//public int[] Classification(Vector[] features, int[] labels)
//{
//}
public double[] Regression(Vector[] features, double[] values)
{
var wMeans = Variable.Vector(Vector.Zero(features[0].Count).ToArray());
var wPrecision = Variable.WishartFromShapeAndRate(100, PositiveDefiniteMatrix.IdentityScaledBy(features[0].Count, 0.01));
var w = Variable.VectorGaussianFromMeanAndPrecision(wMeans, wPrecision).Named("w");
var numItems = Variable.New<int>().Named("numItems");
var i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
var noisePrecision = Variable.New<double>().Named("noisePrecision");
var x = Variable.Array<Vector>(i).Named("x");
var y = Variable.Array<double>(i).Named("y");
using (Variable.ForEach(i))
{
y[i] = Variable.GaussianFromMeanAndPrecision(Variable.InnerProduct(w, x[i]), noisePrecision);
}
numItems.ObservedValue = features.Length;
x.ObservedValue = features;
y.ObservedValue = values;
var engine = new InferenceEngine();
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
var wPosterior = engine.Infer<VectorGaussian>(w);
y.ClearObservedValue();
w.ObservedValue = wPosterior.GetMean();
var inferredValues = engine.Infer<IList<Gaussian>>(y);
return inferredValues.Select(v => v.GetMean()).ToArray();
}
/// <summary>
/// Constructs the model.
/// </summary>
public override void ConstructModel()
{
if (this.HasBeenConstructed)
{
return;
}
this.NumberOfFeatures = Variable.New <int>().Named("numberOfFeatures").Attrib(new DoNotInfer());
this.NumberOfMessages = Variable.New <int>().Named("numberOfMessages").Attrib(new DoNotInfer());
Range features = new Range(this.NumberOfFeatures).Named("features");
Range emails = new Range(this.NumberOfMessages).Named("emails");
// Make sure that the range across messages is handled sequentially
// - this is necessary to ensure the model converges during training
emails.AddAttribute(new Sequential());
// observed data
this.FeatureValue =
Variable.Array(Variable.Array <double>(features), emails).Named("featureValue").Attrib(new DoNotInfer());
// The weights
this.WeightPriors = Variable.New <DistributionStructArray <Gaussian, double> >().Named("WeightPriors").Attrib(new DoNotInfer());
this.Weight = Variable.Array <double>(features).Named("weight");
this.Weight.SetTo(Variable <double[]> .Random(this.WeightPriors));
this.ThresholdPrior = Variable.New <Gaussian>().Named("ThresholdPrior").Attrib(new DoNotInfer());
this.Threshold = Variable.New <double>().Named("threshold");
this.Threshold.SetTo(Variable <double> .Random(this.ThresholdPrior));
// Noise Variance
this.NoiseVariance = Variable.New <double>().Named("NoiseVariance").Attrib(new DoNotInfer());
// Label: is the message replied to?
this.RepliedTo = Variable.Array <bool>(emails).Named("repliedTo");
// Loop over emails
using (Variable.ForEach(emails))
{
var featureScore = Variable.Array <double>(features).Named("featureScore");
featureScore[features] = this.FeatureValue[emails][features] * this.Weight[features];
var score = Variable.Sum(featureScore).Named("score");
this.RepliedTo[emails] = Variable.GaussianFromMeanAndVariance(score, this.NoiseVariance).Named("noisyScore") > this.Threshold;
}
this.InitializeEngine();
this.Engine.OptimiseForVariables = this.Mode == InputMode.Training
? new IVariable[] { this.Weight, this.Threshold }
: this.Engine.OptimiseForVariables = new IVariable[] { this.RepliedTo };
this.HasBeenConstructed = true;
}
public Model()
{
// Classes.
numClasses = Variable.New <int>().Named("numClasses");
c = new Range(numClasses).Named("c");
// Items.
numItems = Variable.New <int>().Named("numItems");
i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
// The prior distribution for weight vector for each class. When
// <see cref="Test"/> is called, this is set to the posterior weight
// distributions from <see cref="Train"/>.
wPrior = Variable.Array <VectorGaussian>(c).Named("wPrior");
// The weight vector for each class.
w = Variable.Array <Vector>(c).Named("w");
w[c] = Variable <Vector> .Random(wPrior[c]);
noisePrecision = Variable.New <double>().Named("noisePrecision");
// Arrays of <see cref="Vector"/>-valued items (feature vectors) and integer labels.
x = Variable.Array <Vector>(i).Named("x");
y = Variable.Array <int>(i).Named("y");
// For all items...
using (Variable.ForEach(i))
{
// ...compute the score of this item across all classes...
score = BPMUtils.ComputeClassScores(w, x[i], noisePrecision);
y[i] = Variable.DiscreteUniform(c);
// ... and constrain the output.
BPMUtils.ConstrainMaximum(y[i], score);
}
// Inference engine settings (EP).
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
}
public void IndexOfMaximumDiverges()
{
int numberOfFeatures = 10;
int numberOfFolders = 3;
int numEmails = 1000;
var featuresCountsObs = Enumerable.Range(0, numEmails).Select(o => Rand.Binomial(numberOfFeatures, 5.0 / numberOfFeatures)).ToArray();
var featureIndicesObs = featuresCountsObs.Select(o => Rand.Perm(numberOfFeatures).ToList().GetRange(0, o).ToArray()).ToArray();
//var trueFeatureWeights = Enumerable.Range(0, numberOfFolders).Select(p=>Enumerable.Range(0, numberOfFeatures).Select(o => Rand.Normal()).ToArray()).ToArray();
//var folders = featureIndicesObs.Select(fi=>fi.Select(p=>trueFeatureWeights.Select(q=>q[p]).Sum()
// random data for now!
var folders = Enumerable.Range(0, numEmails).Select(o => Rand.Int(numberOfFolders)).ToArray();
Range numberOfFeaturesRange = new Range(numberOfFeatures).Named("NumberOfFeaturesRange");
numberOfFeaturesRange.AddAttribute(new Sequential()); // This requires new build of Infer.NET
// creat a range for the number of classes
Range numberOfClases = new Range(numberOfFolders).Named("NumberOfClassesRange");
// Model the total number of items
var numberOfItems = Variable.New <int>().Named("numberOfItems");
numberOfItems.ObservedValue = numEmails;
Range numberOfItemsRange = new Range(numberOfItems).Named("numberOfItemsRange");
numberOfItemsRange.AddAttribute(new Sequential());
// Model the number features present in each item in each class
var featureCounts = Variable.Array <int>(numberOfItemsRange).Named("featureCounts");
Range featureCountItemRange = new Range(featureCounts[numberOfItemsRange]).Named("featureItemCountRange");
featureCounts.ObservedValue = featuresCountsObs;
// Model the features we observe
var featureIndicies = Variable.Array(Variable.Array <int>(featureCountItemRange), numberOfItemsRange).Named("featureIndicies");
featureIndicies.ObservedValue = featureIndicesObs;
// Setup the priors
var FeatureWeights = Variable.Array(Variable.Array <double>(numberOfFeaturesRange), numberOfClases).Named("FeatureWeights");
FeatureWeights[numberOfClases][numberOfFeaturesRange] = Variable.GaussianFromMeanAndPrecision(0, 1).ForEach(numberOfClases, numberOfFeaturesRange);
// Setup the label value (Folder)
var folderValue = Variable.Array <int>(numberOfItemsRange).Named("folderValue");
folderValue.ObservedValue = folders;
var sparseWeightVector =
Variable.Array(Variable.Array(Variable.Array <double>(featureCountItemRange), numberOfClases), numberOfItemsRange).Named("sparseWeightVector");
;
sparseWeightVector[numberOfItemsRange][numberOfClases] = Variable.Subarray <double>(FeatureWeights[numberOfClases], featureIndicies[numberOfItemsRange]);
var scoresWithNoise = Variable.Array(Variable.Array <double>(numberOfClases), numberOfItemsRange).Named("scoresWithNoise");
scoresWithNoise[numberOfItemsRange][numberOfClases] = Variable.GaussianFromMeanAndVariance(Variable.Sum(sparseWeightVector[numberOfItemsRange][numberOfClases]), 1);
folderValue[numberOfItemsRange] = Variable <int> .Factor(MMath.IndexOfMaximumDouble, scoresWithNoise[numberOfItemsRange]);
folderValue.AddAttribute(new MarginalPrototype(Discrete.Uniform(numberOfClases.SizeAsInt)));
var ie = new InferenceEngine();
Console.WriteLine(ie.Infer(FeatureWeights));
}
private void LearnAPIClick5LabelModel(
int numLabels,
bool learnScoreMean,
bool learnScorePrec,
bool learnJudgePrec,
bool learnClickPrec,
bool learnThresholds,
double nominalScoreMean,
double nominalScorePrec,
double nominalJudgePrec,
double nominalClickPrec,
int[] labels,
int[] clicks,
int[] exams,
int chunkSize,
int nPasses,
bool printToConsole,
out Gaussian margScoreMean,
out Gamma margScorePrec,
out Gamma margJudgePrec,
out Gamma margClickPrec,
out Gaussian[] margThresh)
{
//------------------------------------------------------
// Observations
//------------------------------------------------------
Gaussian[][][] allObs = getClickObservations(numLabels, chunkSize, labels, clicks, exams);
int numChunks = allObs.Length;
////-------------------------------------------------------------
//// Prior distributions
////-------------------------------------------------------------
Gaussian priorScoreMean = Gaussian.FromMeanAndVariance(nominalScoreMean, learnScoreMean ? 1 : 0);
Gamma priorScorePrec = Gamma.FromMeanAndVariance(nominalScorePrec, learnScorePrec ? 1 : 0);
Gamma priorJudgePrec = Gamma.FromMeanAndVariance(nominalJudgePrec, learnJudgePrec ? 1 : 0);
Gamma priorClickPrec = Gamma.FromMeanAndVariance(nominalClickPrec, learnClickPrec ? 1 : 0);
Gaussian[] priorThreshMean;
CalculatePriors(learnThresholds, numLabels, out priorThreshMean);
////-----------------------------------------------------
//// Creates shared variables
