include/dai/emalg.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  Charles Vaske  [cvaske at soe dot ucsc dot edu]
   8  *  Copyright (C) 2009  University of California, Santa Cruz
   9  */
  10
  11
  12 #ifndef __defined_libdai_emalg_h
  13 #define __defined_libdai_emalg_h
  14
  15
  16 #include <vector>
  17 #include <map>
  18
  19 #include <dai/factor.h>
  20 #include <dai/daialg.h>
  21 #include <dai/evidence.h>
  22 #include <dai/index.h>
  23 #include <dai/properties.h>
  24
  25
  26 /// \file
  27 /// \brief Defines classes related to Expectation Maximization: EMAlg, ParameterEstimation, CondProbEstimation and SharedParameters
  28 /// \todo Describe EM file format
  29 /// \todo Improve documentation
  30 /// \author Charles Vaske
  31
  32
  33 namespace dai {
  34
  35
  36 /// Interface for a parameter estimation method.
  37 /** This parameter estimation interface is based on sufficient statistics.
  38  *  Implementations are responsible for collecting data from a probability
  39  *  vector passed to it from a SharedParameters container object.
  40  *
  41  *  Implementations of this interface should register a factory function
  42  *  via the static ParameterEstimation::registerMethod function.
  43  *  The default registry only contains CondProbEstimation, named
  44  *  "ConditionalProbEstimation".
  45  *
  46  *  \author Charles Vaske
  47  */
  48 class ParameterEstimation {
  49     public:
  50         /// A pointer to a factory function.
  51         typedef ParameterEstimation* (*ParamEstFactory)( const PropertySet& );
  52
  53         /// Virtual destructor for deleting pointers to derived classes.
  54         virtual ~ParameterEstimation() {}
  55         /// Virtual copy constructor.
  56         virtual ParameterEstimation* clone() const = 0;
  57
  58         /// General factory method for construction of ParameterEstimation subclasses.
  59         static ParameterEstimation* construct( const std::string &method, const PropertySet &p );
  60
  61         /// Register a subclass so that it can be used with ParameterEstimation::construct.
  62         static void registerMethod( const std::string &method, const ParamEstFactory &f ) {
  63             if( _registry == NULL )
  64                 loadDefaultRegistry();
  65             (*_registry)[method] = f;
  66         }
  67
  68         /// Estimate the factor using the accumulated sufficient statistics and reset.
  69         virtual Prob estimate() = 0;
  70
  71         /// Accumulate the sufficient statistics for p.
  72         virtual void addSufficientStatistics( const Prob &p ) = 0;
  73
  74         /// Returns the size of the Prob that should be passed to addSufficientStatistics.
  75         virtual size_t probSize() const = 0;
  76
  77     private:
  78         /// A static registry containing all methods registered so far.
  79         static std::map<std::string, ParamEstFactory> *_registry;
  80
  81         /// Registers default ParameterEstimation subclasses (currently, only CondProbEstimation).
  82         static void loadDefaultRegistry();
  83 };
  84
  85
  86 /// Estimates the parameters of a conditional probability table, using pseudocounts.
  87 /** \author Charles Vaske
  88  */
  89 class CondProbEstimation : private ParameterEstimation {
  90     private:
  91         /// Number of states of the variable of interest
  92         size_t _target_dim;
  93         /// Current pseudocounts
  94         Prob _stats;
  95         /// Initial pseudocounts
  96         Prob _initial_stats;
  97
  98     public:
  99         /// Constructor
 100         /** For a conditional probability \f$ P( X | Y ) \f$,
 101          *  \param target_dimension should equal \f$ | X | \f$
 102          *  \param pseudocounts has length \f$ |X| \cdot |Y| \f$
 103          */
 104         CondProbEstimation( size_t target_dimension, const Prob &pseudocounts );
 105
 106         /// Virtual constructor, using a PropertySet.
 107         /** Some keys in the PropertySet are required.
 108          *  For a conditional probability \f$ P( X | Y ) \f$,
 109          *     - target_dimension should be equal to \f$ | X | \f$
 110          *     - total_dimension should be equal to \f$ |X| \cdot |Y| \f$
 111          *
 112          *  An optional key is:
 113          *     - pseudo_count, which specifies the initial counts (defaults to 1)
 114          */
 115         static ParameterEstimation* factory( const PropertySet &p );
 116
 117         /// Virtual copy constructor
 118         virtual ParameterEstimation* clone() const { return new CondProbEstimation( _target_dim, _initial_stats ); }
 119
 120         /// Virtual destructor
 121         virtual ~CondProbEstimation() {}
 122
 123         /// Returns an estimate of the conditional probability distribution.
 124         /** The format of the resulting Prob keeps all the values for
 125          *  \f$ P(X | Y=y) \f$ in sequential order in the array.
