/// <inheritdoc/>
public Object Read(Stream istream)
{
BayesianNetwork result = new BayesianNetwork();
EncogReadHelper input = new EncogReadHelper(istream);
EncogFileSection section;
String queryType = "";
String queryStr = "";
String contentsStr = "";
while ((section = input.ReadNextSection()) != null)
{
if (section.SectionName.Equals("BAYES-NETWORK") &&
section.SubSectionName.Equals("BAYES-PARAM"))
{
IDictionary <String, String> p = section.ParseParams();
queryType = p["queryType"];
queryStr = p["query"];
contentsStr = p["contents"];
}
if (section.SectionName.Equals("BAYES-NETWORK") &&
section.SubSectionName.Equals("BAYES-TABLE"))
{
result.Contents = contentsStr;
// first, define relationships (1st pass)
foreach (String line in section.Lines)
{
result.DefineRelationship(line);
}
result.FinalizeStructure();
// now define the probabilities (2nd pass)
foreach (String line in section.Lines)
{
result.DefineProbability(line);
}
}
if (section.SectionName.Equals("BAYES-NETWORK") &&
section.SubSectionName.Equals("BAYES-PROPERTIES"))
{
IDictionary <String, String> paras = section.ParseParams();
EngineArray.PutAll(paras, result.Properties);
}
}
// define query, if it exists
if (queryType.Length > 0)
{
IBayesianQuery query = null;
if (queryType.Equals("EnumerationQuery"))
{
query = new EnumerationQuery(result);
}
else
{
query = new SamplingQuery(result);
}
if (query != null && queryStr.Length > 0)
{
result.Query = query;
result.DefineClassificationStructure(queryStr);
}
}
return(result);
}
/// <inheritdoc/>
public void Write(double[] input, double[] ideal, double significance)
{
EngineArray.ArrayCopy(input, _input[_index]);
EngineArray.ArrayCopy(ideal, _ideal[_index]);
_index++;
}
/// <summary>
/// Derive the minimum, using a conjugate gradient method.
/// </summary>
///
/// <param name="maxIterations">The max iterations.</param>
/// <param name="maxError">Stop at this error rate.</param>
/// <param name="eps">The machine's precision.</param>
/// <param name="tol">The convergence tolerance.</param>
/// <param name="network">The network to get the error from.</param>
/// <param name="n">The number of variables.</param>
/// <param name="x">The independent variable.</param>
/// <param name="ystart">The start for y.</param>
/// <param name="bs">Work vector, must have n elements.</param>
/// <param name="direc">Work vector, must have n elements.</param>
/// <param name="g">Work vector, must have n elements.</param>
/// <param name="h">Work vector, must have n elements.</param>
/// <param name="deriv2">Work vector, must have n elements.</param>
/// <returns>The best error.</returns>
public double Calculate(int maxIterations, double maxError,
double eps, double tol,
ICalculationCriteria network, int n, double[] x,
double ystart, double[] bs, double[] direc,
double[] g, double[] h, double[] deriv2)
{
var globalMinimum = new GlobalMinimumSearch();
double fbest = network.CalcErrorWithMultipleSigma(x, direc, deriv2,
true);
double prevBest = 1.0e30d;
for (int i = 0; i < n; i++)
{
direc[i] = -direc[i];
}
EngineArray.ArrayCopy(direc, g);
EngineArray.ArrayCopy(direc, h);
int convergenceCounter = 0;
int poorCj = 0;
// Main loop
for (int iteration = 0; iteration < maxIterations; iteration++)
{
if (fbest < maxError)
{
break;
}
EncogLogging.Log(EncogLogging.LevelInfo,
"Beginning internal Iteration #" + iteration + ", currentError=" + fbest + ",target=" + maxError);
// Check for convergence
double toler;
if (prevBest <= 1.0d)
{
toler = tol;
}
else
{
toler = tol * prevBest;
}
// Stop if there is little improvement
if ((prevBest - fbest) <= toler)
{
if (++convergenceCounter >= 3)
{
break;
}
}
else
{
convergenceCounter = 0;
}
double dot2 = 0;
double dlen = 0;
double dot1 = dot2 = dlen = 0.0d;
double high = 1.0e-4d;
for (int i = 0; i < n; i++)
{
bs[i] = x[i];
if (deriv2[i] > high)
{
high = deriv2[i];
}
dot1 += direc[i] * g[i]; // Directional first derivative
dot2 += direc[i] * direc[i] * deriv2[i]; // and second
dlen += direc[i] * direc[i]; // Length of search vector
}
double scale;
if (Math.Abs(dot2) < EncogFramework.DefaultDoubleEqual)
{
scale = 0;
}
else
{
scale = dot1 / dot2;
}
high = 1.5d / high;
if (high < 1.0e-4d)
{
high = 1.0e-4d;
}
if (scale < 0.0d)
{
scale = high;
}
else if (scale < 0.1d * high)
{
scale = 0.1d * high;
}
else if (scale > 10.0d * high)
{
scale = 10.0d * high;
}
prevBest = fbest;
globalMinimum.Y2 = fbest;
globalMinimum.FindBestRange(0.0d, 2.0d * scale, -3, false, maxError,
network);
if (globalMinimum.Y2 < maxError)
{
if (globalMinimum.Y2 < fbest)
{
for (int i = 0; i < n; i++)
{
x[i] = bs[i] + globalMinimum.Y2 * direc[i];
if (x[i] < 1.0e-10d)
{
x[i] = 1.0e-10d;
}
}
fbest = globalMinimum.Y2;
}
else
{
for (int i = 0; i < n; i++)
{
x[i] = bs[i];
}
}
break;
}
if (convergenceCounter > 0)
{
fbest = globalMinimum.Brentmin(20, maxError, eps, 1.0e-7d,
network, globalMinimum.Y2);
}
else
{
fbest = globalMinimum.Brentmin(10, maxError, 1.0e-6d, 1.0e-5d,
network, globalMinimum.Y2);
}
for (int i = 0; i < n; i++)
{
x[i] = bs[i] + globalMinimum.X2 * direc[i];
if (x[i] < 1.0e-10d)
{
x[i] = 1.0e-10d;
}
}
double improvement = (prevBest - fbest) / prevBest;
if (fbest < maxError)
{
break;
}
for (int i = 0; i < n; i++)
{
direc[i] = -direc[i]; // negative gradient
}
double gam = Gamma(n, g, direc);
if (gam < 0.0d)
{
gam = 0.0d;
}
if (gam > 10.0d)
{
gam = 10.0d;
}
if (improvement < 0.001d)
{
++poorCj;
}
else
{
poorCj = 0;
}
if (poorCj >= 2)
{
if (gam > 1.0d)
{
gam = 1.0d;
}
}
if (poorCj >= 6)
{
poorCj = 0;
gam = 0.0d;
}
FindNewDir(n, gam, g, h, direc);
}
return(fbest);
}
/// <summary>
/// Construct a fold from the specified flat network.
