Examples with DenseVector com.alibaba.alink.common.linalg.DenseVector used on opensource projects

Example 1 with DenseVector

use of com.alibaba.alink.common.linalg.DenseVector in project Alink by alibaba.

the class DeepARModelMapper method predictMultiVar.

@Override
protected Tuple2<Vector[], String> predictMultiVar(Timestamp[] historyTimes, Vector[] historyVals, int predictNum) {
    Timestamp[] predictTimes = TimeSeriesMapper.getPredictTimes(historyTimes, predictNum);
    int window = historyVals.length;
    int series = 0;
    DenseVector[] vectors = new DenseVector[historyVals.length];
    for (int i = 0; i < window; ++i) {
        vectors[i] = VectorUtil.getDenseVector(historyVals[i]);
        if (vectors[i] == null) {
            throw new IllegalArgumentException("history values should not be null.");
        }
        series = vectors[i].size();
    }
    FloatTensor[][] tensors = new FloatTensor[series][window];
    for (int i = 0; i < series; ++i) {
        tensors[i][0] = Tensor.cat(new FloatTensor[] { new FloatTensor(new float[] { 0.0f }), DeepARFeaturesGenerator.generateFromFrequency(calendar.get(), unit, historyTimes[0]) }, -1, null);
        for (int j = 1; j < window; ++j) {
            tensors[i][j] = Tensor.cat(new FloatTensor[] { new FloatTensor(new float[] { (float) vectors[j - 1].get(i) }), DeepARFeaturesGenerator.generateFromFrequency(calendar.get(), unit, historyTimes[j]) }, -1, null);
        }
    }
    FloatTensor[] batch = new FloatTensor[series];
    for (int i = 0; i < series; ++i) {
        batch[i] = Tensor.stack(tensors[i], 0, null);
    }
    Vector[] result = new Vector[predictNum];
    Row[] sigmas = new Row[predictNum];
    for (int i = 0; i < predictNum; ++i) {
        result[i] = new DenseVector(series);
        sigmas[i] = Row.of(new DenseVector(series));
    }
    for (int i = 0; i < series; ++i) {
        float mu = (float) historyVals[window - 1].get(i);
        FloatTensor v = new FloatTensor(new float[] { 0.0f, 0.0f });
        int nonZero = 0;
        for (int j = 0; j < window; ++j) {
            float cell = batch[i].getFloat(j, 0);
            if (cell != 0) {
                nonZero += 1;
            }
            v.setFloat(v.getFloat(0) + cell, 0);
        }
        if (mu != 0) {
            nonZero += 1;
            v.setFloat(v.getFloat(0) + mu, 0);
        }
        if (nonZero == 0) {
            continue;
        }
        v.setFloat(v.getFloat(0) / nonZero + 1.0f, 0);
        for (int j = 0; j < window; ++j) {
            batch[i].setFloat(batch[i].getFloat(j, 0) / v.getFloat(0), j, 0);
        }
        mu = mu / v.getFloat(0);
        for (int j = 0; j < predictNum; ++j) {
            batch[i] = Tensor.cat(new FloatTensor[] { batch[i], Tensor.stack(new FloatTensor[] { Tensor.cat(new FloatTensor[] { new FloatTensor(new float[] { mu }), DeepARFeaturesGenerator.generateFromFrequency(calendar.get(), unit, predictTimes[j]) }, -1, null) }, 0, null) }, 0, null);
            FloatTensor pred;
            try {
                pred = (FloatTensor) tfTableModelPredictModelMapper.map(Row.of(batch[i])).getField(0);
            } catch (Exception e) {
                return Tuple2.of(null, null);
            }
            mu = pred.getFloat(window + j, 0);
            float sigma = pred.getFloat(window + j, 1);
            result[j].set(i, mu * v.getFloat(0) + v.getFloat(1));
            ((Vector) (sigmas[j].getField(0))).set(i, sigma * v.getFloat(0));
        }
    }
    return Tuple2.of(result, new MTable(Arrays.asList(sigmas), new String[] { "sigma" }, new TypeInformation<?>[] { VectorTypes.DENSE_VECTOR }).toString());
}

Example 2 with DenseVector

use of com.alibaba.alink.common.linalg.DenseVector in project Alink by alibaba.

the class ExpandExtendedVarsMapper method map.

