Examples with IntFloatVector com.tencent.angel.ml.math2.vector.IntFloatVector used on opensource projects

Example 81 with IntFloatVector

use of com.tencent.angel.ml.math2.vector.IntFloatVector in project angel by Tencent.

the class RowSplitCombineUtils method combineIntFloatIndexRowSplits.

// //////////////////////////////////////////////////////////////////////////////
// Combine Int key Float value vector
// //////////////////////////////////////////////////////////////////////////////
public static Vector combineIntFloatIndexRowSplits(int matrixId, int rowId, int resultSize, KeyPart[] keyParts, ValuePart[] valueParts, MatrixMeta matrixMeta) {
    IntFloatVector vector = VFactory.sparseFloatVector((int) matrixMeta.getColNum(), resultSize);
    for (int i = 0; i < keyParts.length; i++) {
        mergeTo(vector, keyParts[i], (FloatValuesPart) valueParts[i]);
    }
    vector.setRowId(rowId);
    vector.setMatrixId(matrixId);
    return vector;
}

Example 82 with IntFloatVector

use of com.tencent.angel.ml.math2.vector.IntFloatVector in project angel by Tencent.

the class RegTDataStore method init.

public void init(DataBlock<LabeledData> dataSet) throws IOException {
    numRow = dataSet.size();
    numCol = param.numFeature;
    numNonzero = param.numNonzero;
    instances = new IntFloatVector[numRow];
    labels = new float[numRow];
    preds = new float[numRow];
    weights = new float[numRow];
    baseWeights = new float[numRow];
    // max and min of each feature
    double[] minFeatures = new double[numCol];
    double[] maxFeatures = new double[numCol];
    Arrays.setAll(minFeatures, i -> 0.0f);
    Arrays.setAll(maxFeatures, i -> Float.MAX_VALUE);
    dataSet.resetReadIndex();
    LabeledData data;
    IntFloatVector x = null;
    double y;
    for (int idx = 0; idx < dataSet.size(); idx++) {
        data = dataSet.read();
        if (data.getX() instanceof IntFloatVector) {
            x = (IntFloatVector) data.getX();
        } else if (data.getX() instanceof IntDoubleVector) {
            x = VFactory.sparseFloatVector((int) data.getX().dim(), ((IntDoubleVector) data.getX()).getStorage().getIndices(), Maths.double2Float(((IntDoubleVector) data.getX()).getStorage().getValues()));
        }
        y = data.getY();
        if (y != 1) {
            y = 0;
        }
        int[] indices = x.getStorage().getIndices();
        float[] values = x.getStorage().getValues();
        for (int i = 0; i < indices.length; i++) {
            int fid = indices[i];
            double fvalue = values[i];
            if (fvalue > maxFeatures[fid]) {
                maxFeatures[fid] = fvalue;
            }
            if (fvalue < minFeatures[fid]) {
                minFeatures[fid] = fvalue;
            }
        }
        instances[idx] = x;
        labels[idx] = (float) y;
        preds[idx] = 0.0f;
        weights[idx] = 1.0f;
        baseWeights[idx] = 1.0f;
    }
    featureMeta = new FeatureMeta(numCol, Maths.double2Float(minFeatures), Maths.double2Float(maxFeatures));
}

Example 83 with IntFloatVector

use of com.tencent.angel.ml.math2.vector.IntFloatVector in project angel by Tencent.

the class GradHistHelper method buildHistogram.

