protected IGeneratedAlgorithm InferAll(bool inferOnlySpecifiedVars, IEnumerable <IVariable> vars)
{
IGeneratedAlgorithm ca = GetCompiledInferenceAlgorithm(inferOnlySpecifiedVars, vars);
Execute(ca);
return(ca);
}
/// <summary>
/// Infer the output messages, for the given model and batch number, for all
/// shared variables in this set
/// </summary>
/// <param name="ca">Compiled algorithm</param>
/// <param name="modelNumber">Model id</param>
/// <param name="batchNumber">Batch number</param>
public void InferOutput(IGeneratedAlgorithm ca, Model modelNumber, int batchNumber)
{
foreach (ISharedVariable v in this)
{
v.InferOutput(ca, modelNumber, batchNumber);
}
}
/// <summary>
/// Handles the engine message updated.
/// </summary>
/// <param name="algorithm">The algorithm.</param>
/// <param name="messageEvent">The <see cref="MessageUpdatedEventArgs"/> instance containing the event data.</param>
private void EngineMessageUpdated(IGeneratedAlgorithm algorithm, MessageUpdatedEventArgs messageEvent)
{
if (!this.MessageHistories.ContainsKey(messageEvent.MessageId))
{
this.MessageHistories[messageEvent.MessageId] = new List <Bernoulli>();
}
// Console.WriteLine(messageEvent);
if (messageEvent.Message is Bernoulli)
{
this.MessageHistories[messageEvent.MessageId].Add((Bernoulli)messageEvent.Message);
}
}
/// <summary>
/// Message updated.
/// </summary>
/// <param name="algorithm">The algorithm.</param>
/// <param name="messageEvent">The <see cref="MessageUpdatedEventArgs"/> instance containing the event data.</param>
private void MessageUpdated(IGeneratedAlgorithm algorithm, MessageUpdatedEventArgs messageEvent)
{
if (!this.MessageHistories.ContainsKey(messageEvent.MessageId))
{
this.MessageHistories[messageEvent.MessageId] = new List <Gaussian>();
}
if (messageEvent.Message is Gaussian item)
{
this.MessageHistories[messageEvent.MessageId].Add(item);
}
// Console.WriteLine(messageEvent);
}
private void Execute(IGeneratedAlgorithm ca)
{
// If there is a message update listener, try to add in the engine to listen to messages.
if (this.MessageUpdated != null)
{
DebuggingSupport.TryAddRemoveEventListenerDynamic(ca, OnMessageUpdated, add: true);
}
// Register the ProgressChanged handler only while doing inference within InferenceEngine.
// We do not want the handler to run if the user accesses the GeneratedAlgorithms directly.
ca.ProgressChanged += OnProgressChanged;
try
{
Stopwatch s = null;
if (ShowTimings)
{
s = new Stopwatch();
s.Start();
FileStats.Clear();
}
if (ResetOnObservedValueChanged)
{
ca.Execute(NumberOfIterations);
}
else
{
ca.Update(NumberOfIterations - ca.NumberOfIterationsDone);
}
if (s != null)
{
long elapsed = s.ElapsedMilliseconds;
Console.WriteLine("Inference time was {1}ms (max {0} iterations)",
NumberOfIterations, elapsed);
if (FileStats.ReadCount > 0 || FileStats.WriteCount > 0)
{
Console.WriteLine("{0} file reads {1} file writes", FileStats.ReadCount, FileStats.WriteCount);
}
}
}
finally
{
ca.ProgressChanged -= OnProgressChanged;
if (this.MessageUpdated != null)
{
DebuggingSupport.TryAddRemoveEventListenerDynamic(ca, OnMessageUpdated, add: false);
}
}
}
/// <summary>
/// Tries to dynamically add or remove an event handler to a generated algorithm instance.
/// </summary>
/// <param name="ca">The generated algorithm instance</param>
/// <param name="d">The event handler to add or remove</param>
/// <param name="add">If true will add, otherwise will remove</param>
/// <returns>True if the event handler was added or removed successfully</returns>
internal static bool TryAddRemoveEventListenerDynamic(IGeneratedAlgorithm ca, EventHandler <MessageUpdatedEventArgs> d, bool add)
{
var eventInfo = ca.GetType().GetEvent(MessageEventName);
if (eventInfo == null)
{
return(false);
}
if (add)
{
eventInfo.AddEventHandler(ca, d);
}
else
{
eventInfo.RemoveEventHandler(ca, d);
}
return(true);
}
/// <summary>
/// Compiles the last built model into a CompiledAlgorithm which implements
/// the specified inference algorithm on the model.
