/** Compute y <- alpha*op(a)*x + beta * y (general matrix vector multiplication) */
public static DoubleMatrix gemv(
double alpha, DoubleMatrix a, DoubleMatrix x, double beta, DoubleMatrix y) {
if (false) {
NativeBlas.dgemv(
'N', a.rows, a.columns, alpha, a.data, 0, a.rows, x.data, 0, 1, beta, y.data, 0, 1);
} else {
if (beta == 0.0) {
for (int i = 0; i < y.length; i++) y.data[i] = 0.0;
for (int j = 0; j < a.columns; j++) {
double xj = x.get(j);
if (xj != 0.0) {
for (int i = 0; i < a.rows; i++) y.data[i] += a.get(i, j) * xj;
}
}
} else {
for (int j = 0; j < a.columns; j++) {
double byj = beta * y.data[j];
double xj = x.get(j);
for (int i = 0; i < a.rows; i++) y.data[j] = a.get(i, j) * xj + byj;
}
}
}
return y;
}
private void modifyWeights(
DoubleMatrix trainingExample,
DoubleMatrix result,
DoubleMatrix output,
double learningFactor,
double momentum,
List<DoubleMatrix> previousModifications,
EvaluationContext evalCtx) {
List<DoubleMatrix> errors = countErrors(result, output);
List<Layer> layers = getLayers();
evalCtx.resetContext();
Iterator<ActivationFunction> activationFunctionIter = evalCtx.getActivationFunction();
DoubleMatrix temporalResult = trainingExample;
for (int i = 0; i < errors.size(); i++) {
DoubleMatrix error = errors.get(i);
int layerIndex = i + 1;
Layer layer = layers.get(layerIndex);
DoubleMatrix previousModification = previousModifications.get(i);
ActivationFunction activationFunc = activationFunctionIter.next();
ActivationFunctionDerivative derivative =
DerivativeFactory.getInstance().getDerivative(activationFunc);
if (layer.includeBias()) {
temporalResult = addBiasInput(temporalResult);
}
DoubleMatrix oldVal = temporalResult.dup();
temporalResult = layer.getWeights().mmul(temporalResult);
temporalResult = activationFunc.eval(temporalResult);
// dla kazdego neuronu w warstwie
for (int j = 0; j < layer.getWeights().rows; j++) {
double derVal = derivative.evaluate(temporalResult.get(j));
DoubleMatrix oldDelta = previousModification.getRow(j);
DoubleMatrix delta = oldVal.mul(derVal).mul(learningFactor).mul(error.get(j));
delta = delta.transpose();
delta = delta.add(oldDelta.mul(momentum));
previousModification.putRow(j, delta);
DoubleMatrix oldWeights = layer.getWeights().getRow(j);
DoubleMatrix newWeights = oldWeights.add(delta);
layer.getWeights().putRow(j, newWeights);
}
}
}
/**
* Get Result String
*
* @return result string
*/
public String printComparableResults(boolean newline) {
if (!(resultVector != null)) log.error("Attempted to access null vector.");
String outString = "";
int i = 0;
for (TypeQ question : sortedQuestions) {
Double tempResult = resultVector.get(i);
// if(tempResult.intValue() == 0)
// continue;
if (groundTruthVector != null) {
Double evalInd = groundTruthVector.getSecond().get(i);
if (evalInd.intValue() == 0) {
++i;
continue;
}
}
++i;
if (!newline) outString += "\n";
outString += question + " " + tempResult.intValue();
newline = false;
}
if (log.isDebugEnabled()) {
log.debug(outString);
}
return outString;
}
public static void DFS(
DoubleMatrix A, DoubleMatrix d, DoubleMatrix g, ArrayList[] V, int vertNum) {
int krawedzie = 0;
for (int i = 0; i <= vertNum; i++) {
krawedzie += V[i].size();
}
krawedzie /= 2;
// System.out.println("A wymiar: " + A.columns + " rows: "+A.rows);
int t = 0;
VertexState state[] = new VertexState[(d.rows - 1) * d.rows / 2]; // tyle stanów
for (int i = 0; i < d.rows - 1; i++) {
for (int j = i + 1; j < d.columns; j++) {
for (int x = 0; x < state.length; x++) {
state[x] = VertexState.White; // na początku białe
}
g.put(t, 0, d.get(i, j));
runDFS(i, j, state, d, V, vertNum);
for (int ii = 0; ii < krawedzie; ii++) {
if (state[ii] == VertexState.Black) {
A.put(t, ii, 1);
}
}
t++;
}
}
System.out.println(g);
System.out.println("t:" + t);
return;
}
public static void main(String argv[]) {
modshogun.init_shogun_with_defaults();
int dim = 7;
DoubleMatrix data = DoubleMatrix.rand(dim, dim);
RealFeatures feats = new RealFeatures(data);
DoubleMatrix data_T = data.transpose();
DoubleMatrix symdata = data.add(data_T);
int cols = (1 + dim) * dim / 2;
DoubleMatrix lowertriangle = DoubleMatrix.zeros(1, cols);
int count = 0;
for (int i = 0; i < dim; i++) {
for (int j = 0; j < dim; j++) {
if (j <= i) {
lowertriangle.put(0, count++, symdata.get(i, j));
}
}
}
CustomKernel kernel = new CustomKernel();
kernel.set_triangle_kernel_matrix_from_triangle(lowertriangle);
DoubleMatrix km_triangletriangle = kernel.get_kernel_matrix();
kernel.set_triangle_kernel_matrix_from_full(symdata);
DoubleMatrix km_fulltriangle = kernel.get_kernel_matrix();
kernel.set_full_kernel_matrix_from_full(data);
DoubleMatrix km_fullfull = kernel.get_kernel_matrix();
modshogun.exit_shogun();
}
private static void modi(DoubleMatrix syndrome, int radix) {
for (int i = 0; i < syndrome.getRows(); i++) {
for (int j = 0; j < syndrome.getColumns(); j++) {
syndrome.put(i, j, ((int) syndrome.get(i, j)) % radix);
}
}
}
/**
* 对角线增加值
*
* @param matrix
* @throws Exception
*/
private void addPositiveValue(DoubleMatrix matrix) {
int m = matrix.rows;
if (matrix.columns != m) return;
for (int i = 0; i < m; i++) {
matrix.put(i, i, matrix.get(i, i) + Math.random());
}
}
// Returns value of approximated function
public double approximateFunction(double x, final int NUMBER_OF_EQUATIONS) {
DoubleMatrix X = solveLinearEquation(NUMBER_OF_EQUATIONS);
double result = 0.0;
for (int i = 0; i < NUMBER_OF_EQUATIONS; i++) {
result += X.get(i) * Math.pow(x, i);
}
return result;
}
@Test
public void extractFeature_simpleExpression_correctAnswer() {
// Arrange
setCustomFeatureExpression(" 8 + 14 ");
FormattedSegments segments = getFormattedSegments();
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals(22, features.get(0, 0), errorMargin);
}
@Test
public void extractFeature_gpsSpeedVariableExpression_correctAnswer() {
// Arrange
setCustomFeatureExpression(" 8+ speed1 ");
double[][] gpsData = new double[][] {{0.5}};
FormattedSegments segments = getFormattedSegments(gpsData);
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals(8.5, features.get(0, 0), errorMargin);
}
@Test
public void extractFeature_multipleSegmentsExpression_2ndHasCorrectAnswer() {
// Arrange
setCustomFeatureExpression(" 8+ speed1");
double[][] gpsData = new double[][] {{0.5}, {11.0}};
double[][] xData = new double[][] {{0}, {0}};
FormattedSegments segments = getFormattedSegments(xData, xData, xData, gpsData);
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals(19, features.get(1, 0), errorMargin);
}
@Test
public void extractFeature_meanxVariableExpression_correctAnswer() {
// Arrange
setCustomFeatureExpression("meanx");
double[][] gpsData = new double[][] {{0}};
double[][] xData = new double[][] {{2, 2, 2, 4}}; // mean = 2.5
double[][] yData = new double[][] {{0, 0, 0, 0}};
double[][] zData = new double[][] {{0, 0, 0, 0}};
FormattedSegments segments = getFormattedSegments(xData, yData, zData, gpsData);
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals(2.5, features.get(0, 0), errorMargin);
}
@Test
public void extractFeature_y2VariableExpression_correctAnswer() {
// Arrange
setCustomFeatureExpression("12 * y3+1");
double[][] gpsData = new double[][] {{0}};
double[][] xData = new double[][] {{0, 0, 0}};
double[][] yData = new double[][] {{5, 8, 0.5}};
double[][] zData = new double[][] {{0, 0, 0}};
FormattedSegments segments = getFormattedSegments(xData, yData, zData, gpsData);
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals((12 * 0.5) + 1, features.get(0, 0), errorMargin);
}
@Test
public void extractFeatures_3points_correctResult() {