////-----------------------------------------------------
int numThresholds = numLabels + 1;
SharedVariable <double> scoreMean = SharedVariable <double> .Random(priorScoreMean).Named("scoreMean");
SharedVariable <double> scorePrec = SharedVariable <double> .Random(priorScorePrec).Named("scorePrec");
SharedVariable <double> judgePrec = SharedVariable <double> .Random(priorJudgePrec).Named("judgePrec");
SharedVariable <double> clickPrec = SharedVariable <double> .Random(priorClickPrec).Named("clickPrec");
SharedVariable <double>[] thresholds = new SharedVariable <double> [numThresholds];
for (int t = 0; t < numThresholds; t++)
{
thresholds[t] = SharedVariable <double> .Random(priorThreshMean[t]).Named("threshMeans" + t);
}
//----------------------------------------------------------------------------------
// The model
//----------------------------------------------------------------------------------
Model model = new Model(numChunks);
VariableArray <Gaussian>[] clickObs = new VariableArray <Gaussian> [numLabels];
Variable <int>[] clickObsLength = new Variable <int> [numLabels];
for (int i = 0; i < numLabels; i++)
{
clickObsLength[i] = Variable.New <int>().Named("clickObsLength" + i);
Range r = new Range(clickObsLength[i]).Named("dataCount" + i);
clickObs[i] = Variable.Array <Gaussian>(r).Named("Obs" + i);
VariableArray <double> scores = Variable.Array <double>(r).Named("scores" + i);
VariableArray <double> scoresJ = Variable.Array <double>(r).Named("scoresJ" + i);
VariableArray <double> scoresC = Variable.Array <double>(r).Named("scoresC" + i);
scores[r] = Variable <double> .GaussianFromMeanAndPrecision(scoreMean.GetCopyFor(model), scorePrec.GetCopyFor(model)).ForEach(r);
scoresJ[r] = Variable <double> .GaussianFromMeanAndPrecision(scores[r], judgePrec.GetCopyFor(model));
scoresC[r] = Variable <double> .GaussianFromMeanAndPrecision(scores[r], clickPrec.GetCopyFor(model));
Variable.ConstrainBetween(scoresJ[r], thresholds[i].GetCopyFor(model), thresholds[i + 1].GetCopyFor(model));
Variable.ConstrainEqualRandom <double, Gaussian>(scoresC[r], clickObs[i][r]);
r.AddAttribute(new Sequential());
}
InferenceEngine engine = new InferenceEngine();
//----------------------------------------------------------
// Outer loop iterates over a number of passes
// Inner loop iterates over the unique labels
//----------------------------------------------------------
for (int pass = 0; pass < nPasses; pass++)
{
for (int c = 0; c < numChunks; c++)
{
for (int i = 0; i < numLabels; i++)
{
clickObsLength[i].ObservedValue = allObs[c][i].Length;
clickObs[i].ObservedValue = allObs[c][i];
}
// Infer the output messages
model.InferShared(engine, c);
if (printToConsole)
{
margScoreMean = scoreMean.Marginal <Gaussian>();
margScorePrec = scorePrec.Marginal <Gamma>();
margJudgePrec = judgePrec.Marginal <Gamma>();
margClickPrec = clickPrec.Marginal <Gamma>();
margThresh = new Gaussian[numThresholds];
for (int i = 0; i < numThresholds; i++)
{
margThresh[i] = thresholds[i].Marginal <Gaussian>();
}
Console.WriteLine("****** Pass {0}, chunk {1} ******", pass, c);
Console.WriteLine("----- Marginals -----");
Console.WriteLine("scoreMean = {0}", margScoreMean);
Console.WriteLine("scorePrec = {0}", margScorePrec);
Console.WriteLine("judgePrec = {0}", margJudgePrec);
Console.WriteLine("clickPrec = {0}", margClickPrec);
for (int t = 0; t < numThresholds; t++)
{
Console.WriteLine("threshMean {0} = {1}", t, margThresh[t]);
}
}
}
}
margScoreMean = scoreMean.Marginal <Gaussian>();
margScorePrec = scorePrec.Marginal <Gamma>();
margJudgePrec = judgePrec.Marginal <Gamma>();
margClickPrec = clickPrec.Marginal <Gamma>();
margThresh = new Gaussian[numThresholds];
for (int i = 0; i < numThresholds; i++)
{
margThresh[i] = thresholds[i].Marginal <Gaussian>();
}
}
public Model()
{
// Classes.
numClasses = Variable.New<int>().Named("numClasses");
c = new Range(numClasses).Named("c");
// Items.
numItems = Variable.New<int>().Named("numItems");
i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
// The prior distribution for weight vector for each class. When
// <see cref="Test"/> is called, this is set to the posterior weight
// distributions from <see cref="Train"/>.
wPrior = Variable.Array<VectorGaussian>(c).Named("wPrior");
// The weight vector for each class.
w = Variable.Array<Vector>(c).Named("w");
w[c] = Variable<Vector>.Random(wPrior[c]);
noisePrecision = Variable.New<double>().Named("noisePrecision");
// Arrays of <see cref="Vector"/>-valued items (feature vectors) and integer labels.
x = Variable.Array<Vector>(i).Named("x");
y = Variable.Array<int>(i).Named("y");
// For all items...
using (Variable.ForEach(i))
{
// ...compute the score of this item across all classes...
score = BPMUtils.ComputeClassScores(w, x[i], noisePrecision);
y[i] = Variable.DiscreteUniform(c);
// ... and constrain the output.
BPMUtils.ConstrainMaximum(y[i], score);
}
// Inference engine settings (EP).
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
}
public void Run()
{
InferenceEngine engine = new InferenceEngine();
if (!(engine.Algorithm is ExpectationPropagation))
{
Console.WriteLine("This example only runs with Expectation Propagation");
return;
}
Rand.Restart(0);
int nQuestions = 100;
int nSubjects = 40;
int nChoices = 4;
Gaussian abilityPrior = new Gaussian(0, 1);
Gaussian difficultyPrior = new Gaussian(0, 1);
Gamma discriminationPrior = Gamma.FromShapeAndScale(5, 1);
double[] trueAbility, trueDifficulty, trueDiscrimination;
int[] trueTrueAnswer;
int[][] data = Sample(nSubjects, nQuestions, nChoices, abilityPrior, difficultyPrior, discriminationPrior,
out trueAbility, out trueDifficulty, out trueDiscrimination, out trueTrueAnswer);
Range question = new Range(nQuestions).Named("question");
Range subject = new Range(nSubjects).Named("subject");
Range choice = new Range(nChoices).Named("choice");
var response = Variable.Array(Variable.Array<int>(question), subject).Named("response");
response.ObservedValue = data;
var ability = Variable.Array<double>(subject).Named("ability");
ability[subject] = Variable.Random(abilityPrior).ForEach(subject);
var difficulty = Variable.Array<double>(question).Named("difficulty");
difficulty[question] = Variable.Random(difficultyPrior).ForEach(question);
var discrimination = Variable.Array<double>(question).Named("discrimination");
discrimination[question] = Variable.Random(discriminationPrior).ForEach(question);
var trueAnswer = Variable.Array<int>(question).Named("trueAnswer");
trueAnswer[question] = Variable.DiscreteUniform(nChoices).ForEach(question);
using (Variable.ForEach(subject)) {
using (Variable.ForEach(question)) {
var advantage = (ability[subject] - difficulty[question]).Named("advantage");
var advantageNoisy = Variable.GaussianFromMeanAndPrecision(advantage, discrimination[question]).Named("advantageNoisy");
var correct = (advantageNoisy > 0).Named("correct");
using (Variable.If(correct))
response[subject][question] = trueAnswer[question];
using (Variable.IfNot(correct))
response[subject][question] = Variable.DiscreteUniform(nChoices);
}
}
engine.NumberOfIterations = 5;
subject.AddAttribute(new Sequential()); // needed to get stable convergence
engine.Compiler.UseSerialSchedules = true;
if (false) {
// set this to do majority voting
ability.ObservedValue = Util.ArrayInit(nSubjects, i => 0.0);
difficulty.ObservedValue = Util.ArrayInit(nQuestions, i => 0.0);
discrimination.ObservedValue = Util.ArrayInit(nQuestions, i => 1.0);
}
var trueAnswerPosterior = engine.Infer<IList<Discrete>>(trueAnswer);
int numCorrect = 0;
for (int q = 0; q < nQuestions; q++) {
int bestGuess = trueAnswerPosterior[q].GetMode();
if (bestGuess == trueTrueAnswer[q]) numCorrect++;
}
double pctCorrect = 100.0*numCorrect/nQuestions;
Console.WriteLine("{0}% TrueAnswers correct", pctCorrect.ToString("f0"));
var difficultyPosterior = engine.Infer<IList<Gaussian>>(difficulty);
for (int q = 0; q < Math.Min(nQuestions, 4); q++) {
Console.WriteLine("difficulty[{0}] = {1} (sampled from {2})", q, difficultyPosterior[q], trueDifficulty[q].ToString("g2"));
}
var discriminationPosterior = engine.Infer<IList<Gamma>>(discrimination);
for (int q = 0; q < Math.Min(nQuestions, 4); q++) {
Console.WriteLine("discrimination[{0}] = {1} (sampled from {2})", q, discriminationPosterior[q], trueDiscrimination[q].ToString("g2"));
}
var abilityPosterior = engine.Infer<IList<Gaussian>>(ability);
for (int s = 0; s < Math.Min(nSubjects, 4); s++) {
Console.WriteLine("ability[{0}] = {1} (sampled from {2})", s, abilityPosterior[s], trueAbility[s].ToString("g2"));
}
}
public void Run()
{
InferenceEngine engine = new InferenceEngine();
if (!(engine.Algorithm is ExpectationPropagation))
{
return;
}
int[][] fighter1Data, fighter2Data, outcomeData, winTypeData;
int[] firstYearData;
int[] lastYearData;
int nFighters, nYears, startYear;
LoadData(out fighter1Data, out fighter2Data, out outcomeData, out winTypeData, out firstYearData, out lastYearData,
out nFighters, out nYears, out startYear);
//Skill prior
var skillPrior = new Gaussian(1000, 500 * 500);
var performancePrecisionPrior = Gamma.FromShapeAndRate(2, 26 * 26);
var skillChangePrecisionPrior = Gamma.FromShapeAndRate(2, 26 * 26);
