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
  29
  30 namespace dai {
  31
  32
  33 /// Base class for parameter estimation methods.
  34 /** This class defines the general interface of parameter estimation methods.
  35  *
  36  *  Implementations of this interface (see e.g. CondProbEstimation) should
  37  *  register a factory function (virtual constructor) via the static
  38  *  registerMethod() function.
  39  *  This factory function should return a pointer to a newly constructed
  40  *  object, whose type is a subclass of ParameterEstimation, and gets as
  41  *  input a PropertySet of parameters. After a subclass has been registered,
  42  *  instances of it can be constructed using the construct() method.
  43  *
  44  *  Implementations are responsible for collecting data from a probability
  45  *  vector passed to it from a SharedParameters container object.
  46  *
  47  *  The default registry only contains CondProbEstimation, named
  48  *  "ConditionalProbEstimation".
  49  *
  50  *  \author Charles Vaske
  51  */
  52 class ParameterEstimation {
  53     public:
  54         /// Type of pointer to factory function.
  55         typedef ParameterEstimation* (*ParamEstFactory)( const PropertySet& );
  56
  57         /// Virtual destructor for deleting pointers to derived classes.
  58         virtual ~ParameterEstimation() {}
  59
  60         /// Virtual copy constructor.
  61         virtual ParameterEstimation* clone() const = 0;
  62
  63         /// General factory method that constructs the desired ParameterEstimation subclass
  64         /** \param method Name of the subclass that should be constructed;
  65          *  \param p Parameters passed to constructor of subclass.
  66          *  \note \a method should either be in the default registry or should be registered first using registerMethod().
  67          *  \throw UNKNOWN_PARAMETER_ESTIMATION_METHOD if the requested \a method is not registered.
  68          */
  69         static ParameterEstimation* construct( const std::string &method, const PropertySet &p );
  70
  71         /// Register a subclass so that it can be used with construct().
  72         static void registerMethod( const std::string &method, const ParamEstFactory &f ) {
  73             if( _registry == NULL )
  74                 loadDefaultRegistry();
  75             (*_registry)[method] = f;
  76         }
  77
  78         /// Estimate the factor using the accumulated sufficient statistics and reset.
  79         virtual Prob estimate() = 0;
  80
  81         /// Accumulate the sufficient statistics for \a p.
  82         virtual void addSufficientStatistics( const Prob &p ) = 0;
  83
  84         /// Returns the size of the Prob that should be passed to addSufficientStatistics.
  85         virtual size_t probSize() const = 0;
  86
  87     private:
  88         /// A static registry containing all methods registered so far.
  89         static std::map<std::string, ParamEstFactory> *_registry;
  90
  91         /// Registers default ParameterEstimation subclasses (currently, only CondProbEstimation).
  92         static void loadDefaultRegistry();
  93 };
  94
  95
  96 /// Estimates the parameters of a conditional probability table, using pseudocounts.
  97 /** \author Charles Vaske
  98  */
  99 class CondProbEstimation : private ParameterEstimation {
 100     private:
 101         /// Number of states of the variable of interest
 102         size_t _target_dim;
 103         /// Current pseudocounts
 104         Prob _stats;
 105         /// Initial pseudocounts
 106         Prob _initial_stats;
 107
 108     public:
 109         /// Constructor
 110         /** For a conditional probability \f$ P( X | Y ) \f$,
 111          *  \param target_dimension should equal \f$ | X | \f$
 112          *  \param pseudocounts are the initial pseudocounts, of length \f$ |X| \cdot |Y| \f$
 113          */
 114         CondProbEstimation( size_t target_dimension, const Prob &pseudocounts );
 115
 116         /// Virtual constructor, using a PropertySet.
 117         /** Some keys in the PropertySet are required.
 118          *  For a conditional probability \f$ P( X | Y ) \f$,
 119          *     - \a target_dimension should be equal to \f$ | X | \f$
 120          *     - \a total_dimension should be equal to \f$ |X| \cdot |Y| \f$
 121          *
 122          *  An optional key is:
 123          *     - \a pseudo_count which specifies the initial counts (defaults to 1)
 124          */
 125         static ParameterEstimation* factory( const PropertySet &p );
 126
 127         /// Virtual copy constructor
 128         virtual ParameterEstimation* clone() const { return new CondProbEstimation( _target_dim, _initial_stats ); }
 129
 130         /// Virtual destructor
 131         virtual ~CondProbEstimation() {}
 132
 133         /// Returns an estimate of the conditional probability distribution.