// Arrange
DoubleMatrix x, y, z, gpsSpeed;
x = new DoubleMatrix(new double[][] {{1, 1, 1}});
y = new DoubleMatrix(new double[][] {{1, 1, 1}});
z = new DoubleMatrix(new double[][] {{7, 3, 5}});
gpsSpeed = new DoubleMatrix(x.rows, 1);
FormattedSegments input = createFormattedSegments(x, y, z, gpsSpeed);
// Act
DoubleMatrix features = featureExtractor.extractFeatures(input);
// Assert
assertEquals(3, features.get(0, 0), errorMargin);
}
@Test
public void extractFeatures_onePointPerSequence_result0() {
// Arrange
DoubleMatrix x, y, z, gpsSpeed;
x = new DoubleMatrix(new double[][] {{1}});
y = new DoubleMatrix(new double[][] {{1}});
z = new DoubleMatrix(new double[][] {{1}});
gpsSpeed = new DoubleMatrix(x.rows, 1);
FormattedSegments input = createFormattedSegments(x, y, z, gpsSpeed);
// Act
DoubleMatrix features = featureExtractor.extractFeatures(input);
// Assert
assertEquals(0, features.get(0, 0), errorMargin);
}
@Test
public void extractFeature_stdyVariableExpression_correctAnswer() {
// Arrange
setCustomFeatureExpression("stdy");
double[][] gpsData = new double[][] {{0}};
double[][] xData = new double[][] {{0, 0, 0}};
double[][] yData = new double[][] {{1, 2, 3}};
double[][] zData = new double[][] {{0, 0, 0}};
FormattedSegments segments = getFormattedSegments(xData, yData, zData, gpsData);
double expected = Math.sqrt((1 + 0 + 1) / (3 - 1));
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals(expected, features.get(0, 0), errorMargin);
}
@Test
public void extractFeature_altitude_correctAnswer() {
// Arrange
setCustomFeatureExpression("meanalt");
double[][] gpsData = new double[][] {{0}};
double[][] accData = new double[][] {{0}};
double[][] altitudeData = new double[][] {{2, 3}};
DateTime[][] timeData = new DateTime[1][1];
FormattedSegments segments =
getFormattedSegments(
accData, accData, accData, gpsData, gpsData, gpsData, altitudeData, timeData);
double expected = 2.5;
// Act
DoubleMatrix features = featureExtractor.extractFeatures(segments);
// Assert
assertEquals(expected, features.get(0, 0), errorMargin);
}
public static DoubleMatrix conv2d(DoubleMatrix input, DoubleMatrix kernel, Type type) {
DoubleMatrix xShape = new DoubleMatrix(1, 2);
xShape.put(0, input.rows);
xShape.put(1, input.columns);
DoubleMatrix yShape = new DoubleMatrix(1, 2);
yShape.put(0, kernel.rows);
yShape.put(1, kernel.columns);
DoubleMatrix zShape = xShape.addi(yShape).subi(1);
int retRows = (int) zShape.get(0);
int retCols = (int) zShape.get(1);
ComplexDoubleMatrix fftInput = complexDisceteFourierTransform(input, retRows, retCols);
ComplexDoubleMatrix fftKernel = complexDisceteFourierTransform(kernel, retRows, retCols);
ComplexDoubleMatrix mul = fftKernel.muli(fftInput);
ComplexDoubleMatrix retComplex = complexInverseDisceteFourierTransform(mul);
DoubleMatrix ret = retComplex.getReal();
if (type == Type.VALID) {
DoubleMatrix validShape = xShape.subi(yShape).addi(1);
DoubleMatrix start = zShape.subi(validShape).divi(2);
DoubleMatrix end = start.addi(validShape);
if (start.get(0) < 1 || start.get(1) < 1)
throw new IllegalStateException("Illegal row index " + start);
if (end.get(0) < 1 || end.get(1) < 1)
throw new IllegalStateException("Illegal column index " + end);
ret =
ret.get(
RangeUtils.interval((int) start.get(0), (int) end.get(0)),
RangeUtils.interval((int) start.get(1), (int) end.get(1)));
}
return ret;
}
/**
* --- B(l,p,h) = a000 + a100*l + a010*p + a001*h + a110*l*p + a101*l*h + a011*p*h + a200*l^2 +
* a020*p^2 + a002*h^2
*/
private double polyVal(
final DoubleMatrix C, final double line, final double pixel, final double height)
throws Exception {
logger.trace("polyVal for baseline modeling");
if (C.length != 10) {
throw new Exception();
} else {
return C.get(0, 0)
+ C.get(1, 0) * line
+ C.get(2, 0) * pixel
+ C.get(3, 0) * height
+ C.get(4, 0) * line * pixel
+ C.get(5, 0) * line * height
+ C.get(6, 0) * pixel * height
+ C.get(7, 0) * sqr(line)
+ C.get(8, 0) * sqr(pixel)
+ C.get(9, 0) * sqr(height);
}
}
public void model(
final SLCImage master, final SLCImage slave, Orbit masterOrbit, Orbit slaveOrbit)
throws Exception {
if (!masterOrbit.isInterpolated()) {
logger.debug("Baseline cannot be computed, master orbit not initialized.");
throw new Exception("Baseline.model_parameters: master orbit not initialized");
} else if (!slaveOrbit.isInterpolated()) {
logger.debug("Baseline cannot be computed, slave orbit not initialized.");
throw new Exception("Baseline.model_parameters: slave orbit not initialized");
}
if (isInitialized) {
logger.warn("baseline already isInitialized??? (returning)");
return;
}
masterWavelength = master.getRadarWavelength();
// Model r = nearRange + drange_dp*p -- p starts at 1
nearRange = master.pix2range(1.0);
drange_dp = master.pix2range(2.0) - master.pix2range(1.0);
nearRange -= drange_dp; // -- p starts at 1
// Set min/maxima for normalization
linMin = master.currentWindow.linelo; // also used during polyval...
linMax = master.currentWindow.linehi;
pixMin = master.currentWindow.pixlo;
pixMax = master.currentWindow.pixhi;
hMin = 0.0;
hMax = 5000.0;
// Loop counters and sampling
int cnt = 0; // matrix index
final double deltaPixels = master.currentWindow.pixels() / N_POINTS_RNG;
final double deltaLines = master.currentWindow.lines() / N_POINTS_AZI;
final double deltaHeight = (hMax - hMin) / N_HEIGHTS;
// Declare matrices for modeling Bperp
// Note: for stability of normalmatrix, fill aMatrix with normalized line, etc.
// perpendicular baseline
DoubleMatrix bPerpMatrix = new DoubleMatrix(N_POINTS_AZI * N_POINTS_RNG * N_HEIGHTS, 1);
// parallel baseline
DoubleMatrix bParMatrix = new DoubleMatrix(N_POINTS_AZI * N_POINTS_RNG * N_HEIGHTS, 1);
// viewing angle
DoubleMatrix thetaMatrix = new DoubleMatrix(N_POINTS_AZI * N_POINTS_RNG * N_HEIGHTS, 1);
// incidence angle
DoubleMatrix thetaIncMatrix = new DoubleMatrix(N_POINTS_AZI * N_POINTS_RNG * N_HEIGHTS, 1);
// design matrix
DoubleMatrix aMatrix = new DoubleMatrix(N_POINTS_AZI * N_POINTS_RNG * N_HEIGHTS, numCoeffs);
// Loop over heights(k), lines(i), pixels(j) to estimate baseline
// height levels
for (long k = 0; k < N_HEIGHTS; ++k) {
final double height = hMin + k * deltaHeight;
// azimuth direction
for (long i = 0; i < N_POINTS_AZI; ++i) {
final double line = master.currentWindow.linelo + i * deltaLines;
Point pointOnEllips; // point, returned by lp2xyz
double sTazi, sTrange;
// Azimuth time for this line
final double mTazi = master.line2ta(line);
// xyz for master satellite from time
final Point pointOnMasterOrb = masterOrbit.getXYZ(mTazi);
// Continue looping in range direction
for (long j = 0; j < N_POINTS_RNG; ++j) {
final double pixel = master.currentWindow.pixlo + j * deltaPixels;
// ______ Range time for this pixel ______
// final double m_trange = master.pix2tr(pixel);
pointOnEllips = masterOrbit.lph2xyz(line, pixel, height, master);
// Compute xyz for slave satellite from pointOnEllips
Point pointTime = slaveOrbit.xyz2t(pointOnEllips, slave);
sTazi = pointTime.y;
sTrange = pointTime.x;
// Slave position
final Point pointOnSlaveOrb = slaveOrbit.getXYZ(sTazi);
// Compute angle between near parallel orbits
final Point velOnMasterOrb = masterOrbit.getXYZDot(mTazi);
final Point velOnSlaveOrb = slaveOrbit.getXYZDot(sTazi);
final double angleOrbits = velOnMasterOrb.angle(velOnSlaveOrb);
logger.debug(
"Angle between orbits master-slave (at l,p= "
+ line
+ ","
+ pixel
+ ") = "
+ rad2deg(angleOrbits)
+ " [deg]");
// Note: convergence assumed constant!
orbitConvergence = angleOrbits;
// final heading = angle(velOnMasterOrb,[1 0 0]) //?
