/// <summary>Generate model based on a set of examples.</summary>
/// <param name="x">The Matrix to process.</param>
/// <param name="y">The Vector to process.</param>
/// <returns>Model.</returns>
public override IModel Generate(Matrix x, Vector y)
{
// because I said so...
if (this.MaxIterations == -1)
{
this.MaxIterations = x.Rows * 1000;
}
var network = Network.Default(this.Descriptor, x, y, this.Activation);
var model = new NeuralNetworkModel { Descriptor = this.Descriptor, Network = network };
this.OnModelChanged(this, ModelEventArgs.Make(model, "Initialized"));
for (var i = 0; i < this.MaxIterations; i++)
{
var idx = i % x.Rows;
network.Forward(x[idx, VectorType.Row]);
// OnModelChanged(this, ModelEventArgs.Make(model, "Forward"));
network.Back(y[idx], this.LearningRate);
var output = string.Format("Run ({0}/{1})", i, this.MaxIterations);
this.OnModelChanged(this, ModelEventArgs.Make(model, output));
}
return model;
}
/// <summary>Estimates.</summary>
/// <param name="X">The Matrix to process.</param>
/// <param name="type">(Optional) the type.</param>
public void Estimate(Matrix X, VectorType type = VectorType.Row)
{
var n = type == VectorType.Row ? X.Rows : X.Cols;
var s = type == VectorType.Row ? X.Cols : X.Rows;
this.Mu = X.Sum(type) / n;
this.Sigma = Matrix.Zeros(s);
for (var i = 0; i < n; i++)
{
var x = X[i, type] - this.Mu;
this.Sigma += x.Outer(x);
}
this.Sigma *= 1d / (n - 1d);
}
/// <summary>
/// Compute the error cost of the given Theta parameter for the training and label sets
/// </summary>
/// <param name="theta">Learning Theta parameters</param>
/// <param name="X">Training set</param>
/// <param name="y">Training labels</param>
/// <param name="lambda">Regularization constant</param>
/// <param name="regularizer">Regularization term function.</param>
/// <returns></returns>
public double ComputeCost(Vector theta, Matrix X, Vector y, double lambda, IRegularizer regularizer)
{
var m = X.Rows;
var j = 0.0;
var s = (X * theta).ToVector();
j = 1.0 / (2.0 * m) * ((s - y) ^ 2.0).Sum();
if (lambda != 0)
{
j = regularizer.Regularize(j, theta, m, lambda);
}
return j;
}
/// <summary>
/// Compute the error cost of the given Theta parameter for the training and label sets
/// </summary>
/// <param name="theta">Learning Theta parameters</param>
/// <param name="X">Training set</param>
/// <param name="y">Training labels</param>
/// <param name="lambda">Regularisation constant</param>
/// <param name="regularizer">Regularization term function.</param>
/// <returns></returns>
public Vector ComputeGradient(Vector theta, Matrix X, Vector y, double lambda, IRegularizer regularizer)
{
var m = X.Rows;
var gradient = Vector.Zeros(theta.Length);
var s = (X * theta).ToVector();
for (var i = 0; i < theta.Length; i++)
{
gradient[i] = 1.0 / m * ((s - y) * X[i, VectorType.Col]).Sum();
}
if (lambda != 0)
{
gradient = regularizer.Regularize(theta, gradient, m, lambda);
}
return gradient;
}
/// <summary>Generate model based on a set of examples.</summary>
/// <param name="x">The Matrix to process.</param>
/// <param name="y">The Vector to process.</param>
/// <returns>Model.</returns>
public override IModel Generate(Matrix x, Vector y)
{
var N = y.Length;
var a = Vector.Zeros(N);
// compute kernel
var K = this.Kernel.Compute(x);
var n = 1;
// hopefully enough to converge right? ;)
// need to be smarter about storing SPD kernels...
var found_error = true;
while (n < 500 && found_error)
{
found_error = false;
for (var i = 0; i < N; i++)
{
found_error = y[i] * a.Dot(K[i]) <= 0;
if (found_error)
{
a[i] += y[i];
}
}
n++;
}
// anything that *matters*
// i.e. support vectors
var indices = a.Indices(d => d != 0);
// slice up examples to contain
// only support vectors
return new KernelPerceptronModel
{
Kernel = this.Kernel, A = a.Slice(indices), Y = y.Slice(indices), X = x.Slice(indices)
};
}
/// <summary>Generate Linear Regression model based on a set of examples.</summary>
/// <param name="x">The Matrix to process.</param>
/// <param name="y">The Vector to process.</param>
/// <returns>Model.</returns>
public override IModel Generate(Matrix x, Vector y)
{
// create initial theta
var theta = Vector.Ones(x.Cols + 1);
var copy = x.Copy();
// normalise features
for (var i = 0; i < copy.Cols; i++)
{
var j = FeatureNormalizer.FeatureScale(copy[i, VectorType.Col]);
for (var k = 0; k < copy.Rows; k++)
{
copy[k, i] = j[k];
}
}
// add intercept term
copy = copy.Insert(Vector.Ones(copy.Rows), 0, VectorType.Col);
// run gradient descent
var run = GradientDescent.Run(
theta,
copy,
y,
this.MaxIterations,
this.LearningRate,
new LinearCostFunction(),
this.Lambda,
new Regularization());
// once converged create model and apply theta
var model = new LinearRegressionModel(x.Mean(VectorType.Row), x.StdDev(VectorType.Row))
{
Descriptor = this.Descriptor, Theta = run.Item2
};
return model;
}
/// <summary>
/// Initialization method for performing custom actions prior to being optimized.
/// </summary>
public override void Initialize()
{
base.Initialize();
YReformed = Y.Reshape(X.Rows, VectorType.Row);
}