 134         /** The format of the resulting Prob keeps all the values for
 135          *  \f$ P(X | Y=y) \f$ in sequential order in the array.
 136          */
 137         virtual Prob estimate();
 138
 139         /// Accumulate sufficient statistics from the expectations in \a p
 140         virtual void addSufficientStatistics( const Prob &p );
 141
 142         /// Returns the required size for arguments to addSufficientStatistics().
 143         virtual size_t probSize() const { return _stats.size(); }
 144 };
 145
 146
 147 /// Represents a single factor or set of factors whose parameters should be estimated.
 148 /** To ensure that parameters can be shared between different factors during
 149  *  EM learning, each factor's values are reordered to match a desired variable
 150  *  ordering. The ordering of the variables in a factor may therefore differ
 151  *  from the canonical ordering used in libDAI. The SharedParameters
 152  *  class combines one or more factors (together with the specified orderings
 153  *  of the variables) with a ParameterEstimation object, taking care of the
 154  *  necessary permutations of the factor entries / parameters.
 155  *
 156  *  \author Charles Vaske
 157  */
 158 class SharedParameters {
 159     public:
 160         /// Convenience label for an index of a factor in a FactorGraph.
 161         typedef size_t FactorIndex;
 162         /// Convenience label for a grouping of factor orientations.
 163         typedef std::map<FactorIndex, std::vector<Var> > FactorOrientations;
 164
 165     private:
 166         /// Maps factor indices to the corresponding VarSets
 167         std::map<FactorIndex, VarSet> _varsets;
 168         /// Maps factor indices to the corresponding Permute objects that permute the canonical ordering into the desired ordering
 169         std::map<FactorIndex, Permute> _perms;
 170         /// Maps factor indices to the corresponding desired variable orderings
 171         FactorOrientations _varorders;
 172         /// Parameter estimation method to be used
 173         ParameterEstimation *_estimation;
 174         /// Indicates whether \c *this gets ownership of _estimation
 175         bool _ownEstimation;
 176
 177         /// Calculates the permutation that permutes the canonical ordering into the desired ordering
 178         /** \param varOrder Desired ordering of variables
 179          *  \param outVS Contains variables in \a varOrder represented as a VarSet
 180          *  \return Permute object for permuting variables in varOrder from the canonical libDAI ordering into the desired ordering
 181          */
 182         static Permute calculatePermutation( const std::vector<Var> &varOrder, VarSet &outVS );
 183
 184         /// Initializes _varsets and _perms from _varorders and checks whether their state spaces correspond with _estimation.probSize()
 185         void setPermsAndVarSetsFromVarOrders();
 186
 187     public:
 188         /// Constructor
 189         /** \param varorders  all the factor orientations for this parameter
 190          *  \param estimation a pointer to the parameter estimation method
 191          *  \param ownPE whether the constructed object gets ownership of \a estimation
 192          */
 193         SharedParameters( const FactorOrientations &varorders, ParameterEstimation *estimation, bool ownPE=false );
 194
 195         /// Construct a SharedParameters object from an input stream \a is and a factor graph \a fg
 196         /** \see \ref fileformats-emalg-sharedparameters
 197          *  \throw INVALID_EMALG_FILE if the input stream is not valid
 198          */
 199         SharedParameters( std::istream &is, const FactorGraph &fg );
 200
 201         /// Copy constructor
 202         SharedParameters( const SharedParameters &sp ) : _varsets(sp._varsets), _perms(sp._perms), _varorders(sp._varorders), _estimation(sp._estimation), _ownEstimation(sp._ownEstimation) {
 203             // If sp owns its _estimation object, we should clone it instead of copying the pointer
 204             if( _ownEstimation )
 205                 _estimation = _estimation->clone();
 206         }
 207
 208         /// Destructor
 209         ~SharedParameters() {
 210             // If we own the _estimation object, we should delete it now
 211             if( _ownEstimation )
 212                 delete _estimation;
 213         }
 214
 215         /// Collect the sufficient statistics from expected values (beliefs) according to \a alg
 216         /** For each of the relevant factors (that shares the parameters we are interested in),
 217          *  the corresponding belief according to \a alg is obtained and its entries are permuted
 218          *  such that their ordering corresponds with the shared parameters that we are estimating.