// orbitHeading = 0.0; // not yet used
// The baseline parameters, derived from the positions (x,y,z)
// alpha is angle counterclockwize(b, plane with normal=rho1=rho2)
// theta is angle counterclockwize(rho1 = pointOnMasterOrb, r1 = pointOnMasterOrb -
// pointOnEllips, r2 = pointOnSlaveOrb - pointOnEllips)
// construct helper class
BaselineComponents baselineComponents = new BaselineComponents().invoke();
compute_B_Bpar_Bperp_Theta(
baselineComponents, pointOnEllips, pointOnMasterOrb, pointOnSlaveOrb);
final double b = baselineComponents.getB();
final double bPar = baselineComponents.getBpar();
final double bPerp = baselineComponents.getBperp();
final double theta = baselineComponents.getTheta();
final double thetaInc = computeIncAngle(pointOnMasterOrb, pointOnEllips); // [rad]!!!
// Modelling of bPerp(l,p) = a00 + a10*l + a01*p
bPerpMatrix.put(cnt, 0, bPerp);
bParMatrix.put(cnt, 0, bPar);
thetaMatrix.put(cnt, 0, theta);
thetaIncMatrix.put(cnt, 0, thetaInc);
// --- b(l,p,h) = a000 +
// a100*l + a010*p + a001*h +
// a110*l*p + a101*l*h + a011*p*h +
// a200*l^2 + a020*p^2 + a002*h^2
aMatrix.put(cnt, 0, 1.0);
aMatrix.put(cnt, 1, normalize2(line, linMin, linMax));
aMatrix.put(cnt, 2, normalize2(pixel, pixMin, pixMax));
aMatrix.put(cnt, 3, normalize2(height, hMin, hMax));
aMatrix.put(cnt, 4, normalize2(line, linMin, linMax) * normalize2(pixel, pixMin, pixMax));
aMatrix.put(cnt, 5, normalize2(line, linMin, linMax) * normalize2(height, hMin, hMax));
aMatrix.put(cnt, 6, normalize2(pixel, pixMin, pixMax) * normalize2(height, hMin, hMax));
aMatrix.put(cnt, 7, sqr(normalize2(line, linMin, linMax)));
aMatrix.put(cnt, 8, sqr(normalize2(pixel, pixMin, pixMax)));
aMatrix.put(cnt, 9, sqr(normalize2(height, hMin, hMax)));
cnt++;
// b/alpha representation of baseline
final double alpha =
(bPar == 0 && bPerp == 0)
? Double.NaN
: theta - Math.atan2(bPar, bPerp); // sign ok atan2
// horizontal/vertical representation of baseline
final double bH = b * Math.cos(alpha); // sign ok
final double bV = b * Math.sin(alpha); // sign ok
// TODO: check sign of infinity!!!
// Height ambiguity: [h] = -lambda/4pi * (r1sin(theta)/bPerp) * phi==2pi
final double hAmbiguity =
(bPerp == 0)
? Double.POSITIVE_INFINITY
: -master.getRadarWavelength()
* (pointOnMasterOrb.min(pointOnEllips)).norm()
* Math.sin(theta)
/ (2.0 * bPerp);
// Some extra info if in DEBUG unwrapMode
logger.debug(
"The baseline parameters for (l,p,h) = " + line + ", " + pixel + ", " + height);
logger.debug("\talpha (deg), BASELINE: \t" + rad2deg(alpha) + " \t" + b);
logger.debug("\tbPar, bPerp: \t" + bPar + " \t" + bPerp);
logger.debug("\tbH, bV: \t" + bH + " \t" + bV);
logger.debug("\tHeight ambiguity: \t" + hAmbiguity);
logger.debug("\ttheta (deg): \t" + rad2deg(theta));
logger.debug("\tthetaInc (deg): \t" + rad2deg(thetaInc));
logger.debug("\tpointOnMasterOrb (x,y,z) = " + pointOnMasterOrb.toString());
logger.debug("\tpointOnSlaveOrb (x,y,z) = " + pointOnSlaveOrb.toString());
logger.debug("\tpointOnEllips (x,y,z) = " + pointOnEllips.toString());
} // loop pixels
} // loop lines
} // loop heights
// Model all Baselines as 2d polynomial of degree 1
DoubleMatrix nMatrix = matTxmat(aMatrix, aMatrix);
DoubleMatrix rhsBperp = matTxmat(aMatrix, bPerpMatrix);
DoubleMatrix rhsBpar = matTxmat(aMatrix, bParMatrix);
DoubleMatrix rhsTheta = matTxmat(aMatrix, thetaMatrix);
DoubleMatrix rhsThetaInc = matTxmat(aMatrix, thetaIncMatrix);
// DoubleMatrix Qx_hat = nMatrix;
final DoubleMatrix Qx_hat =
LinearAlgebraUtils.invertChol(Decompose.cholesky(nMatrix).transpose());
// TODO: refactor to _internal_ cholesky decomposition
// choles(Qx_hat); // Cholesky factorisation normalmatrix
// solvechol(Qx_hat,rhsBperp); // Solution Bperp coefficients in rhsB
// solvechol(Qx_hat,rhsBpar); // Solution Theta coefficients in rhsTheta
// solvechol(Qx_hat,rhsTheta); // Solution Theta coefficients in rhsTheta
// solvechol(Qx_hat,rhsThetaInc); // Solution Theta_inc coefficients in rhsThetaInc
// invertchol(Qx_hat); // Covariance matrix of normalized unknowns
rhsBperp = Solve.solvePositive(nMatrix, rhsBperp);
rhsBpar = Solve.solvePositive(nMatrix, rhsBpar);
rhsTheta = Solve.solvePositive(nMatrix, rhsTheta);
rhsThetaInc = Solve.solvePositive(nMatrix, rhsThetaInc);
// Info on inversion, normalization is ok______
final DoubleMatrix yHatBperp = aMatrix.mmul(rhsBperp);
final DoubleMatrix eHatBperp = bPerpMatrix.sub(yHatBperp);
// DoubleMatrix Qe_hat = Qy - Qy_hat;
// DoubleMatrix y_hatT = aMatrix * rhsTheta;
// DoubleMatrix e_hatT = thetaMatrix - y_hatT;
// Copy estimated coefficients to private fields
bperpCoeffs = rhsBperp;
bparCoeffs = rhsBpar;
thetaCoeffs = rhsTheta;
thetaIncCoeffs = rhsThetaInc;
// Test inverse -- repair matrix!!!
