public override bool Equals(object obj)
{
MessageInitializer other = obj as MessageInitializer;
if (other == null ||
_hashCode != other._hashCode ||
_distribution.DependsOnMoreThanOneVariable != other._distribution.DependsOnMoreThanOneVariable ||
_distribution.ToString() != other._distribution.ToString()
)
{
return(false);
}
foreach (Leaf leaf in _fullLeafCollection)
{
if (IsMissing(leaf) != other.IsMissing(leaf) ||
LeafToTargetStatistics(leaf) != other.LeafToTargetStatistics(leaf))
{
return(false);
}
// if these distributions depend on the predictor variables, then make sure they all match up.
if (_distribution.DependsOnMoreThanOneVariable)
{
foreach (KeyValuePair <Converter <Leaf, SufficientStatistics>, Converter <Leaf, SufficientStatistics> > predMapPair in SpecialFunctions.EnumerateTwo(LeafToPredictorStatisticsList, other.LeafToPredictorStatisticsList))
{
if (predMapPair.Key(leaf) != predMapPair.Value(leaf))
{
return(false);
}
}
}
}
return(true);
}
public double ComputeLogLikelihoodModelGivenData(MessageInitializer messageInitializer, OptimizationParameterList paramList)
{
double loglikelihood = ComputeLogLikelihoodModelGivenData(messageInitializer, paramList, false);
if (double.IsNegativeInfinity(loglikelihood))
{
loglikelihood = ComputeLogLikelihoodModelGivenData(messageInitializer, paramList, true);
}
return(loglikelihood);
}
/// <summary>
/// Learns the optimal parameters for the data contained in the messageInitializer and returns the corresponding Score.
/// </summary>
public Score MaximizeLikelihood(MessageInitializer messageInitializer)
{
Score score;
#region Caching details
//Key aKey = Key.GetInstance(Tree, messageInitializer);
if (_cache.ContainsKey(messageInitializer))
{
CacheHits++;
score = _cache[messageInitializer];
#if (DEBUG)
Score scoreLive = MaximizeLikelihoodInternal(messageInitializer);
if (Math.Abs(score.Loglikelihood - scoreLive.Loglikelihood) >= 10e-7)
{
double diff = scoreLive.Loglikelihood - score.Loglikelihood;
Debug.WriteLine("Cache differs from computed score by " + diff);
}
// note: minute (10E-14) differences sometimes arise. The original explanation was that these were differences in rounding
// errors caused when missing data was caught in different places. I have tried to localize the catch of missing data and
// throw errors elsewhere but still have the same rounding errors. Not sure what else the cause could be.
SpecialFunctions.CheckCondition(//score.Loglikelihood == scoreLive.Loglikelihood,
Math.Abs(score.Loglikelihood - scoreLive.Loglikelihood) < 10e-7,
"Cached score " + score.Loglikelihood + " doesn't match live score " + scoreLive.Loglikelihood);
#endif
}
else
{
CacheMisses++;
score = MaximizeLikelihoodInternal(messageInitializer);
if (_cache.Count > MAX_CACHE_SIZE)
{
//_cache.Clear();
ClearCache();
}
_cache.Add(messageInitializer, score);
}
#endregion
return(score);
}
protected override double ComputeLLR(ModelScorer modelScorer, PhyloTree phyloTree, StringBuilder stringBuilder, double targetMarginal, double predictorMarginal,
Converter <Leaf, SufficientStatistics> predictorDistributionClassFunction, Converter <Leaf, SufficientStatistics> targetDistributionClassFunction)
{
NullModelDistribution.EmpiricalEquilibrium = targetMarginal;
NullModelDistribution.InitialParamVals = null;
MessageInitializer messageInitializer = modelScorer.CreateMessageInitializer(predictorDistributionClassFunction, targetDistributionClassFunction, NullModelDistribution);
List <double> logLikelihoodList = new List <double>();
