cc.factorie.infer.Summary.scala Maven / Gradle / Ivy

/* Copyright (C) 2008-2016 University of Massachusetts Amherst.
   This file is part of "FACTORIE" (Factor graphs, Imperative, Extensible)
   http://factorie.cs.umass.edu, http://github.com/factorie
   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at
    http://www.apache.org/licenses/LICENSE-2.0
   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License. */

package cc.factorie.infer

import cc.factorie.la.{DenseTensor1, Tensor}
import cc.factorie.model.{DotFamily, Factor, Family}
import cc.factorie.variable._

/** The result of inference: a collection of Marginal objects.
    @author Andrew McCallum */
trait Summary {
  /** The collection of all univariate Marginals available in this Summary */
  def marginals: Iterable[Marginal1]
  /** All the variables for which this Summary has a univariate Marginal. */
  def variables: Iterable[Var] = marginals.map(_._1)
  /** If this Summary has a univariate Marginal for variable v, return it; otherwise return null. */
  def marginal(v:Var): Marginal1
  /** If this Summary has a Marginal that touches all or a subset of the neighbors of this factor
      return the Marginal with the maximally-available subset. */
  def marginal(factor:Factor): FactorMarginal
  def factorMarginals: Iterable[FactorMarginal]
  def logZ: Double
  /** If this summary has a univariate Marginal for variable v, return it in an Option; otherwise return None. */
  def getMarginal(v:Var): Option[Marginal1] = { val m = marginal(v); if (m eq null) None else Some(m) }
  def setToMaximize(implicit d:DiffList): Unit = marginals.foreach(_.setToMaximize(d)) // Note that order may matter here if Marginals are not Marginal1 and overlap with each other!
}
/*
trait IncrementableInfer[-A,-B] extends Infer[A,B] {
  override def infer(variables: A, model: B): Option[WarmStartSummary]
}

trait WarmStartSummary {
  def updateScores(v: DiscreteVar, t: Tensor): Unit
  def updateSummary(): Unit
}

case class ModelWithInference[M](model: M, variables: Seq[DiscreteVar], inference: IncrementableInfer[Seq[DiscreteVar], M])

object InferByDualDecomposition extends Infer[Seq[ModelWithInference[_]],Seq[(Any,Any,DiscreteVar,DiscreteVar)]] {
  def infer(variables: Seq[ModelWithInference[_]], model: Seq[(Any, Any, DiscreteVar, DiscreteVar)]) = null
}
*/

/** A Summary that can be used to gather weighted samples into its Marginals. */
// TODO Consider the relationship between this and Accumulator
// TODO Consider removing this
trait IncrementableSummary extends Summary {
  def incrementCurrentValues(weight:Double): Unit
}

/** A Summary that contains multiple Marginals of type M, each a marginal distribution over a single variable. */
class Summary1[V<:Var,M<:Marginal1] extends Summary {
  protected val _marginals = new scala.collection.mutable.HashMap[V,M]
  protected val _factorMarginals = new scala.collection.mutable.HashMap[Factor,FactorMarginal]
  def marginals = _marginals.values
  def factorMarginals = _factorMarginals.values
  def logZ = throw new Error("logZ is not defined")
  def marginal(v:Var): M = _marginals(v.asInstanceOf[V]) // We don't actually care that this type check does nothing because only Vs could be added to the HashMap
  def marginal(f:Factor): FactorMarginal = _factorMarginals(f)
  def +=(marginal:M) = {
    val vars = marginal.variables
    require(vars.size == 1)
    val v = vars.head.asInstanceOf[V]
    if (_marginals.contains(v)) throw new Error("Marginal already present for variable "+v)
    _marginals(v) = marginal
  }
  def +=(marginal:FactorMarginal) = {
    if (_factorMarginals.contains(marginal.factor)) throw new Error("Marginal already present for factor "+marginal.factor)
    _factorMarginals(marginal.factor) = marginal
  }
}

/** A Summary containing only one Marginal. */
class SingletonSummary[M<:Marginal1](val marginal:M) extends Summary {
  def marginals = Seq(marginal)
  def marginal(v:Var): M = if (Seq(v) == marginal.variables) marginal else null.asInstanceOf[M]
  def marginal(f:Factor) = null
  def logZ = throw new Error("logZ not definable for SingletonSummary")
  def factorMarginals = Nil
}



