Added logProbability/evaluate in conditionals/Bayes net

gtsam/inference/BayesNet-inst.h

@@ -88,6 +88,22 @@ void BayesNet<CONDITIONAL>::saveGraph(const std::string& filename,
   of.close();
 }
 
+/* ************************************************************************* */
+template <class CONDITIONAL>
+double BayesNet<CONDITIONAL>::logProbability(const HybridValues& x) const {
+  double sum = 0.;
+  for (const auto& gc : *this) {
+    if (gc) sum += gc->logProbability(x);
+  }
+  return sum;
+}
+
+/* ************************************************************************* */
+template <class CONDITIONAL>
+double BayesNet<CONDITIONAL>::evaluate(const HybridValues& x) const {
+  return exp(-logProbability(x));
+}
+
 /* ************************************************************************* */
 
 }  // namespace gtsam

gtsam/inference/BayesNet.h

@@ -25,6 +25,8 @@
 
 namespace gtsam {
 
+class HybridValues;
+
 /**
  * A BayesNet is a tree of conditionals, stored in elimination order.
  * @ingroup inference
@@ -68,7 +70,6 @@ class BayesNet : public FactorGraph<CONDITIONAL> {
       const KeyFormatter& formatter = DefaultKeyFormatter) const override;
 
   /// @}
-
   /// @name Graph Display
   /// @{
 
@@ -86,6 +87,13 @@ class BayesNet : public FactorGraph<CONDITIONAL> {
                  const KeyFormatter& keyFormatter = DefaultKeyFormatter,
                  const DotWriter& writer = DotWriter()) const;
 
+  /// @}
+  /// @name HybridValues methods
+  /// @{
+
+  double logProbability(const HybridValues& x) const;
+  double evaluate(const HybridValues& c) const;
+
   /// @}
 };
 

gtsam/inference/Conditional-inst.h

@@ -18,30 +18,42 @@
 // \callgraph
 #pragma once
 
-#include <iostream>
-
 #include <gtsam/inference/Conditional.h>
 
+#include <cmath>
+#include <iostream>
+
 namespace gtsam {
 
 /* ************************************************************************* */
-  template<class FACTOR, class DERIVEDFACTOR>
+template <class FACTOR, class DERIVEDCONDITIONAL>
-  void Conditional<FACTOR,DERIVEDFACTOR>::print(const std::string& s, const KeyFormatter& formatter) const {
+void Conditional<FACTOR, DERIVEDCONDITIONAL>::print(
+    const std::string& s, const KeyFormatter& formatter) const {
   std::cout << s << " P(";
-    for(Key key: frontals())
+  for (Key key : frontals()) std::cout << " " << formatter(key);
-      std::cout << " " << formatter(key);
+  if (nrParents() > 0) std::cout << " |";
-    if (nrParents() > 0)
+  for (Key parent : parents()) std::cout << " " << formatter(parent);
-      std::cout << " |";
-    for(Key parent: parents())
-      std::cout << " " << formatter(parent);
   std::cout << ")" << std::endl;
 }
 
 /* ************************************************************************* */
-  template<class FACTOR, class DERIVEDFACTOR>
+template <class FACTOR, class DERIVEDCONDITIONAL>
-  bool Conditional<FACTOR,DERIVEDFACTOR>::equals(const This& c, double tol) const
+bool Conditional<FACTOR, DERIVEDCONDITIONAL>::equals(const This& c,
-  {
+                                                     double tol) const {
   return nrFrontals_ == c.nrFrontals_;
 }
 
+/* ************************************************************************* */
+template <class FACTOR, class DERIVEDCONDITIONAL>
+double Conditional<FACTOR, DERIVEDCONDITIONAL>::logProbability(
+    const HybridValues& c) const {
+  throw std::runtime_error("Conditional::logProbability is not implemented");
 }
+
+/* ************************************************************************* */
+template <class FACTOR, class DERIVEDCONDITIONAL>
+double Conditional<FACTOR, DERIVEDCONDITIONAL>::evaluate(
+    const HybridValues& c) const {
+  return exp(static_cast<const DERIVEDCONDITIONAL*>(this)->logProbability(c));
+}
+}  // namespace gtsam