/// </summary>
///
/// <param name="flat">THe flat network.</param>
public NetworkFold(FlatNetwork flat)
{
_weights = EngineArray.ArrayCopy(flat.Weights);
_output = EngineArray.ArrayCopy(flat.LayerOutput);
}
/// <summary>
/// Copy the weights and output from the network.
/// </summary>
///
/// <param name="source">The network to copy from.</param>
public void CopyFromNetwork(FlatNetwork source)
{
EngineArray.ArrayCopy(source.Weights, _weights);
EngineArray.ArrayCopy(source.LayerOutput, _output);
}
internal static void arraycopy(float[] source, int sourceIndex, float[] target, int targetIndex, int size)
{
EngineArray.ArrayCopy(source, sourceIndex, target, targetIndex, size);
}
/// <summary>
/// Construct a truth table line.
/// </summary>
/// <param name="prob">The probability.</param>
/// <param name="result">The result.</param>
/// <param name="args">The arguments.</param>
public TableLine(double prob, int result, int[] args)
{
Probability = prob;
_result = result;
_arguments = EngineArray.ArrayCopy(args);
}
/// <summary>
/// Set the current state.
/// </summary>
/// <param name="s">The new current state.</param>
public void SetCurrentState(double[] s)
{
_currentState = new BiPolarMLData(s.Length);
EngineArray.ArrayCopy(s, _currentState.Data);
}
private void ProcessCalc()
{
AnalystField firstOutputField = null;
int barsNeeded = Math.Abs(Analyst.DetermineMinTimeSlice());
IndentLevel = 2;
AddLine("if( _inputCount>0 && CurrentBar>=" + barsNeeded + " )");
AddLine("{");
IndentIn();
AddLine("double[] input = new double[_inputCount];");
AddLine("double[] output = new double[_outputCount];");
int idx = 0;
foreach (AnalystField field in Analyst.Script.Normalize
.NormalizedFields)
{
if (field.Input)
{
String str;
DataField df = Analyst.Script
.FindDataField(field.Name);
switch (field.Action)
{
case NormalizationAction.PassThrough:
str = EngineArray.Replace(df.Source, "##", "" + (-field.TimeSlice));
AddLine("input[" + idx + "]=" + str + ";");
idx++;
break;
case NormalizationAction.Normalize:
str = EngineArray.Replace(df.Source, "##", "" + (-field.TimeSlice));
AddLine("input[" + idx + "]=Norm(" + str + ","
+ field.NormalizedHigh + ","
+ field.NormalizedLow + ","
+ field.ActualHigh + ","
+ field.ActualLow + ");");
idx++;
break;
case NormalizationAction.Ignore:
break;
default:
throw new AnalystCodeGenerationError(
"Can't generate Ninjascript code, unsupported normalizatoin action: "
+ field.Action.ToString());
}
}
if (field.Output)
{
if (firstOutputField == null)
{
firstOutputField = field;
}
}
}
if (firstOutputField != null)
{
AddLine("Compute(input,output);");
AddLine("Output.Set(DeNorm(output[0]" + ","
+ firstOutputField.NormalizedHigh + ","
+ firstOutputField.NormalizedLow + ","
+ firstOutputField.ActualHigh + ","
+ firstOutputField.ActualLow + "));");
IndentOut();
}
AddLine("}");
IndentLevel = 2;
}
/// <summary>
///
/// </summary>
///
public void Reset(int seed)
{
CurrentState.Clear();
EngineArray.Fill(_weights, 0.0d);
}
/// <summary>
/// Clear any connection weights.
/// </summary>
///
public void Clear()
{
EngineArray.Fill(_weights, 0);
}
/// <summary>
/// Read an object.
