private void OnClassifyImage(object sender, EventArgs e)
{
try
{
_loggingService.Log("Classify drawing has started");
var imagePreprocessor = new ImagePreprocessor();
double[] pixels = imagePreprocessor.Preprocess(_uploadImageView.Image);
IPredictionModel predictionModel = Global.PredictionModel;
double[] prediction = predictionModel.Predict(pixels);
_uploadImageView.ProcessPrediction(prediction);
_loggingService.Log("Classify drawing has completed");
}
catch (NullReferenceException exception)
{
_loggingService.Log(exception);
_messageService.ShowMessage("No image was uploaded. Please upload an image and try again.", "Upload error", icon: MessageBoxIcon.Information);
}
catch (Exception exception)
{
_loggingService.Log(exception);
_messageService.ShowMessage("An error ocurred while classyfing the drawing. Please try again.", "Classification error", icon: MessageBoxIcon.Information);
}
}
protected virtual void ValidateModel(IPredictionModel model)
{
if (!(model is IKnnPredictionModel <TPredictionResult>))
{
throw new ArgumentException("Invalid prediction model passed for KNN predictor!");
}
}
public IList <TPredictionResult> Predict(IDataFrame queryDataFrame, IPredictionModel model, int dependentFeatureIndex)
{
ValidateModel(model);
var knnModel = model as IKnnPredictionModel <TPredictionResult>;
var results = new ConcurrentBag <RowIndexDistanceDto <TPredictionResult> >();
var normalizedData = NormalizeData(queryDataFrame, knnModel, dependentFeatureIndex);
var normalizedTrainingData = normalizedData.Item1;
var queryMatrix = normalizedData.Item2;
Parallel.For(0, queryDataFrame.RowCount, queryRowIdx =>
{
var rowVector = queryMatrix.Row(queryRowIdx);
var distances = new ConcurrentBag <RowIndexDistanceDto <TPredictionResult> >();
for (int trainingRowIdx = 0; trainingRowIdx < normalizedTrainingData.RowCount; trainingRowIdx++)
{
var trainingRow = normalizedTrainingData.Row(trainingRowIdx);
TPredictionResult dependentFeatureValue = knnModel.ExpectedTrainingOutcomes[trainingRowIdx];
double distance = DistanceMeasure.Distance(rowVector, trainingRow);
var distanceDto = new RowIndexDistanceDto <TPredictionResult>(trainingRowIdx, distance, dependentFeatureValue);
distances.Add(distanceDto);
}
var sortedDistances = distances.OrderBy(distDto => distDto.Distance).Take(knnModel.KNeighbors);
var result = new RowIndexDistanceDto <TPredictionResult>(queryRowIdx, 0, _resultHandler(sortedDistances, WeightingFunction));
results.Add(result);
});
return(results.OrderBy(res => res.RowIndex).Select(res => res.DependentFeatureValue).ToList());
}
private async void OnRunBenchmark(object sender, EventArgs e)
{
IPredictionModel predictionModel = Global.PredictionModel;
if (predictionModel == null)
{
_messageService.ShowMessage("The prediction model must be loaded first.", "Prediction model", icon: MessageBoxIcon.Information);
return;
}
_benchmarkView.ResetView();
_cancellationTokenSource = new CancellationTokenSource();
CancellationToken token = _cancellationTokenSource.Token;
token.Register(() => { _benchmarkView.IsBenchmarkRunning = false; });
try
{
await Task.Run(() => {
RunBenchmark(predictionModel);
}, token);
}
catch (Exception exception)
{
_loggingService.Log(exception);
_messageService.ShowMessage("An error ocurred while running the benchmark. Please try again.", "Benchmark error", icon: MessageBoxIcon.Information);
}
}
private void RunBenchmark(IPredictionModel model)
{
_loggingService.Log("Running benchmark has started");