/// </summary>
/// <returns></returns>
private IGeneratedAlgorithm Compile()
{
mb.SetModelName(ModelNamespace, ModelName);
if (ShowMsl)
{
Console.WriteLine(mb.ModelString());
}
if (ShowFactorGraph || SaveFactorGraphToFolder != null)
{
if (SaveFactorGraphToFolder != null && Visualizer?.GraphWriter != null)
{
Directory.CreateDirectory(SaveFactorGraphToFolder);
Visualizer.GraphWriter.WriteGraph(mb, SaveFactorGraphToFolder + @"\" + ModelName);
}
if (ShowFactorGraph && Visualizer?.FactorGraphVisualizer != null)
{
Visualizer.FactorGraphVisualizer.VisualizeFactorGraph(mb);
}
}
Stopwatch s = null;
if (ShowTimings)
{
s = new Stopwatch();
s.Start();
}
IGeneratedAlgorithm compiledAlgorithm = Compiler.CompileWithoutParams(mb.modelType, null, mb.Attributes);
if (ShowTimings)
{
s.Stop();
Console.WriteLine("Compilation time was " + s.ElapsedMilliseconds + "ms.");
}
CompiledAlgorithmInfo info = new CompiledAlgorithmInfo(compiledAlgorithm, mb.observedVars);
compiledAlgorithms.Push(info);
foreach (IVariable v in mb.variablesToInfer)
{
compiledAlgorithmForVariable[v] = info;
}
SetObservedValues(info);
return(info.exec);
}
/// <summary>
/// Initializes a new instance of the <see cref="BinaryModel"/> class.
/// </summary>
/// <param name="trainModel">If set to <c>true</c> train model.</param>
/// <param name="showFactorGraph">If set to <c>true</c> show factor graph.</param>
/// <param name="debug">If set to <c>true</c> debug.</param>
public BinaryModel(bool trainModel, bool showFactorGraph = false, bool debug = false)
{
evidence = Variable.Bernoulli(0.5).Named("evidence");
using (Variable.If(evidence))
{
numberOfResidents = Variable.New<int>().Named("numberOfResidents").Attrib(new DoNotInfer());
numberOfFeatures = Variable.New<int>().Named("numberOfFeatures").Attrib(new DoNotInfer());
var resident = new Range(numberOfResidents).Named("resident");
var feature = new Range(numberOfFeatures).Named("feature");
numberOfExamples = Variable.Array<int>(resident).Named("numberOfExamples").Attrib(new DoNotInfer());
var example = new Range(numberOfExamples[resident]).Named("example").Attrib(new Sequential());
noisePrecision = Variable.New<double>().Named("noisePrecision").Attrib(new DoNotInfer());
weightPriorMeans = Variable.New<GaussianArray>().Named("weightPriorMeans").Attrib(new DoNotInfer());
weightPriorPrecisions = Variable.New<GammaArray>().Named("weightPriorPrecisions").Attrib(new DoNotInfer());
weightMeans = Variable.Array<double>(feature).Named("weightMeans");
weightPrecisions = Variable.Array<double>(feature).Named("weightPrecisions");
weightMeans.SetTo(Variable<double[]>.Random(weightPriorMeans));
weightPrecisions.SetTo(Variable<double[]>.Random(weightPriorPrecisions));
weights = Variable.Array(Variable.Array<double>(feature), resident).Named("weights");
featureValues = Variable.Array(Variable.Array(Variable.Array<double>(feature), example), resident).Named("featureValues").Attrib(new DoNotInfer());
// if (!useBias)
// {
// thresholdPriors = Variable.New<GaussianArray>().Named("thresholdPrior").Attrib(new DoNotInfer());
// thresholds = Variable.Array<double>(resident).Named("threshold");
// thresholds.SetTo(Variable<double[]>.Random(thresholdPriors));
// }
activities = Variable.Array(Variable.Array<bool>(example), resident).Named("activities");
// activities[resident][example].AddAttribute(new MarginalPrototype(new Bernoulli()));
using (Variable.ForEach(resident))
{
var products = Variable.Array(Variable.Array<double>(feature), example).Named("products");
var scores = Variable.Array<double>(example).Named("scores");
var scoresPlusNoise = Variable.Array<double>(example).Named("scoresPlusNoise");
weights[resident][feature] = Variable.GaussianFromMeanAndPrecision(weightMeans[feature], weightPrecisions[feature]);
using (Variable.ForEach(example))
{
using (Variable.ForEach(feature))
{