for (int i = 0; i < Qx_hat.rows; i++) {
for (int j = 0; j < i; j++) {
Qx_hat.put(j, i, Qx_hat.get(i, j));
}
}
final double maxDev = abs(nMatrix.mmul(Qx_hat).sub(DoubleMatrix.eye(Qx_hat.rows))).max();
logger.debug("BASELINE: max(abs(nMatrix*inv(nMatrix)-I)) = " + maxDev);
if (maxDev > .01) {
logger.warn(
"BASELINE: max. deviation nMatrix*inv(nMatrix) from unity = "
+ maxDev
+ ". This is larger than .01: do not use this!");
} else if (maxDev > .001) {
logger.warn(
"BASELINE: max. deviation nMatrix*inv(nMatrix) from unity = "
+ maxDev
+ ". This is between 0.01 and 0.001 (maybe not use it)");
}
// Output solution and give max error
// --- B(l,p,h) = a000 +
// a100*l + a010*p + a001*h +
// a110*l*p + a101*l*h + a011*p*h +
// a200*l^2 + a020*p^2 + a002*h^2
logger.debug("--------------------");
logger.debug("Result of modeling: Bperp(l,p) = a000 + a100*l + a010*p + a001*h + ");
logger.debug(" a110*l*p + a101*l*h + a011*p*h + a200*l^2 + a020*p^2 + a002*h^2");
logger.debug("l,p,h in normalized coordinates [-2:2].");
logger.debug("Bperp_a000 = " + rhsBperp.get(0, 0));
logger.debug("Bperp_a100 = " + rhsBperp.get(1, 0));
logger.debug("Bperp_a010 = " + rhsBperp.get(2, 0));
logger.debug("Bperp_a001 = " + rhsBperp.get(3, 0));
logger.debug("Bperp_a110 = " + rhsBperp.get(4, 0));
logger.debug("Bperp_a101 = " + rhsBperp.get(5, 0));
logger.debug("Bperp_a011 = " + rhsBperp.get(6, 0));
logger.debug("Bperp_a200 = " + rhsBperp.get(7, 0));
logger.debug("Bperp_a020 = " + rhsBperp.get(8, 0));
logger.debug("Bperp_a002 = " + rhsBperp.get(9, 0));
double maxerr = (abs(eHatBperp)).max();
if (maxerr > 2.00) //
{
logger.warn("Max. error bperp modeling at 3D datapoints: " + maxerr + "m");
} else {
logger.info("Max. error bperp modeling at 3D datapoints: " + maxerr + "m");
}
logger.debug("--------------------");
logger.debug("Range: r(p) = r0 + dr*p");
logger.debug("l and p in un-normalized, absolute, coordinates (1:nMatrix).");
final double range1 = master.pix2range(1.0);
final double range5000 = master.pix2range(5000.0);
final double drange = (range5000 - range1) / 5000.0;
logger.debug("range = " + (range1 - drange) + " + " + drange + "*p");
// orbit initialized
isInitialized = true;
}
/**
* 测试模型
*
* @param filename
*/
public void test_model(String filename, int testCount) {
try {
this.test_data_Count = testCount;
test_set = loadMatrix(filename, test_data_Count, totalColCount);
} catch (Exception e) {
e.printStackTrace();
}
printInfo("测试数据载入完毕…………");
// 测试数据输入矩阵初始化 横向即每列为一组输入
testP = new DoubleMatrix(NumberofInputNeurons, test_data_Count);
// 测试机输出矩阵 横向
testT = new DoubleMatrix(NumberofOutputNeurons, test_data_Count);
// 初始化
for (int i = 0; i < test_data_Count; i++) {
for (int j = 0; j < totalColCount; j++) {
if (j < NumberofInputNeurons) {
// 输入部分
testP.put(j, i, test_set.get(i, j));
} else {
testT.put(j - NumberofInputNeurons, i, test_set.get(i, j));
}
}
}
// end 初始化
printInfo("测试数据初始化完毕…………");
long start_time_test = System.currentTimeMillis();
// 计算输入权重与阈值部分
DoubleMatrix tmpH = new DoubleMatrix(NumberofHiddenNeurons, test_data_Count);
tmpH = InputWeight.mmul(testP);
printInfo("权重计算完毕…………");
// 加阈值
// 注意横向问题
for (int i = 0; i < NumberofHiddenNeurons; i++) {
for (int j = 0; j < test_data_Count; j++) {
tmpH.put(i, j, tmpH.get(i, j) + BiasofHiddenNeurons.get(i, 0));
}
}
printInfo("阈值计算完毕…………");
// 输出矩阵
DoubleMatrix H = new DoubleMatrix(NumberofHiddenNeurons, test_data_Count);
if (func.startsWith("sig")) {
for (int i = 0; i < NumberofHiddenNeurons; i++) {
for (int j = 0; j < test_data_Count; j++) {
double tmp = tmpH.get(i, j);
tmp = 1.0f / (1 + Math.exp(-tmp));
H.put(i, j, tmp);
}
}
} else {
throw new IllegalArgumentException("不支持的激活函数类型");
}
printInfo("激活函数计算完毕…………");
// 将H转置
DoubleMatrix Ht = H.transpose();
// 测试误差
DoubleMatrix Yt = new DoubleMatrix(test_data_Count, NumberofOutputNeurons);
Yt = Ht.mmul(OutputWeight);
printInfo("测试完毕…………");
long end_time_test = System.currentTimeMillis();
TestingTime = (end_time_test - start_time_test) * 1.0f / 1000;
double MSE = 0;
printInfo("共有" + NumberofOutputNeurons + "个输出值");
for (int out = 0; out < NumberofOutputNeurons; out++) {
printInfo("计算第" + (out + 1) + "个输出值");
double mseTem = 0.0;
for (int i = 0; i < test_data_Count; i++) {
mseTem += (Yt.get(i, out) - testT.get(out, i)) * (Yt.get(i, out) - testT.get(out, i));
}
printInfo(
"第" + NumberofOutputNeurons + "个输出值计算完毕,MSE=" + Math.sqrt(mseTem / train_data_Count));
MSE += mseTem;
}
TestingAccuracy = Math.sqrt(MSE / test_data_Count);
printInfo("计算MSE完毕…………");
}
// TODO: DOUBLECHECK EVERYTHING
@Override
public void map(Text key, jBLASArrayWritable input, GibbsSampling.Context context)
throws IOException, InterruptedException {
/* *******************************************************************/
/* initialize all memory we're going to use during the process */
long start_time = System.nanoTime();
ArrayList<DoubleMatrix> data = input.getData();
label = data.get(4);
v_data = data.get(5);
// check to see if we are in the first layer or there are layers beneath us we must sample
// from
if (data.size() > 6) {
int prelayer = (data.size() - 6) / 3;
DoubleMatrix[] preWeights = new DoubleMatrix[prelayer],
preHbias = new DoubleMatrix[prelayer],
preVbias = new DoubleMatrix[prelayer];
for (int i = 0; i < prelayer; i++) {
preWeights[i] = data.get(6 + i * 3);
preHbias[i] = data.get(7 + i * 3);
preVbias[i] = data.get(8 + i * 3);
}
DoubleMatrix vnew = null;
for (int i = 0; i < prelayer; i++) {
weights = preWeights[i];
vbias = preVbias[i];
hbias = preHbias[i];
vnew = sample_h_from_v(i == 0 ? v_data : vnew);
}
v_data = vnew;
}
weights = data.get(0);
hbias = data.get(1);
hiddenChain = data.get(2);
vbias = data.get(3);
// check if we need to attach labels to the observed variables
if (vbias.columns != v_data.columns) {
DoubleMatrix labels = DoubleMatrix.zeros(1, classCount);
int labelNum = (new Double(label.get(0))).intValue();
labels.put(labelNum, 1.0);
v_data = DoubleMatrix.concatHorizontally(v_data, labels);
}
w1 = DoubleMatrix.zeros(weights.rows, weights.columns);
hb1 = DoubleMatrix.zeros(hbias.rows, hbias.columns);
vb1 = DoubleMatrix.zeros(vbias.rows, vbias.columns);