var performancePrecision = Variable.Random(performancePrecisionPrior).Named("performancePrecision");
var skillChangePrecision = Variable.Random(skillChangePrecisionPrior).Named("skillChangePrecision");
var matchupThresholdPrior = new Gaussian(200, 50 * 50);
var matchupThreshold = Variable.Random(matchupThresholdPrior).Named("matchupThreshold");
var finishThresholdPrior = new Gaussian(20, 10 * 10);
var finishThreshold = Variable.Random(finishThresholdPrior).Named("finishThreshold");
var decicionThresholdPrior = new Gaussian(10, 10 * 10);
var decisionThreshold = Variable.Random(decicionThresholdPrior).Named("decisionThreshold");
Range fighter = new Range(nFighters).Named("fighter");
Range year = new Range(nYears).Named("year");
VariableArray <int> firstYear = Variable.Array <int>(fighter).Named("firstYear");
var skill = Variable.Array(Variable.Array <double>(fighter), year).Named("skill");
using (var yearBlock = Variable.ForEach(year))
{
var y = yearBlock.Index;
using (Variable.If(y == 0))
{
skill[year][fighter] = Variable.Random(skillPrior).ForEach(fighter);
}
using (Variable.If(y > 0))
{
using (Variable.ForEach(fighter))
{
Variable <bool> isFirstYear = (firstYear[fighter] >= y).Named("isFirstYear");
using (Variable.If(isFirstYear))
{
skill[year][fighter] = Variable.Random(skillPrior);
}
using (Variable.IfNot(isFirstYear))
{
skill[year][fighter] = Variable.GaussianFromMeanAndPrecision(skill[y - 1][fighter], skillChangePrecision);
}
}
}
}
firstYear.ObservedValue = firstYearData;
int[] nMatchesData = Util.ArrayInit(nYears, y => outcomeData[y].Length);
var nMatches = Variable.Observed(nMatchesData, year).Named("nMatches");
Range match = new Range(nMatches[year]).Named("match");
var fighter1 = Variable.Observed(fighter1Data, year, match).Named("fighter1");
var fighter2 = Variable.Observed(fighter2Data, year, match).Named("fighter2");
var outcome = Variable.Observed(outcomeData, year, match).Named("outcome");
var winType = Variable.Observed(winTypeData, year, match).Named("winType");
Variable.ConstrainTrue(finishThreshold > decisionThreshold);
Variable.ConstrainTrue(decisionThreshold > 0);
using (Variable.ForEach(year))
{
using (Variable.ForEach(match))
{
var w = fighter1[year][match];
var b = fighter2[year][match];
Variable <double> fighter1_performance = Variable.GaussianFromMeanAndPrecision(skill[year][w], performancePrecision).Named("fighter1_performance");
Variable <double> fighter2_performance = Variable.GaussianFromMeanAndPrecision(skill[year][b], performancePrecision).Named("fighter2_performance");
Variable.ConstrainFalse(skill[year][w] - skill[year][b] > matchupThreshold);
Variable.ConstrainFalse(skill[year][b] - skill[year][w] > matchupThreshold);
using (Variable.Case(outcome[year][match], 0))
{ // fighter2 wins
Variable.ConstrainTrue((fighter2_performance - fighter1_performance) > decisionThreshold);
//Finish
using (Variable.Case(winType[year][match], 1))
{
Variable.ConstrainTrue((fighter2_performance - fighter1_performance) > finishThreshold);
}
}
using (Variable.Case(outcome[year][match], 1))
{ // draw
Variable.ConstrainFalse((fighter2_performance - fighter1_performance) < decisionThreshold);
Variable.ConstrainFalse((fighter1_performance - fighter2_performance) < decisionThreshold);
}
using (Variable.Case(outcome[year][match], 2))
{ // fighter1 wins
Variable.ConstrainTrue((fighter1_performance - fighter2_performance) > decisionThreshold);
//Finish
using (Variable.Case(winType[year][match], 1))
{
Variable.ConstrainTrue((fighter1_performance - fighter2_performance) > finishThreshold);
}
}
}
}
year.AddAttribute(new Sequential()); // helps inference converge faster
engine.Compiler.UseSerialSchedules = false;
engine.NumberOfIterations = 200;
var skillPost = engine.Infer <Gaussian[][]>(skill);
var matchupThresholdPost = engine.Infer <Gaussian>(matchupThreshold);
var decisionThresholdPost = engine.Infer <Gaussian>(decisionThreshold);
var finishThresholdPost = engine.Infer <Gaussian>(finishThreshold);
var skillChangePrecisionPost = engine.Infer <Gamma>(skillChangePrecision);
Console.WriteLine("Matchup Threshold prec {0}", matchupThresholdPost);
Console.WriteLine("Decision Threshold {0}", decisionThresholdPost);
Console.WriteLine("Finish threshold {0}", finishThresholdPost);
Console.WriteLine("Skilll change prec {0}", skillChangePrecisionPost);
using (System.IO.StreamWriter file =
new System.IO.StreamWriter(@"outfile.txt"))
{
file.WriteLine("key year name skillmean variance");
for (int i = 0; i < nFighters; i++)
{
for (int y = 0; y < nYears; y++)
{
if (y >= firstYearData[i] && y <= lastYearData[i])
{
file.WriteLine("{0} {1} {2} {3} {4}", i, startYear + y, keyToFighter[i], skillPost[y][i].GetMean(), skillPost[y][i].GetVariance());
}
}
}
}
}
public void Run()
{
InferenceEngine engine = new InferenceEngine();
if (!(engine.Algorithm is ExpectationPropagation))
{
Console.WriteLine("This example only runs with Expectation Propagation");
return;
}
Rand.Restart(0);
int nQuestions = 100;
int nSubjects = 40;
int nChoices = 4;
Gaussian abilityPrior = new Gaussian(0, 1);
Gaussian difficultyPrior = new Gaussian(0, 1);
Gamma discriminationPrior = Gamma.FromMeanAndVariance(1, 0.01);
double[] trueAbility, trueDifficulty, trueDiscrimination;
int[] trueTrueAnswer;
int[][] data = Sample(nSubjects, nQuestions, nChoices, abilityPrior, difficultyPrior, discriminationPrior,
out trueAbility, out trueDifficulty, out trueDiscrimination, out trueTrueAnswer);
Range question = new Range(nQuestions).Named("question");
Range subject = new Range(nSubjects).Named("subject");
Range choice = new Range(nChoices).Named("choice");
var response = Variable.Array(Variable.Array <int>(question), subject).Named("response");
response.ObservedValue = data;
var ability = Variable.Array <double>(subject).Named("ability");
ability[subject] = Variable.Random(abilityPrior).ForEach(subject);
var difficulty = Variable.Array <double>(question).Named("difficulty");
difficulty[question] = Variable.Random(difficultyPrior).ForEach(question);
var discrimination = Variable.Array <double>(question).Named("discrimination");
discrimination[question] = Variable.Random(discriminationPrior).ForEach(question);
var trueAnswer = Variable.Array <int>(question).Named("trueAnswer");
trueAnswer[question] = Variable.DiscreteUniform(choice).ForEach(question);
using (Variable.ForEach(subject)) {
using (Variable.ForEach(question)) {
var advantage = (ability[subject] - difficulty[question]).Named("advantage");
var advantageNoisy = Variable.GaussianFromMeanAndPrecision(advantage, discrimination[question]).Named("advantageNoisy");
var correct = (advantageNoisy > 0).Named("correct");
using (Variable.If(correct))
response[subject][question] = trueAnswer[question];
using (Variable.IfNot(correct))
response[subject][question] = Variable.DiscreteUniform(choice);
}
}
engine.NumberOfIterations = 5;
subject.AddAttribute(new Sequential()); // needed to get stable convergence
question.AddAttribute(new Sequential()); // needed to get stable convergence
bool doMajorityVoting = false; // set this to 'true' to do majority voting
if (doMajorityVoting)
{
ability.ObservedValue = Util.ArrayInit(nSubjects, i => 0.0);
difficulty.ObservedValue = Util.ArrayInit(nQuestions, i => 0.0);
discrimination.ObservedValue = Util.ArrayInit(nQuestions, i => 1.0);
}
var trueAnswerPosterior = engine.Infer <IList <Discrete> >(trueAnswer);
int numCorrect = 0;
for (int q = 0; q < nQuestions; q++)
{
int bestGuess = trueAnswerPosterior[q].GetMode();
if (bestGuess == trueTrueAnswer[q])
{
numCorrect++;
}
}
double pctCorrect = 100.0 * numCorrect / nQuestions;
Console.WriteLine("{0}% TrueAnswers correct", pctCorrect.ToString("f0"));
var difficultyPosterior = engine.Infer <IList <Gaussian> >(difficulty);
for (int q = 0; q < Math.Min(nQuestions, 4); q++)
{
Console.WriteLine("difficulty[{0}] = {1} (sampled from {2})", q, difficultyPosterior[q], trueDifficulty[q].ToString("g2"));
}
var discriminationPosterior = engine.Infer <IList <Gamma> >(discrimination);
for (int q = 0; q < Math.Min(nQuestions, 4); q++)
{
Console.WriteLine("discrimination[{0}] = {1} (sampled from {2})", q, discriminationPosterior[q], trueDiscrimination[q].ToString("g2"));
}
var abilityPosterior = engine.Infer <IList <Gaussian> >(ability);
for (int s = 0; s < Math.Min(nSubjects, 4); s++)
{
Console.WriteLine("ability[{0}] = {1} (sampled from {2})", s, abilityPosterior[s], trueAbility[s].ToString("g2"));
}
}
public Model(SharedVariableArray<Vector> w, Range c, int numChunks)
{
// Items.
numItems = Variable.New<int>().Named("numItems");
i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
// The model identifier for the shared variables.
model = new MicrosoftResearch.Infer.Models.Model(numChunks).Named("model");
// The weight vector for each submodel.
wModel = w.GetCopyFor(model).Named("wModel");
noisePrecision = Variable.New<double>().Named("noisePrecision");
// Arrays of <see cref="Vector"/>-valued items (feature vectors) and integer labels.
x = Variable.Array<Vector>(i).Named("x");
y = Variable.Array<int>(i).Named("y");
// For all items...
using (Variable.ForEach(i))
{
// ...compute the score of this item across all classes...
score = BPMUtils.ComputeClassScores(wModel, x[i], noisePrecision);
y[i] = Variable.DiscreteUniform(c);
// ... and constrain the output.
