/// <summary>
/// Predict the index of each classifer that applies to each row of X using trained
/// weights of a neural network.
/// </summary>
/// <param name="theta_1">The first set of trained weights between the input layer
/// and the hidden layer.</param>
/// <param name="theta_2">The second set of trained weights between the hidden layer
/// and the output layer.</param>
/// <param name="X">A Matrix of example rows.</param>
/// <returns>A Matrix (column vector, m * 1) containing the zero-based indices of
/// the most probable classification prediction for each input row in X.</returns>
/// <remarks>This version only has one hidden layer, which will suffice for many
/// problems. However, it should probably be updated to allow for an arbitrary
/// number of layers.</remarks>
public static Matrix Predict(Matrix theta_1, Matrix theta_2, Matrix X)
{
Matrix A1 = Matrix.AddIdentityColumn(X);
// Calculate the second (hidden) layer.
Matrix Z2 = Matrix.MultiplyByTranspose(theta_1, A1);
Matrix A2 = LogisticRegression.Sigmoid(Z2);
A2 = Matrix.AddIdentityColumn(A2.Transpose);
// Calculate 3rd layer (output)
Matrix Z3 = Matrix.MultiplyByTranspose(theta_2, A2);
Matrix A3 = LogisticRegression.Sigmoid(Z3);
return(Matrix.MaxIndex(A3).Transpose);
}
static void TitanicDemo()
{
WriteH1("Testing");
#region Dummy data importer
//WriteH2("Dummy Data");
//DataImporterDummy di = new DataImporterDummy();
//DataFrame df = new DataFrame(di);
//df.Load(String.Empty, true, true);
//Console.WriteLine(df.TotalColumns);
//foreach(string h in df.Headers)
//{
// Console.WriteLine(h);
//}
#endregion
#region CSV data importer
WriteH2("CSV Data (Titanic)");
DataImporterCSV di_csv = new DataImporterCSV();
DataExporterCSV de_csv = new DataExporterCSV();
DataFrame df_train = new DataFrame(di_csv, de_csv);
DataFrame df_test = new DataFrame(di_csv, de_csv);
df_train.Load(@"c:\temp\titanic.csv", true, "Survived");
df_test.Load(@"c:\temp\titanic_test.csv", true);
Console.WriteLine($"Total Columns (training data): {df_train.TotalColumns}");
Console.WriteLine($"Total Columns (testing data): {df_test.TotalColumns}");
// Change the type of some of the training columns
df_train.SetColumnType("pclass", DataFrameColumnType.Factors);
df_train.SetColumnType("sex", DataFrameColumnType.Factors);
df_train.SetColumnType("age", DataFrameColumnType.Bins);
//df_train["age"].SetBins(new double[] { 0.0, 18.0, 100.0 });
df_train["age"].SetBins(new double[] { 0.0, 15.0, 25.0, 30.0, 40.0, 50.0, 55.0, 65.0, 75.0, 100.0 });
df_train["age"].EmptyValue = 30.27; // Average value of known ages
df_train.SetColumnType("fare", DataFrameColumnType.Double);
df_train.SetColumnType("sibsp", DataFrameColumnType.Double);
df_train.SetColumnType("parch", DataFrameColumnType.Double);
df_train.CreateDataColumn("CabinLetter", GetCabinLetter);
df_test.CreateDataColumn("CabinLetter", GetCabinLetter); // For now, we have to explicitly set both training set
// and test set separately.
df_train.SetColumnType("CabinLetter", DataFrameColumnType.Factors);
df_train.SetColumnType("survived", DataFrameColumnType.Double);
// Try and match the types in the testing set
df_test.MatchColumns(df_train);
Console.WriteLine($"df_train hasResults? {df_train.HasResults}. df_test hasResults? {df_test.HasResults}");
// Start calculations
Matrix Xtrain = df_train.ExportFeatures();
Matrix ytrain = df_train.ExportResults();
Matrix Xtest = df_test.ExportFeatures();
// Try Logistic Regression
double[] labels = new double[] { 0.0, 1.0 };
Matrix lr_theta = LogisticRegression.OneVsAll(Xtrain, ytrain, labels, 0.1, 1000);
Matrix lr_prediction = LogisticRegression.PredictOneVsAll(lr_theta, Xtest);
int input_layer_size = Xtrain.Columns;
int output_layer_size = labels.Length;
int hidden_layer_size = (input_layer_size + output_layer_size) / 2;
Matrix[] nn_theta = NeuralNetwork.Train(Xtrain, ytrain, input_layer_size, hidden_layer_size, labels, 0.1, 1000);
Matrix nn_prediction = NeuralNetwork.Predict(nn_theta[0], nn_theta[1], Xtest);