/* ********************************************************************/
// sample hidden state to get positive phase
// if empty, use it as the start of the chain
// or use persistent hidden state from pCD
DoubleMatrix phaseSample = sample_h_from_v(v_data);
h1_data = new DoubleMatrix();
v1_data = new DoubleMatrix();
if (hiddenChain == null) {
data.set(2, new DoubleMatrix(hbias.rows, hbias.columns));
hiddenChain = data.get(2);
hiddenChain.copy(phaseSample);
h1_data.copy(phaseSample);
} else {
h1_data.copy(hiddenChain);
}
// run Gibbs chain for k steps
for (int j = 0; j < gibbsSteps; j++) {
v1_data.copy(sample_v_from_h(h1_data));
h1_data.copy(sample_h_from_v(v1_data));
}
DoubleMatrix hprob = propup(v1_data);
weight_contribution(hiddenChain, v_data, hprob, v1_data);
hiddenChain.copy(h1_data);
data.get(0).copy(w1);
data.get(1).copy(hb1);
data.get(2).copy(hiddenChain);
data.get(3).copy(vb1);
jBLASArrayWritable outputmatrix = new jBLASArrayWritable(data);
context.write(key, outputmatrix);
log.info("Job completed in: " + (System.nanoTime() - start_time) / (1E6) + " ms");
}
public static void main(String[] args) {
Scanner s = new Scanner(System.in);
int n = s.nextInt();
int nn = n;
DoubleMatrix d = DoubleMatrix.zeros(2 * n, 2 * n);
DoubleMatrix d2 = DoubleMatrix.zeros(n, n);
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
d.put(i, j, s.nextDouble());
d2.put(i, j, d.get(i, j));
}
}
// d2 = new DoubleMatrix(d.data);
System.out.println(d);
System.out.println(d);
List<Integer> L = new ArrayList<Integer>(2 * n);
List<Double> R = new ArrayList<Double>(2 * n);
for (int i = 0; i < 2 * n; i++) {
L.add(0);
R.add(0.0);
}
// inicjalizacja lisci - jako etykiety kolejne liczby od 0
for (int i = 0; i < n; i++) {
L.set(i, i);
}
// V - drzewo addytywne, które tworzymy
ArrayList[] V = new ArrayList[2 * n];
for (int i = 0; i < V.length; i++) {
V[i] = new ArrayList<Integer>();
}
double suma, rmin, rr;
int i, j, vertNum = n;
while (n > 3) {
// wyznaczanie r dla każdego liścia
for (int a = 0; a < n; a++) {
suma = 0;
for (int b = 0; b < n; b++) {
suma = suma + d.get(L.get(a), L.get(b));
}
suma = suma / (n - 2);
R.set(a, suma);
}
// wyznaczania sąsiadów na podstawie r
i = 0;
j = 1;
rmin = d.get(L.get(0), L.get(1)) - (R.get(0) + R.get(1));
for (int a = 0; a < n - 1; a++) {
for (int b = a + 1; b < n; b++) {
rr = d.get(L.get(a), L.get(b)) - (R.get(a) + R.get(b));
if (rr < rmin) {
rmin = rr;
i = a;
j = b;
}
}
}
// usuniecie ze zbioru lisci i,j oraz dodanie k
L.set(n, vertNum);
vertNum++;
i = L.remove(i);
j = L.remove(j - 1);
n = n - 1;
// uaktualnienie d dla każdego pozostałego liścia
for (int l = 0; l < n - 1; l++) {
double value = (d.get(i, L.get(l)) + d.get(j, L.get(l)) - d.get(i, j)) / 2;
d.put(L.get(n - 1), L.get(l), value);
d.put(L.get(l), L.get(n - 1), value);
}
// dodanie odpowiednich krawędzi do tworzonego drzewa
V[i].add((vertNum - 1));
V[j].add((vertNum - 1));
V[vertNum - 1].add(i);
V[vertNum - 1].add(j);
// wyznaczenie odległości między nowym wierzchołkiem oraz i,j
double value = (d.get(i, j) + d.get(i, L.get(0)) - d.get(j, L.get(0))) / 2;
d.put(i, vertNum - 1, value);
d.put(vertNum - 1, i, value);
d.put(j, vertNum - 1, d.get(i, j) - d.get(i, vertNum - 1));
d.put(vertNum - 1, j, d.get(i, j) - d.get(i, vertNum - 1));
}
// 3 elementowe drzewo
double value;
value = (d.get(L.get(0), L.get(1)) + d.get(L.get(0), L.get(2)) - d.get(L.get(1), L.get(2))) / 2;
d.put(L.get(0), vertNum, value);
d.put(vertNum, L.get(0), value);
value = (d.get(L.get(0), L.get(1)) + d.get(L.get(1), L.get(2)) - d.get(L.get(0), L.get(2))) / 2;
d.put(L.get(1), vertNum, value);
d.put(vertNum, L.get(1), value);
value = (d.get(L.get(0), L.get(2)) + d.get(L.get(1), L.get(2)) - d.get(L.get(0), L.get(1))) / 2;
d.put(L.get(2), vertNum, value);
d.put(vertNum, L.get(2), value);
V[vertNum].add(L.get(0));
V[vertNum].add(L.get(1));
V[vertNum].add(L.get(2));
V[L.get(0)].add(vertNum);
V[L.get(1)].add(vertNum);
V[L.get(2)].add(vertNum);
// wypisanie wyników
System.out.println(d);
// DoubleMatrix w2 = DoubleMatrix.zeros(2*n, 2*n);
ArrayList w = new ArrayList<Integer>();
for (int a = 0; a <= vertNum; a++) {
System.out.print(a);
System.out.print(" : ");
for (int b = 0; b < V[a].size(); b++) {
System.out.print(V[a].get(b));
System.out.print(" ");
// w2.put(a,b,Integer.parseInt(V[a].get(b).toString()));
w.add(V[a].get(b));
}
System.out.println("");
}
DoubleMatrix A = DoubleMatrix.zeros((nn * (nn - 1)) / 2, vertNum);
DoubleMatrix g = DoubleMatrix.zeros((nn * (nn - 1)) / 2, 1);
double blad = nk(A, d2, g, V, vertNum); // wrzucam to do siebie - mkd
System.out.println(A);
DoubleMatrix p = (new DoubleMatrix(A.rows, A.columns, A.data)).transpose();
System.out.println(p.rows + " " + p.columns);
DoubleMatrix p2 =
(new DoubleMatrix(p.rows, p.columns, p.data))
.mmul((new DoubleMatrix(A.rows, A.columns, A.data)));
System.out.println("p2: " + p2);
DoubleMatrix p3 = MatrixFunctions.pow(p2, -1);
System.out.println("p3: " + p3);
DoubleMatrix p4 = p3.mmul(p);
DoubleMatrix b = p4.mmul(g);
// DoubleMatrix b = MatrixFunctions.pow(A.transpose().mmul(A), -1).mmul(A.transpose()).mmul(g);
System.out.println(g);
System.out.println(b);
System.out.println("Kwadrat bledu wynosi " + blad);
}
public static double[] polyFit2D(
final DoubleMatrix x, final DoubleMatrix y, final DoubleMatrix z, final int degree)
throws IllegalArgumentException {
logger.setLevel(Level.INFO);
if (x.length != y.length || !x.isVector() || !y.isVector()) {
logger.error("polyfit: require same size vectors.");
throw new IllegalArgumentException("polyfit: require same size vectors.");
}
final int numOfObs = x.length;
final int numOfUnkn = numberOfCoefficients(degree) + 1;
DoubleMatrix A = new DoubleMatrix(numOfObs, numOfUnkn); // designmatrix
DoubleMatrix mul;
/** Set up design-matrix */
for (int p = 0; p <= degree; p++) {
for (int q = 0; q <= p; q++) {
mul = pow(y, (p - q)).mul(pow(x, q));
if (q == 0 && p == 0) {
A = mul;
} else {
A = DoubleMatrix.concatHorizontally(A, mul);
}
}
}
// Fit polynomial
logger.debug("Solving lin. system of equations with Cholesky.");
DoubleMatrix N = A.transpose().mmul(A);
DoubleMatrix rhs = A.transpose().mmul(z);
// solution seems to be OK up to 10^-09!