_benchmarkView.IsBenchmarkRunning = true;
var provider = new BatchDataProvider(DirectoryHelper.TestLabelsPath, DirectoryHelper.TestImagesPath, 100);
var acc = 0;
for (var i = 0; i < 100; i++)
{
if (!_benchmarkView.IsBenchmarkRunning)
{
break;
}
MnistImageBatch data = provider.GetData();
int[] predictions = data.Pixels.Select(model.Predict).Select(x => x.ArgMax()).ToArray();
acc += data.Labels.Where((lbl, pred) => lbl == predictions[pred]).Count();
_benchmarkView.PerformProgressStep();
_benchmarkView.DrawGrid(new ImageGridModel(data, predictions));
}
_benchmarkView.SetAccuracy(acc);
_benchmarkView.IsBenchmarkRunning = false;
_loggingService.Log("Running benchmark has completed");
}
protected override void ValidateModel(IPredictionModel model)
{
if (!(model is IBackwardsEliminationKnnModel <TPredictionResult>))
{
throw new ArgumentException("Invalid model passed to Backwards Elimination KNN Predictor!");
}
}
public double computeRMSE(Users validationUsers, Items validationItems, IPredictionModel model)
{
double sse = 0;
double actualRating;
double predictedRating;
int n = 0;
foreach (User user in validationUsers)
{
foreach (string itemId in user.GetRatedItems())
{
Item item = validationItems.GetItemById(itemId);
actualRating = user.GetRating(itemId);
predictedRating = model.Predict(user, item);
if (predictedRating == -1) //in case the user and item are not exist the prediction is the average rating
{
predictedRating = avgRating;
}
sse += Math.Pow(actualRating - predictedRating, 2);
n++;
}
}
return(Math.Sqrt(sse / n));
}
public void addModel(RecommenderSystem.PredictionMethod method, IPredictionModel model)
{
if (!predictionMethodsDictionary.ContainsKey(method))
{
predictionMethodsDictionary.Add(method, model);
}
}
private static IPredictionModel LoadModel()
{
var loader = new PredictionModelLoader();
IPredictionModel model = loader.Load();
return(model);
}
public IPredictionModel Load()
{
if (DialogResult.OK != _openFileDialog.ShowDialog() || _openFileDialog.FileNames.Length == 0 || _openFileDialog.FileNames.Any(string.IsNullOrWhiteSpace))
{
return(null);
}
IPredictionModel model = ClusterPredictionModel.FromFiles(_openFileDialog.FileNames);
return(model);
}
private List <string> GetTopItems(IPredictionModel predictionModel, string sUserId, int cRecommendations)
{
var currentUser = testUsers.getUserById(sUserId);
var currentItems = trainUsers.getUserById(sUserId).GetRatedItems(); //in case the user is also in the training set we want to filter out those rated items from train set
var candidateItems = trainItems.GetAllItemsIds().Except(currentItems); //select items that current user is not yet rated
var candidateItemsDic = candidateItems.ToDictionary(item => item, item => predictionModel.Predict(currentUser, trainItems.GetItemById(item)));
var orderByPrediction = candidateItemsDic.OrderByDescending(item => item.Value);
return(orderByPrediction.Select(item => item.Key).Take(cRecommendations).ToList());
}
public IList<TPredictionVal> Predict(IDataFrame queryDataFrame, IPredictionModel model, string dependentFeatureName)
{
if (!(model is IRandomForestModel))
{
throw new ArgumentException("Invalid model passed to Random Forest Predictor");
}
var randomForestModel = (IRandomForestModel)model;
var weightedPredictons = new Tuple<IList<TPredictionVal>, double>[randomForestModel.DecisionTrees.Count];
Parallel.For(
0,
weightedPredictons.Length,
i =>
{
var predictions = decisionTreePredictor.Predict(