/// </summary>
///
public Object Read(Stream mask0)
{
var result = new BasicNetwork();
var flat = new FlatNetwork();
var ins0 = new EncogReadHelper(mask0);
EncogFileSection section;
while ((section = ins0.ReadNextSection()) != null)
{
if (section.SectionName.Equals("BASIC") &&
section.SubSectionName.Equals("PARAMS"))
{
IDictionary <String, String> paras = section.ParseParams();
EngineArray.PutAll(paras, result.Properties);
}
if (section.SectionName.Equals("BASIC") &&
section.SubSectionName.Equals("NETWORK"))
{
IDictionary <String, String> p = section.ParseParams();
flat.BeginTraining = EncogFileSection.ParseInt(p,
BasicNetwork.TagBeginTraining);
flat.ConnectionLimit = EncogFileSection.ParseDouble(p,
BasicNetwork.TagConnectionLimit);
flat.ContextTargetOffset = EncogFileSection.ParseIntArray(
p, BasicNetwork.TagContextTargetOffset);
flat.ContextTargetSize = EncogFileSection.ParseIntArray(
p, BasicNetwork.TagContextTargetSize);
flat.EndTraining = EncogFileSection.ParseInt(p,
BasicNetwork.TagEndTraining);
flat.HasContext = EncogFileSection.ParseBoolean(p,
BasicNetwork.TagHasContext);
flat.InputCount = EncogFileSection.ParseInt(p,
PersistConst.InputCount);
flat.LayerCounts = EncogFileSection.ParseIntArray(p,
BasicNetwork.TagLayerCounts);
flat.LayerFeedCounts = EncogFileSection.ParseIntArray(p,
BasicNetwork.TagLayerFeedCounts);
flat.LayerContextCount = EncogFileSection.ParseIntArray(
p, BasicNetwork.TagLayerContextCount);
flat.LayerIndex = EncogFileSection.ParseIntArray(p,
BasicNetwork.TagLayerIndex);
flat.LayerOutput = section.ParseDoubleArray(p, PersistConst.Output);
flat.LayerSums = new double[flat.LayerOutput.Length];
flat.OutputCount = EncogFileSection.ParseInt(p,
PersistConst.OutputCount);
flat.WeightIndex = EncogFileSection.ParseIntArray(p,
BasicNetwork.TagWeightIndex);
flat.Weights = section.ParseDoubleArray(p, PersistConst.Weights);
flat.BiasActivation = section.ParseDoubleArray(p, BasicNetwork.TagBiasActivation);
}
else if (section.SectionName.Equals("BASIC") &&
section.SubSectionName.Equals("ACTIVATION"))
{
int index = 0;
flat.ActivationFunctions = new IActivationFunction[flat.LayerCounts.Length];
foreach (String line in section.Lines)
{
IActivationFunction af;
IList <String> cols = EncogFileSection
.SplitColumns(line);
String name = ReflectionUtil.AfPath
+ cols[0];
try
{
af = (IActivationFunction)ReflectionUtil.LoadObject(name);
}
catch (TypeLoadException e)
{
throw new PersistError(e);
}
catch (TargetException e)
{
throw new PersistError(e);
}
catch (MemberAccessException e)
{
throw new PersistError(e);
}
for (int i = 0; i < af.ParamNames.Length; i++)
{
af.Params[i] =
CSVFormat.EgFormat.Parse(cols[i + 1]);
}
flat.ActivationFunctions[index++] = af;
}
}
}
result.Structure.Flat = flat;
return(result);
}
/// <summary>
/// Perform a SVD fit.
/// </summary>
/// <param name="x">The X matrix.</param>
/// <param name="y">The Y matrix.</param>
/// <param name="a">The A matrix.</param>
/// <param name="funcs">The RBF functions.</param>
/// <returns>The fit.</returns>
public static double Svdfit(double[][] x, double[][] y, double[][] a,
IRadialBasisFunction[] funcs)
{
int i, j, k;
double wmax, tmp, thresh, sum, TOL = 1e-13d;
//Allocated memory for svd matrices
double[][] u = EngineArray.AllocateDouble2D(x.Length, funcs.Length);
double[][] v = EngineArray.AllocateDouble2D(funcs.Length, funcs.Length);
var w = new double[funcs.Length];
//Fill input matrix with values based on fitting functions and input coordinates
for (i = 0; i < x.Length; i++)
{
for (j = 0; j < funcs.Length; j++)
{
u[i][j] = funcs[j].Calculate(x[i]);
}
}
//Perform decomposition
Svdcmp(u, w, v);
//Check for w values that are close to zero and replace them with zeros such that they are ignored in backsub
wmax = 0;
for (j = 0; j < funcs.Length; j++)
{
if (w[j] > wmax)
{
wmax = w[j];
}
}
thresh = TOL * wmax;
for (j = 0; j < funcs.Length; j++)
{
if (w[j] < thresh)
{
w[j] = 0;
}
}
//Perform back substitution to get result
Svdbksb(u, w, v, y, a);
//Calculate chi squared for the fit
double chisq = 0;
for (k = 0; k < y[0].Length; k++)
{
for (i = 0; i < y.Length; i++)
{
sum = 0.0d;
for (j = 0; j < funcs.Length; j++)
{
sum += a[j][k] * funcs[j].Calculate(x[i]);
}
tmp = (y[i][k] - sum);
chisq += tmp * tmp;
}
}
return(Math.Sqrt(chisq / (y.Length * y[0].Length)));
}
/// <inheritdoc/>
public void Write(double[] input, double[] ideal)
{
EngineArray.ArrayCopy(input, this.input[index]);
EngineArray.ArrayCopy(ideal, this.ideal[index]);
index++;
}
/// <inheritdoc />
private void InternalCompute(int outputNeuron)
{
int row = 0;
var error = new ErrorCalculation();
var derivative = new double[_weightCount];
// Loop over every training element
foreach (IMLDataPair pair in _training)
{
EngineArray.Fill(derivative, 0);
IMLData networkOutput = _network.Compute(pair.Input);
double e = pair.Ideal[outputNeuron] - networkOutput[outputNeuron];
error.UpdateError(networkOutput[outputNeuron], pair.Ideal[outputNeuron]);
int currentWeight = 0;
// loop over the output weights
int outputFeedCount = _network.GetLayerTotalNeuronCount(_network.LayerCount - 2);
for (int i = 0; i < _network.OutputCount; i++)
{
for (int j = 0; j < outputFeedCount; j++)
{
double jc;
if (i == outputNeuron)
{
jc = ComputeDerivative(pair.Input, outputNeuron,
currentWeight, _dStep,
networkOutput[outputNeuron], row);
}
else
{
jc = 0;
}
_gradients[currentWeight] += jc * e;
derivative[currentWeight] = jc;
currentWeight++;
}
}
// Loop over every weight in the neural network
while (currentWeight < _network.Flat.Weights.Length)
{
double jc = ComputeDerivative(
pair.Input, outputNeuron, currentWeight,
_dStep,
networkOutput[outputNeuron], row);
derivative[currentWeight] = jc;
_gradients[currentWeight] += jc * e;
currentWeight++;
}
row++;
UpdateHessian(derivative);
}
_sse += error.CalculateSSE();
}
/// <summary>
/// Construct the singular value decomposition
/// </summary>
/// <param name="Arg">Rectangular matrix</param>
public SingularValueDecomposition(Matrix Arg)
{
// Derived from LINPACK code.