@Override
protected void map(SlicedSelectedSample selection, SlicedResult result) throws Exception {
    final DenseVector vector = VectorUtil.getDenseVector(VectorUtil.getVector(selection.get(0)));
    final DenseVector cVector = VectorUtil.getDenseVector(VectorUtil.getVector(selection.get(1)));
    if (vector == null) {
        result.set(0, null);
        return;
    }
    if (cVector == null) {
        result.set(0, vector);
        return;
    }
    final int size = vector.size();
    final int cSize = cVector.size();
    final int gSize = size / numVars;
    final int cPlusSize = cSize + numVars;
    final int finalSize = gSize * cPlusSize;
    final double[] array = vector.getData();
    final double[] cArray = cVector.getData();
    final double[] buf = new double[finalSize];
    for (int i = 0; i < gSize; ++i) {
        final int s = i * cPlusSize;
        for (int j = i * numVars, j1 = s; j < (i + 1) * numVars; ++j, ++j1) {
            buf[j1] = array[j];
        }
        for (int j = 0, j1 = s + numVars; j < cSize; ++j, ++j1) {
            buf[j1] = cArray[j];
        }
    }
    result.set(0, new DenseVector(buf));
}

Example 3 with DenseVector

use of com.alibaba.alink.common.linalg.DenseVector in project Alink by alibaba.

the class LookupVectorInTimeSeriesMapper method map.

@Override
protected void map(SlicedSelectedSample selection, SlicedResult result) throws Exception {
    MTable mTable = null;
    if (selection.get(1) == null) {
        result.set(0, null);
        return;
    }
    if (selection.get(1) instanceof MTable) {
        mTable = (MTable) selection.get(1);
    } else {
        mTable = new MTable((String) selection.get(1));
    }
    if (mTable.getNumRow() == 0) {
        result.set(0, null);
        return;
    }
    Timestamp lookupTime = (Timestamp) selection.get(0);
    TableSchema schema = mTable.getTableSchema();
    String timeCol = null;
    String vectorCol = null;
    int timeIdx = -1;
    TypeInformation<?>[] colTypes = schema.getFieldTypes();
    for (int i = 0; i < colTypes.length; i++) {
        if (colTypes[i] == Types.SQL_TIMESTAMP) {
            timeCol = schema.getFieldNames()[i];
            timeIdx = i;
        }
        if (colTypes[i] == VectorTypes.VECTOR || colTypes[i] == VectorTypes.DENSE_VECTOR || colTypes[i] == VectorTypes.SPARSE_VECTOR) {
            vectorCol = schema.getFieldNames()[i];
        }
    }
    if (null != timeCol && null != vectorCol) {
        List<Object> times = MTableUtils.getColumn(mTable, timeCol);
        int idxRow = times.indexOf(lookupTime);
        int idxCol = TableUtil.findColIndex(schema, vectorCol);
        if (idxRow >= 0) {
            result.set(0, mTable.getEntry(idxRow, idxCol));
            return;
        } else {
            mTable.orderBy(timeIdx);
            Timestamp[] timesArr = MTableUtils.getColumn(mTable, timeCol).toArray(new Timestamp[] {});
            int pos = Arrays.binarySearch(timesArr, lookupTime);
            if (pos == -1) {
                result.set(0, mTable.getEntry(0, idxCol));
            } else if (-pos == timesArr.length + 1) {
                result.set(0, mTable.getEntry(timesArr.length - 1, idxCol));
            } else {
                int pos0 = -2 - pos;
                int pos1 = -1 - pos;
                long time0 = timesArr[pos0].getTime();
                long time1 = timesArr[pos1].getTime();
                double scale = (double) (lookupTime.getTime() - time0) / (double) (time1 - time0);
                DenseVector inter = ((DenseVector) mTable.getEntry(pos0, idxCol)).scale(1 - scale);
                inter.plusEqual(((DenseVector) mTable.getEntry(pos1, idxCol)).scale(scale));
                result.set(0, inter);
            }
            return;
        }
    }
    result.set(0, null);
}