gtsam/inference/Conditional.h

@@ -24,13 +24,37 @@
 
 namespace gtsam {
 
+  class HybridValues;  // forward declaration.
+
   /**
-   * Base class for conditional densities.  This class iterators and
+   * This is the base class for all conditional distributions/densities, 
-   * access to the frontal and separator keys.
+   * which are implemented as specialized factors.  This class does not store any
+   * data other than its keys.  Derived classes store data such as matrices and
+   * probability tables.
+   *
+   * The `evaluate` method is used to evaluate the factor, and together with 
+   * `logProbability` is the main methods that need to be implemented in derived
+   * classes. These two methods relate to the `error` method in the factor by:
+   *   probability(x) = k exp(-error(x))
+   * where k is a normalization constant making \int probability(x) == 1.0, and
+   *   logProbability(x) = K - error(x)
+   * i.e., K = log(K).
+   * 
+   * There are four broad classes of conditionals that derive from Conditional:
+   *
+   * - \b Gaussian conditionals, implemented in \class GaussianConditional, a 
+   *   Gaussian density over a set of continuous variables.
+   * - \b Discrete conditionals, implemented in \class DiscreteConditional, which
+   *   represent a discrete conditional distribution over discrete variables.
+   * - \b Hybrid conditional densities, such as \class GaussianMixture, which is 
+   *   a density over continuous variables given discrete/continuous parents.
+   * - \b Symbolic factors, used to represent a graph structure, implemented in 
+   *   \class SymbolicConditional. Only used for symbolic elimination etc.
    *
    * Derived classes *must* redefine the Factor and shared_ptr typedefs to refer
    * to the associated factor type and shared_ptr type of the derived class.  See
    * SymbolicConditional and GaussianConditional for examples.
+   * 
    * \nosubgrouping
    */
   template<class FACTOR, class DERIVEDCONDITIONAL>
@@ -78,6 +102,8 @@ namespace gtsam {
     /// @name Standard Interface
     /// @{
 
+    virtual ~Conditional() {}
+
     /** return the number of frontals */
     size_t nrFrontals() const { return nrFrontals_; }
 
@@ -98,6 +124,27 @@ namespace gtsam {
     /** return a view of the parent keys */
     Parents parents() const { return boost::make_iterator_range(beginParents(), endParents()); }
 
+    /**
+     * All conditional types need to implement a `logProbability` function, for which
+     *   exp(logProbability(x)) = evaluate(x).
+     */
+    virtual double logProbability(const HybridValues& c) const;
+
+    /**
+     * All conditional types need to implement an `evaluate` function, that yields
+     * a true probability. The default implementation just exponentiates logProbability.
+     */
+    virtual double evaluate(const HybridValues& c) const;
+
+    /// Evaluate probability density, sugar.
+    double operator()(const HybridValues& x) const {
+      return evaluate(x);
+    }
+
+    /// @}
+    /// @name Advanced Interface
+    /// @{
+
     /** Iterator pointing to first frontal key. */
     typename FACTOR::const_iterator beginFrontals() const { return asFactor().begin(); }
 
@@ -110,10 +157,6 @@ namespace gtsam {
     /** Iterator pointing past the last parent key. */
     typename FACTOR::const_iterator endParents() const { return asFactor().end(); }
 
-    /// @}
-    /// @name Advanced Interface
-    /// @{
-
     /** Mutable version of nrFrontals */
     size_t& nrFrontals() { return nrFrontals_; }
 

gtsam/inference/Factor.cpp

@@ -43,4 +43,10 @@ namespace gtsam {
     return keys_ == other.keys_;
   }
 
+  /* ************************************************************************* */
+  double Factor::error(const HybridValues& c) const {
+    throw std::runtime_error("Factor::error is not implemented");
+  }
+
+
 }

gtsam/inference/Factor.h

@@ -44,10 +44,7 @@ namespace gtsam {
    *
    * The `error` method is used to evaluate the factor, and is the only method
    * that is required to be implemented in derived classes, although it has a 
-   * default implementation that throws an exception. The meaning of the error
+   * default implementation that throws an exception.
-   * is slightly different for factors and conditionals: in the former it is the
-   * negative log-likelihood, and in the latter it is the negative log of the 
-   * properly normalized conditional distribution or density.
    * 
    * There are five broad classes of factors that derive from Factor:
    *
@@ -55,15 +52,12 @@ namespace gtsam {
    *   represent a nonlinear likelihood function over a set of variables.
    * - \b Gaussian factors, such as \class JacobianFactor and \class HessianFactor, which
    *   represent a Gaussian likelihood over a set of variables.
-   *   A \class GaussianConditional, which represent a Gaussian density over a set of
+   * - \b Discrete factors, such as \class DiscreteFactor and \class DecisionTreeFactor, which
-   *   variables conditioned on another set of variables.
-   * - \b Discrete factors, such as \class DiscreteFactor and \class DiscreteConditional, which
    *   represent a discrete distribution over a set of variables.
    * - \b Hybrid factors, such as \class HybridFactor, which represent a mixture of
    *   Gaussian and discrete distributions over a set of variables.
    * - \b Symbolic factors, used to represent a graph structure, such as
-   *   \class SymbolicFactor and \class SymbolicConditional. They do not override the
+   *   \class SymbolicFactor, only used for symbolic elimination etc.
-   *  `error` method, and are used only for symbolic elimination etc.
    *
    * Note that derived classes must also redefine the `This` and `shared_ptr`
    * typedefs. See JacobianFactor, etc. for examples.
@@ -154,11 +148,8 @@ namespace gtsam {
   /**
    * All factor types need to implement an error function.
    * In factor graphs, this is the negative log-likelihood.
-   * In Bayes nets, it is the negative log density, i.e., properly normalized.
    */
-  virtual double error(const HybridValues& c) const {
+  virtual double error(const HybridValues& c) const;
-    throw std::runtime_error("Factor::error is not implemented");
-  }
 
    /**
     * @return the number of variables involved in this factor