// Initialize.
double[][] A = Arg.GetArrayCopy();
m = Arg.Rows;
n = Arg.Cols;
/*
* Apparently the failing cases are only a proper subset of (m<n), so
* let's not throw error. Correct fix to come later? if (m<n) { throw
* new IllegalArgumentException("Jama SVD only works for m >= n"); }
*/
int nu = Math.Min(m, n);
s = new double[Math.Min(m + 1, n)];
umatrix = EngineArray.AllocateDouble2D(m, nu);
vmatrix = EngineArray.AllocateDouble2D(n, n);
var e = new double[n];
var work = new double[m];
bool wantu = true;
bool wantv = true;
// Reduce A to bidiagonal form, storing the diagonal elements
// in s and the super-diagonal elements in e.
int nct = Math.Min(m - 1, n);
int nrt = Math.Max(0, Math.Min(n - 2, m));
for (int k = 0; k < Math.Max(nct, nrt); k++)
{
if (k < nct)
{
// Compute the transformation for the k-th column and
// place the k-th diagonal in s[k].
// Compute 2-norm of k-th column without under/overflow.
s[k] = 0;
for (int i = k; i < m; i++)
{
s[k] = EncogMath.Hypot(s[k], A[i][k]);
}
if (s[k] != 0.0)
{
if (A[k][k] < 0.0)
{
s[k] = -s[k];
}
for (int i = k; i < m; i++)
{
A[i][k] /= s[k];
}
A[k][k] += 1.0;
}
s[k] = -s[k];
}
for (int j = k + 1; j < n; j++)
{
if ((k < nct) & (s[k] != 0.0))
{
// Apply the transformation.
double t = 0;
for (int i = k; i < m; i++)
{
t += A[i][k] * A[i][j];
}
t = -t / A[k][k];
for (int i = k; i < m; i++)
{
A[i][j] += t * A[i][k];
}
}
// Place the k-th row of A into e for the
// subsequent calculation of the row transformation.
e[j] = A[k][j];
}
if (wantu & (k < nct))
{
// Place the transformation in U for subsequent back
// multiplication.
for (int i = k; i < m; i++)
{
umatrix[i][k] = A[i][k];
}
}
if (k < nrt)
{
// Compute the k-th row transformation and place the
// k-th super-diagonal in e[k].
// Compute 2-norm without under/overflow.
e[k] = 0;
for (int i = k + 1; i < n; i++)
{
e[k] = EncogMath.Hypot(e[k], e[i]);
}
if (e[k] != 0.0)
{
if (e[k + 1] < 0.0)
{
e[k] = -e[k];
}
for (int i = k + 1; i < n; i++)
{
e[i] /= e[k];
}
e[k + 1] += 1.0;
}
e[k] = -e[k];
if ((k + 1 < m) & (e[k] != 0.0))
{
// Apply the transformation.
for (int i = k + 1; i < m; i++)
{
work[i] = 0.0;
}
for (int j = k + 1; j < n; j++)
{
for (int i = k + 1; i < m; i++)
{
work[i] += e[j] * A[i][j];
}
}
for (int j = k + 1; j < n; j++)
{
double t = -e[j] / e[k + 1];
for (int i = k + 1; i < m; i++)
{
A[i][j] += t * work[i];
}
}
}
if (wantv)
{
// Place the transformation in V for subsequent
// back multiplication.
for (int i = k + 1; i < n; i++)
{
vmatrix[i][k] = e[i];
}
}
}
}
// Set up the final bidiagonal matrix or order p.
int p = Math.Min(n, m + 1);
if (nct < n)
{
s[nct] = A[nct][nct];
}
if (m < p)
{
s[p - 1] = 0.0;
}
if (nrt + 1 < p)
{
e[nrt] = A[nrt][p - 1];
}
e[p - 1] = 0.0;
// If required, generate U.
if (wantu)
{
for (int j = nct; j < nu; j++)
{
for (int i = 0; i < m; i++)
{
umatrix[i][j] = 0.0;
}
umatrix[j][j] = 1.0;
}
for (int k = nct - 1; k >= 0; k--)
{
if (s[k] != 0.0)
{
for (int j = k + 1; j < nu; j++)
{
double t = 0;
for (int i = k; i < m; i++)
{
t += umatrix[i][k] * umatrix[i][j];
}
t = -t / umatrix[k][k];
for (int i = k; i < m; i++)
{
umatrix[i][j] += t * umatrix[i][k];
}
}
for (int i = k; i < m; i++)
{
umatrix[i][k] = -umatrix[i][k];
}
umatrix[k][k] = 1.0 + umatrix[k][k];
for (int i = 0; i < k - 1; i++)
{
umatrix[i][k] = 0.0;
}
}
else
{
for (int i = 0; i < m; i++)
{
umatrix[i][k] = 0.0;
}
umatrix[k][k] = 1.0;
}
}
}
// If required, generate V.
if (wantv)
{
for (int k = n - 1; k >= 0; k--)
{
if ((k < nrt) & (e[k] != 0.0))
{
for (int j = k + 1; j < nu; j++)
{
double t = 0;
for (int i = k + 1; i < n; i++)
{
t += vmatrix[i][k] * vmatrix[i][j];
}
t = -t / vmatrix[k + 1][k];
for (int i = k + 1; i < n; i++)
{
vmatrix[i][j] += t * vmatrix[i][k];
}
}
}
for (int i = 0; i < n; i++)
{
vmatrix[i][k] = 0.0;
}
vmatrix[k][k] = 1.0;
}
}
// Main iteration loop for the singular values.
int pp = p - 1;
int iter = 0;
double eps = Math.Pow(2.0, -52.0);
double tiny = Math.Pow(2.0, -966.0);
while (p > 0)
{
int k, kase;
// Here is where a test for too many iterations would go.
// This section of the program inspects for
// negligible elements in the s and e arrays. On
// completion the variables kase and k are set as follows.
// kase = 1 if s(p) and e[k-1] are negligible and k<p
// kase = 2 if s(k) is negligible and k<p
// kase = 3 if e[k-1] is negligible, k<p, and
// s(k), ..., s(p) are not negligible (qr step).