foreach (bool useParameter in new bool[] { false, true })
{
Score score = modelScorer.ScoreModel(messageInitializer, useParameter);
stringBuilder.Append(SpecialFunctions.CreateTabString(score.ToString(useParameter ? AlternativeModelDistribution : NullModelDistribution), ""));
logLikelihoodList.Add(score.Loglikelihood);
AltModelDistribution.InitialParamVals = score.OptimizationParameters;
Debug.WriteLine(SpecialFunctions.CreateTabString("AltModelDistribution.InitialParamVals = score.OptimizationParameters", score.OptimizationParameters));
}
double diff = logLikelihoodList[1] - logLikelihoodList[0];
return(diff);
}
public override Dictionary <string, SufficientStatistics> GenerateRandomMapping(Dictionary <string, SufficientStatistics> realCaseIdToNonMissingValue, ref Random random)
{
//!!!!put check in to make sure ISufficientSTatistics is reall BooleanStatistics
Converter <Leaf, SufficientStatistics> leafToDistnClassFunction = PhyloDDriver.CreateSufficientStatisticsMap(realCaseIdToNonMissingValue);
PhyloTree tree = _modelScorer.PhyloTree;
MessageInitializer messageInitializer = MessageInitializerDiscrete.GetInstance(leafToDistnClassFunction, _discreteDistribution, new int[] { 1, 1 }, tree.LeafCollection);
Score score = _modelScorer.MaximizeLikelihood(messageInitializer);
double percentNonMissing = (double)tree.CountOfNonMissingLeaves(realCaseIdToNonMissingValue) /
(double)SpecialFunctions.Count(tree.LeafCollection);
double equilibrium = score.OptimizationParameters[(int)DistributionDiscreteConditional.ParameterIndex.Equilibrium].Value;
double lambda = score.OptimizationParameters[(int)DistributionDiscreteConditional.ParameterIndex.Lambda].Value;
Dictionary <string, BooleanStatistics> randomCaseIdToNonMissingValue = tree.EvolveBinaryTree(equilibrium, lambda, 1 - percentNonMissing, ref random);
Dictionary <string, SufficientStatistics> converted;
SpecialFunctions.ConvertDictionaryToBaseClasses(randomCaseIdToNonMissingValue, out converted);
return(converted);
}
protected Score MaximizeLikelihoodInternal(MessageInitializer messageInitializer)
{
OptimizationParameterList paramsToOptimize = messageInitializer.GetOptimizationParameters();
int functionEvaluationCount = 0;
bool useLogMethod = false;
Converter <OptimizationParameterList, double> functionToOptimize =
delegate(OptimizationParameterList paramList)
{
FuncCalls++;
++functionEvaluationCount;
//Debug.WriteLine("EvalCount " + functionEvaluationCount.ToString());
double loglikelihood = ComputeLogLikelihoodModelGivenData(messageInitializer, paramList, useLogMethod);
if (!useLogMethod && double.IsNegativeInfinity(loglikelihood))
{
useLogMethod = true;
loglikelihood = ComputeLogLikelihoodModelGivenData(messageInitializer, paramList, useLogMethod);
}
//SpecialFunctions.CheckCondition(!double.IsNaN(loglikelihood), "for debugging: got a NaN from ComputeLogLikelihoodModelGivenData");
//if (double.IsNaN(loglikelihood))
//{
// return double.NegativeInfinity;
//}
return(loglikelihood);
};
double loglikelihoodExternal = GridSearch.Optimize(functionToOptimize, paramsToOptimize, 10, 5);
Score score = Score.GetInstance(loglikelihoodExternal, paramsToOptimize, messageInitializer.PropogationDistribution);
Debug.WriteLine(SpecialFunctions.CreateTabString(GridSearch.DebugCount, score, functionEvaluationCount));
return(score);
}
public override double ComputeLogLikelihoodModelGivenData(MessageInitializer messageInitializer, OptimizationParameterList paramList, bool useLogMethod)