rhs = Solve.solveSymmetric(N, rhs);
DoubleMatrix Qx_hat = Solve.solveSymmetric(N, DoubleMatrix.eye(N.getRows()));
double maxDeviation = (N.mmul(Qx_hat).sub(DoubleMatrix.eye(Qx_hat.rows))).normmax();
logger.debug("polyfit orbit: max(abs(N*inv(N)-I)) = " + maxDeviation);
// ___ report max error... (seems sometimes this can be extremely large) ___
if (maxDeviation > 1e-6) {
logger.warn("polyfit orbit: max(abs(N*inv(N)-I)) = {}", maxDeviation);
logger.warn("polyfit orbit interpolation unstable!");
}
// work out residuals
DoubleMatrix y_hat = A.mmul(rhs);
DoubleMatrix e_hat = y.sub(y_hat);
if (e_hat.normmax() > 0.02) {
logger.warn(
"WARNING: Max. polyFit2D approximation error at datapoints (x,y,or z?): {}",
e_hat.normmax());
} else {
logger.info(
"Max. polyFit2D approximation error at datapoints (x,y,or z?): {}", e_hat.normmax());
}
if (logger.isDebugEnabled()) {
logger.debug("REPORTING POLYFIT LEAST SQUARES ERRORS");
logger.debug(" time \t\t\t y \t\t\t yhat \t\t\t ehat");
for (int i = 0; i < numOfObs; i++) {
logger.debug(
" ("
+ x.get(i)
+ ","
+ y.get(i)
+ ") :"
+ "\t"
+ y.get(i)
+ "\t"
+ y_hat.get(i)
+ "\t"
+ e_hat.get(i));
}
}
return rhs.toArray();
}
public static DoubleMatrix polyval(
final DoubleMatrix x, final DoubleMatrix y, final DoubleMatrix coeff, int degree) {
if (!x.isColumnVector()) {
logger.warn("polyValGrid: require (x) standing data vectors!");
throw new IllegalArgumentException("polyval functions require (x) standing data vectors!");
}
if (!y.isColumnVector()) {
logger.warn("polyValGrid: require (y) standing data vectors!");
throw new IllegalArgumentException("polyval functions require (y) standing data vectors!");
}
if (!coeff.isColumnVector()) {
logger.warn("polyValGrid: require (coeff) standing data vectors!");
throw new IllegalArgumentException(
"polyval functions require (coeff) standing data vectors!");
}
if (degree < -1) {
logger.warn("polyValGrid: degree < -1 ????");
}
if (x.length > y.length) {
logger.warn("polValGrid: x larger than y, while optimized for y larger x");
}
if (degree == -1) {
degree = degreeFromCoefficients(coeff.length);
}
// evaluate polynomial //
DoubleMatrix result = new DoubleMatrix(x.length, y.length);
int i;
int j;
double c00,
c10,
c01,
c20,
c11,
c02,
c30,
c21,
c12,
c03,
c40,
c31,
c22,
c13,
c04,
c50,
c41,
c32,
c23,
c14,
c05;
switch (degree) {
case 0:
result.put(0, 0, coeff.get(0, 0));
break;
case 1:
c00 = coeff.get(0, 0);
c10 = coeff.get(1, 0);
c01 = coeff.get(2, 0);
for (j = 0; j < result.columns; j++) {
double c00pc01y1 = c00 + c01 * y.get(j, 0);
for (i = 0; i < result.rows; i++) {
result.put(i, j, c00pc01y1 + c10 * x.get(i, 0));
}
}
break;
case 2:
c00 = coeff.get(0, 0);
c10 = coeff.get(1, 0);
c01 = coeff.get(2, 0);
c20 = coeff.get(3, 0);
c11 = coeff.get(4, 0);
c02 = coeff.get(5, 0);
for (j = 0; j < result.columns; j++) {
double y1 = y.get(j, 0);
double c00pc01y1 = c00 + c01 * y1;
double c02y2 = c02 * Math.pow(y1, 2);
double c11y1 = c11 * y1;
for (i = 0; i < result.rows; i++) {
double x1 = x.get(i, 0);
result.put(i, j, c00pc01y1 + c10 * x1 + c20 * Math.pow(x1, 2) + c11y1 * x1 + c02y2);
}
}
break;
case 3:
c00 = coeff.get(0, 0);
c10 = coeff.get(1, 0);
c01 = coeff.get(2, 0);
c20 = coeff.get(3, 0);
c11 = coeff.get(4, 0);
c02 = coeff.get(5, 0);
c30 = coeff.get(6, 0);
c21 = coeff.get(7, 0);
c12 = coeff.get(8, 0);
c03 = coeff.get(9, 0);
for (j = 0; j < result.columns; j++) {
double y1 = y.get(j, 0);
double y2 = Math.pow(y1, 2);
double c00pc01y1 = c00 + c01 * y1;
double c02y2 = c02 * y2;
double c11y1 = c11 * y1;
double c21y1 = c21 * y1;
double c12y2 = c12 * y2;
double c03y3 = c03 * y1 * y2;
for (i = 0; i < result.rows; i++) {
double x1 = x.get(i, 0);
double x2 = Math.pow(x1, 2);
result.put(
i,
j,
c00pc01y1
+ c10 * x1
+ c20 * x2
+ c11y1 * x1
+ c02y2
+ c30 * x1 * x2
+ c21y1 * x2
+ c12y2 * x1
+ c03y3);
}
}
break;
case 4:
c00 = coeff.get(0, 0);
c10 = coeff.get(1, 0);
c01 = coeff.get(2, 0);
c20 = coeff.get(3, 0);
c11 = coeff.get(4, 0);
c02 = coeff.get(5, 0);
c30 = coeff.get(6, 0);
c21 = coeff.get(7, 0);
c12 = coeff.get(8, 0);
c03 = coeff.get(9, 0);
c40 = coeff.get(10, 0);
c31 = coeff.get(11, 0);
c22 = coeff.get(12, 0);
c13 = coeff.get(13, 0);
c04 = coeff.get(14, 0);
for (j = 0; j < result.columns; j++) {
double y1 = y.get(j, 0);
double y2 = Math.pow(y1, 2);
double c00pc01y1 = c00 + c01 * y1;
double c02y2 = c02 * y2;
double c11y1 = c11 * y1;
double c21y1 = c21 * y1;
double c12y2 = c12 * y2;
double c03y3 = c03 * y1 * y2;
double c31y1 = c31 * y1;
double c22y2 = c22 * y2;
double c13y3 = c13 * y2 * y1;
double c04y4 = c04 * y2 * y2;
for (i = 0; i < result.rows; i++) {
double x1 = x.get(i, 0);
double x2 = Math.pow(x1, 2);
result.put(
i,
j,
c00pc01y1
+ c10 * x1
+ c20 * x2
+ c11y1 * x1
+ c02y2
+ c30 * x1 * x2
+ c21y1 * x2
+ c12y2 * x1
+ c03y3
+ c40 * x2 * x2
+ c31y1 * x2 * x1
+ c22y2 * x2
+ c13y3 * x1
+ c04y4);
}
}
break;
case 5:
c00 = coeff.get(0, 0);
c10 = coeff.get(1, 0);
c01 = coeff.get(2, 0);
c20 = coeff.get(3, 0);
c11 = coeff.get(4, 0);
c02 = coeff.get(5, 0);
c30 = coeff.get(6, 0);
c21 = coeff.get(7, 0);
c12 = coeff.get(8, 0);
c03 = coeff.get(9, 0);
c40 = coeff.get(10, 0);
c31 = coeff.get(11, 0);
c22 = coeff.get(12, 0);
c13 = coeff.get(13, 0);
c04 = coeff.get(14, 0);
c50 = coeff.get(15, 0);
c41 = coeff.get(16, 0);
c32 = coeff.get(17, 0);
c23 = coeff.get(18, 0);
c14 = coeff.get(19, 0);
c05 = coeff.get(20, 0);
for (j = 0; j < result.columns; j++) {
double y1 = y.get(j, 0);
double y2 = Math.pow(y1, 2);
double y3 = y2 * y1;
double c00pc01y1 = c00 + c01 * y1;
double c02y2 = c02 * y2;
double c11y1 = c11 * y1;
double c21y1 = c21 * y1;
double c12y2 = c12 * y2;
double c03y3 = c03 * y3;
double c31y1 = c31 * y1;
double c22y2 = c22 * y2;
double c13y3 = c13 * y3;
double c04y4 = c04 * y2 * y2;
double c41y1 = c41 * y1;
double c32y2 = c32 * y2;
double c23y3 = c23 * y3;
double c14y4 = c14 * y2 * y2;
double c05y5 = c05 * y3 * y2;
for (i = 0; i < result.rows; i++) {
double x1 = x.get(i, 0);
double x2 = Math.pow(x1, 2);
double x3 = x1 * x2;
result.put(
i,
j,
c00pc01y1
+ c10 * x1
+ c20 * x2
+ c11y1 * x1
+ c02y2
+ c30 * x3
+ c21y1 * x2
+ c12y2 * x1
+ c03y3
+ c40 * x2 * x2
+ c31y1 * x3
+ c22y2 * x2
+ c13y3 * x1
+ c04y4
+ c50 * x3 * x2
+ c41y1 * x2 * x2
+ c32y2 * x3
+ c23y3 * x2
+ c14y4 * x1
+ c05y5);
}
}
break;
// TODO: solve up to 5 efficiently, do rest in loop
default:
for (j = 0; j < result.columns; j++) {
double yy = y.get(j, 0);