// kase = 4 if e(p-1) is negligible (convergence).
for (k = p - 2; k >= -1; k--)
{
if (k == -1)
{
break;
}
if (Math.Abs(e[k]) <= tiny + eps
* (Math.Abs(s[k]) + Math.Abs(s[k + 1])))
{
e[k] = 0.0;
break;
}
}
if (k == p - 2)
{
kase = 4;
}
else
{
int ks;
for (ks = p - 1; ks >= k; ks--)
{
if (ks == k)
{
break;
}
double t = (ks != p ? Math.Abs(e[ks]) : 0.0)
+ (ks != k + 1 ? Math.Abs(e[ks - 1]) : 0.0);
if (Math.Abs(s[ks]) <= tiny + eps * t)
{
s[ks] = 0.0;
break;
}
}
if (ks == k)
{
kase = 3;
}
else if (ks == p - 1)
{
kase = 1;
}
else
{
kase = 2;
k = ks;
}
}
k++;
// Perform the task indicated by kase.
switch (kase)
{
// Deflate negligible s(p).
case 1:
{
double f = e[p - 2];
e[p - 2] = 0.0;
for (int j = p - 2; j >= k; j--)
{
double t = EncogMath.Hypot(s[j], f);
double cs = s[j] / t;
double sn = f / t;
s[j] = t;
if (j != k)
{
f = -sn * e[j - 1];
e[j - 1] = cs * e[j - 1];
}
if (wantv)
{
for (int i = 0; i < n; i++)
{
t = cs * vmatrix[i][j] + sn * vmatrix[i][p - 1];
vmatrix[i][p - 1] = -sn * vmatrix[i][j] + cs * vmatrix[i][p - 1];
vmatrix[i][j] = t;
}
}
}
}
break;
// Split at negligible s(k).
case 2:
{
double f = e[k - 1];
e[k - 1] = 0.0;
for (int j = k; j < p; j++)
{
double t = EncogMath.Hypot(s[j], f);
double cs = s[j] / t;
double sn = f / t;
s[j] = t;
f = -sn * e[j];
e[j] = cs * e[j];
if (wantu)
{
for (int i = 0; i < m; i++)
{
t = cs * umatrix[i][j] + sn * umatrix[i][k - 1];
umatrix[i][k - 1] = -sn * umatrix[i][j] + cs * umatrix[i][k - 1];
umatrix[i][j] = t;
}
}
}
}
break;
// Perform one qr step.
case 3:
{
// Calculate the shift.
double scale = Math.Max(Math.Max(Math
.Max(Math.Max(Math.Abs(s[p - 1]), Math.Abs(s[p - 2])),
Math.Abs(e[p - 2])), Math.Abs(s[k])), Math
.Abs(
e[k]));
double sp = s[p - 1] / scale;
double spm1 = s[p - 2] / scale;
double epm1 = e[p - 2] / scale;
double sk = s[k] / scale;
double ek = e[k] / scale;
double b = ((spm1 + sp) * (spm1 - sp) + epm1 * epm1) / 2.0;
double c = (sp * epm1) * (sp * epm1);
double shift = 0.0;
if ((b != 0.0) | (c != 0.0))
{
shift = Math.Sqrt(b * b + c);
if (b < 0.0)
{
shift = -shift;
}
shift = c / (b + shift);
}
double f = (sk + sp) * (sk - sp) + shift;
double g = sk * ek;
// Chase zeros.
for (int j = k; j < p - 1; j++)
{
double t = EncogMath.Hypot(f, g);
double cs = f / t;
double sn = g / t;
if (j != k)
{
e[j - 1] = t;
}
f = cs * s[j] + sn * e[j];
e[j] = cs * e[j] - sn * s[j];
g = sn * s[j + 1];
s[j + 1] = cs * s[j + 1];
if (wantv)
{
for (int i = 0; i < n; i++)
{
t = cs * vmatrix[i][j] + sn * vmatrix[i][j + 1];
vmatrix[i][j + 1] = -sn * vmatrix[i][j] + cs * vmatrix[i][j + 1];
vmatrix[i][j] = t;
}
}
t = EncogMath.Hypot(f, g);
cs = f / t;
sn = g / t;
s[j] = t;
f = cs * e[j] + sn * s[j + 1];
s[j + 1] = -sn * e[j] + cs * s[j + 1];
g = sn * e[j + 1];
e[j + 1] = cs * e[j + 1];
if (wantu && (j < m - 1))
{
for (int i = 0; i < m; i++)
{
t = cs * umatrix[i][j] + sn * umatrix[i][j + 1];
umatrix[i][j + 1] = -sn * umatrix[i][j] + cs * umatrix[i][j + 1];
umatrix[i][j] = t;
}
}
}
e[p - 2] = f;
iter = iter + 1;
}
break;
// Convergence.
case 4:
{
// Make the singular values positive.
if (s[k] <= 0.0)
{
s[k] = (s[k] < 0.0 ? -s[k] : 0.0);
if (wantv)
{
for (int i = 0; i <= pp; i++)
{
vmatrix[i][k] = -vmatrix[i][k];
}
}
}
// Order the singular values.
while (k < pp)
{
if (s[k] >= s[k + 1])
{
break;
}
double t = s[k];
s[k] = s[k + 1];
s[k + 1] = t;
if (wantv && (k < n - 1))
{
for (int i = 0; i < n; i++)
{
t = vmatrix[i][k + 1];
vmatrix[i][k + 1] = vmatrix[i][k];
vmatrix[i][k] = t;
}
}
if (wantu && (k < m - 1))
{
for (int i = 0; i < m; i++)
{
t = umatrix[i][k + 1];
umatrix[i][k + 1] = umatrix[i][k];
umatrix[i][k] = t;
}
}
k++;
}
iter = 0;
p--;
}
break;
}
}
}
/// <inheritdoc/>
public int Classify(IMLData input)
{
IMLData output = Compute(input);
return(EngineArray.MaxIndex(output));
}
/// <summary>
/// Construct a network analyze class. Analyze the specified network.