 219          *  Then, the parameter estimation subclass method addSufficientStatistics() is called with
 220          *  this vector of expected values of the parameters as input.
 221          */
 222         void collectSufficientStatistics( InfAlg &alg );
 223
 224         /// Estimate and set the shared parameters
 225         /** Based on the sufficient statistics collected so far, the shared parameters are estimated
 226          *  using the parameter estimation subclass method estimate(). Then, each of the relevant
 227          *  factors in \a fg (that shares the parameters we are interested in) is set according
 228          *  to those parameters (permuting the parameters accordingly).
 229          */
 230         void setParameters( FactorGraph &fg );
 231 };
 232
 233
 234 /// A MaximizationStep groups together several parameter estimation tasks (SharedParameters objects) into a single unit.
 235 /** \author Charles Vaske
 236  */
 237 class MaximizationStep {
 238     private:
 239         /// Vector of parameter estimation tasks of which this maximization step consists
 240         std::vector<SharedParameters> _params;
 241
 242     public:
 243         /// Default constructor
 244         MaximizationStep() : _params() {}
 245
 246         /// Construct MaximizationStep from a vector of parameter estimation tasks
 247         MaximizationStep( std::vector<SharedParameters> &maximizations ) : _params(maximizations) {}
 248
 249         /// Constructor from an input stream and a corresponding factor graph
 250         /** \see \ref fileformats-emalg-maximizationstep
 251          */
 252         MaximizationStep( std::istream &is, const FactorGraph &fg_varlookup );
 253
 254         /// Collect the beliefs from this InfAlg as expectations for the next Maximization step
 255         void addExpectations( InfAlg &alg );
 256
 257         /// Using all of the currently added expectations, make new factors with maximized parameters and set them in the FactorGraph.
 258         void maximize( FactorGraph &fg );
 259
 260     /// \name Iterator interface
 261     //@{
 262         /// Iterator over the parameter estimation tasks
 263         typedef std::vector<SharedParameters>::iterator iterator;
 264         /// Constant iterator over the parameter estimation tasks
 265         typedef std::vector<SharedParameters>::const_iterator const_iterator;
 266
 267         /// Returns iterator that points to the first parameter estimation task
 268         iterator begin() { return _params.begin(); }
 269         /// Returns constant iterator that points to the first parameter estimation task
 270         const_iterator begin() const { return _params.begin(); }
 271         /// Returns iterator that points beyond the last parameter estimation task
 272         iterator end() { return _params.end(); }
 273         /// Returns constant iterator that points beyond the last parameter estimation task
 274         const_iterator end() const { return _params.end(); }
 275     //@}
 276 };
 277
 278
 279 /// EMAlg performs Expectation Maximization to learn factor parameters.
 280 /** This requires specifying:
 281  *     - Evidence (instances of observations from the graphical model),
 282  *     - InfAlg for performing the E-step (which includes the factor graph),
 283  *     - a vector of MaximizationStep 's steps to be performed.
 284  *
 285  *  This implementation can perform incremental EM by using multiple
 286  *  MaximizationSteps. An expectation step is performed between execution
 287  *  of each MaximizationStep. A call to iterate() will cycle through all
 288  *  MaximizationStep 's. A call to run() will call iterate() until the
 289  *  termination criteria have been met.
 290  *
 291  *  Having multiple and separate maximization steps allows for maximizing some
 292  *  parameters, performing another E-step, and then maximizing separate
 293  *  parameters, which may result in faster convergence in some cases.