queryDataFrame,
randomForestModel.DecisionTrees[i],
dependentFeatureName);
var weight = 1 - randomForestModel.OutOfBagErrors[i];
weightedPredictons[i] = new Tuple<IList<TPredictionVal>, double>(predictions, weight);
});
var predictionVotes = new Dictionary<int, IDictionary<TPredictionVal, double>>();
foreach (var weightedPrediction in weightedPredictons)
{
for (int rowIdx = 0; rowIdx < queryDataFrame.RowCount; rowIdx++)
{
var predictedVal = weightedPrediction.Item1[rowIdx];
var weight = useVotingWeightedByOob ? weightedPrediction.Item2 : 1.0;
if (!predictionVotes.ContainsKey(rowIdx))
{
predictionVotes.Add(rowIdx, new Dictionary<TPredictionVal, double>());
}
if (!predictionVotes[rowIdx].ContainsKey(predictedVal))
{
predictionVotes[rowIdx].Add(predictedVal, 0.0);
}
predictionVotes[rowIdx][predictedVal] += weight;
}
}
var results = predictionVotes.Select(
rowVotes => rowVotes.Value.OrderByDescending(weightedPredictions => weightedPredictions.Value).First().Key)
.ToList();
return results;
}
private void OnClassifyDrawing(object sender, EventArgs e)
{
try
{
_loggingService.Log("Classify drawing has started");
var imagePreprocessor = new ImagePreprocessor();
IPredictionModel predictionModel = Global.PredictionModel;
Image img = _slidingWindowView.Drawing;
foreach (Size windowSize in WindowSizes)
{
foreach (BoundingBox boundingBox in ImageUtilities.SlidingWindow(img, windowSize, 112))
{
try
{
double[] pixels = imagePreprocessor.Preprocess(boundingBox.Image);
double[] prediction = predictionModel.Predict(pixels);
// If classification is over 99% draw a bounding box at this location
int predicted = prediction.ArgMax();
double predictedAccuracy = prediction[prediction.ArgMax()];
if (predictedAccuracy >= 0.95)
{
_slidingWindowView.DrawBoundingBox(boundingBox, predicted, predictedAccuracy);
}
}
catch (Exception exception)
{
_loggingService.Log(exception);
}
}
}
_loggingService.Log("Classify drawing has completed");
}
catch (Exception exception)
{
_loggingService.Log(exception);
_messageService.ShowMessage("An error ocurred while classyfing the drawing. Please try again.", "Classification error", icon: MessageBoxIcon.Information);
}
}
public Dictionary <PredictionMethod, double> ComputeRMSE(List <PredictionMethod> lMethods, int cTrials = 0)
{
Console.WriteLine("***************** Model Evaluation *********************");
Dictionary <RecommenderSystem.PredictionMethod, double> results = new Dictionary <PredictionMethod, double>();
foreach (var method in lMethods)
{
IPredictionModel model = predictionEngine.getModel(method);
if (model != null)
{
var rmse = evaluationEngine.computeRMSE(testUsers, testItems, model);
Console.WriteLine(String.Format("Model: {0}, RMSE: {1}", method, rmse));
results.Add(method, rmse);
}
}
Console.WriteLine("*************************************************************");
return(results);
}
public IList<double> Predict(IDataFrame queryDataFrame, IPredictionModel model, string dependentFeatureName)
{
if (!(model is ILinearRegressionModel))
{
throw new ArgumentException("Invalid model passed to Linear Regression predictor!");
}
var linearRegressionModel = model as ILinearRegressionModel;
var xMatrix = queryDataFrame.GetSubsetByColumns(
queryDataFrame.ColumnNames.Except(new[] { dependentFeatureName }).ToList())
.GetAsMatrixWithIntercept();
var results = new List<double>();
for (int rowIdx = 0; rowIdx < xMatrix.RowCount; rowIdx++)
{
var queryRow = xMatrix.Row(rowIdx);