/** A Summary with all its probability on one variable-value Assignment. */
class AssignmentSummary(val assignment:Assignment) extends Summary {
  lazy val _marginals = assignment.variables.map(v=> v -> new Marginal1 {
    def _1 = v
    def setToMaximize(implicit d: DiffList) = v match { case vv:MutableVar => vv.set(assignment(vv)) }
  }).toMap
  def marginals = _marginals.values
  def marginal(v:Var): Marginal1 = _marginals.getOrElse(v, null)
  def marginal(f:Factor): FactorMarginal = null.asInstanceOf[FactorMarginal]
  override def setToMaximize(implicit d:DiffList): Unit = assignment.setVariables(d)
  def logZ = throw new Error("AssignmentSummary does not define logZ")
  def factorMarginals = Nil
}

/** A Summary with all its probability on one variable-value Assignment,
    and which can also be used for learning because it includes factors. */
class MAPSummary(val mapAssignment: Assignment, factors: Seq[Factor]) extends Summary {
  /** The collection of all Marginals available in this Summary */
  class SingletonMarginal(val _1: Var) extends Marginal1 {
    def setToMaximize(implicit d: DiffList) { _1 match { case v: MutableVar => v.set(mapAssignment(v)) } }
  }
  val marginals = mapAssignment.variables.map(new SingletonMarginal(_))
  def marginal(v: Var) = mapAssignment.get(v) match {
    case Some(_) => new SingletonMarginal(v)
    case None => null
  }
  trait FactorMarginalWithScore extends FactorMarginal { val score: Double }
  class SingletonFactorMarginal(val factor: Factor) extends FactorMarginalWithScore {
    val tensorStatistics = factor.assignmentStatistics(mapAssignment).asInstanceOf[Tensor]
    def variables = factor.variables
    val score = factor.assignmentScore(mapAssignment)
  }
  class NonMarginalFactorMarginal(val factor: Factor) extends FactorMarginalWithScore {
    val tensorStatistics = new DenseTensor1(1)
    def variables = factor.variables
    val score = factor.assignmentScore(mapAssignment)
  }
  def marginal(factor: Factor): FactorMarginalWithScore =
    if (factor.isInstanceOf[Family#Factor] && factor.asInstanceOf[Family#Factor].family.isInstanceOf[DotFamily])
     new SingletonFactorMarginal(factor)
    else new NonMarginalFactorMarginal(factor)
  def factorMarginals = factors.map(marginal)
  def logZ = factors.map(marginal(_).score).sum
}


/** A summary with a separate Proportions distribution for each of its DiscreteVars */
// TODO Consider renaming FullyFactorizedDiscreteSummary or IndependentDiscreteSummary or PerVariableDiscreteSummary
// TODO Consider making this inherit from Summary1
class DiscreteSummary1[V<:DiscreteVar] extends IncrementableSummary {
  def this(vs:Iterable[V]) = { this(); ++=(vs) }
  //val variableClass = m.erasure
  val _marginals1 = new scala.collection.mutable.HashMap[V,SimpleDiscreteMarginal1[V]]
  def marginals = _marginals1.values
  override def variables = _marginals1.keys
  lazy val variableSet = variables.toSet
  def marginal(v1:Var): SimpleDiscreteMarginal1[V] = _marginals1.getOrElse(v1.asInstanceOf[V], null)
  def marginal2(vs:Var*): DiscreteMarginal = vs match {
    case Seq(v:V @unchecked) => _marginals1(v) // Note, this doesn't actually check for a type match on V, because of erasure, but it shoudn't matter
    // case Seq(v:V @unchecked, w:V @unchecked) => new DiscreteMarginal2[V,V](v, w, new DenseTensorProportions2(new Outer1Tensor2(_marginals1(v).proportions,_marginals1(w).proportions), false))
    case _ => null
  }
  def marginal(f:Factor): FactorMarginal = null
  def +=(m:SimpleDiscreteMarginal1[V]): Unit = _marginals1(m._1) = m
  def +=(v:V): Unit = this += new SimpleDiscreteMarginal1(v, null) // but not yet initialized marginal proportions
  def ++=(vs:Iterable[V]): Unit = vs.foreach(+=(_))
  //def ++=(ms:Iterable[DiscreteMarginal1[V]]): Unit = ms.foreach(+=(_))
  def incrementCurrentValues(weight:Double): Unit = for (m <- marginals) m.incrementCurrentValue(weight)
  //def maximize(implicit d:DiffList): Unit = for (m <- marginals) m._1.asInstanceOf[DiscreteVariable].set(m.proportions.maxIndex)
  def factorMarginals = Nil
  def logZ = throw new Error("DiscreteSummary1 does not define logZ")
}