/// </summary>
///
/// <param name="network">The network to analyze.</param>
public AnalyzeNetwork(BasicNetwork network)
{
int assignDisabled = 0;
int assignedTotal = 0;
IList <Double> biasList = new List <Double>();
IList <Double> weightList = new List <Double>();
IList <Double> allList = new List <Double>();
for (int layerNumber = 0; layerNumber < network.LayerCount - 1; layerNumber++)
{
int fromCount = network.GetLayerNeuronCount(layerNumber);
int fromBiasCount = network
.GetLayerTotalNeuronCount(layerNumber);
int toCount = network.GetLayerNeuronCount(layerNumber + 1);
// weights
for (int fromNeuron = 0; fromNeuron < fromCount; fromNeuron++)
{
for (int toNeuron = 0; toNeuron < toCount; toNeuron++)
{
double v = network.GetWeight(layerNumber, fromNeuron,
toNeuron);
if (network.Structure.ConnectionLimited)
{
if (Math.Abs(v) < network.Structure.ConnectionLimit)
{
assignDisabled++;
}
}
weightList.Add(v);
allList.Add(v);
assignedTotal++;
}
}
// bias
if (fromCount != fromBiasCount)
{
int biasNeuron = fromCount;
for (int toNeuron = 0; toNeuron < toCount; toNeuron++)
{
double v = network.GetWeight(layerNumber, biasNeuron,
toNeuron);
if (network.Structure.ConnectionLimited)
{
if (Math.Abs(v) < network.Structure.ConnectionLimit)
{
assignDisabled++;
}
}
biasList.Add(v);
allList.Add(v);
assignedTotal++;
}
}
}
_disabledConnections = assignDisabled;
_totalConnections = assignedTotal;
_weights = new NumericRange(weightList);
_bias = new NumericRange(biasList);
_weightsAndBias = new NumericRange(allList);
_weightValues = EngineArray.ListToDouble(weightList);
_allValues = EngineArray.ListToDouble(allList);
_biasValues = EngineArray.ListToDouble(biasList);
}
/// <inheritdoc/>
public void Clear()
{
EngineArray.Fill(_gradients, 0);
_hessianMatrix.Clear();
}
/// <summary>
/// Compute the output from this network.
/// </summary>
///
/// <param name="input">The input to the network.</param>
/// <returns>The output from the network.</returns>
public override sealed IMLData Compute(IMLData input)
{
var xout = new double[OutputCount];
double psum = 0.0d;
int r = -1;
foreach (IMLDataPair pair in _samples)
{
r++;
if (r == Exclude)
{
continue;
}
double dist = 0.0d;
for (int i = 0; i < InputCount; i++)
{
double diff = input[i] - pair.Input[i];
diff /= _sigma[i];
dist += diff * diff;
}
if (Kernel == PNNKernelType.Gaussian)
{
dist = Math.Exp(-dist);
}
else if (Kernel == PNNKernelType.Reciprocal)
{
dist = 1.0d / (1.0d + dist);
}
if (dist < 1.0e-40d)
{
dist = 1.0e-40d;
}
if (OutputMode == PNNOutputMode.Classification)
{
var pop = (int)pair.Ideal[0];
xout[pop] += dist;
}
else if (OutputMode == PNNOutputMode.Unsupervised)
{
for (int i = 0; i < InputCount; i++)
{
xout[i] += dist * pair.Input[i];
}
psum += dist;
}
else if (OutputMode == PNNOutputMode.Regression)
{
for (int i = 0; i < OutputCount; i++)
{
xout[i] += dist * pair.Ideal[i];
}
psum += dist;
}
}
if (OutputMode == PNNOutputMode.Classification)
{
psum = 0.0d;
for (int i = 0; i < OutputCount; i++)
{
if (_priors[i] >= 0.0d)
{
xout[i] *= _priors[i] / _countPer[i];
}
psum += xout[i];
}
if (psum < 1.0e-40d)
{
psum = 1.0e-40d;
}
for (int i = 0; i < OutputCount; i++)
{
xout[i] /= psum;
}
IMLData result = new BasicMLData(1);
result[0] = EngineArray.MaxIndex(xout);
return(result);
}
else if (OutputMode == PNNOutputMode.Unsupervised)
{
for (int i = 0; i < InputCount; i++)
{
xout[i] /= psum;
}
}
else if (OutputMode == PNNOutputMode.Regression)
{
for (int i = 0; i < OutputCount; i++)
{
xout[i] /= psum;
}
}
return(new BasicMLData(xout));
}
/// <summary>
/// Read an object.
/// </summary>
public Object Read(Stream mask0)
{
var ins0 = new EncogReadHelper(mask0);
EncogFileSection section;
var samples = new BasicMLDataSet();
IDictionary <String, String> networkParams = null;
PNNKernelType kernel = default(PNNKernelType) /* was: null */;
PNNOutputMode outmodel = default(PNNOutputMode) /* was: null */;
int inputCount = 0;
int outputCount = 0;
double error = 0;
double[] sigma = null;
while ((section = ins0.ReadNextSection()) != null)
{
if (section.SectionName.Equals("PNN") &&
section.SubSectionName.Equals("PARAMS"))
{
networkParams = section.ParseParams();
}
if (section.SectionName.Equals("PNN") &&
section.SubSectionName.Equals("NETWORK"))
{
IDictionary <String, String> paras = section.ParseParams();
inputCount = EncogFileSection.ParseInt(paras,
PersistConst.InputCount);
outputCount = EncogFileSection.ParseInt(paras,
PersistConst.OutputCount);
kernel = StringToKernel(paras[PersistConst.Kernel]);
outmodel = StringToOutputMode(paras[PropertyOutputMode]);
error = EncogFileSection
.ParseDouble(paras, PersistConst.Error);
sigma = section.ParseDoubleArray(paras, PersistConst.Sigma);
}
if (section.SectionName.Equals("PNN") &&
section.SubSectionName.Equals("SAMPLES"))
{
foreach (String line in section.Lines)
{
IList <String> cols = EncogFileSection
.SplitColumns(line);
int index = 0;
var inputData = new BasicMLData(inputCount);
for (int i = 0; i < inputCount; i++)
{
inputData[i] =
CSVFormat.EgFormat.Parse(cols[index++]);
}
var idealData = new BasicMLData(inputCount);
idealData[0] = CSVFormat.EgFormat.Parse(cols[index++]);
IMLDataPair pair = new BasicMLDataPair(inputData,
idealData);
samples.Add(pair);
}
}
}
var result = new BasicPNN(kernel, outmodel, inputCount,
outputCount);
if (networkParams != null)
{
EngineArray.PutAll(networkParams, result.Properties);
}
result.Samples = samples;
result.Error = error;
if (sigma != null)
{
EngineArray.ArrayCopy(sigma, result.Sigma);
}
return(result);
}
/// <summary>
/// Perform a revert.