for (i = 0; i < result.rows; i++) {
double xx = x.get(i, 0);
result.put(i, j, polyval(xx, yy, coeff, degree));
}
}
} // switch degree
return result;
}
public static DoubleMatrix normalize(DoubleMatrix t) {
return t.sub(t.get(t.length / 2)).div(10.0);
}
/** 训练模型 */
private void train() {
printInfo("训练数据载入完毕，开始训练模型………………训练数据规模" + train_set.rows + "x" + train_set.columns);
// 输入矩阵初始化 横向即每列为一组输入
/*
* 4 8 …… N
*
* 1 5 …… N
* 2 6 …… N
* 3 7 …… N
*/
P = new DoubleMatrix(NumberofInputNeurons, train_data_Count);
// 测试机输出矩阵 横向
T = new DoubleMatrix(NumberofOutputNeurons, train_data_Count);
// 初始化
for (int i = 0; i < train_data_Count; i++) {
for (int j = 0; j < totalColCount; j++) {
if (j < NumberofInputNeurons) {
// 输入部分
P.put(j, i, train_set.get(i, j));
} else {
T.put(j - NumberofInputNeurons, i, train_set.get(i, j));
}
}
}
printInfo("训练矩阵初始化完毕………………开始隐含层神经元权重与阈值");
// end 初始化
long start_time_train = System.currentTimeMillis();
// 随机初始化输入权值矩阵
// 行数为隐含层神经元个数，列数为输入层神经元
/**
* W11 W12 W13(第一个隐含神经元对应各输入的权值) W21 W22 W23(第二个隐含神经元对应各输入的权值) ………………………………(第N个隐含神经元对应各输入的权值)
*/
InputWeight = DoubleMatrix.rand(NumberofHiddenNeurons, NumberofInputNeurons);
// 初始化阈值
BiasofHiddenNeurons = DoubleMatrix.rand(NumberofHiddenNeurons, 1);
printInfo("隐含层神经元权重与阈值初始化完毕………………开始计算输入权重");
DoubleMatrix tmpH = new DoubleMatrix(NumberofHiddenNeurons, train_data_Count);
tmpH = InputWeight.mmul(P);
printInfo("输入矩阵计算权重完毕………………开始计算输入权重");
// 加阈值
// 注意横向问题
for (int i = 0; i < NumberofHiddenNeurons; i++) {
for (int j = 0; j < train_data_Count; j++) {
tmpH.put(i, j, tmpH.get(i, j) + BiasofHiddenNeurons.get(i, 0));
}
}
printInfo("输入矩阵计算阈值完毕………………开始计算输入激活函数");
// 输出矩阵
DoubleMatrix H = new DoubleMatrix(NumberofHiddenNeurons, train_data_Count);
if (func.startsWith("sig")) {
for (int i = 0; i < NumberofHiddenNeurons; i++) {
for (int j = 0; j < train_data_Count; j++) {
double tmp = tmpH.get(i, j);
tmp = 1.0f / (1 + Math.exp(-tmp));
H.put(i, j, tmp);
}
}
} else {
throw new IllegalArgumentException("不支持的激活函数类型");
}
printInfo("输出矩阵计算激活函数完毕，开始求解广义逆………………矩阵规模" + H.columns + "x" + H.rows);
// 将H转置
DoubleMatrix Ht = H.transpose();
// 求广义逆
DoubleMatrix Ht_MP = getMPMatrix(Ht);
printInfo("输出矩阵广义逆求解完毕………………开始求解输出矩阵权重");
// 隐含层输出权重矩阵= Ht_MP * Tt
OutputWeight = Ht_MP.mmul(T.transpose());
printInfo("输出矩阵权重求解完毕………………");
long end_time_train = System.currentTimeMillis();
TrainingTime = (end_time_train - start_time_train) * 1.0f / 1000;
printInfo("模型训练完毕………………开始评估模型");
// 误差评估
DoubleMatrix Yt = new DoubleMatrix(train_data_Count, NumberofOutputNeurons);
Yt = Ht.mmul(OutputWeight);
double MSE = 0;
printInfo("共有" + NumberofOutputNeurons + "个输出值");
for (int out = 0; out < NumberofOutputNeurons; out++) {
printInfo("计算第" + (out + 1) + "个输出值");
double mseTem = 0.0;
for (int i = 0; i < train_data_Count; i++) {
System.out.println(Yt.get(i, out) + " " + T.get(out, i));
mseTem += (Yt.get(i, out) - T.get(out, i)) * (Yt.get(i, out) - T.get(out, i));
}
printInfo(
"第" + NumberofOutputNeurons + "个输出值计算完毕,MSE=" + Math.sqrt(mseTem / train_data_Count));
MSE += mseTem;
}
TrainingAccuracy = Math.sqrt(MSE / train_data_Count);
printInfo("模型评估完毕………………");
}
private void costantiniUnwrap() throws LPException {
final int ny = wrappedPhase.rows - 1; // start from Zero!
final int nx = wrappedPhase.columns - 1; // start from Zero!
if (wrappedPhase.isVector()) throw new IllegalArgumentException("Input must be 2D array");
if (wrappedPhase.rows < 2 || wrappedPhase.columns < 2)
throw new IllegalArgumentException("Size of input must be larger than 2");
// Default weight
DoubleMatrix w1 = DoubleMatrix.ones(ny + 1, 1);
w1.put(0, 0.5);
w1.put(w1.length - 1, 0.5);
DoubleMatrix w2 = DoubleMatrix.ones(1, nx + 1);
w2.put(0, 0.5);
w2.put(w2.length - 1, 0.5);
DoubleMatrix weight = w1.mmul(w2);
DoubleMatrix i, j, I_J, IP1_J, I_JP1;
DoubleMatrix Psi1, Psi2;
DoubleMatrix[] ROWS;
// Compute partial derivative Psi1, eqt (1,3)
i = intRangeDoubleMatrix(0, ny - 1);
j = intRangeDoubleMatrix(0, nx);
ROWS = grid2D(i, j);
I_J = JblasUtils.sub2ind(wrappedPhase.rows, ROWS[0], ROWS[1]);
IP1_J = JblasUtils.sub2ind(wrappedPhase.rows, ROWS[0].add(1), ROWS[1]);
Psi1 =
JblasUtils.getMatrixFromIdx(wrappedPhase, IP1_J)
.sub(JblasUtils.getMatrixFromIdx(wrappedPhase, I_J));
Psi1 = UnwrapUtils.wrapDoubleMatrix(Psi1);
// Compute partial derivative Psi2, eqt (2,4)
i = intRangeDoubleMatrix(0, ny);
j = intRangeDoubleMatrix(0, nx - 1);
ROWS = grid2D(i, j);
I_J = JblasUtils.sub2ind(wrappedPhase.rows, ROWS[0], ROWS[1]);
I_JP1 = JblasUtils.sub2ind(wrappedPhase.rows, ROWS[0], ROWS[1].add(1));
Psi2 =
JblasUtils.getMatrixFromIdx(wrappedPhase, I_JP1)
.sub(JblasUtils.getMatrixFromIdx(wrappedPhase, I_J));
Psi2 = UnwrapUtils.wrapDoubleMatrix(Psi2);
// Compute beq
DoubleMatrix beq = DoubleMatrix.zeros(ny, nx);
i = intRangeDoubleMatrix(0, ny - 1);
j = intRangeDoubleMatrix(0, nx - 1);
ROWS = grid2D(i, j);
I_J = JblasUtils.sub2ind(Psi1.rows, ROWS[0], ROWS[1]);
I_JP1 = JblasUtils.sub2ind(Psi1.rows, ROWS[0], ROWS[1].add(1));
beq.addi(JblasUtils.getMatrixFromIdx(Psi1, I_JP1).sub(JblasUtils.getMatrixFromIdx(Psi1, I_J)));
I_J = JblasUtils.sub2ind(Psi2.rows, ROWS[0], ROWS[1]);
I_JP1 = JblasUtils.sub2ind(Psi2.rows, ROWS[0].add(1), ROWS[1]);
beq.subi(JblasUtils.getMatrixFromIdx(Psi2, I_JP1).sub(JblasUtils.getMatrixFromIdx(Psi2, I_J)));
beq.muli(-1 / (2 * Constants._PI));
for (int k = 0; k < beq.length; k++) {
beq.put(k, Math.round(beq.get(k)));
}
beq.reshape(beq.length, 1);
logger.debug("Constraint matrix");
i = intRangeDoubleMatrix(0, ny - 1);
j = intRangeDoubleMatrix(0, nx - 1);
ROWS = grid2D(i, j);
DoubleMatrix ROW_I_J = JblasUtils.sub2ind(i.length, ROWS[0], ROWS[1]);
int nS0 = nx * ny;
// Use by S1p, S1m
DoubleMatrix[] COLS;
COLS = grid2D(i, j);
DoubleMatrix COL_IJ_1 = JblasUtils.sub2ind(i.length, COLS[0], COLS[1]);
COLS = grid2D(i, j.add(1));
DoubleMatrix COL_I_JP1 = JblasUtils.sub2ind(i.length, COLS[0], COLS[1]);
int nS1 = (nx + 1) * (ny);
// SOAPBinding.Use by S2p, S2m
COLS = grid2D(i, j);
DoubleMatrix COL_IJ_2 = JblasUtils.sub2ind(i.length + 1, COLS[0], COLS[1]);
COLS = grid2D(i.add(1), j);
DoubleMatrix COL_IP1_J = JblasUtils.sub2ind(i.length + 1, COLS[0], COLS[1]);
int nS2 = nx * (ny + 1);
// Equality constraint matrix (Aeq)
/*
S1p = + sparse(ROW_I_J, COL_I_JP1,1,nS0,nS1) ...
- sparse(ROW_I_J, COL_IJ_1,1,nS0,nS1);
S1m = -S1p;
S2p = - sparse(ROW_I_J, COL_IP1_J,1,nS0,nS2) ...