BPMUtils.ConstrainMaximum(y[i], score);
}
// Inference engine settings (EP).
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
}
private static void DistributedScheduleTestProcess(
ICommunicator comm,
IList <Gaussian>[] xResults,
Gaussian[] shiftResults,
Compiler.Graphs.DistributedCommunicationInfo distributedCommunicationInfo,
int[][] scheduleForProcess,
// [distributedStage][thread][block][i]
int[][][][] schedulePerThreadForProcess)
{
var nodeCountVar = Variable.Observed(0).Named("nodeCount");
Range node = new Range(nodeCountVar).Named("node");
node.AddAttribute(new Sequential()
{
BackwardPass = true
});
var itemCountVar = Variable.Observed(0).Named("itemCount");
Range item = new Range(itemCountVar).Named("item");
var x = Variable.Array <double>(item).Named("x");
x[item] = Variable.GaussianFromMeanAndPrecision(0, 1).ForEach(item);
var parentCount = Variable.Observed(default(int[]), node).Named("parentCount");
Range parent = new Range(parentCount[node]).Named("parent");
var indices = Variable.Observed(default(int[][]), node, parent).Named("indices");
indices.SetValueRange(item);
var shift = Variable.GaussianFromMeanAndPrecision(0, 1).Named("shift");
shift.AddAttribute(new PointEstimate());
using (Variable.ForEach(node))
{
var subArray = Variable.Subarray(x, indices[node]).Named("subArray");
using (Variable.If(parentCount[node] == 1))
{
Variable.ConstrainEqualRandom(subArray[0], Gaussian.FromMeanAndVariance(0, 1));
}
using (Variable.If(parentCount[node] == 2))
{
Variable.ConstrainEqual(subArray[0], subArray[1] + 1);
}
Variable.ConstrainEqualRandom(shift, new Gaussian(1, 2));
}
// this dummy part of the model causes depth cloning to occur
Range item2 = new Range(0);
var indices2 = Variable.Observed(new int[0], item2).Named("indices2");
var subArray2 = Variable.Subarray(x, indices2);
Variable.ConstrainEqual(subArray2[item2], 0.0);
var distributedStageCount = Variable.Observed(0).Named("distributedStageCount");
Range distributedStage = new Range(distributedStageCount).Named("stage");
var commVar = Variable.Observed(default(ICommunicator)).Named("comm");
if (schedulePerThreadForProcess != null)
{
var threadCount = Variable.Observed(0).Named("threadCount");
Range thread = new Range(threadCount).Named("thread");
var blockCountOfDistributedStage = Variable.Observed(default(int[]), distributedStage).Named("blockCount");
Range gameBlock = new Range(blockCountOfDistributedStage[distributedStage]).Named("block");
var gameCountInBlockOfDistributedStage = Variable.Observed(default(int[][][]), distributedStage, thread, gameBlock).Named("GameCountInBlock");
Range gameInBlock = new Range(gameCountInBlockOfDistributedStage[distributedStage][thread][gameBlock]).Named("gameInBlock");
var gamesInBlockOfDistributedStage = Variable.Array(Variable.Array(Variable.Array(Variable.Array <int>(gameInBlock), gameBlock), thread), distributedStage).Named("GamesInBlock");
gamesInBlockOfDistributedStage.ObservedValue = default(int[][][][]);
node.AddAttribute(new DistributedSchedule(commVar, gamesInBlockOfDistributedStage));
threadCount.ObservedValue = schedulePerThreadForProcess[0].Length;
blockCountOfDistributedStage.ObservedValue = Util.ArrayInit(schedulePerThreadForProcess.Length, stageIndex => schedulePerThreadForProcess[stageIndex][0].Length);
gameCountInBlockOfDistributedStage.ObservedValue = Util.ArrayInit(schedulePerThreadForProcess.Length, stageIndex =>
Util.ArrayInit(schedulePerThreadForProcess[stageIndex].Length, t =>
Util.ArrayInit(schedulePerThreadForProcess[stageIndex][t].Length, b =>
schedulePerThreadForProcess[stageIndex][t][b].Length)));
gamesInBlockOfDistributedStage.ObservedValue = schedulePerThreadForProcess;
}
else
{
var gameCountInLocalBlock = Variable.Observed(new int[0], distributedStage).Named("gameCountInLocalBlock");
Range gameInLocalBlock = new Range(gameCountInLocalBlock[distributedStage]).Named("gameInLocalBlock");
var nodesInLocalBlock = Variable.Observed(new int[0][], distributedStage, gameInLocalBlock).Named("nodesInLocalBlock");
node.AddAttribute(new DistributedSchedule(commVar, nodesInLocalBlock));
gameCountInLocalBlock.ObservedValue = Util.ArrayInit(scheduleForProcess.Length, stageIndex =>
scheduleForProcess[stageIndex].Length);
nodesInLocalBlock.ObservedValue = scheduleForProcess;
}
var processCount = Variable.Observed(0).Named("processCount");
Range sender = new Range(processCount);
var arrayIndicesToSendCount = Variable.Observed(default(int[][]), distributedStage, sender).Named("arrayIndicesToSendCount");
Range arrayIndexToSend = new Range(arrayIndicesToSendCount[distributedStage][sender]);
var arrayIndicesToSendVar = Variable.Observed(default(int[][][]), distributedStage, sender, arrayIndexToSend).Named("arrayIndicesToSend");
var arrayIndicesToReceiveCount = Variable.Observed(default(int[][]), distributedStage, sender).Named("arrayIndicesToReceiveCount");
Range arrayIndexToReceive = new Range(arrayIndicesToReceiveCount[distributedStage][sender]);
var arrayIndicesToReceiveVar = Variable.Observed(default(int[][][]), distributedStage, sender, arrayIndexToReceive).Named("arrayIndexToReceive");
indices.AddAttribute(new DistributedCommunication(arrayIndicesToSendVar, arrayIndicesToReceiveVar));
distributedStageCount.ObservedValue = scheduleForProcess.Length;
commVar.ObservedValue = comm;
processCount.ObservedValue = comm.Size;
nodeCountVar.ObservedValue = distributedCommunicationInfo.indices.Length;
itemCountVar.ObservedValue = distributedCommunicationInfo.arrayLength;
parentCount.ObservedValue = distributedCommunicationInfo.indicesCount;
indices.ObservedValue = distributedCommunicationInfo.indices;
arrayIndicesToSendCount.ObservedValue = distributedCommunicationInfo.arrayIndicesToSendCount;
arrayIndicesToSendVar.ObservedValue = distributedCommunicationInfo.arrayIndicesToSend;
arrayIndicesToReceiveCount.ObservedValue = distributedCommunicationInfo.arrayIndicesToReceiveCount;
arrayIndicesToReceiveVar.ObservedValue = distributedCommunicationInfo.arrayIndicesToReceive;
InferenceEngine engine = new InferenceEngine();
//engine.Compiler.UseExistingSourceFiles = true;
engine.ModelName = "DistributedScheduleTest" + comm.Rank;
engine.ShowProgress = false;
engine.NumberOfIterations = 2;
engine.OptimiseForVariables = new IVariable[] { x, shift };
var xActual = engine.Infer <IList <Gaussian> >(x);
xResults[comm.Rank] = xActual;
var shiftActual = engine.Infer <Gaussian>(shift);
shiftResults[comm.Rank] = shiftActual;
}
private int[][][] parallelScheduleTest(int numThreads, int nodeCount, out IReadOnlyList <Gaussian> xMarginal, out Gaussian shiftMarginal, out IReadOnlyList <int[]> variablesUsedByNode)
{
int maxParentCount = 1;
variablesUsedByNode = GenerateVariablesUsedByNode(nodeCount, maxParentCount);
ParallelScheduler ps = new ParallelScheduler();
ps.CreateGraph(variablesUsedByNode);
var schedule = ps.GetScheduleWithBarriers(numThreads);
ParallelScheduler.WriteSchedule(schedule, true);
var schedulePerThread = ps.ConvertToSchedulePerThread(schedule, numThreads);
var nodeCountVar = Variable.Observed(nodeCount).Named("nodeCount");
Range node = new Range(nodeCountVar).Named("node");
node.AddAttribute(new Sequential()
{
BackwardPass = true
});
var x = Variable.Array <double>(node).Named("x");
x[node] = Variable.GaussianFromMeanAndPrecision(0, 1).ForEach(node);
var parentCount = Variable.Observed(variablesUsedByNode.Select(a => a.Length).ToArray(), node).Named("parentCount");
Range parent = new Range(parentCount[node]).Named("parent");
var indices = Variable.Observed(variablesUsedByNode.ToArray(), node, parent).Named("indices");
var shift = Variable.GaussianFromMeanAndPrecision(0, 1).Named("shift");
shift.AddAttribute(new PointEstimate());
using (Variable.ForEach(node))
{
var subArray = Variable.Subarray(x, indices[node]).Named("subArray");
using (Variable.If(parentCount[node] == 1))
{
Variable.ConstrainEqualRandom(subArray[0], Gaussian.FromMeanAndVariance(0, 1));
}
using (Variable.If(parentCount[node] == 2))
{
Variable.ConstrainEqual(subArray[0], subArray[1] + 1);
}
Variable.ConstrainEqualRandom(shift, new Gaussian(1, 2));
}
InferenceEngine engine = new InferenceEngine();
engine.NumberOfIterations = 2;
engine.OptimiseForVariables = new IVariable[] { x, shift };
var xExpected = engine.Infer(x);
//Console.WriteLine(xExpected);
var shiftExpected = engine.Infer(shift);
var threadCount = Variable.Observed(0).Named("threadCount");
Range thread = new Range(threadCount).Named("thread");
var blockCount = Variable.Observed(0).Named("blockCount");
Range gameBlock = new Range(blockCount).Named("block");
var gameCountInBlock = Variable.Observed(default(int[][]), thread, gameBlock).Named("GameCountInBlock");
Range gameInBlock = new Range(gameCountInBlock[thread][gameBlock]).Named("gameInBlock");
var gamesInBlock = Variable.Observed(default(int[][][]), thread, gameBlock, gameInBlock).Named("GamesInBlock");
node.AddAttribute(new ParallelSchedule(gamesInBlock));
threadCount.ObservedValue = schedulePerThread.Length;
blockCount.ObservedValue = (schedulePerThread.Length == 0) ? 0 : schedulePerThread[0].Length;
gameCountInBlock.ObservedValue = Util.ArrayInit(schedulePerThread.Length, t =>
Util.ArrayInit(schedulePerThread[t].Length, b => schedulePerThread[t][b].Length));
gamesInBlock.ObservedValue = schedulePerThread;
var xActual = engine.Infer <IReadOnlyList <Gaussian> >(x);
//Debug.WriteLine(xActual);
Assert.True(xExpected.Equals(xActual));
var shiftActual = engine.Infer <Gaussian>(shift);
Assert.True(shiftExpected.Equals(shiftActual));
xMarginal = xActual;
shiftMarginal = shiftActual;
return(schedulePerThread);
}
public Model()
{
// Classes.