{
return(PhyloTree.ComputeLogLikelihoodModelGivenDataGaussian(messageInitializer, paramList));
}
public abstract double ComputeLogLikelihoodModelGivenData(MessageInitializer messageInitializer, OptimizationParameterList paramList, bool useLogMethod);
protected override string CreateReportLine(
ModelScorer modelScorer,
PhyloTree phyloTree,
RowData rowAndTargetData,
UniversalWorkList workList,
int rowIndex, int workListCount, int workIndex)
{
//!!!there is very similar code in ModelTesterDiscrete.cs
Dictionary <string, string> row = rowAndTargetData.Row;
string predictorVariable = row[PhyloTree.PredictorVariableColumnName];
string targetVariable = row[PhyloTree.TargetVariableColumnName]; // e.g. A@182 (amino acid "A" at position 182)
int nullIndex = int.Parse(row[PhyloTree.NullIndexColumnName]);
//Dictionary<string, bool> caseIdToNonNullPredictorValue = workList.NullIndexToPredictorToCaseIdToNonMissingValue[nullIndex][predictorVariable];
Dictionary <string, SufficientStatistics> caseIdToNonNullPredictorValue = rowAndTargetData.PredictorData; //workList.GetCaseIdToNonMissingValueForNullIndexAndPredictorVariable(nullIndex, predictorVariable);
Dictionary <string, SufficientStatistics> caseIdToNonMissingTargetValue = rowAndTargetData.TargetData;
Converter <Leaf, SufficientStatistics> targetDistributionMap = CreateSufficientStatisticsMap(caseIdToNonMissingTargetValue);
Converter <Leaf, SufficientStatistics> predictorDistributionClassFunction = CreateSufficientStatisticsMap(caseIdToNonNullPredictorValue);
int[] predictorCounts = phyloTree.CountsOfLeaves(predictorDistributionClassFunction);
int predictorFalseNameCount = predictorCounts[(int)DistributionDiscreteBinary.DistributionClass.False];
int predictorTrueNameCount = predictorCounts[(int)DistributionDiscreteBinary.DistributionClass.True];
int targetNonMissingCount = phyloTree.CountOfNonMissingLeaves(caseIdToNonMissingTargetValue);
int globalNonMissingCount = phyloTree.GlobalNonMissingCount(predictorDistributionClassFunction, targetDistributionMap);
StringBuilder stringBuilder = new StringBuilder(
SpecialFunctions.CreateTabString(
this, rowIndex, workListCount, workIndex, nullIndex, predictorVariable,
predictorFalseNameCount,
predictorTrueNameCount,
predictorTrueNameCount + predictorFalseNameCount,
targetVariable,
targetNonMissingCount,
globalNonMissingCount,
""));
bool ignoreRow = false;
foreach (int count in predictorCounts)
{
if (count == 0)
{
ignoreRow = true;
}
}
if (ignoreRow)
{
CompleteRowWithNaN(stringBuilder);
}
else
{
List <double> logLikelihoodList = new List <double>();
MessageInitializer messageInitializer =
modelScorer.CreateMessageInitializer(predictorDistributionClassFunction, targetDistributionMap, NullModelDistribution);
NullModelDistribution.InitialParamVals = null;
foreach (bool useParameter in new bool[] { false, true })
{
Score score = modelScorer.ScoreModel(messageInitializer, useParameter);
stringBuilder.Append(SpecialFunctions.CreateTabString(score, ""));
Debug.Write(SpecialFunctions.CreateTabString(score, ""));
logLikelihoodList.Add(score.Loglikelihood);
AltModelDistribution.InitialParamVals = score.OptimizationParameters;
}
double diff = logLikelihoodList[1] - logLikelihoodList[0];
double pValue = SpecialFunctions.LogLikelihoodRatioTest(Math.Max(diff, 0), ChiSquareDegreesOfFreedom);
stringBuilder.Append(SpecialFunctions.CreateTabString(diff, pValue));
Debug.WriteLine(SpecialFunctions.CreateTabString(diff, pValue));
}
return(stringBuilder.ToString());
}