// Exporting
DataFrame df_lr_export = df_test;
DataFrame df_nn_export = df_test;
DataFrameColumn col_lr_results = new DataFrameColumn(df_lr_export, lr_prediction, 0);
DataFrameColumn col_nn_results = new DataFrameColumn(df_nn_export, nn_prediction, 0);
col_lr_results.Header = col_nn_results.Header = "Survived";
df_lr_export.Save(@"c:\temp\lr_results.csv");
df_nn_export.Save(@"c:\temp\nn_results.csv");
#endregion
}
static void LogisticRegressionDemo()
{
WriteH1("Logistic Regression");
#region Sigmoid Function
WriteH2("Sigmoid Function");
Matrix m1 = new Matrix(new double[, ] {
{ 1200000 }
});
Matrix sigmoid1 = LogisticRegression.Sigmoid(m1);
Console.Write("Target: 1.0 Actual: {0}", sigmoid1);
m1[0, 0] = -25000;
sigmoid1 = LogisticRegression.Sigmoid(m1);
Console.Write("Target: 0.0 Actual: {0}", sigmoid1);
m1[0, 0] = 0;
sigmoid1 = LogisticRegression.Sigmoid(m1);
Console.Write("Target: 0.5 Actual: {0}", sigmoid1);
m1 = new Matrix(new double[, ] {
{ 4, 5, 6 }
});
sigmoid1 = LogisticRegression.Sigmoid(m1);
Console.Write("Target: 0.98 0.99 0.997 Actual: {0}", sigmoid1);
#endregion
#region Predict
WriteH2("Prediction");
m1 = new Matrix(new double[, ] {
{ 1, 1 }, { 1, 2.5 }, { 1, 3 }, { 1, 4 }
});
Matrix theta = new Matrix(new double[, ] {
{ -3.5 }, { 1.3 }
});
Matrix prediction = LogisticRegression.Predict(m1, theta);
Console.WriteLine("Target: 0.0 ; 0.0 ; 1.0 ; 1.0 ; Actual: {0}", prediction.ToString().Replace("\n", "; "));
m1 = Matrix.Magic(3);
theta = new Matrix(new double[, ] {
{ 4 }, { 3 }, { -8 }
});
prediction = LogisticRegression.Predict(m1, theta);
Console.WriteLine("Target: 0.0 ; 0.0 ; 1.0 ; Actual: {0}", prediction.ToString().Replace("\n", "; "));
#endregion
#region Cost Function
WriteH2("Cost Function");
Matrix X = Matrix.AddIdentityColumn(Matrix.Magic(3));
Matrix y = new Matrix(new double[, ] {
{ 1 }, { 0 }, { 1 }
});
theta = new Matrix(new double[, ] {
{ -2 }, { -1 }, { 1 }, { 2 }
});
Tuple <double, Matrix> cost = LogisticRegression.CostFunction(X, y, theta);
Console.WriteLine("Target: 4.6832 ; Actual: {0}", cost.Item1);
#endregion
#region Regularized Cost Function
WriteH2("Regularized Cost Function");
MinimizeOptions options = new MinimizeOptions();
options.RegularizationParameter = 3;
cost = LogisticRegression.CostFunction(X, y, theta, options);
Console.WriteLine("Target: 7.6832 ; Actual: {0}", cost.Item1);
X = new Matrix(new double[, ] {
{ 1.0, 0.1, 0.6, 1.1 },
{ 1.0, 0.2, 0.7, 1.2 },
{ 1.0, 0.3, 0.8, 1.3 },
{ 1.0, 0.4, 0.9, 1.4 },
{ 1.0, 0.5, 1.0, 1.5 }
});
y = new Matrix(new double[, ] {
{ 1.0 },
{ 0.0 },
{ 1.0 },
{ 0.0 },
{ 1.0 }
});
theta = new Matrix(new double[, ] {
{ -2 }, { -1 }, { 1 }, { 2 }
});
cost = LogisticRegression.CostFunction(X, y, theta, options);
Console.WriteLine("Target: 2.5348 ; Actual: {0}", cost.Item1);
#endregion
#region OneVsAll
WriteH2("One vs All");
X = new Matrix(new double[, ] {
{ 8.0, 1.0, 6.0 },
{ 3.0, 5.0, 7.0 },
{ 4.0, 9.0, 2.0 },
{ 0.84147, 0.90930, 0.14112 },
{ 0.54030, -0.41615, -0.98999 }
});
y = new Matrix(new double[, ] {
{ 1.0 },
{ 2.0 },
{ 2.0 },
{ 1.0 },
{ 3.0 }
});
//Matrix testTheta = new Matrix(4, 1);
//Matrix X0 = Matrix.Join(Matrix.Ones(5, 1), X, MatrixDimensions.Columns);
//cost = LogisticRegression.CostFunction(X0, y==1, testTheta, 0.1);
//Console.WriteLine(cost.Item1);
//Console.WriteLine(cost.Item2);
double[] labels = new double[] { 1.0, 2.0, 3.0 };
Matrix all_theta = LogisticRegression.OneVsAll(X, y, labels, 0.1);
Console.WriteLine(all_theta);
#endregion
#region PredictOneVsAll
WriteH2("Predict One vs All");
X = new Matrix(new double[, ] {
{ 1.0, 7.0 },
{ 4.0, 5.0 },
{ 7.0, 8.0 },
{ 1.0, 4.0 }
});
all_theta = new Matrix(new double[, ] {
{ 1.0, -6.0, 3.0 },
{ -2.0, 4.0, -3.0 }
});
prediction = LogisticRegression.PredictOneVsAll(all_theta, X);
Console.WriteLine("Target: 0; 1; 1; 0; Actual: {0}", prediction.ToString().Replace("\n", "; "));
#endregion
}
/// <summary>
/// The Neural Network cost function for a two layer classification Neural Network.