numClasses = Variable.New<int>().Named("numClasses");
c = new Range(numClasses).Named("c");
// Features.
numFeatures = Variable.New<int>().Named("numFeatures");
f = new Range(numFeatures).Named("f");
// Items.
numItems = Variable.New<int>().Named("numItems");
i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
// Features per item.
numFeaturesPerItem = Variable.Array<int>(i).Named("numFeaturesPerItem");
fItem = new Range(numFeaturesPerItem[i]).Named("fItem");
// The prior distribution for weight vector for each class. When
// <see cref="Test"/> is called, this is set to the posterior weight
// distributions from <see cref="Train"/>.
wPrior = Variable.Array(Variable.Array<Gaussian>(f), c).Named("wPrior");
// The weight vector for each class.
w = Variable.Array(Variable.Array<double>(f), c).Named("w");
w[c][f] = Variable<double>.Random(wPrior[c][f]);
noisePrecision = Variable.New<double>().Named("noisePrecision");
// Jagged array of feature values - each item is an array of data values
// whose indices are given by the corresponding indices[i].
values = Variable.Array(Variable.Array<double>(fItem), i).Named("values");
// Jagged array of indices for the items.
indices = Variable.Array(Variable.Array<int>(fItem), i).Named("indices");
// Labels.
y = Variable.Array<int>(i).Named("y");
// For all items...
using (Variable.ForEach(i))
{
// ...compute the score of this item across all classes...
score = BPMUtils.ComputeClassScores(w, values[i], indices[i], fItem, noisePrecision);
y[i] = Variable.DiscreteUniform(c);
// ...and constrain the output.
BPMUtils.ConstrainMaximum(y[i], score);
}
// Inference engine settings (EP).
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
}
/// <summary>
/// Constructs the model.
/// </summary>
public override void ConstructModel()
{
if (this.HasBeenConstructed)
{
return;
}
this.NumberOfPeople = Variable.New <int>().Named("NumberOfPeople").Attrib(new DoNotInfer());
this.NumberOfFeatures = Variable.New <int>().Named("NumberOfFeatures").Attrib(new DoNotInfer());
Range buckets = new Range(this.NumberOfFeatures).Named("buckets");
Range users = new Range(this.NumberOfPeople).Named("users");
this.NumberOfMessages = Variable.Array <int>(users).Named("numberOfMessages").Attrib(new DoNotInfer());
Range emails = new Range(this.NumberOfMessages[users]).Named("emails");
// Make sure that the range across messages is handled sequentially
// - this is necessary to ensure the model converges during training
emails.AddAttribute(new Sequential());
this.FeatureCounts = Variable.Array(Variable.Array <int>(emails), users).Named("FeatureCounts").Attrib(new DoNotInfer());
Range indices = new Range(this.FeatureCounts[users][emails]).Named("indices");
// observed data
this.FeatureValue =
Variable.Array(Variable.Array(Variable.Array <double>(indices), emails), users)
.Named("featureValue")
.Attrib(new DoNotInfer());
this.FeatureIndices =
Variable.Array(Variable.Array(Variable.Array <int>(indices), emails), users)
.Named("featureIndices")
.Attrib(new DoNotInfer());
// The weights
this.Weight = Variable.Array(Variable.Array <double>(buckets), users).Named("weight");
// The priors on the weights
this.WeightMean = Variable.Array <double>(buckets).Named("weightMean");
this.WeightPrecision = Variable.Array <double>(buckets).Named("weightPrecision");
this.WeightMeanPriors = Variable.New <DistributionStructArray <Gaussian, double> >().Named("WeightMeanPriors").Attrib(new DoNotInfer());
this.WeightPrecisionPriors = Variable.New <DistributionStructArray <Gamma, double> >().Named("WeightPrecisionPriors").Attrib(new DoNotInfer());
this.WeightMean.SetTo(Variable <double[]> .Random(this.WeightMeanPriors));
this.WeightPrecision.SetTo(Variable <double[]> .Random(this.WeightPrecisionPriors));
// Noise Variance
this.NoiseVariance = Variable.New <double>().Named("NoiseVariance").Attrib(new DoNotInfer());
// Label: is the message replied to?
this.RepliedTo = Variable.Array(Variable.Array <bool>(emails), users).Named("repliedTo");
// Loop over people
using (Variable.ForEach(users))
{
// Loop over features
using (Variable.ForEach(buckets))
{
this.Weight[users][buckets] = Variable.GaussianFromMeanAndPrecision(
this.WeightMean[buckets],
this.WeightPrecision[buckets]);
}
// Loop over emails
using (Variable.ForEach(emails))
{
var featureWeight = Variable.Subarray(this.Weight[users], this.FeatureIndices[users][emails]).Named("featureWeight");
var featureScore = Variable.Array <double>(indices).Named("featureScore");
featureScore[indices] = this.FeatureValue[users][emails][indices]
* featureWeight[indices];
var score = Variable.Sum(featureScore).Named("score");
this.RepliedTo[users][emails] = Variable.GaussianFromMeanAndVariance(score, this.NoiseVariance).Named("noisyScore") > 0;
}
}
this.InitializeEngine();
// if during personalisation we update the weights rather than the weight mean and precision, then this will need to change
switch (this.Mode)
{
case InputMode.CommunityTraining:
case InputMode.Training:
this.Engine.OptimiseForVariables = new IVariable[] { this.Weight, this.WeightMean, this.WeightPrecision };
break;
default:
this.Engine.OptimiseForVariables = this.Engine.OptimiseForVariables = new IVariable[] { this.RepliedTo };
break;
}
this.HasBeenConstructed = true;
}
public InferenceResult<ClusterRegressionWeights[]> Regression(Vector[] features, double[] values, int clusters)
{
var dimensions = features[0].Count;
var evidence = Variable.Bernoulli(0.5).Named("evidence");
var evidenceBlock = Variable.If(evidence);
var clustersRange = new Range(clusters).Named("clustersRange");
var wMeans = Variable.Array<Vector>(clustersRange).Named("wMeans");
var wPrecision = Variable.Array<PositiveDefiniteMatrix>(clustersRange).Named("wPrecision");
var noise = Variable.Array<double>(clustersRange).Named("noise");
var w = Variable.Array<Vector>(clustersRange).Named("w");
using(Variable.ForEach(clustersRange))
{
wMeans[clustersRange] = Variable
.VectorGaussianFromMeanAndPrecision(
Vector.Zero(dimensions),
PositiveDefiniteMatrix.IdentityScaledBy(dimensions, 0.01));
wPrecision[clustersRange] = Variable.WishartFromShapeAndRate(100, PositiveDefiniteMatrix.IdentityScaledBy(dimensions, 0.01));
w[clustersRange] = Variable.VectorGaussianFromMeanAndPrecision(wMeans[clustersRange], wPrecision[clustersRange]);
noise[clustersRange] = Variable.GammaFromMeanAndVariance(2, 5);
}
var numItems = Variable.New<int>().Named("numItems");
var i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
var initialWeights = Enumerable.Range(0, clusters).Select(_ => 1.0).ToArray();
var mixtureWeightsPrior = Variable.Dirichlet(clustersRange, initialWeights).Named("mixtureWeightsPrior");
var latentIndex = Variable.Array<int>(i).Named("latentIndex");
var x = Variable.Array<Vector>(i).Named("x");
var y = Variable.Array<double>(i).Named("y");
using (Variable.ForEach(i))
{
latentIndex[i] = Variable.Discrete(mixtureWeightsPrior);
using(Variable.Switch(latentIndex[i]))
{
y[i] = Variable.GaussianFromMeanAndPrecision(Variable.InnerProduct(w[latentIndex[i]], x[i]), noise[latentIndex[i]]);
}
}
numItems.ObservedValue = features.Length;
var zinit = new Discrete[features.Length];
for (var j = 0; j < zinit.Length; j++)
zinit[j] = Discrete.PointMass(Rand.Int(clustersRange.SizeAsInt), clustersRange.SizeAsInt);
latentIndex.InitialiseTo(Distribution<int>.Array(zinit));
evidenceBlock.CloseBlock();
x.ObservedValue = features;
y.ObservedValue = values;
var engine = new InferenceEngine(new VariationalMessagePassing());
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
var wPosterior = engine.Infer<IList<VectorGaussian>>(w);
var mixtureWeightsPosterior = engine.Infer<Dirichlet>(mixtureWeightsPrior);
var noisePosterior = engine.Infer<IList<Gamma>>(noise);
var bEvidence = engine.Infer<Bernoulli>(evidence);
var clusterList = new List<ClusterRegressionWeights>();
for (var c = 0; c < clusters; c++ )