 126          */
 127         virtual Prob estimate();
 128
 129         /// Accumulate sufficient statistics from the expectations in p.
 130         virtual void addSufficientStatistics( const Prob &p );
 131
 132         /// Returns the required size for arguments to addSufficientStatistics.
 133         virtual size_t probSize() const { return _stats.size(); }
 134 };
 135
 136
 137 /// A single factor or set of factors whose parameters should be estimated.
 138 /** To ensure that parameters can be shared between different factors during
 139  *  EM learning, each factor's values are reordered to match a desired variable
 140  *  ordering. The ordering of the variables in a factor may therefore differ
 141  *  from the canonical ordering used in libDAI. The SharedParameters
 142  *  class couples one or more factors (together with the specified orderings
 143  *  of the variables) with a ParameterEstimation object, taking care of the
 144  *  necessary permutations of the factor entries / parameters.
 145  *
 146  *  \author Charles Vaske
 147  */
 148 class SharedParameters {
 149     public:
 150         /// Convenience label for an index into a factor in a FactorGraph.
 151         typedef size_t FactorIndex;
 152         /// Convenience label for a grouping of factor orientations.
 153         typedef std::map<FactorIndex, std::vector<Var> > FactorOrientations;
 154
 155     private:
 156         /// Maps factor indices to the corresponding VarSets
 157         std::map<FactorIndex, VarSet> _varsets;
 158         /// Maps factor indices to the corresponding Permute objects that permute the desired ordering into the canonical ordering
 159         std::map<FactorIndex, Permute> _perms;
 160         /// Maps factor indices to the corresponding desired variable orderings
 161         FactorOrientations _varorders;
 162         /// Parameter estimation method to be used
 163         ParameterEstimation *_estimation;
 164         /// Indicates whether the object pointed to by _estimation should be deleted upon destruction
 165         bool _deleteEstimation;
 166
 167         /// Calculates the permutation that permutes the variables in varorder into the canonical ordering
 168         static Permute calculatePermutation( const std::vector<Var> &varorder, VarSet &outVS );
 169
 170         /// Initializes _varsets and _perms from _varorders
 171         void setPermsAndVarSetsFromVarOrders();
 172
 173     public:
 174         /// Copy constructor
 175         SharedParameters( const SharedParameters &sp );
 176
 177         /// Constructor
 178         /** \param varorders  all the factor orientations for this parameter
 179          *  \param estimation a pointer to the parameter estimation method
 180          *  \param deletePE whether the parameter estimation object should be deleted in the destructor
 181          */
 182         SharedParameters( const FactorOrientations &varorders, ParameterEstimation *estimation, bool deletePE=false );
 183
 184         /// Constructor for making an object from a stream and a factor graph
 185         SharedParameters( std::istream &is, const FactorGraph &fg_varlookup );
 186
 187         /// Destructor
 188         ~SharedParameters() {
 189             if( _deleteEstimation )
 190                 delete _estimation;
 191         }
 192
 193         /// Collect the necessary statistics from expected values
 194         void collectSufficientStatistics( InfAlg &alg );
 195
 196         /// Estimate and set the shared parameters
 197         void setParameters( FactorGraph &fg );
 198
 199         /// Returns the parameters
 200         void collectParameters( const FactorGraph &fg, std::vector<Real> &outVals, std::vector<Var> &outVarOrder );
 201 };
 202
 203
 204 /// A MaximizationStep groups together several parameter estimation tasks into a single unit.
 205 /** \author Charles Vaske
 206  */
 207 class MaximizationStep {
 208     private:
 209         std::vector<SharedParameters> _params;
 210
 211     public:
 212         /// Default constructor
 213         MaximizationStep() : _params() {}
 214
 215         /// Constructor from a vector of SharedParameters objects
 216         MaximizationStep( std::vector<SharedParameters> &maximizations ) : _params(maximizations) {}
 217
 218         /// Constructor from an input stream and a corresponding factor graph
 219         MaximizationStep( std::istream &is, const FactorGraph &fg_varlookup );
 220
 221         /// Collect the beliefs from this InfAlg as expectations for the next Maximization step.
 222         void addExpectations( InfAlg &alg );
 223
 224         /// Using all of the currently added expectations, make new factors with maximized parameters and set them in the FactorGraph.
 225         void maximize( FactorGraph &fg );
 226
 227     /// \name Iterator interface
 228     //@{
 229         typedef std::vector<SharedParameters>::iterator iterator;
 230         typedef std::vector<SharedParameters>::const_iterator const_iterator;
 231         iterator begin() { return _params.begin(); }
 232         const_iterator begin() const { return _params.begin(); }
 233         iterator end() { return _params.end(); }
 234         const_iterator end() const { return _params.end(); }
 235     //@}
 236 };
 237
 238
 239 /// EMAlg performs Expectation Maximization to learn factor parameters.