/// </summary>
/// <param name="nn_parameters">The unrolled parameter vector that contains all the weights.</param>
/// <param name="input_layer_size">The number of nodes in the input layer.</param>
/// <param name="hidden_layer_size">The number of nodes in the hidden layer.</param>
/// <param name="labels">A list of classification labels.</param>
/// <param name="X">The feature set Matrix.</param>
/// <param name="y">The result set Matrix.</param>
/// <param name="lambda">The regularization parameter which helps reduce overfitting.
/// Note that using values that are too high will lead to underfitting.</param>
/// <returns>The cost of using the given value of theta, and the gradient of
/// the cost (useful for iterative minimization functions)</returns>
public static Tuple <double, Matrix> NNCostFunction(Matrix X, Matrix y, Matrix nn_parameters, MinimizeOptions options)
{
double lambda = options.RegularizationParameter;
int input_layer_size = options.InputLayerSize;
int hidden_layer_size = options.HiddenLayerSize;
double[] labels = options.Labels;
double costFunction = 0;
int num_labels = labels.Length;
List <Matrix> output_gradient = new List <Matrix>();
Matrix Theta1 = Matrix.Reshape(nn_parameters, 0, hidden_layer_size, input_layer_size + 1);
Matrix Theta2 = Matrix.Reshape(nn_parameters, (hidden_layer_size * (input_layer_size + 1)), num_labels, hidden_layer_size + 1);
// y_matrix has the following attributes:
// Rows: same as the number of rows in Y -- one for each example result.
// Columns: one for each label.
// Values: Each row consists of zeros, except for one, which matches the
// value of y in that row to the index of the label. For example, if there
// are three labels (3, 6, 8), and y contains 2 rows (8, 3), then y_matrix
// would be:
// 0 0 1
// 1 0 0
Matrix y_matrix = AssignLabels(y, labels);
// Add ones to the X Matrix
Matrix a1 = Matrix.AddIdentityColumn(X);
Matrix z2 = a1 * Theta1.Transpose;
Matrix a2 = LogisticRegression.Sigmoid(z2);
a2 = Matrix.AddIdentityColumn(a2);
Matrix z3 = a2 * Theta2.Transpose;
Matrix a3 = LogisticRegression.Sigmoid(z3);
Matrix log1 = Matrix.ElementLog(a3);
Matrix log2 = Matrix.ElementLog(1 - a3);
Matrix part1 = Matrix.ElementMultiply(-y_matrix, log1);
Matrix part2 = Matrix.ElementMultiply((1 - y_matrix), log2);
Matrix t0 = Theta1.RemoveColumn(0);
Matrix t1 = Theta2.RemoveColumn(0);
// Calculate regularization component
double multiplier = lambda / (2 * X.Rows);
double reg1 = Matrix.ElementPower(t0, 2).SumAllElements;
double reg2 = Matrix.ElementPower(t1, 2).SumAllElements;
double r = multiplier * (reg1 + reg2);
// Calculate cost
costFunction = (1.0 / X.Rows) * (part1 - part2).SumAllElements + r;
// Back Propogation
Matrix d3 = a3 - y_matrix;
Matrix d2 = Matrix.ElementMultiply(
(t1.Transpose * d3.Transpose).Transpose,
SigmoidGradient(z2)
);
Matrix Delta1 = d2.Transpose * a1;
Matrix Delta2 = d3.Transpose * a2;
Theta1 = Matrix.Join(new Matrix(t0.Rows, 1), t0, MatrixDimensions.Columns);
Theta2 = Matrix.Join(new Matrix(t1.Rows, 1), t1, MatrixDimensions.Columns);
double scale_value = lambda / X.Rows;
Matrix Theta1_scaled = Theta1 * scale_value;
Matrix Theta2_scaled = Theta2 * scale_value;
Matrix Theta1_grad = ((Delta1 / X.Rows) + Theta1_scaled).Unrolled;
Matrix Theta2_grad = ((Delta2 / X.Rows) + Theta2_scaled).Unrolled;
return(new Tuple <double, Matrix>(costFunction, Matrix.Join(Theta1_grad, Theta2_grad, MatrixDimensions.Rows)));
}