{
clusterList.Add(new ClusterRegressionWeights(wPosterior[c].GetMean(), noisePosterior[c], mixtureWeightsPosterior.GetMean()[c]));
}
return new InferenceResult<ClusterRegressionWeights[]>(bEvidence, clusterList.ToArray());
}
public void Run()
{
// This example requires EP
InferenceEngine engine = new InferenceEngine();
if (!(engine.Algorithm is Algorithms.ExpectationPropagation))
{
Console.WriteLine("This example only runs with Expectation Propagation");
return;
}
int nPlayers = 10;
int nYears = 10;
Rand.Restart(1);
var skillPrior = new Gaussian(1200, 800 * 800);
var drawMarginMeanPrior = new Gaussian(700, 500 * 500);
var drawMarginPrecisionPrior = Gamma.FromShapeAndRate(2, 500 * 500);
var performancePrecisionPrior = Gamma.FromShapeAndRate(2, 800 * 800);
var skillChangePrecisionPrior = Gamma.FromShapeAndRate(2, 26 * 26);
var drawMarginChangePrecisionPrior = Gamma.FromShapeAndRate(2, 10 * 10);
var whiteAdvantagePrior = new Gaussian(0, 200 * 200);
var drawMarginMean = Variable.Random(drawMarginMeanPrior).Named("drawMarginMean");
var drawMarginPrecision = Variable.Random(drawMarginPrecisionPrior).Named("drawMarginPrecision");
var performancePrecision = Variable.Random(performancePrecisionPrior).Named("performancePrecision");
var skillChangePrecision = Variable.Random(skillChangePrecisionPrior).Named("skillChangePrecision");
var drawMarginChangePrecision = Variable.Random(drawMarginChangePrecisionPrior).Named("drawMarginChangePrecision");
var whiteAdvantage = Variable.Random(whiteAdvantagePrior).Named("whiteAdvantage");
Range player = new Range(nPlayers).Named("player");
Range year = new Range(nYears).Named("year");
VariableArray <int> firstYear = Variable.Array <int>(player).Named("firstYear");
var skill = Variable.Array(Variable.Array <double>(player), year).Named("skill");
var drawMargin = Variable.Array(Variable.Array <double>(player), year).Named("drawMargin");
using (var yearBlock = Variable.ForEach(year))
{
var y = yearBlock.Index;
using (Variable.If(y == 0))
{
skill[year][player] = Variable.Random(skillPrior).ForEach(player);
drawMargin[year][player] = Variable.GaussianFromMeanAndPrecision(drawMarginMean, drawMarginPrecision).ForEach(player);
}
using (Variable.If(y > 0))
{
using (Variable.ForEach(player))
{
Variable <bool> isFirstYear = (firstYear[player] >= y).Named("isFirstYear");
using (Variable.If(isFirstYear))
{
skill[year][player] = Variable.Random(skillPrior);
drawMargin[year][player] = Variable.GaussianFromMeanAndPrecision(drawMarginMean, drawMarginPrecision);
}
using (Variable.IfNot(isFirstYear))
{
skill[year][player] = Variable.GaussianFromMeanAndPrecision(skill[y - 1][player], skillChangePrecision);
drawMargin[year][player] = Variable.GaussianFromMeanAndPrecision(drawMargin[y - 1][player], drawMarginChangePrecision);
}
}
}
}
// Sample parameter values according to the above model
firstYear.ObservedValue = Util.ArrayInit(nPlayers, i => Rand.Int(nYears));
Parameters parameters = new Parameters();
parameters.drawMarginMean = drawMarginMeanPrior.Sample();
parameters.drawMarginPrecision = drawMarginPrecisionPrior.Sample();
parameters.performancePrecision = performancePrecisionPrior.Sample();
parameters.skillChangePrecision = skillChangePrecisionPrior.Sample();
parameters.drawMarginChangePrecision = drawMarginChangePrecisionPrior.Sample();
parameters.whiteAdvantage = whiteAdvantagePrior.Sample();
parameters.skill = Util.ArrayInit(nYears, y => Util.ArrayInit(nPlayers, i => skillPrior.Sample()));
parameters.drawMargin = Util.ArrayInit(nYears, y => Util.ArrayInit(nPlayers, i => Gaussian.Sample(parameters.drawMarginMean, parameters.drawMarginPrecision)));
for (int y = 0; y < nYears; y++)
{
for (int i = 0; i < nPlayers; i++)
{
if (y > firstYear.ObservedValue[i])
{
parameters.skill[y][i] = Gaussian.Sample(parameters.skill[y - 1][i], parameters.skillChangePrecision);
parameters.drawMargin[y][i] = Gaussian.Sample(parameters.drawMargin[y - 1][i], parameters.drawMarginChangePrecision);
}
}
}
// Sample game outcomes
int[][] whiteData, blackData, outcomeData;
GenerateData(parameters, firstYear.ObservedValue, out whiteData, out blackData, out outcomeData);
bool inferParameters = false; // make this true to infer additional parameters
if (!inferParameters)
{
// fix the true parameters
drawMarginMean.ObservedValue = parameters.drawMarginMean;
drawMarginPrecision.ObservedValue = parameters.drawMarginPrecision;
performancePrecision.ObservedValue = parameters.performancePrecision;
skillChangePrecision.ObservedValue = parameters.skillChangePrecision;
drawMarginChangePrecision.ObservedValue = parameters.drawMarginChangePrecision;
}
// Learn the skills from the data
int[] nGamesData = Util.ArrayInit(nYears, y => outcomeData[y].Length);
var nGames = Variable.Observed(nGamesData, year).Named("nGames");
Range game = new Range(nGames[year]).Named("game");
var whitePlayer = Variable.Observed(whiteData, year, game).Named("whitePlayer");
var blackPlayer = Variable.Observed(blackData, year, game).Named("blackPlayer");
var outcome = Variable.Observed(outcomeData, year, game).Named("outcome");
using (Variable.ForEach(year))
{
using (Variable.ForEach(game))
{
var w = whitePlayer[year][game];
var b = blackPlayer[year][game];
Variable <double> white_performance = Variable.GaussianFromMeanAndPrecision(skill[year][w], performancePrecision).Named("white_performance");
Variable <double> black_performance = Variable.GaussianFromMeanAndPrecision(skill[year][b], performancePrecision).Named("black_performance");
Variable <double> white_drawMargin = Variable.Copy(drawMargin[year][w]).Named("white_drawMargin");
Variable <double> black_drawMargin = Variable.Copy(drawMargin[year][b]).Named("black_drawMargin");
Variable <double> white_delta = (white_performance - black_performance + whiteAdvantage).Named("white_delta");
using (Variable.Case(outcome[year][game], 0))
{ // black wins
Variable.ConstrainTrue(white_delta + white_drawMargin < 0);
}
using (Variable.Case(outcome[year][game], 1))
{ // draw
Variable.ConstrainBetween(white_delta, -white_drawMargin, black_drawMargin);
}
using (Variable.Case(outcome[year][game], 2))
{ // white wins
Variable.ConstrainTrue(white_delta - black_drawMargin > 0);
}
}
}
year.AddAttribute(new Models.Attributes.Sequential()); // helps inference converge faster
engine.NumberOfIterations = 10;
var skillPost = engine.Infer <Gaussian[][]>(skill);
var drawMarginPost = engine.Infer <Gaussian[][]>(drawMargin);
// compare estimates to the true values
if (inferParameters)
{
Console.WriteLine("drawMargin mean = {0} (truth = {1})", engine.Infer <Gaussian>(drawMarginMean), parameters.drawMarginMean);
Console.WriteLine("drawMargin precision = {0} (truth = {1})", engine.Infer <Gamma>(drawMarginPrecision).GetMean(), parameters.drawMarginPrecision);
Console.WriteLine("performancePrecision = {0} (truth = {1})", engine.Infer <Gamma>(performancePrecision).GetMean(), parameters.performancePrecision);
Console.WriteLine("skillChangePrecision = {0} (truth = {1})", engine.Infer <Gamma>(skillChangePrecision).GetMean(), parameters.skillChangePrecision);
Console.WriteLine("drawMarginChangePrecision = {0} (truth = {1})", engine.Infer <Gamma>(drawMarginChangePrecision).GetMean(), parameters.drawMarginChangePrecision);
}
Console.WriteLine("white advantage = {0} (truth = {1})", engine.Infer <Gaussian>(whiteAdvantage), parameters.whiteAdvantage);
int countPrinted = 0;
for (int y = 0; y < nYears; y++)
{
for (int p = 0; p < nPlayers; p++)
{
if (y >= firstYear.ObservedValue[p])
{
if (++countPrinted > 3)
{
break;
}
Console.WriteLine("skill[{0}][{1}] = {2} (truth = {3:g4})", y, p, skillPost[y][p], parameters.skill[y][p]);
Console.WriteLine("drawMargin[{0}][{1}] = {2} (truth = {3:g4})", y, p, drawMarginPost[y][p], parameters.drawMargin[y][p]);
}
}
}
}
/// <summary>
/// Constructs the Matchbox model with the specified settings.
