/// <summary>
/// Runs the training algorithm on multiple batches of instances.
/// </summary>
/// <param name="instanceSource">The source of instances.</param>
/// <param name="labelSource">An optional source of labels.</param>
private void TrainOnMultipleBatches(TInstanceSource instanceSource, TLabelSource labelSource)
{
double logEvidence = 0.0;
// Retrieve data properties through the mapping
bool isSparse = this.Mapping.IsSparse(instanceSource);
int featureCount = this.Mapping.GetFeatureCountSafe(instanceSource);
// Create the inference algorithms, if required
if (!this.IsTrained)
{
this.InferenceAlgorithms = this.CreateInferenceAlgorithms(isSparse, featureCount);
}
// Set the number of training data batches and reset per-batch constraint distributions
this.InferenceAlgorithms.SetBatchCount(this.Settings.Training.BatchCount);
for (int iteration = 1; iteration <= this.Settings.Training.IterationCount; iteration++)
{
for (int batch = 0; batch < this.Settings.Training.BatchCount; batch++)
{
// Retrieve the feature values, indexes and labels through the mapping
double[][] featureValues = this.Mapping.GetAllFeatureVectorValuesSafe(instanceSource, batch);
int[][] featureIndexes = this.Mapping.GetAllFeatureVectorIndexesSafe(instanceSource, batch);
TLabel[] labels = this.Mapping.GetAllLabelsSafe(instanceSource, labelSource, batch);
// Check training data is self-consistent
this.CheckDataConsistency(isSparse, featureCount, featureValues, featureIndexes, labels);
// Check features are consistent with the inference algorithms
this.CheckFeatureAlgorithmConsistency(isSparse, featureCount);
// Run the training algorithm on the specified batch
int batchedIterationCount = iteration < MaxBatchedIterationCount ? iteration : MaxBatchedIterationCount;
double logBatchEvidence = this.InferenceAlgorithms.Train(
featureValues, featureIndexes, labels, batchedIterationCount, batch);
// Last iteration: compute evidence on the specified batch
if (iteration == this.Settings.Training.IterationCount)
{
logEvidence += logBatchEvidence;
}
}
// Raise IterationChanged event
this.OnBatchedIterationChanged(iteration);
}
this.logModelEvidence += logEvidence;
// Mark classifier as trained
if (this.IsTrained)
{
this.isIncrementallyTrained = true;
}
this.IsTrained = true;
}
/// <summary>
/// Runs the training algorithm on a single batch of instances.
/// </summary>
/// <param name="instanceSource">The source of instances.</param>
/// <param name="labelSource">An optional source of labels.</param>
private void TrainOnSingleBatch(TInstanceSource instanceSource, TLabelSource labelSource)
{
// Retrieve the data through the mapping
bool isSparse = this.Mapping.IsSparse(instanceSource);
int featureCount = this.Mapping.GetFeatureCountSafe(instanceSource);
double[][] featureValues = this.Mapping.GetAllFeatureVectorValuesSafe(instanceSource);
int[][] featureIndexes = this.Mapping.GetAllFeatureVectorIndexesSafe(instanceSource);
TLabel[] labels = this.Mapping.GetAllLabelsSafe(instanceSource, labelSource);
// Check training data is self-consistent
this.CheckDataConsistency(isSparse, featureCount, featureValues, featureIndexes, labels);
// Create the inference algorithms, if required
if (!this.IsTrained)
{
this.InferenceAlgorithms = this.CreateInferenceAlgorithms(isSparse, featureCount);
}
// Set the number of training data batches and reset per-batch output messages
this.InferenceAlgorithms.SetBatchCount(this.Settings.Training.BatchCount);
// Check features are consistent with the inference algorithms
this.CheckFeatureAlgorithmConsistency(isSparse, featureCount);
// Subscribe to the IterationChanged event on the training algorithm
bool subscribed = this.TrySubscribeToIterationChangedEvent();
// Run the training algorithm
this.logModelEvidence = this.InferenceAlgorithms.Train(
featureValues, featureIndexes, labels, this.Settings.Training.IterationCount);
// Unsubscribe from IterationChanged event
this.UnsubscribeFromIterationChangedEvent(subscribed);
// Mark classifier as trained
if (this.IsTrained)
{
this.isIncrementallyTrained = true;
}
this.IsTrained = true;
}