 240 /** This requires specifying:
 241  *     - Evidence (instances of observations from the graphical model),
 242  *     - InfAlg for performing the E-step, which includes the factor graph,
 243  *     - a vector of MaximizationSteps steps to be performed.
 244  *
 245  *  This implementation can perform incremental EM by using multiple
 246  *  MaximizationSteps.  An expectation step is performed between execution
 247  *  of each MaximizationStep.  A call to iterate() will cycle through all
 248  *  MaximizationSteps.
 249  *
 250  *  Having multiple and separate maximization steps allows for maximizing some
 251  *  parameters, performing another E step, and then maximizing separate
 252  *  parameters, which may result in faster convergence in some cases.
 253  *
 254  *  \author Charles Vaske
 255  */
 256 class EMAlg {
 257     private:
 258         /// All the data samples used during learning
 259         const Evidence &_evidence;
 260
 261         /// How to do the expectation step
 262         InfAlg &_estep;
 263
 264         /// The maximization steps to take
 265         std::vector<MaximizationStep> _msteps;
 266
 267         /// Number of iterations done
 268         size_t _iters;
 269
 270         /// History of likelihoods
 271         std::vector<Real> _lastLogZ;
 272
 273         /// Maximum number of iterations
 274         size_t _max_iters;
 275
 276         /// Convergence tolerance
 277         Real _log_z_tol;
 278
 279     public:
 280         /// Key for setting maximum iterations @see setTermConditions
 281         static const std::string MAX_ITERS_KEY;
 282         /// Default maximum iterations @see setTermConditions
 283         static const size_t MAX_ITERS_DEFAULT;
 284         /// Key for setting likelihood termination condition @see setTermConditions
 285         static const std::string LOG_Z_TOL_KEY;
 286         /// Default likelihood tolerance @see setTermConditions
 287         static const Real LOG_Z_TOL_DEFAULT;
 288
 289         /// Construct an EMAlg from all these objects
 290         EMAlg( const Evidence &evidence, InfAlg &estep, std::vector<MaximizationStep> &msteps, const PropertySet &termconditions )
 291           : _evidence(evidence), _estep(estep), _msteps(msteps), _iters(0), _lastLogZ(), _max_iters(MAX_ITERS_DEFAULT), _log_z_tol(LOG_Z_TOL_DEFAULT)
 292         {
 293               setTermConditions( termconditions );
 294         }
 295
 296         /// Construct an EMAlg from an Evidence object, an InfAlg object, and an input stream
 297         EMAlg( const Evidence &evidence, InfAlg &estep, std::istream &mstep_file );
 298
 299         /// Change the conditions for termination
 300         /** There are two possible parameters in the PropertySet
 301          *    - max_iters maximum number of iterations
 302          *    - log_z_tol proportion of increase in logZ
 303          *
 304          *  \see hasSatisifiedTermConditions()
 305          */
 306         void setTermConditions( const PropertySet &p );
 307
 308         /// Determine if the termination conditions have been met.
 309         /** There are two sufficient termination conditions:
 310          *    -# the maximum number of iterations has been performed
 311          *    -# the ratio of logZ increase over previous logZ is less than the
 312          *       tolerance, i.e.,
 313          *       \f$ \frac{\log(Z_t) - \log(Z_{t-1})}{| \log(Z_{t-1}) | } < \mathrm{tol} \f$.
 314          */
 315         bool hasSatisfiedTermConditions() const;
 316
 317         /// Return the last calculated log likelihood
 318         Real getLogZ() const { return _lastLogZ.back(); }
 319
 320         /// Returns number of iterations done so far
 321         size_t getCurrentIters() const { return _iters; }
 322
 323         /// Get the iteration method used
 324         const InfAlg& eStep() const { return _estep; }
 325
 326         /// Perform an iteration over all maximization steps
 327         Real iterate();
 328
 329         /// Perform an iteration over a single MaximizationStep
 330         Real iterate( MaximizationStep &mstep );
 331
 332         /// Iterate until termination conditions are satisfied
 333         void run();
 334
 335     /// \name Iterator interface
 336     //@{
 337         typedef std::vector<MaximizationStep>::iterator s_iterator;
 338         typedef std::vector<MaximizationStep>::const_iterator const_s_iterator;
 339         s_iterator s_begin() { return _msteps.begin(); }
 340         const_s_iterator s_begin() const { return _msteps.begin(); }
 341         s_iterator s_end() { return _msteps.end(); }
 342         const_s_iterator s_end() const { return _msteps.end(); }
 343     //@}
 344 };
 345
 346
 347 } // end of namespace dai
 348
 349
 350 #endif