adj_psi_F ) {
 232             _adj_b_V = adj_b_V;
 233             _adj_b_F = adj_b_F;
 234             _init_adj_psi_V = adj_psi_V;
 235             _init_adj_psi_F = adj_psi_F;
 236             Regenerate();
 237         }
 238
 239         /// Initializes belief adjoints and with zero initial factor adjoints and regenerates
 240         void init( const std::vector<Prob> &adj_b_V, const std::vector<Prob> &adj_b_F ) {
 241             init( adj_b_V, adj_b_F, getZeroAdjV(*_fg), getZeroAdjF(*_fg) );
 242         }
 243
 244         /// Initializes variable belief adjoints (and sets factor belief adjoints to zero) and with zero initial factor adjoints and regenerates
 245         void init( const std::vector<Prob> &adj_b_V ) {
 246             init(adj_b_V, getZeroAdjF(*_fg));
 247         }
 248
 249         /// Run until change is less than given tolerance
 250         void run();
 251
 252         /// Return number of iterations done so far
 253         size_t doneIters() { return _iters; }
 254
 255         /// Returns reference to variable factor adjoint
 256         Prob& adj_psi_V(size_t i) { return _adj_psi_V[i]; }
 257         /// Returns constant reference to variable factor adjoint
 258         const Prob& adj_psi_V(size_t i) const { return _adj_psi_V[i]; }
 259         /// Returns reference to factor adjoint
 260         Prob& adj_psi_F(size_t I) { return _adj_psi_F[I]; }
 261         /// Returns constant reference to factor adjoint
 262         const Prob& adj_psi_F(size_t I) const { return _adj_psi_F[I]; }
 263         /// Returns reference to variable belief adjoint
 264         Prob& adj_b_V(size_t i) { return _adj_b_V[i]; }
 265         /// Returns constant reference to variable belief adjoint
 266         const Prob& adj_b_V(size_t i) const { return _adj_b_V[i]; }
 267         /// Returns reference to factor belief adjoint
 268         Prob& adj_b_F(size_t I) { return _adj_b_F[I]; }
 269         /// Returns constant reference to factor belief adjoint
 270         const Prob& adj_b_F(size_t I) const { return _adj_b_F[I]; }
 271
 272      protected:
 273         /// Returns reference to variable->factor message adjoint
 274         Prob& adj_n(size_t i, size_t _I) { return _adj_n[i][_I]; }
 275         /// Returns constant reference to variable->factor message adjoint
 276         const Prob& adj_n(size_t i, size_t _I) const { return _adj_n[i][_I]; }
 277         /// Returns reference to factor->variable message adjoint
 278         Prob& adj_m(size_t i, size_t _I) { return _adj_m[i][_I]; }
 279         /// Returns constant reference to factor->variable message adjoint
 280         const Prob& adj_m(size_t i, size_t _I) const { return _adj_m[i][_I]; }
 281
 282      public:
 283         /// Parameters of this algorithm
 284         /* PROPERTIES(props,BBP) {
 285            /// Enumeration of possible update schedules
 286            DAI_ENUM(UpdateType,SEQ_FIX,SEQ_MAX,SEQ_BP_REV,SEQ_BP_FWD,PAR);
 287
 288            /// Verbosity
 289            size_t verbose;
 290
 291            /// Maximum number of iterations
 292            size_t maxiter;
 293
 294            /// Tolerance (not used for updates = SEQ_BP_REV, SEQ_BP_FWD)
 295            Real tol;
 296
 297            /// Damping constant (0 for none); damping = 1 - lambda where lambda is the damping constant used in [\ref EaG09]
 298            Real damping;
 299
 300            /// Update schedule
 301            UpdateType updates;
 302
 303            // DISABLED BECAUSE IT IS BUGGY:
 304            // bool clean_updates;
 305         }
 306         */
 307 /* {{{ GENERATED CODE: DO NOT EDIT. Created by
 308     ./scripts/regenerate-properties include/dai/bbp.h src/bbp.cpp
 309 */
 310         struct Properties {
 311             /// Enumeration of possible update schedules
 312             DAI_ENUM(UpdateType,SEQ_FIX,SEQ_MAX,SEQ_BP_REV,SEQ_BP_FWD,PAR);
 313             /// Verbosity
 314             size_t verbose;
 315             /// Maximum number of iterations
 316             size_t maxiter;
 317             /// Tolerance (not used for updates = SEQ_BP_REV, SEQ_BP_FWD)
 318             Real tol;
 319             /// Damping constant (0 for none); damping = 1 - lambda where lambda is the damping constant used in [\ref EaG09]
 320             Real damping;
 321             /// Update schedule
 322             UpdateType updates;
 323
 324             /// Set members from PropertySet
 325             void set(const PropertySet &opts);
 326             /// Get members into PropertySet
 327             PropertySet get() const;
 328             /// Convert to a string which can be parsed as a PropertySet
 329             std::string toString() const;
 330         } props;
 331 /* }}} END OF GENERATED CODE */
 332 };
 333
 334
 335 /// Enumeration of several cost functions that can be used with BBP.
 336 DAI_ENUM(bbp_cfn_t,CFN_GIBBS_B,CFN_GIBBS_B2,CFN_GIBBS_EXP,CFN_GIBBS_B_FACTOR,CFN_GIBBS_B2_FACTOR,CFN_GIBBS_EXP_FACTOR,CFN_VAR_ENT,CFN_FACTOR_ENT,CFN_BETHE_ENT);
 337
 338 /// Initialise BBP using InfAlg, cost function, and stateP
 339 /** Calls bbp.init with adjoints calculated from ia.beliefV and
 340  *  ia.beliefF. stateP is a Gibbs state and can be NULL, it will be
 341  *  initialised using a Gibbs run of 2*fg.Iterations() iterations.
 342  */
 343 void initBBPCostFnAdj( BBP &bbp, const InfAlg &ia, bbp_cfn_t cfn_type, const std::vector<size_t> *stateP );
 344
 345 /// Answers question: does the given cost function depend on having a Gibbs state?
 346 bool needGibbsState( bbp_cfn_t cfn );
 347
 348 /// Calculate actual value of cost function (cfn_type, stateP)
 349 /** This function returns the actual value of the cost function whose
 350  *  gradient with respect to singleton beliefs is given by
 351  *  gibbsToB1Adj on the same arguments
 352  */
 353 Real getCostFn( const InfAlg &fg, bbp_cfn_t cfn_type, const std::vector<size_t> *stateP );
 354
 355 /// Function to test the validity of adjoints computed by BBP given a state for each variable using numerical derivatives.
 356 /** Factors containing a variable are multiplied by psi_1 adjustments to verify accuracy of _adj_psi_V.
 357  *  \param bp BP object.
 358  *  \param state Global state of all variables.
 359  *  \param bbp_props BBP Properties.
 360  *  \param cfn Cost function to be used.
 361  *  \param h controls size of perturbation.
 362  */
 363 Real numericBBPTest( const InfAlg &bp, const std::vector<size_t> *state, const PropertySet &bbp_props, bbp_cfn_t cfn, Real h );
 364
 365
 366 } // end of namespace dai
 367
 368
 369 #endif