public TraModel()
{
this.PropertyChanged += TraModel_PropertyChanged;
trained = false;
_selectionTrainingPackage = new TrainingPackage();
_selectedTrainingSets = new ObservableCollection<DataSet>();
_progressLog = new ProgressLog();
_selectedTrainingSets.CollectionChanged += _selectedTrainingSets_CollectionChanged;
}
public override void Run()
{
double[] itemToClassify = new double[2] {2.81,5.46};
double[] result = new double[27];
_trainingPackage = MakeTrainingPackage();
//Now we inintialize and train the LDAPatternRecognizer.
//Input dimension is 2 because we have 2 features per feature vector. Output dimension is 27 because
//on the final application there will be 27 possible movements including rest (movement 0)
_patternRecognizer = GenericFactory<PatternRecognizer>.Instance.CreateProduct("LDA");//new LDAPatternRecognizer(_trainingPackage, 2, 27);
//_patternRecognizer.inputDim = 2;
//_patternRecognizer.outputDim = 27;
_patternRecognizer.trainingPackage = _trainingPackage;
_patternRecognizer.activationFunctionIdx = 0;
_patternRecognizer.normalizerIdx = 0;
_patternRecognizer.RunTraining();
result = (double[])_patternRecognizer.Classify(itemToClassify);
}
/// <summary>
/// Starting from data recorded or loaded at the acquisition stage, this method produces a training set for the
/// training stage
/// </summary>
public void ProcessRecording()
{
running = true;
_pretreatedRecording.parameters = _acqRecording.parameters;
_treatmentConfig.features = selectedFeatures.ToList<string>();
_treatmentConfig.features.Sort();
_featureExtractor.selectedFeatures = _treatmentConfig.features;
_trainingPackage = new TrainingPackage();
Pretreat();
_pipeline.Init();
_pipeline.Start();
foreach (object generic in _windowMaker.lockingCollection.GetConsumingEnumerable()) ;
//A dirty way of waiting for the WindowMaker stage to finish!
_pipeline.Stop();
_treatedWindows = _featureExtractor.outputData;
_trainingPackage.recordingConfig = _acqRecording.parameters;
//The following line is intended to make the DataProviders work as intended with online data.
_trainingPackage.recordingConfig.scheduleWarmupItems = 0;
_trainingPackage.treatmentConfig = _treatmentConfig;
BuildTrainingPackage();
treated = true;
running = false;
//And hopefully we have now a training package :)
}
/// <summary>
/// Initializes a new TrainingPackage, effectively dereferencing the existing one
/// </summary>
public void ClearTreatment()
{
_trainingPackage = new TrainingPackage();
treated = false;
}
public LDAPatternRecognizer(TrainingPackage trainingPackage)
: base(trainingPackage)
{
Init();
}
/// <summary>
/// Used to remove elements from the feature vectors belonging to channels that are not active.
/// Frames stored in each DataWindow are also discarded.
/// </summary>
/// <param name="inputDataSet"></param>
/// <returns>A DataSet object containing no frames and only features for the active channels.</returns>
private DataSet CreateLightDataSet(DataSet inputDataSet, TrainingPackage trainigPackage)
{
DataSet outputDataSet = new DataSet(inputDataSet.movementCode, inputDataSet.movementComposition);
foreach (DataWindow window in inputDataSet.set)
{
DataWindow lightWindow = new DataWindow();
foreach (string featureName in window.features.Keys)
{
double[] channelVector;
object placeholder;
double[] cleanVector = new double[trainingPackage.recordingConfig.activeChannels];
window.features.TryGetValue(featureName,out placeholder);
channelVector = (double[])placeholder;
int pos = 0;
for (int i = 0; i < channelVector.Length; i++)
{
if(trainingPackage.recordingConfig.channelMask[i])
{
cleanVector[pos] = channelVector[i];
pos++;
}
}
lightWindow.features.Add(featureName,cleanVector);
}
outputDataSet.set.Add(lightWindow);
}
return outputDataSet;
}
/// <summary>
/// Performs the training of the selected PatternRecognizer
/// </summary>
public void RunTraining()
{
trainingFailed = false;
if (_selectionTrainingPackage.trainingSets.Count > 0 && _patternRecognizer != null)
{
progressLog.logItems.Add(new ProgressLogItem(ProgressLogItem.Info, String.Format("Training {0} pattern recognizer", _patternRecognizer.GetType().GetProperty("ID").GetValue(null, null))));
//We shouldn't need this anymore. inputDim and outputDim are calculated ad the PatternRecognizer class
//when a new trainingPackage is assigned to the PatternRecognizer.