/// </summary>
/// <param name="buildTrainingModel">Specifies whether the method should build a training or a prediction model.</param>
/// <param name="breakTraitSymmetry">Specifies whether the constraints for breaking trait symmetry should be specified.</param>
/// <param name="usePreAdjustedUserParameters">Specifies whether user parameters should be sampled from custom priors (without using features).</param>
/// <param name="usePreAdjustedItemParameters">Specifies whether item parameters should be sampled from custom priors (without using features).</param>
/// <returns>The variables to infer from the built model.</returns>
public static IVariable[] BuildModel(
bool buildTrainingModel,
bool breakTraitSymmetry,
bool usePreAdjustedUserParameters,
bool usePreAdjustedItemParameters)
{
// Counts
var userCount = Variable.Observed(default(int)).Named("UserCount");
var itemCount = Variable.Observed(default(int)).Named("ItemCount");
var userThresholdCount = Variable.Observed(default(int)).Named("UserThresholdCount");
var observationCount = Variable.Observed(default(int)).Named("ObservationCount");
var traitCount = Variable.Observed(default(int)).Named("TraitCount");
// Ranges
var user = new Range(userCount).Named("user");
var item = new Range(itemCount).Named("item");
var trait = new Range(traitCount).Named("trait");
var observation = new Range(observationCount).Named("observation");
var userThreshold = new Range(userThresholdCount).Named("userThreshold");
var ratingValue = new Range(userThresholdCount - 1).Named("ratingValue");
if (buildTrainingModel)
{
// Use a sequential schedule
user.AddAttribute(new Sequential());
item.AddAttribute(new Sequential());
observation.AddAttribute(new Sequential());
}
// Latent variables
var userTraits = Variable.Array(Variable.Array <double>(trait), user).Named("UserTraits");
var itemTraits = Variable.Array(Variable.Array <double>(trait), item).Named("ItemTraits");
var userBias = Variable.Array <double>(user).Named("UserBias");
var itemBias = Variable.Array <double>(item).Named("ItemBias");
var userThresholds = Variable.Array(Variable.Array <double>(userThreshold), user).Named("UserThresholds");
// User latent variable definitions
var userFeatureCount = Variable.Observed(default(int)).Named("UserFeatureCount");
var userFeature = new Range(userFeatureCount).Named("userFeature");
var userTraitFeatureWeights = Variable.Array(Variable.Array <double>(userFeature), trait).Named("UserTraitFeatureWeights");
var userBiasFeatureWeights = Variable.Array <double>(userFeature).Named("UserBiasFeatureWeights");
if (usePreAdjustedUserParameters)
{
var userTraitsPrior = Variable.Observed(default(GaussianMatrix)).Named("UserTraitsPrior");
var userBiasPrior = Variable.Observed(default(GaussianArray)).Named("UserBiasPrior");
var userThresholdsPrior = Variable.Observed(default(GaussianMatrix)).Named("UserThresholdsPrior");
userTraits.SetTo(Variable <double[][]> .Random(userTraitsPrior));
userBias.SetTo(Variable <double[]> .Random(userBiasPrior));
userThresholds.SetTo(Variable <double[][]> .Random(userThresholdsPrior));
}
else
{
// Features
var nonZeroUserFeatureCounts = Variable.Observed(default(IList <int>), user).Named("NonZeroUserFeatureCounts");
var nonZeroUserFeature = new Range(nonZeroUserFeatureCounts[user]).Named("nonZeroUserFeature");
var nonZeroUserFeatureIndices = Variable.Observed(default(IList <IList <int> >), user, nonZeroUserFeature).Named("NonZeroUserFeatureIndices");
var nonZeroUserFeatureValues = Variable.Observed(default(IList <IList <double> >), user, nonZeroUserFeature).Named("NonZeroUserFeatureValues");
// Feature weights
var userTraitFeatureWeightVariance = Variable.Observed(default(double)).Named("UserTraitFeatureWeightPriorVariance");
var userBiasFeatureWeightVariance = Variable.Observed(default(double)).Named("UserBiasFeatureWeightPriorVariance");
userTraitFeatureWeights[trait][userFeature] =
Variable.GaussianFromMeanAndVariance(0.0, userTraitFeatureWeightVariance).ForEach(userFeature).ForEach(trait);
userBiasFeatureWeights[userFeature] =
Variable.GaussianFromMeanAndVariance(0.0, userBiasFeatureWeightVariance).ForEach(userFeature);
// Traits, biases and thresholds
var userTraitVariance = Variable.Observed(default(double)).Named("UserTraitVariance");
var userBiasVariance = Variable.Observed(default(double)).Named("UserBiasVariance");
var userThresholdPriorMean = Variable.Observed(default(double[]), userThreshold).Named("UserThresholdPriorMean");
var userThresholdPriorVariance = Variable.Observed(default(double)).Named("UserThresholdPriorVariance");
var middleUserThresholdIndex = Variable.Observed(default(int)).Named("MiddleUserThresholdIndex");
using (var userThresholdIterationBlock = Variable.ForEach(userThreshold))
{
var middleThreshold = userThresholdIterationBlock.Index == middleUserThresholdIndex;
var userThresholdPriorVarianceForThreshold = Variable.New <double>().Named("UserThresholdPriorVarianceForThreshold");
using (Variable.If(middleThreshold))
{
// Break symmetry between the user thresholds and the user bias
userThresholdPriorVarianceForThreshold.SetTo(0.0);
}
using (Variable.IfNot(middleThreshold))
{
userThresholdPriorVarianceForThreshold.SetTo(userThresholdPriorVariance);
}
userThresholds[user][userThresholdIterationBlock.Index] = Variable.GaussianFromMeanAndVariance(
userThresholdPriorMean[userThresholdIterationBlock.Index], userThresholdPriorVarianceForThreshold).ForEach(user);
}
using (Variable.ForEach(user))
{
var nonZeroUserTraitFeatureWeights = Variable.Subarray(userTraitFeatureWeights[trait], nonZeroUserFeatureIndices[user]);
nonZeroUserTraitFeatureWeights.Name = "nonZeroUserTraitFeatureWeights";
var nonZeroUserTraitFeatureWeightProducts = Variable.Array(Variable.Array <double>(nonZeroUserFeature), trait).Named("nonZeroUserTraitFeatureWeightProducts");
nonZeroUserTraitFeatureWeightProducts[trait][nonZeroUserFeature] =
nonZeroUserTraitFeatureWeights[nonZeroUserFeature] * nonZeroUserFeatureValues[user][nonZeroUserFeature];
var userTraitMean = Variable.Sum(nonZeroUserTraitFeatureWeightProducts[trait]).Named("userTraitMean");
userTraits[user][trait] = Variable.GaussianFromMeanAndVariance(userTraitMean, userTraitVariance);
var nonZeroUserBiasFeatureWeights = Variable.Subarray(userBiasFeatureWeights, nonZeroUserFeatureIndices[user]);
nonZeroUserBiasFeatureWeights.Name = "nonZeroUserBiasFeatureWeights";
var nonZeroUserBiasFeatureWeightProducts = Variable.Array <double>(nonZeroUserFeature).Named("nonZeroUserBiasFeatureWeightProducts");
nonZeroUserBiasFeatureWeightProducts[nonZeroUserFeature] =
nonZeroUserBiasFeatureWeights[nonZeroUserFeature] * nonZeroUserFeatureValues[user][nonZeroUserFeature];
var userBiasMean = Variable.Sum(nonZeroUserBiasFeatureWeightProducts).Named("userBiasMean");
userBias[user] = Variable.GaussianFromMeanAndVariance(userBiasMean, userBiasVariance);
}
// Trait, bias and threshold messages
var userTraitsMessage = Variable.Observed(default(GaussianMatrix)).Named("UserTraitsMessage");
var userBiasMessage = Variable.Observed(default(GaussianArray)).Named("UserBiasMessage");
var userThresholdsMessage = Variable.Observed(default(GaussianMatrix)).Named("UserThresholdsMessage");
Variable.ConstrainEqualRandom <double[][], GaussianMatrix>(userTraits, userTraitsMessage);
Variable.ConstrainEqualRandom <double[], GaussianArray>(userBias, userBiasMessage);
Variable.ConstrainEqualRandom <double[][], GaussianMatrix>(userThresholds, userThresholdsMessage);
userTraits.AddAttribute(QueryTypes.MarginalDividedByPrior);
userBias.AddAttribute(QueryTypes.MarginalDividedByPrior);
userThresholds.AddAttribute(QueryTypes.MarginalDividedByPrior);
userTraits.AddAttribute(QueryTypes.Marginal);
userBias.AddAttribute(QueryTypes.Marginal);
userThresholds.AddAttribute(QueryTypes.Marginal);
//// User thresholds do not have the initialize backward attribute because we want the
//// schedule to send their backward message before the affinity backward message.