+ sparse(ROW_I_J, COL_IJ_2,1,nS0,nS2);
S2m = -S2p;
*/
// ToDo: Aeq matrix should be sparse from it's initialization, look into JblasMatrix factory for
// howto
// ...otherwise even a data set of eg 40x40 pixels will exhaust heap:
// ... dimension of Aeq (equality constraints) matrix for 30x30 input is 1521x6240 matrix
// ... dimension of Aeq ( ) matrix for 50x50 input is 2401x9800
// ... dimension of Aeq ( ) matrix for 512x512 input is 261121x1046528
DoubleMatrix S1p =
JblasUtils.setUpMatrixFromIdx(nS0, nS1, ROW_I_J, COL_I_JP1)
.sub(JblasUtils.setUpMatrixFromIdx(nS0, nS1, ROW_I_J, COL_IJ_1));
DoubleMatrix S1m = S1p.neg();
DoubleMatrix S2p =
JblasUtils.setUpMatrixFromIdx(nS0, nS2, ROW_I_J, COL_IP1_J)
.neg()
.add(JblasUtils.setUpMatrixFromIdx(nS0, nS2, ROW_I_J, COL_IJ_2));
DoubleMatrix S2m = S2p.neg();
DoubleMatrix Aeq =
concatHorizontally(concatHorizontally(S1p, S1m), concatHorizontally(S2p, S2m));
final int nObs = Aeq.columns;
final int nUnkn = Aeq.rows;
DoubleMatrix c1 = JblasUtils.getMatrixFromRange(0, ny, 0, weight.columns, weight);
DoubleMatrix c2 = JblasUtils.getMatrixFromRange(0, weight.rows, 0, nx, weight);
c1.reshape(c1.length, 1);
c2.reshape(c2.length, 1);
DoubleMatrix cost =
DoubleMatrix.concatVertically(
DoubleMatrix.concatVertically(c1, c1), DoubleMatrix.concatVertically(c2, c2));
logger.debug("Minimum network flow resolution");
StopWatch clockLP = new StopWatch();
LinearProgram lp = new LinearProgram(cost.data);
lp.setMinProblem(true);
boolean[] integerBool = new boolean[nObs];
double[] lowerBound = new double[nObs];
double[] upperBound = new double[nObs];
for (int k = 0; k < nUnkn; k++) {
lp.addConstraint(new LinearEqualsConstraint(Aeq.getRow(k).toArray(), beq.get(k), "cost"));
}
for (int k = 0; k < nObs; k++) {
integerBool[k] = true;
lowerBound[k] = 0;
upperBound[k] = 99999;
}
// setup bounds and integer nature
lp.setIsinteger(integerBool);
lp.setUpperbound(upperBound);
lp.setLowerbound(lowerBound);
LinearProgramSolver solver = SolverFactory.newDefault();
// double[] solution;
// solution = solver.solve(lp);
DoubleMatrix solution = new DoubleMatrix(solver.solve(lp));
clockLP.stop();
logger.debug("Total GLPK time: {} [sec]", (double) (clockLP.getElapsedTime()) / 1000);
// Displatch the LP solution
int offset;
int[] idx1p = JblasUtils.intRangeIntArray(0, nS1 - 1);
DoubleMatrix x1p = solution.get(idx1p);
x1p.reshape(ny, nx + 1);
offset = idx1p[nS1 - 1] + 1;
int[] idx1m = JblasUtils.intRangeIntArray(offset, offset + nS1 - 1);
DoubleMatrix x1m = solution.get(idx1m);
x1m.reshape(ny, nx + 1);
offset = idx1m[idx1m.length - 1] + 1;
int[] idx2p = JblasUtils.intRangeIntArray(offset, offset + nS2 - 1);
DoubleMatrix x2p = solution.get(idx2p);
x2p.reshape(ny + 1, nx);
offset = idx2p[idx2p.length - 1] + 1;
int[] idx2m = JblasUtils.intRangeIntArray(offset, offset + nS2 - 1);
DoubleMatrix x2m = solution.get(idx2m);
x2m.reshape(ny + 1, nx);
// Compute the derivative jumps, eqt (20,21)
DoubleMatrix k1 = x1p.sub(x1m);
DoubleMatrix k2 = x2p.sub(x2m);
// (?) Round to integer solution
if (roundK == true) {
for (int idx = 0; idx < k1.length; idx++) {
k1.put(idx, FastMath.floor(k1.get(idx)));
}
for (int idx = 0; idx < k2.length; idx++) {
k2.put(idx, FastMath.floor(k2.get(idx)));
}
}
// Sum the jumps with the wrapped partial derivatives, eqt (10,11)
k1.reshape(ny, nx + 1);
k2.reshape(ny + 1, nx);
k1.addi(Psi1.div(Constants._TWO_PI));
k2.addi(Psi2.div(Constants._TWO_PI));
// Integrate the partial derivatives, eqt (6)
// cumsum() method in JblasTester -> see cumsum_demo() in JblasTester.cumsum_demo()
DoubleMatrix k2_temp = DoubleMatrix.concatHorizontally(DoubleMatrix.zeros(1), k2.getRow(0));
k2_temp = JblasUtils.cumsum(k2_temp, 1);
DoubleMatrix k = DoubleMatrix.concatVertically(k2_temp, k1);
k = JblasUtils.cumsum(k, 1);
// Unwrap - final solution
unwrappedPhase = k.mul(Constants._TWO_PI);
}
public static double[] polyFit(DoubleMatrix t, DoubleMatrix y, final int degree)
throws IllegalArgumentException {
logger.setLevel(Level.INFO);
if (t.length != y.length || !t.isVector() || !y.isVector()) {
logger.error("polyfit: require same size vectors.");
throw new IllegalArgumentException("polyfit: require same size vectors.");
}
// Normalize _posting_ for numerical reasons
final int numOfPoints = t.length;
// Check redundancy
final int numOfUnknowns = degree + 1;
logger.debug("Degree of interpolating polynomial: {}", degree);
logger.debug("Number of unknowns: {}", numOfUnknowns);
logger.debug("Number of data points: {}", numOfPoints);
if (numOfPoints < numOfUnknowns) {
logger.error("Number of points is smaller than parameters solved for.");
throw new IllegalArgumentException("Number of points is smaller than parameters solved for.");
}
// Set up system of equations to solve coeff :: Design matrix
logger.debug("Setting up linear system of equations");
DoubleMatrix A = new DoubleMatrix(numOfPoints, numOfUnknowns);
// work with columns
for (int j = 0; j <= degree; j++) {
A.putColumn(j, pow(t, j));
}
// Fit polynomial through computed vector of phases
logger.debug("Solving lin. system of equations with Cholesky.");
DoubleMatrix N = A.transpose().mmul(A);
DoubleMatrix rhs = A.transpose().mmul(y);
// solution seems to be OK up to 10^-09!
DoubleMatrix x = Solve.solveSymmetric(N, rhs);
DoubleMatrix Qx_hat = Solve.solveSymmetric(N, DoubleMatrix.eye(N.getRows()));
double maxDeviation = (N.mmul(Qx_hat).sub(DoubleMatrix.eye(Qx_hat.rows))).normmax();
logger.debug("polyfit orbit: max(abs(N*inv(N)-I)) = " + maxDeviation);
// ___ report max error... (seems sometimes this can be extremely large) ___
if (maxDeviation > 1e-6) {
logger.warn("polyfit orbit: max(abs(N*inv(N)-I)) = {}", maxDeviation);
logger.warn("polyfit orbit interpolation unstable!");
}
// work out residuals
DoubleMatrix y_hat = A.mmul(x);
DoubleMatrix e_hat = y.sub(y_hat);
// 0.05 is already 1 wavelength! (?)
if (e_hat.normmax() > 0.02) {
logger.warn(
"WARNING: Max. approximation error at datapoints (x,y,or z?): {}", e_hat.normmax());
} else {
logger.debug("Max. approximation error at datapoints (x,y,or z?): {}", e_hat.normmax());
}
if (logger.isDebugEnabled()) {
logger.debug("REPORTING POLYFIT LEAST SQUARES ERRORS");
logger.debug(" time \t\t\t y \t\t\t yhat \t\t\t ehat");
for (int i = 0; i < numOfPoints; i++) {
logger.debug(" " + t.get(i) + "\t" + y.get(i) + "\t" + y_hat.get(i) + "\t" + e_hat.get(i));
}
for (int i = 0; i < numOfPoints - 1; i++) {
// ___ check if dt is constant, not necessary for me, but may ___
// ___ signal error in header data of SLC image ___
double dt = t.get(i + 1) - t.get(i);
logger.debug("Time step between point " + i + 1 + " and " + i + "= " + dt);
if (Math.abs(dt - (t.get(1) - t.get(0))) > 0.001) // 1ms of difference we allow...
logger.warn("WARNING: Orbit: data does not have equidistant time interval?");
}
}
return x.toArray();
}