products[example][feature] = weights[resident][feature] * featureValues[resident][example][feature];
}
scores[example] = Variable.Sum(products[example]).Named("score");
scoresPlusNoise[example] = Variable.GaussianFromMeanAndPrecision(scores[example], noisePrecision).Named("scorePlusNoise");
// if (useBias)
{
activities[resident][example] = scoresPlusNoise[example] > 0;
}
// else
// {
// var diff = (scoresPlusNoise[example] - thresholds[resident]).Named("diff");
// activities[example][resident] = diff > 0;
// }
}
}
}
engine = new InferenceEngine
{
Algorithm = new ExpectationPropagation { DefaultNumberOfIterations = trainModel ? 10 : 1 },
ShowFactorGraph = showFactorGraph,
ShowProgress = false,
// BrowserMode = BrowserMode.Never, // debug ? BrowserMode.OnError : BrowserMode.Never,
ShowWarnings = debug
};
if (debug)
{
engine.Compiler.GenerateInMemory = false;
engine.Compiler.WriteSourceFiles = true;
engine.Compiler.IncludeDebugInformation = true;
engine.Compiler.CatchExceptions = true;
}
#if USE_PRECOMPILED_ALGORITHM
numberOfResidents.ObservedValue = default(int);
numberOfExamples.ObservedValue = default(int);
numberOfFeatures.ObservedValue = default(int);
noisePrecision.ObservedValue = default(double);
featureValues.ObservedValue = default(double[][][]);
weightPriorMeans.ObservedValue = default(DistributionStructArray<Gaussian, double>); // (DistributionStructArray<Gaussian, double>)Distribution<double>.Array(default(Gaussian[]));
weightPriorPrecisions.ObservedValue = default(DistributionStructArray<Gamma, double>); // (DistributionStructArray<Gamma, double>)Distribution<double>.Array(default(Gamma[]));
activities.ObservedValue = default(bool[][]);
if (trainModel)
{
activities.AddAttribute(new DoNotInfer());
algorithm = engine.GetCompiledInferenceAlgorithm(new IVariable[] { weights, weightMeans, weightPrecisions });
}
else
{
activities.AddAttribute(QueryTypes.Marginal);
algorithm = engine.GetCompiledInferenceAlgorithm(new IVariable[] { activities });
}
#endif
}
/// <summary>
/// Infers the marginal distribution for the specified variable.
/// </summary>
/// <typeparam name="TReturn">Desired return type which may be a distribution type or an array type if the argument is a VariableArray</typeparam>
/// <param name="var">The variable whose marginal is to be inferred</param>
/// <returns>The marginal distribution (or an approximation to it)</returns>
public TReturn Infer <TReturn>(IVariable var)
{
IGeneratedAlgorithm ca = InferAll(false, var);
return(ca.Marginal <TReturn>(var.Name));
}
/// <summary>
/// Infers the marginal distribution for the specified variable.
/// </summary>
/// <param name="var">The variable whose marginal is to be inferred</param>
/// <returns>The marginal distribution (or an approximation to it)</returns>
public object Infer(IVariable var)
{
IGeneratedAlgorithm ca = InferAll(false, var);
return(ca.Marginal(var.Name));
}
public CompiledAlgorithmInfo(IGeneratedAlgorithm exec, IEnumerable <Variable> observedVars)
{
this.exec = exec;
this.observedVarsInOrder.AddRange(observedVars);
this.observedVars.AddRange(observedVars);
}
/// <summary>
/// Update all the SharedVariables registered with this model.
/// </summary>
/// <param name="engine"></param>
/// <param name="batchNumber">A number from 0 to BatchCount-1</param>
public void InferShared(IGeneratedAlgorithm engine, int batchNumber)
{
SharedVariables.SetInput(this, batchNumber);
SharedVariables.InferOutput(engine, this, batchNumber);
}
/// <summary> /// Infer the shared variable's output message for the given model and batch number. /// </summary> /// <param name="ca">The compiled algorithm.</param> /// <param name="modelNumber">The model id.</param> /// <param name="batchNumber">The batch number.</param> public abstract void InferOutput(IGeneratedAlgorithm ca, Model modelNumber, int batchNumber);
/// <summary>
/// Initializes a new instance of the <see cref="BinaryModel"/> class.