if (buildTrainingModel)
{
// Trait, bias and threshold initializers
var userTraitsInitializer = Variable.Observed(default(GaussianMatrix)).Named("UserTraitsInitializer");
var userBiasInitializer = Variable.Observed(default(GaussianArray)).Named("UserBiasInitializer");
var userThresholdsInitializer = Variable.Observed(default(GaussianMatrix)).Named("UserThresholdsInitializer");
userTraits.InitialiseTo(userTraitsInitializer);
userBias.InitialiseTo(userBiasInitializer);
userThresholds.InitialiseTo(userThresholdsInitializer);
}
}
// Item latent variable definitions
var itemFeatureCount = Variable.Observed(default(int)).Named("ItemFeatureCount");
var itemFeature = new Range(itemFeatureCount).Named("itemFeature");
var itemTraitFeatureWeights = Variable.Array(Variable.Array <double>(itemFeature), trait).Named("ItemTraitFeatureWeights");
var itemBiasFeatureWeights = Variable.Array <double>(itemFeature).Named("ItemBiasFeatureWeights");
if (usePreAdjustedItemParameters)
{
// Define using custom priors
var itemTraitsPrior = Variable.Observed(default(GaussianMatrix)).Named("ItemTraitsPrior");
var itemBiasPrior = Variable.Observed(default(GaussianArray)).Named("ItemBiasPrior");
itemTraits.SetTo(Variable <double[][]> .Random(itemTraitsPrior));
itemBias.SetTo(Variable <double[]> .Random(itemBiasPrior));
}
else
{
// Features
var nonZeroItemFeatureCounts = Variable.Observed(default(IList <int>), item).Named("NonZeroItemFeatureCounts");
var nonZeroItemFeature = new Range(nonZeroItemFeatureCounts[item]).Named("nonZeroItemFeature");
var nonZeroItemFeatureIndices = Variable.Observed(default(IList <IList <int> >), item, nonZeroItemFeature).Named("NonZeroItemFeatureIndices");
var nonZeroItemFeatureValues = Variable.Observed(default(IList <IList <double> >), item, nonZeroItemFeature).Named("NonZeroItemFeatureValues");
// Feature weights
var itemTraitFeatureWeightVariance = Variable.Observed(default(double)).Named("ItemTraitFeatureWeightPriorVariance");
var itemBiasFeatureWeightVariance = Variable.Observed(default(double)).Named("ItemBiasFeatureWeightPriorVariance");
itemTraitFeatureWeights[trait][itemFeature] =
Variable.GaussianFromMeanAndVariance(0.0, itemTraitFeatureWeightVariance).ForEach(itemFeature).ForEach(trait);
itemBiasFeatureWeights[itemFeature] =
Variable.GaussianFromMeanAndVariance(0.0, itemBiasFeatureWeightVariance).ForEach(itemFeature);
// Traits and biases
var itemTraitVariance = Variable.Observed(default(double)).Named("ItemTraitVariance");
var itemBiasVariance = Variable.Observed(default(double)).Named("ItemBiasVariance");
using (Variable.ForEach(item))
{
var nonZeroItemTraitFeatureWeights = Variable.Subarray(itemTraitFeatureWeights[trait], nonZeroItemFeatureIndices[item]);
nonZeroItemTraitFeatureWeights.Name = "nonZeroItemTraitFeatureWeights";
var nonZeroItemTraitFeatureWeightProducts = Variable.Array(Variable.Array <double>(nonZeroItemFeature), trait);
nonZeroItemTraitFeatureWeightProducts[trait][nonZeroItemFeature] =
nonZeroItemTraitFeatureWeights[nonZeroItemFeature] * nonZeroItemFeatureValues[item][nonZeroItemFeature];
var itemTraitMean = Variable.Sum(nonZeroItemTraitFeatureWeightProducts[trait]).Named("itemTraitMean");
itemTraits[item][trait] = Variable.GaussianFromMeanAndVariance(itemTraitMean, itemTraitVariance);
var nonZeroItemBiasFeatureWeights = Variable.Subarray(itemBiasFeatureWeights, nonZeroItemFeatureIndices[item]);
nonZeroItemBiasFeatureWeights.Name = "nonZeroItemBiasFeatureWeights";
var nonZeroItemBiasFeatureWeightProducts = Variable.Array <double>(nonZeroItemFeature);
nonZeroItemBiasFeatureWeightProducts[nonZeroItemFeature] =
nonZeroItemBiasFeatureWeights[nonZeroItemFeature] * nonZeroItemFeatureValues[item][nonZeroItemFeature];
var itemBiasMean = Variable.Sum(nonZeroItemBiasFeatureWeightProducts).Named("itemBiasMean");
itemBias[item] = Variable.GaussianFromMeanAndVariance(itemBiasMean, itemBiasVariance);
}
// Trait and bias messages
var itemTraitsMessage = Variable.Observed(default(GaussianMatrix)).Named("ItemTraitsMessage");
var itemBiasMessage = Variable.Observed(default(GaussianArray)).Named("ItemBiasMessage");
Variable.ConstrainEqualRandom <double[][], GaussianMatrix>(itemTraits, itemTraitsMessage);
Variable.ConstrainEqualRandom <double[], GaussianArray>(itemBias, itemBiasMessage);
itemTraits.AddAttribute(QueryTypes.MarginalDividedByPrior);
itemBias.AddAttribute(QueryTypes.MarginalDividedByPrior);
itemTraits.AddAttribute(QueryTypes.Marginal);
itemBias.AddAttribute(QueryTypes.Marginal);
if (buildTrainingModel)
{
// Trait and bias initializers
var itemTraitsInitializer = Variable.Observed(default(GaussianMatrix)).Named("ItemTraitsInitializer");
var itemBiasInitializer = Variable.Observed(default(GaussianArray)).Named("ItemBiasInitializer");
itemTraits.InitialiseTo(itemTraitsInitializer);
itemBias.InitialiseTo(itemBiasInitializer);
}
}
// Initialize first min(itemCount, traitCount) item traits to identity matrix to break symmetry
if (breakTraitSymmetry)
{
// Note that this code affects evidence computation
using (var traitIterationBlock1 = Variable.ForEach(trait))
{
var traitCopy = trait.Clone();
using (var traitIterationBlock2 = Variable.ForEach(traitCopy))
{
using (Variable.If(traitIterationBlock1.Index < itemCount & traitIterationBlock2.Index < itemCount))
{
var diagonalElement = traitIterationBlock1.Index == traitIterationBlock2.Index;
using (Variable.If(diagonalElement))
{
Variable.ConstrainEqual(itemTraits[traitIterationBlock1.Index][traitIterationBlock2.Index], 1);
}
using (Variable.IfNot(diagonalElement))
{
Variable.ConstrainEqual(itemTraits[traitIterationBlock1.Index][traitIterationBlock2.Index], 0);
}
}
}
}
}
// Observation data
var userIds = Variable.Observed(default(IList <int>), observation).Named("UserIds");
var itemIds = Variable.Observed(default(IList <int>), observation).Named("ItemIds");
var ratings = Variable.IList <int>(observation).Named("Ratings");
if (buildTrainingModel)
{
ratings.ObservedValue = default(IList <int>);
ratings.SetValueRange(ratingValue);
}
else
{
ratings[observation] = Variable.DiscreteUniform(ratingValue).ForEach(observation);
}
// Noise variances
var affinityNoiseVariance = Variable.Observed(default(double)).Named("AffinityNoiseVariance");
var thresholdNoiseVariance = Variable.Observed(default(double)).Named("UserThresholdNoiseVariance");
// Connect everything to observations
// Send VMP messages here for correct message reconstruction in batching
using (Variable.ForEach(observation))
{
var userId = userIds[observation];
var itemId = itemIds[observation];
// Products of user and item traits (as if running VMP)
// Make a copy so that we can attach attributes to the argument of Factor.Product_SHG09, not the original itemTraits array.
var itemTrait = Variable.Copy(itemTraits[itemId][trait]).Named("itemTrait");
var userTrait = Variable.Copy(userTraits[userId][trait]).Named("userTrait");
var traitProducts = Variable.Array <double>(trait);
traitProducts[trait] = Variable <double> .Factor(Factor.Product_SHG09, userTrait, itemTrait);
// Affinity
var userBiasObs = Variable.Copy(userBias[userId]).Named("userBiasObs");
var itemBiasObs = Variable.Copy(itemBias[itemId]).Named("itemBiasObs");
var affinity =
Variable <double> .Factor(Factor.Product_SHG09, userBiasObs, 1.0)
+ Variable <double> .Factor(Factor.Product_SHG09, itemBiasObs, 1.0)
+ Variable.Sum(traitProducts);
var noisyAffinity = Variable.GaussianFromMeanAndVariance(affinity, affinityNoiseVariance);
// During prediction, we don't want to update the parameters. Variable.Cut ensures this.
if (!buildTrainingModel)
{
noisyAffinity = Variable.Cut(noisyAffinity);
}
// User thresholds
var useSharedUserThresholds = Variable.Observed(default(bool)).Named("UseSharedUserThresholds");
var userThresholdsObs = Variable.Array <double>(userThreshold).Named("UserThresholdsObs");
using (Variable.If(useSharedUserThresholds))
{
userThresholdsObs.SetTo(Variable.Copy(userThresholds[0]));
}
using (Variable.IfNot(useSharedUserThresholds))
{
userThresholdsObs.SetTo(Variable.Copy(userThresholds[userId]));
}
var noisyUserThresholds = Variable.Array <double>(userThreshold).Named("NoisyUserThresholds");
var noisyUserThreshold =
Variable.GaussianFromMeanAndVariance(
Variable <double> .Factor(Factor.Product_SHG09, userThresholdsObs[userThreshold], 1.0),
thresholdNoiseVariance);
if (!buildTrainingModel)
{
noisyUserThreshold = Variable.Cut(noisyUserThreshold);
}
noisyUserThresholds[userThreshold] = noisyUserThreshold;
if (buildTrainingModel)
{
// Every argument to Factor.Product_SHG09 should have an InitialiseBackward attribute, for the special first iteration.
// They should also be initialized to the current marginal distribution.
// This ensures that the factor receives the marginal during the special first iteration.
itemTrait.AddAttribute(new InitialiseBackward());
userTrait.AddAttribute(new InitialiseBackward());
userBiasObs.AddAttribute(new InitialiseBackward());
itemBiasObs.AddAttribute(new InitialiseBackward());
userThresholdsObs.AddAttribute(new InitialiseBackward());
}
// Rating
var rating = ratings[observation];
using (Variable.Switch(rating))
{
// This hack allows indexing the thresholds with the ratingValue range instead of the userThreshold range
var currentRating = (rating + 0).Named("CurrentRating");
var nextRating = (rating + 1).Named("NextRating");
Variable.ConstrainBetween(
noisyAffinity, noisyUserThresholds[currentRating], noisyUserThresholds[nextRating]);
}
}
IVariable[] result;
if (buildTrainingModel)
{
// Community training
result = new IVariable[]
{
userTraits, userBias, userThresholds, itemTraits, itemBias, userTraitFeatureWeights,
userBiasFeatureWeights, itemTraitFeatureWeights, itemBiasFeatureWeights
};
}
else
{
// Prediction
result = new IVariable[] { ratings };
}
return(result);
}
public Model(ISharedVariableArray<VariableArray<double>, double[][]> w, Range c, int numChunks)
{
// Items.
numItems = Variable.New<int>().Named("numItems");
i = new Range(numItems).Named("i");
i.AddAttribute(new Sequential());
// Features per item.
numFeaturesPerItem = Variable.Array<int>(i).Named("numFeaturesPerItem");
fItem = new Range(numFeaturesPerItem[i]).Named("fItem");
// The model identifier for the shared variables.
model = new MicrosoftResearch.Infer.Models.Model(numChunks).Named("model");
// The weight vector for each submodel.
wModel = w.GetCopyFor(model).Named("wModel");
noisePrecision = Variable.New<double>().Named("noisePrecision");
// Jagged array of feature values - each item is an array of data values
// whose indices are given by the corresponding indices[i].
values = Variable.Array(Variable.Array<double>(fItem), i).Named("values");
// Jagged array of indices for the items.
indices = Variable.Array(Variable.Array<int>(fItem), i).Named("indices");
// Labels.
y = Variable.Array<int>(i).Named("y");
// For all items...
using (Variable.ForEach(i))
{
// ...compute the score of this item across all classes...
score = BPMUtils.ComputeClassScores(wModel, values[i], indices[i], fItem, noisePrecision);
y[i] = Variable.DiscreteUniform(c);
// ... and constrain the output.
BPMUtils.ConstrainMaximum(y[i], score);
}
// Inference engine settings (EP).
engine.Compiler.UseSerialSchedules = true;
engine.ShowProgress = false;
}