/*
_patternRecognizer.inputDim = _selectionTrainingPackage.trainingSets.ElementAt(0).set.ElementAt(0).features.Count * (int)_selectionTrainingPackage.recordingConfig.nChannels;
if (_patternRecognizer.multipleActivationEnabled) _patternRecognizer.outputDim = totalSingleMovements;
else _patternRecognizer.outputDim = totalMovements; //This must be taken from the ViewModel
*/
_patternRecognizer.trainingPackage = _selectionTrainingPackage;
_patternRecognizer.RunTraining();
progressLog.logItems.Add(new ProgressLogItem(ProgressLogItem.Info, String.Format("Training complete.")));
InstanceManager<PatternRecognizer>.Instance.Register(_patternRecognizer);
progressLog.logItems.Add(new ProgressLogItem(ProgressLogItem.Info, String.Format("List of trained pattern recognizers updated.")));
trained = true;
//We do this because from now on the used training package belongs to the trained PatternRecognizer only!
//So we initialize a new _selectionTrainingPackage for the next training
_selectionTrainingPackage = new TrainingPackage();
UpdateSelectionTrainingPackage();
// Training packages are no longer needed in this PatterRecognizer, so we clear them
_patternRecognizer.trainingPackage.trainingSets.Clear();
}
else
{
progressLog.logItems.Add(new ProgressLogItem(ProgressLogItem.Info, String.Format("Training: nothing to do!")));
trainingFailed = true;
}
}
public PatternRecognizer()
{
Init();
_trainingPackage = null;
_inputDim = 0;
_outputDim = 0;
}
/// <summary>
///
/// </summary>
/// <param name="newTrainingPackage">Reference to the TrainingPackage instance that the PatternRecognizer will use for training and verification.</param>
public PatternRecognizer(TrainingPackage newTrainingPackage)
{
Init();
_trainingPackage = newTrainingPackage;
}
/// <summary>
/// Performs a copy of the configuration values only, not copying any data.
/// </summary>
/// <param name="source"></param>
public void Copy(TrainingPackage source)
{
recordingConfig = source.recordingConfig;
treatmentConfig = source.treatmentConfig;
movementCodes.Clear();
foreach (int item in source.movementCodes)
movementCodes.Add(item);
}
/// <summary>
///Prepare the training package with a training set based on the example solved in
///http://people.revoledu.com/kardi/tutorial/LDA/Numerical%20Example.html
///One set of vectors will be assigned to movement code 1, the other to movement code 2.
/// </summary>
/// <returns></returns>
private TrainingPackage MakeTrainingPackage()
{
DataWindow tempWindow;
TrainingPackage trainingPackage;
List<DataWindow> windowList1 = new List<DataWindow>();
List<DataWindow> windowList2 = new List<DataWindow>();
//Filling the window lists with feature vectors
//List 1
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 2.95);
tempWindow.features.Add("diameter", 6.63);
windowList1.Add(tempWindow);
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 2.53);
tempWindow.features.Add("diameter", 7.79);
windowList1.Add(tempWindow);
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 3.57);
tempWindow.features.Add("diameter", 5.65);
windowList1.Add(tempWindow);
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 3.16);
tempWindow.features.Add("diameter", 5.47);
windowList1.Add(tempWindow);
//List 2
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 2.58);
tempWindow.features.Add("diameter", 4.46);
windowList2.Add(tempWindow);
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 2.16);
tempWindow.features.Add("diameter", 6.22);
windowList2.Add(tempWindow);
tempWindow = new DataWindow();
tempWindow.features.Add("curvature", 3.27);
tempWindow.features.Add("diameter", 3.52);
windowList2.Add(tempWindow);
trainingPackage = new TrainingPackage();
trainingPackage.trainingSets.Add(new DataSet(1, windowList1));
trainingPackage.trainingSets.Add(new DataSet(2, windowList2));
return trainingPackage;
}