/// </summary>
/// <param name="trainModel">If set to <c>true</c> train model.</param>
/// <param name="showFactorGraph">If set to <c>true</c> show factor graph.</param>
/// <param name="debug">If set to <c>true</c> debug.</param>
public BinaryModel(bool trainModel, bool showFactorGraph = false, bool debug = false)
{
evidence = Variable.Bernoulli(0.5).Named("evidence");
using (Variable.If(evidence))
{
numberOfResidents = Variable.New <int>().Named("numberOfResidents").Attrib(new DoNotInfer());
numberOfFeatures = Variable.New <int>().Named("numberOfFeatures").Attrib(new DoNotInfer());
var resident = new Range(numberOfResidents).Named("resident");
var feature = new Range(numberOfFeatures).Named("feature");
numberOfExamples = Variable.Array <int>(resident).Named("numberOfExamples").Attrib(new DoNotInfer());
var example = new Range(numberOfExamples[resident]).Named("example").Attrib(new Sequential());
noisePrecision = Variable.New <double>().Named("noisePrecision").Attrib(new DoNotInfer());
weightPriorMeans = Variable.New <GaussianArray>().Named("weightPriorMeans").Attrib(new DoNotInfer());
weightPriorPrecisions = Variable.New <GammaArray>().Named("weightPriorPrecisions").Attrib(new DoNotInfer());
weightMeans = Variable.Array <double>(feature).Named("weightMeans");
weightPrecisions = Variable.Array <double>(feature).Named("weightPrecisions");
weightMeans.SetTo(Variable <double[]> .Random(weightPriorMeans));
weightPrecisions.SetTo(Variable <double[]> .Random(weightPriorPrecisions));
weights = Variable.Array(Variable.Array <double>(feature), resident).Named("weights");
featureValues = Variable.Array(Variable.Array(Variable.Array <double>(feature), example), resident).Named("featureValues").Attrib(new DoNotInfer());
// if (!useBias)
// {
// thresholdPriors = Variable.New<GaussianArray>().Named("thresholdPrior").Attrib(new DoNotInfer());
// thresholds = Variable.Array<double>(resident).Named("threshold");
// thresholds.SetTo(Variable<double[]>.Random(thresholdPriors));
// }
activities = Variable.Array(Variable.Array <bool>(example), resident).Named("activities");
// activities[resident][example].AddAttribute(new MarginalPrototype(new Bernoulli()));
using (Variable.ForEach(resident))
{
var products = Variable.Array(Variable.Array <double>(feature), example).Named("products");
var scores = Variable.Array <double>(example).Named("scores");
var scoresPlusNoise = Variable.Array <double>(example).Named("scoresPlusNoise");
weights[resident][feature] = Variable.GaussianFromMeanAndPrecision(weightMeans[feature], weightPrecisions[feature]);
using (Variable.ForEach(example))
{
using (Variable.ForEach(feature))
{
products[example][feature] = weights[resident][feature] * featureValues[resident][example][feature];
}
scores[example] = Variable.Sum(products[example]).Named("score");
scoresPlusNoise[example] = Variable.GaussianFromMeanAndPrecision(scores[example], noisePrecision).Named("scorePlusNoise");
// if (useBias)
{
activities[resident][example] = scoresPlusNoise[example] > 0;
}
// else
// {
// var diff = (scoresPlusNoise[example] - thresholds[resident]).Named("diff");
// activities[example][resident] = diff > 0;
// }
}
}
}
engine = new InferenceEngine
{
Algorithm = new ExpectationPropagation {
DefaultNumberOfIterations = trainModel ? 10 : 1
},
ShowFactorGraph = showFactorGraph,
ShowProgress = false,
// BrowserMode = BrowserMode.Never, // debug ? BrowserMode.OnError : BrowserMode.Never,
ShowWarnings = debug
};
if (debug)
{
engine.Compiler.GenerateInMemory = false;
engine.Compiler.WriteSourceFiles = true;
engine.Compiler.IncludeDebugInformation = true;
engine.Compiler.CatchExceptions = true;
}
#if USE_PRECOMPILED_ALGORITHM
numberOfResidents.ObservedValue = default(int);
numberOfExamples.ObservedValue = default(int);
numberOfFeatures.ObservedValue = default(int);
noisePrecision.ObservedValue = default(double);
featureValues.ObservedValue = default(double[][][]);
weightPriorMeans.ObservedValue = default(DistributionStructArray <Gaussian, double>); // (DistributionStructArray<Gaussian, double>)Distribution<double>.Array(default(Gaussian[]));
weightPriorPrecisions.ObservedValue = default(DistributionStructArray <Gamma, double>); // (DistributionStructArray<Gamma, double>)Distribution<double>.Array(default(Gamma[]));
activities.ObservedValue = default(bool[][]);
if (trainModel)
{
activities.AddAttribute(new DoNotInfer());
algorithm = engine.GetCompiledInferenceAlgorithm(new IVariable[] { weights, weightMeans, weightPrecisions });
}
else
{
activities.AddAttribute(QueryTypes.Marginal);
algorithm = engine.GetCompiledInferenceAlgorithm(new IVariable[] { activities });
}
#endif
}