Example 4 with DenseVector

use of com.alibaba.alink.common.linalg.DenseVector in project Alink by alibaba.

the class MleEstimate method bfgsEstimateParams.

public void bfgsEstimateParams(double[] data, int aOrder, int bOrder) {
    double mean = TsMethod.mean(data);
    for (int i = 0; i < data.length; i++) {
        data[i] -= mean;
    }
    this.aOrder = aOrder;
    this.bOrder = bOrder;
    // initiation of parameters
    double[] initAlpha = new double[aOrder];
    double[] initBeta = new double[bOrder];
    for (int i = 0; i < aOrder; i++) {
        initAlpha[i] = 0.1;
    }
    for (int i = 0; i < bOrder; i++) {
        initBeta[i] = 0.1;
    }
    double var = 0;
    for (double aData : data) {
        var += aData * aData;
    }
    var = var / data.length;
    double sumAlpha = 0;
    double sumBeta = 0;
    for (double alpha : initAlpha) {
        sumAlpha += alpha * alpha;
    }
    for (double beta : initBeta) {
        sumBeta += beta * beta;
    }
    double initC = Math.sqrt(var * (1 - sumAlpha - sumBeta));
    double[] ch = new double[data.length];
    for (int i = 0; i < bOrder; i++) {
        ch[i] = var;
    }
    // ///////-----------------------
    // int m = this.aOrder + this.bOrder + 1 + 1;
    int m = this.aOrder + this.bOrder + 1;
    DenseVector lowerBound = DenseVector.zeros(m);
    DenseVector upperBound = DenseVector.ones(m);
    lowerBound.set(m - 1, var * 1e-8);
    upperBound.set(m - 1, var * 10);
    // lowerBound.set(m - 2, var * 1e-8);
    // upperBound.set(m - 2, var * 10);
    // lowerBound.set(m - 2, Double.NEGATIVE_INFINITY);
    // upperBound.set(m - 2, Double.POSITIVE_INFINITY);
    GarchGradientTarget garchGradientTarget = new GarchGradientTarget();
    garchGradientTarget.fit(initAlpha, initBeta, initC, data, ch);
    DenseVector result = TimeSeriesUtils.calcGarchLBFGSB(lowerBound, upperBound, garchGradientTarget, data, 500, 1e-8);
    this.alpha = new double[aOrder];
    this.beta = new double[bOrder];
    for (int i = 0; i < aOrder; i++) {
        alpha[i] = result.get(i);
    }
    for (int i = 0; i < bOrder; i++) {
        beta[i] = result.get(i + aOrder) * result.get(i + aOrder);
    }
    this.c = result.get(aOrder + bOrder) * result.get(aOrder + bOrder);
    this.h = garchGradientTarget.residual;
// /////////-----------------------
// //find parameters
// GarchGradientTarget garchProblem = new GarchGradientTarget();
// garchProblem.fit(initAlpha, initBeta, initC, data, ch);
// AbstractGradientTarget problem = BFGS.solve(
// garchProblem, 1000, 0.00001,
// 0.000001, new int[] {2, 3}, -1);
// 
// this.alpha = new double[aOrder];
// this.beta = new double[bOrder];
// DenseMatrix result = problem.getFinalCoef();
// for (int i = 0; i < aOrder; i++) {
// alpha[i] = result.get(i, 0) * result.get(i, 0);
// }
// for (int i = 0; i < bOrder; i++) {
// beta[i] = result.get(i + aOrder, 0) * result.get(i + aOrder, 0);
// }
// this.c = result.get(aOrder + bOrder, 0) * result.get(aOrder + bOrder, 0);
// this.h = problem.getResidual();
// this.loglike = -problem.getMinValue();
// this.warn = problem.getWarn();
// 
// DenseMatrix obsInformation = problem.getH();
// this.alphaStdError = new double[aOrder];
// this.betaStdError = new double[bOrder];
// for (int i = 0; i < aOrder; i++) {
// this.alphaStdError[i] = Math.sqrt(obsInformation.get(i, i));
// if (Double.isNaN(this.alphaStdError[i])) {
// this.alphaStdError[i] = -99;
// }
// 
// }
// for (int i = 0; i < bOrder; i++) {
// this.betaStdError[i] = Math.sqrt(obsInformation.get(i + aOrder, i + aOrder));
// if (Double.isNaN(this.betaStdError[i])) {
// this.betaStdError[i] = -99;
// }
// }
// this.cStdError = Math.sqrt(obsInformation.get(aOrder + bOrder, aOrder + bOrder));
// if (Double.isNaN(this.cStdError)) {
// this.cStdError = -99;
// }
// 
// sumAlpha = 0;
// sumBeta = 0;
// for (int i = 0; i < this.aOrder; i++) {
// sumAlpha += this.alpha[i];
// }
// for (int i = 0; i < this.bOrder; i++) {
// sumBeta += this.beta[i];
// }
// 
// this.unconSigma2 = this.c / (1 - sumAlpha - sumBeta);
// this.residual = new double[this.h.length];
// for (int i = 0; i < h.length; i++) {
// this.residual[i] = data[i] / h[i];
// }
}