/// </summary>
/// <param name="revertedData">The source data to revert from.</param>
public void PerformRevert(IDictionary <String, String> revertedData)
{
_data.Clear();
EngineArray.PutAll(revertedData, _data);
}
/// <summary>
/// Copy weights and output to the network.
/// </summary>
///
/// <param name="target">The network to copy to.</param>
public void CopyToNetwork(FlatNetwork target)
{
EngineArray.ArrayCopy(_weights, target.Weights);
EngineArray.ArrayCopy(_output, target.LayerOutput);
}
/// <summary>
/// Perform one iteration.
/// </summary>
///
public override void Iteration()
{
if (_mustInit)
{
Init();
}
int numWeights = _weights.Length;
// Storage space for previous iteration values.
if (_restart)
{
// First time through, set initial values for SCG parameters.
_lambda = FirstLambda;
_lambda2 = 0;
_k = 1;
_success = true;
_restart = false;
}
// If an error reduction is possible, calculate 2nd order info.
if (_success)
{
// If the search direction is small, stop.
_magP = EngineArray.VectorProduct(_p, _p);
double sigma = FirstSigma
/ Math.Sqrt(_magP);
// In order to compute the new step, we need a new gradient.
// First, save off the old data.
EngineArray.ArrayCopy(Gradients, _oldGradient);
EngineArray.ArrayCopy(_weights, _oldWeights);
_oldError = Error;
// Now we move to the new point in weight space.
for (int i = 0; i < numWeights; ++i)
{
_weights[i] += sigma * _p[i];
}
EngineArray.ArrayCopy(_weights, Network.Weights);
// And compute the new gradient.
CalculateGradients();
// Now we have the new gradient, and we continue the step
// computation.
_delta = 0;
for (int i = 0; i < numWeights; ++i)
{
double step = (Gradients[i] - _oldGradient[i])
/ sigma;
_delta += _p[i] * step;
}
}
// Scale delta.
_delta += (_lambda - _lambda2) * _magP;
// If delta <= 0, make Hessian positive definite.
if (_delta <= 0)
{
_lambda2 = 2 * (_lambda - _delta / _magP);
_delta = _lambda * _magP - _delta;
_lambda = _lambda2;
}
// Calculate step size.
double mu = EngineArray.VectorProduct(_p, _r);
double alpha = mu / _delta;
// Calculate the comparison parameter.
// We must compute a new gradient, but this time we do not
// want to keep the old values. They were useful only for
// approximating the Hessian.
for (int i = 0; i < numWeights; ++i)
{
_weights[i] = _oldWeights[i] + alpha * _p[i];
}
EngineArray.ArrayCopy(_weights, Network.Weights);
CalculateGradients();
double gdelta = 2 * _delta * (_oldError - Error)
/ (mu * mu);
// If gdelta >= 0, a successful reduction in error is possible.
if (gdelta >= 0)
{
// Product of r(k+1) by r(k)
double rsum = 0;
// Now r = r(k+1).
for (int i = 0; i < numWeights; ++i)
{
double tmp = -Gradients[i];
rsum += tmp * _r[i];
_r[i] = tmp;
}
_lambda2 = 0;
_success = true;
// Do we need to restart?
if (_k >= numWeights)
{
_restart = true;
EngineArray.ArrayCopy(_r, _p);
}
else
{
// Compute new conjugate direction.
double beta = (EngineArray.VectorProduct(_r, _r) - rsum)
/ mu;
// Update direction vector.
for (int i = 0; i < numWeights; ++i)
{
_p[i] = _r[i] + beta * _p[i];
}
_restart = false;
}
if (gdelta >= 0.75D)
{
_lambda *= 0.25D;
}
}
else
{
// A reduction in error was not possible.
// under_tolerance = false;
// Go back to w(k) since w(k) + alpha*p(k) is not better.
EngineArray.ArrayCopy(_oldWeights, _weights);
CurrentError = _oldError;
_lambda2 = _lambda;
_success = false;
}
if (gdelta < 0.25D)
{
_lambda += _delta * (1 - gdelta) / _magP;
}
_lambda = BoundNumbers.Bound(_lambda);
++_k;
EngineArray.ArrayCopy(_weights, Network.Weights);
}
/// <inheritdoc/>
public Object Read(Stream istream)
{
int states = 0;
int[] items;
double[] pi = null;
Matrix transitionProbability = null;
IDictionary <String, String> properties = null;
IList <IStateDistribution> distributions = new List <IStateDistribution>();
EncogReadHelper reader = new EncogReadHelper(istream);
EncogFileSection section;
while ((section = reader.ReadNextSection()) != null)
{
if (section.SectionName.Equals("HMM") &&
section.SubSectionName.Equals("PARAMS"))
{
properties = section.ParseParams();
}
if (section.SectionName.Equals("HMM") &&
section.SubSectionName.Equals("CONFIG"))
{
IDictionary <String, String> p = section.ParseParams();
states = EncogFileSection.ParseInt(p, HiddenMarkovModel.TAG_STATES);
if (p.ContainsKey(HiddenMarkovModel.TAG_ITEMS))
{
items = EncogFileSection.ParseIntArray(p, HiddenMarkovModel.TAG_ITEMS);
}
pi = section.ParseDoubleArray(p, HiddenMarkovModel.TAG_PI);
transitionProbability = EncogFileSection.ParseMatrix(p, HiddenMarkovModel.TAG_TRANSITION);
}
else if (section.SectionName.Equals("HMM") &&
section.SubSectionName.StartsWith("DISTRIBUTION-"))
{
IDictionary <String, String> p = section.ParseParams();
String t = p[HiddenMarkovModel.TAG_DIST_TYPE];
if ("ContinousDistribution".Equals(t))
{
double[] mean = section.ParseDoubleArray(p, HiddenMarkovModel.TAG_MEAN);
Matrix cova = EncogFileSection.ParseMatrix(p, HiddenMarkovModel.TAG_COVARIANCE);
ContinousDistribution dist = new ContinousDistribution(mean, cova.Data);
distributions.Add(dist);
}
else if ("DiscreteDistribution".Equals(t))
{
Matrix prob = EncogFileSection.ParseMatrix(p, HiddenMarkovModel.TAG_PROBABILITIES);
DiscreteDistribution dist = new DiscreteDistribution(prob.Data);
distributions.Add(dist);
}
}
}
HiddenMarkovModel result = new HiddenMarkovModel(states);
EngineArray.PutAll(properties, result.Properties);
result.TransitionProbability = transitionProbability.Data;
result.Pi = pi;
int index = 0;
foreach (IStateDistribution dist in distributions)
{
result.StateDistributions[index++] = dist;
}
return(result);
}
/// <summary>
/// Clear to zero.