 294  *
 295  *  \author Charles Vaske
 296  */
 297 class EMAlg {
 298     private:
 299         /// All the data samples used during learning
 300         const Evidence &_evidence;
 301
 302         /// How to do the expectation step
 303         InfAlg &_estep;
 304
 305         /// The maximization steps to take
 306         std::vector<MaximizationStep> _msteps;
 307
 308         /// Number of iterations done
 309         size_t _iters;
 310
 311         /// History of likelihoods
 312         std::vector<Real> _lastLogZ;
 313
 314         /// Maximum number of iterations
 315         size_t _max_iters;
 316
 317         /// Convergence tolerance
 318         Real _log_z_tol;
 319
 320     public:
 321         /// Key for setting maximum iterations
 322         static const std::string MAX_ITERS_KEY;
 323         /// Default maximum iterations
 324         static const size_t MAX_ITERS_DEFAULT;
 325         /// Key for setting likelihood termination condition
 326         static const std::string LOG_Z_TOL_KEY;
 327         /// Default likelihood tolerance
 328         static const Real LOG_Z_TOL_DEFAULT;
 329
 330         /// Construct an EMAlg from several objects
 331         /** \param evidence Specifies the observed evidence
 332          *  \param estep Inference algorithm to be used for the E-step
 333          *  \param msteps Vector of maximization steps, each of which is a group of parameter estimation tasks
 334          *  \param termconditions Termination conditions @see setTermConditions()
 335          */
 336         EMAlg( const Evidence &evidence, InfAlg &estep, std::vector<MaximizationStep> &msteps, const PropertySet &termconditions )
 337           : _evidence(evidence), _estep(estep), _msteps(msteps), _iters(0), _lastLogZ(), _max_iters(MAX_ITERS_DEFAULT), _log_z_tol(LOG_Z_TOL_DEFAULT)
 338         {
 339               setTermConditions( termconditions );
 340         }
 341
 342         /// Construct an EMAlg from Evidence \a evidence, an InfAlg \a estep, and an input stream \a mstep_file
 343         /** \see \ref fileformats-emalg
 344          */
 345         EMAlg( const Evidence &evidence, InfAlg &estep, std::istream &mstep_file );
 346
 347         /// Change the conditions for termination
 348         /** There are two possible parameters in the PropertySet \a p:
 349          *    - \a max_iters maximum number of iterations
 350          *    - \a log_z_tol critical proportion of increase in logZ
 351          *
 352          *  \see hasSatisifiedTermConditions()
 353          */
 354         void setTermConditions( const PropertySet &p );
 355
 356         /// Determine if the termination conditions have been met.
 357         /** There are two sufficient termination conditions:
 358          *    -# the maximum number of iterations has been performed
 359          *    -# the ratio of logZ increase over previous logZ is less than the
 360          *       tolerance, i.e.,
 361          *       \f$ \frac{\log(Z_t) - \log(Z_{t-1})}{| \log(Z_{t-1}) | } < \mathrm{tol} \f$.
 362          */
 363         bool hasSatisfiedTermConditions() const;
 364
 365         /// Return the last calculated log likelihood
 366         Real logZ() const { return _lastLogZ.back(); }
 367
 368         /// Returns number of iterations done so far
 369         size_t Iterations() const { return _iters; }
 370
 371         /// Get the E-step method used
 372         const InfAlg& eStep() const { return _estep; }
 373
 374         /// Iterate once over all maximization steps
 375         /** \return Log-likelihood after iteration
 376          */
 377         Real iterate();
 378
 379         /// Iterate over a single MaximizationStep
 380         Real iterate( MaximizationStep &mstep );
 381
 382         /// Iterate until termination conditions are satisfied
 383         void run();
 384
 385     /// \name Iterator interface
 386     //@{
 387         /// Iterator over the maximization steps
 388         typedef std::vector<MaximizationStep>::iterator s_iterator;
 389         /// Constant iterator over the maximization steps
 390         typedef std::vector<MaximizationStep>::const_iterator const_s_iterator;
 391
 392         /// Returns iterator that points to the first maximization step
 393         s_iterator s_begin() { return _msteps.begin(); }
 394         /// Returns constant iterator that points to the first maximization step
 395         const_s_iterator s_begin() const { return _msteps.begin(); }
 396         /// Returns iterator that points beyond the last maximization step
 397         s_iterator s_end() { return _msteps.end(); }
 398         /// Returns constant iterator that points beyond the last maximization step
 399         const_s_iterator s_end() const { return _msteps.end(); }
 400     //@}
 401 };
 402
 403
 404 } // end of namespace dai
 405
 406
 407 #endif