var result = linearRegressionModel.Weights.DotProduct(queryRow);
results.Add(result);
}
return results;
}
public IList <TDecisionValue> Predict(IDataFrame queryDataFrame, IPredictionModel model, string dependentFeatureName)
{
if (!(model is IDecisionTreeNode))
{
throw new ArgumentException("Invalid model passed to Decision Tree Predictor");
}
var results = new ConcurrentBag <Tuple <int, TDecisionValue> >();
var queryDataFrameWithoutDependentFeature =
queryDataFrame.GetSubsetByColumns(
queryDataFrame.ColumnNames.Except(new[] { dependentFeatureName }).ToList());
for (int rowIdx = 0; rowIdx < queryDataFrameWithoutDependentFeature.RowCount; rowIdx++)
{
IDataVector <TDecisionValue> dataVector = queryDataFrameWithoutDependentFeature.GetRowVector <TDecisionValue>(rowIdx);
Tuple <TDecisionValue, double> predictionResults = ProcessInstance(dataVector, (IDecisionTreeNode)model, 1.0);
results.Add(new Tuple <int, TDecisionValue>(rowIdx, predictionResults.Item1));
}
return(results.OrderBy(tpl => tpl.Item1).Select(tpl => tpl.Item2).ToList());
}
public IList <double> Predict(IDataFrame queryDataFrame, IPredictionModel model, string dependentFeatureName)
{
if (!(model is ILinearRegressionModel))
{
throw new ArgumentException("Invalid model passed to Linear Regression predictor!");
}
var linearRegressionModel = model as ILinearRegressionModel;
var xMatrix = queryDataFrame.GetSubsetByColumns(
queryDataFrame.ColumnNames.Except(new[] { dependentFeatureName }).ToList())
.GetAsMatrixWithIntercept();
var results = new List <double>();
for (int rowIdx = 0; rowIdx < xMatrix.RowCount; rowIdx++)
{
var queryRow = xMatrix.Row(rowIdx);
var result = linearRegressionModel.Weights.DotProduct(queryRow);
results.Add(result);
}
return(results);
}
public IList <IDataQualityReport <TPredictionResult> > CrossValidate(
IPredictionModelBuilder modelBuilder,
IModelBuilderParams modelBuilderParams,
IPredictor <TPredictionResult> predictor,
IDataQualityMeasure <TPredictionResult> qualityMeasure,
IDataFrame dataFrame,
string dependentFeatureName,
double percetnagOfTrainData,
int folds)
{
var trainingDataCount = (int)Math.Round(percetnagOfTrainData * dataFrame.RowCount);
var testDataCount = dataFrame.RowCount - trainingDataCount;
var shuffledAllIndices = dataFrame.RowIndices.Shuffle(_randomizer);
var maxWindowsCount = dataFrame.RowCount / testDataCount;
var iterationAccuracies = new List <IDataQualityReport <TPredictionResult> >();
var currentWindowNo = 0;
for (var i = 0; i < folds; i++)
{
if (currentWindowNo == maxWindowsCount)
{
currentWindowNo = 0;
shuffledAllIndices = shuffledAllIndices.Shuffle();
}
var offset = currentWindowNo * testDataCount;
var trainingIndices = shuffledAllIndices.Skip(offset).Take(trainingDataCount).ToList();
var trainingData = dataFrame.GetSubsetByRows(trainingIndices);
var testIndices = shuffledAllIndices.Except(trainingIndices).ToList();
var testData = dataFrame.GetSubsetByRows(testIndices);
IPredictionModel model = modelBuilder.BuildModel(trainingData, dependentFeatureName, modelBuilderParams);
IList <TPredictionResult> predictions = predictor.Predict(testData, model, dependentFeatureName);
IList <TPredictionResult> expected = testData.GetColumnVector <TPredictionResult>(dependentFeatureName);
IDataQualityReport <TPredictionResult> qualityReport = qualityMeasure.GetReport(expected, predictions);
iterationAccuracies.Add(qualityReport);
currentWindowNo++;
}
return(iterationAccuracies);
}