Example 5 with DenseVector

use of com.alibaba.alink.common.linalg.DenseVector in project Alink by alibaba.

the class HoltWinters method fit.

/**
 * holt winters fit.
 *
 * @param data          time series data.
 * @param frequency     time series frequency
 * @param alpha         alpha
 * @param beta          beta
 * @param gamma         gamma
 * @param doTrend       doTrend
 * @param doSeasonal    doSeasonal
 * @param seasonalType  seasonalType
 * @param levelStart    levelStart
 * @param trendStart    trendStart
 * @param seasonalStart seasonalStart
 * @return HoltWintersModel
 */
public static HoltWintersModel fit(double[] data, int frequency, double alpha, double beta, double gamma, boolean doTrend, boolean doSeasonal, SeasonalType seasonalType, Double levelStart, Double trendStart, double[] seasonalStart) {
    if (data == null) {
        throw new RuntimeException("Data is null.");
    }
    for (double val : data) {
        if (Double.isNaN(val) || Double.isInfinite(val)) {
            return null;
        }
    }
    Tuple3<Double, Double, double[]> starts = initStart(data, frequency, doTrend, doSeasonal, seasonalType, levelStart, trendStart, seasonalStart);
    levelStart = starts.f0;
    trendStart = starts.f1;
    seasonalStart = starts.f2;
    DenseVector coefs;
    try {
        coefs = train(data, frequency, alpha, beta, gamma, doTrend, doSeasonal, seasonalType, levelStart, trendStart, seasonalStart);
    } catch (Throwable ex) {
        for (int i = 0; i < data.length; i++) {
            data[i] += 1e-10;
        }
        coefs = train(data, frequency, alpha, beta, gamma, doTrend, doSeasonal, seasonalType, levelStart, trendStart, seasonalStart);
    }
    boolean isAddType = seasonalType == SeasonalType.ADDITIVE;
    Tuple3<Double, Double, double[]> initialData = HoltWintersUtil.calculateInitialData(data, coefs.getData(), isAddType, frequency, levelStart, trendStart, seasonalStart);
    HoltWintersModel model = new HoltWintersModel();
    model.levelStart = initialData.f0;
    model.trendStart = initialData.f1;
    model.seasonalStart = initialData.f2;
    model.coefs = coefs.getData();
    model.seasonalType = seasonalType;
    model.frequency = frequency;
    return model;
}