/// </summary>
public void Clear()
{
EngineArray.Fill(_data, 0);
}
/// <summary>
/// Determine the winner for the specified input. This is the number of the
/// winning neuron.
/// </summary>
///
/// <param name="input">The input patter to present to the neural network.</param>
/// <returns>The winning neuron.</returns>
public int Winner(IMLData input)
{
IMLData output = Compute(input);
return(EngineArray.MaxIndex(output));
}
public void CopyTo(double[] target, int targetIndex, int count)
{
EngineArray.ArrayCopy(_data, 0, target, targetIndex, count);
}
/// <summary>
/// dst = src
///
/// Copy a vector.
/// </summary>
/// <param name="dst">an array of doubles</param>
/// <param name="src">an array of doubles</param>
public void Copy(double[] dst, double[] src)
{
EngineArray.ArrayCopy(src, dst);
}
/// <summary>
/// Perform one iteration.
/// </summary>
///
public override void Iteration()
{
if (shouldInit)
{
Init();
}
int numWeights = this.weights.Length;
// Storage space for previous iteration values.
if (this.restart)
{
// First time through, set initial values for SCG parameters.
this.lambda = TrainFlatNetworkSCG.FIRST_LAMBDA;
this.lambda2 = 0;
this.k = 1;
this.success = true;
this.restart = false;
}
// If an error reduction is possible, calculate 2nd order info.
if (this.success)
{
// If the search direction is small, stop.
this.magP = EngineArray.VectorProduct(this.p, this.p);
double sigma = TrainFlatNetworkSCG.FIRST_SIGMA
/ Math.Sqrt(this.magP);
// In order to compute the new step, we need a new gradient.
// First, save off the old data.
EngineArray.ArrayCopy(this.gradients, this.oldGradient);
EngineArray.ArrayCopy(this.weights, this.oldWeights);
this.oldError = Error;
// Now we move to the new point in weight space.
for (int i = 0; i < numWeights; ++i)
{
this.weights[i] += sigma * this.p[i];
}
EngineArray.ArrayCopy(this.weights, this.network.Weights);
// And compute the new gradient.
CalculateGradients();
// Now we have the new gradient, and we continue the step
// computation.
this.delta = 0;
for (int i_0 = 0; i_0 < numWeights; ++i_0)
{
double step = (this.gradients[i_0] - this.oldGradient[i_0])
/ sigma;
this.delta += this.p[i_0] * step;
}
}
// Scale delta.
this.delta += (this.lambda - this.lambda2) * this.magP;
// If delta <= 0, make Hessian positive definite.
if (this.delta <= 0)
{
this.lambda2 = 2 * (this.lambda - this.delta / this.magP);
this.delta = this.lambda * this.magP - this.delta;
this.lambda = this.lambda2;
}
// Calculate step size.
double mu = EngineArray.VectorProduct(this.p, this.r);
double alpha = mu / this.delta;
// Calculate the comparison parameter.
// We must compute a new gradient, but this time we do not
// want to keep the old values. They were useful only for
// approximating the Hessian.
for (int i_1 = 0; i_1 < numWeights; ++i_1)
{
this.weights[i_1] = this.oldWeights[i_1] + alpha * this.p[i_1];
}
EngineArray.ArrayCopy(this.weights, this.network.Weights);
CalculateGradients();
double gdelta = 2 * this.delta * (this.oldError - Error)
/ (mu * mu);
// If gdelta >= 0, a successful reduction in error is possible.
if (gdelta >= 0)
{
// Product of r(k+1) by r(k)
double rsum = 0;
// Now r = r(k+1).
for (int i_2 = 0; i_2 < numWeights; ++i_2)
{
double tmp = -this.gradients[i_2];
rsum += tmp * this.r[i_2];
this.r[i_2] = tmp;
}
this.lambda2 = 0;
this.success = true;
// Do we need to restart?
if (this.k >= numWeights)
{
this.restart = true;
EngineArray.ArrayCopy(this.r, this.p);
}
else
{
// Compute new conjugate direction.
double beta = (EngineArray.VectorProduct(this.r, this.r) - rsum)
/ mu;
// Update direction vector.
for (int i_3 = 0; i_3 < numWeights; ++i_3)
{
this.p[i_3] = this.r[i_3] + beta * this.p[i_3];
}
this.restart = false;
}
if (gdelta >= 0.75D)
{
this.lambda *= 0.25D;
}
}
else
{
// A reduction in error was not possible.
// under_tolerance = false;
// Go back to w(k) since w(k) + alpha*p(k) is not better.
EngineArray.ArrayCopy(this.oldWeights, this.weights);
this.currentError = this.oldError;
this.lambda2 = this.lambda;
this.success = false;
}
if (gdelta < 0.25D)
{
this.lambda += this.delta * (1 - gdelta) / this.magP;
}
this.lambda = BoundNumbers.Bound(this.lambda);
++this.k;
EngineArray.ArrayCopy(this.weights, this.network.Weights);
}