public Dictionary <PredictionMethod, double> ComputeRMSE(List <PredictionMethod> lMethods, out Dictionary <PredictionMethod, Dictionary <PredictionMethod, double> > dConfidence)
{
// compute RMSE
Dictionary <PredictionMethod, double> results = new Dictionary <PredictionMethod, double>();
foreach (var method in lMethods)
{
IPredictionModel model = predictionEngine.getModel(method);
if (model != null)
{
var rmse = evaluationEngine.computeRMSE(testUsers, testItems, model);
results.Add(method, rmse);
}
}
// compute dConfidence
dConfidence = new Dictionary <PredictionMethod, Dictionary <PredictionMethod, double> >();
foreach (var method in lMethods)
{
dConfidence.Add(method, new Dictionary <PredictionMethod, double>());
}
List <Tuple <PredictionMethod, PredictionMethod> > methodPairs = DataUtils.getAllPairedCombinations(lMethods);
foreach (var methodPair in methodPairs)
{
PredictionMethod method1 = methodPair.Item1;
PredictionMethod method2 = methodPair.Item2;
Tuple <double, double> pApB = evaluationEngine.computeConfidence(testUsers, testItems, predictionEngine.getModel(method1), predictionEngine.getModel(method2));
dConfidence[method1].Add(method2, pApB.Item1);
dConfidence[method2].Add(method1, pApB.Item2);
}
return(results);
}
public IList<double> Predict(IDataFrame queryDataFrame, IPredictionModel model, int dependentFeatureIndex)
{
return Predict(queryDataFrame, model, queryDataFrame.ColumnNames[dependentFeatureIndex]);
}
public BaseModelMethod(IPredictionModel model)
{
this.model = (MatrixFactorizationModel)model;
}
public IList <TDecisionValue> Predict(IDataFrame queryDataFrame, IPredictionModel model, int dependentFeatureIndex)
{
return(this.Predict(queryDataFrame, model, queryDataFrame.ColumnNames[dependentFeatureIndex]));
}
public IList <double> Predict(IDataFrame queryDataFrame, IPredictionModel model, int dependentFeatureIndex)
{
return(Predict(queryDataFrame, model, queryDataFrame.ColumnNames[dependentFeatureIndex]));
}
public IList <TPredictionResult> Predict(IDataFrame queryDataFrame, IPredictionModel model, string dependentFeatureName)
{
return(Predict(queryDataFrame, model, queryDataFrame.ColumnNames.IndexOf(dependentFeatureName)));
}
public Tuple <double, double> computeConfidence(Users validationUsers, Items validationItems, IPredictionModel modelA, IPredictionModel modelB)
{
double aCounter = 0;
double bCounter = 0;
double aPrediction;
double bPrediction;
double aError;
double bError;
double actualRating;
// calcualte number wins for each model
foreach (User user in validationUsers)
{
foreach (string itemId in user.GetRatedItems())
{
Item item = validationItems.GetItemById(itemId);
actualRating = user.GetRating(itemId);
aPrediction = modelA.Predict(user, item);
bPrediction = modelB.Predict(user, item);
aError = Math.Abs(actualRating - aPrediction);
bError = Math.Abs(actualRating - bPrediction);
if (aError < bError)
{
aCounter++;
}
else if (aError > bError)
{
bCounter++;
}
else
{
aCounter += 0.5;
bCounter += 0.5;
}
}
}
int n = (int)(aCounter + bCounter);
// calcualte pA
double sum = 0;
for (int i = (int)aCounter; i < n; i++)
{
sum += MathUtils.Factorial(n) / (MathUtils.Factorial(n - i) * MathUtils.Factorial(i));
}
double pA = (1 - Math.Pow(0.5, n) * sum);
// calculate pB
sum = 0;
for (int i = (int)bCounter; i < n; i++)
{
sum += MathUtils.Factorial(n) / (MathUtils.Factorial(n - i) * MathUtils.Factorial(i));
}
double pB = (1 - Math.Pow(0.5, n) * sum);
return(Tuple.Create(pA, pB));
}