public IntDoubleVector buildHistogram(int insStart, int insEnd) {
    // 1. new feature's histogram (grad + hess)
    // size: sampled_featureNum * (2 * splitNum)
    // in other words, concatenate each feature's histogram
    int featureNum = this.controller.fSet.length;
    int splitNum = this.controller.param.numSplit;
    IntDoubleVector histogram = new IntDoubleVector(featureNum * 2 * splitNum, new IntDoubleDenseVectorStorage(new double[featureNum * 2 * splitNum]));
    // 2. get the span of this node
    int nodeStart = insStart;
    // inclusive
    int nodeEnd = insEnd;
    LOG.debug(String.format("Build histogram of node[%d]: size[%d] instance span [%d - %d]", this.nid, histogram.getDim(), nodeStart, nodeEnd));
    // ------ 3. using sparse-aware method to build histogram ---
    // first add grads of all instances to the zero bin of all features, then loop the non-zero entries of all the instances
    float gradSum = 0.0f;
    float hessSum = 0.0f;
    long parseInstanceTime = 0;
    long startTime = System.currentTimeMillis();
    for (int idx = nodeStart; idx <= nodeEnd; idx++) {
        // 3.1. get the instance index
        int insIdx = this.controller.instancePos[idx];
        // 3.2. get the grad and hess of the instance
        GradPair gradPair = this.controller.gradPairs[insIdx];
        // 3.3. add to the sum
        gradSum += gradPair.getGrad();
        hessSum += gradPair.getHess();
        IntFloatVector instance = this.controller.trainDataStore.instances[insIdx];
        int numNnz = instance.getStorage().getIndices().length;
        long tmpTime = System.currentTimeMillis();
        int[] indices = instance.getStorage().getIndices();
        float[] values = instance.getStorage().getValues();
        parseInstanceTime += System.currentTimeMillis() - tmpTime;
        // 3.4. loop the non-zero entries
        for (int i = 0; i < numNnz; i++) {
            int fid = indices[i];
            // 3.4.1. get feature value
            float fv = values[i];
            // 3.4.2. current feature's position in the sampled feature set
            // int fPos = findFidPlace(this.controller.fSet, fid);
            int fPos = this.controller.fPos[fid];
            if (fPos == -1) {
                continue;
            }
            // 3.4.3. find the position of feature value in a histogram
            // the search area in the sketch is [fid * #splitNum, (fid+1) * #splitNum - 1]
            int start = fid * splitNum;
            // inclusive
            int end;
            if (this.controller.cateFeatNum.containsKey(fid)) {
                end = start + this.controller.cateFeatNum.get(fid) - 1;
            } else {
                end = start + splitNum - 1;
            }
            int fValueIdx = findFvaluePlace(this.controller.sketches, fv, start, end);
            assert fValueIdx >= 0 && fValueIdx < splitNum;
            int gradIdx = 2 * splitNum * fPos + fValueIdx;
            int hessIdx = gradIdx + splitNum;
            // 3.4.4. add the grad and hess to the corresponding bin
            histogram.set(gradIdx, histogram.get(gradIdx) + gradPair.getGrad());
            histogram.set(hessIdx, histogram.get(hessIdx) + gradPair.getHess());
            // 3.4.5. add the reverse to the bin that contains 0.0f
            int fZeroValueIdx = findFvaluePlace(this.controller.sketches, 0.0f, start, end);
            assert fZeroValueIdx >= 0 && fZeroValueIdx < splitNum;
            int gradZeroIdx = 2 * splitNum * fPos + fZeroValueIdx;
            int hessZeroIdx = gradZeroIdx + splitNum;
            double curGrad = histogram.get(gradZeroIdx);
            double curHess = histogram.get(hessZeroIdx);
            histogram.set(gradZeroIdx, curGrad - gradPair.getGrad());
            histogram.set(hessZeroIdx, curHess - gradPair.getHess());
        }
    }
    // 4. add the grad and hess sum to the zero bin of all features
    for (int fid = 0; fid < featureNum; fid++) {
        int fPos = findFidPlace(this.controller.fSet, fid);
        if (fPos == -1) {
            continue;
        }
        int start = fPos * splitNum;
        int end;
        if (this.controller.cateFeatNum.containsKey(fid)) {
            end = start + this.controller.cateFeatNum.get(fid) - 1;
        } else {
            end = start + splitNum - 1;
        }
        int fZeroValueIdx = findFvaluePlace(this.controller.sketches, 0.0f, start, end);
        int gradZeroIdx = 2 * splitNum * fPos + fZeroValueIdx;
        int hessZeroIdx = 2 * splitNum * fPos + fZeroValueIdx + splitNum;
        histogram.set(gradZeroIdx, histogram.get(gradZeroIdx) + gradSum);
        histogram.set(hessZeroIdx, histogram.get(hessZeroIdx) + hessSum);
    }
    LOG.debug(String.format("Build histogram cost %d ms, parse instance cost %d ms", System.currentTimeMillis() - startTime, parseInstanceTime));
    return histogram;
}

Example 84 with IntFloatVector

use of com.tencent.angel.ml.math2.vector.IntFloatVector in project angel by Tencent.

the class MixedBinaryOutNonZAExecutor method apply.

private static Vector apply(CompIntFloatVector v1, IntLongVector v2, Binary op) {
    IntFloatVector[] parts = v1.getPartitions();
    Storage[] resParts = StorageSwitch.applyComp(v1, v2, op);
    if (v2.isDense()) {
        long[] v2Values = v2.getStorage().getValues();
        int base = 0, k = 0;
        for (IntFloatVector part : parts) {
            IntFloatVectorStorage resPart = (IntFloatVectorStorage) resParts[k];
            float[] newValues = resPart.getValues();
            if (part.isDense()) {
                float[] partValue = part.getStorage().getValues();
                for (int i = 0; i < partValue.length; i++) {
                    int idx = i + base;
                    newValues[i] = op.apply(partValue[i], v2Values[idx]);
                }
            } else if (part.isSparse()) {
                if (part.size() < Constant.denseLoopThreshold * part.getDim()) {
                    for (int i = 0; i < part.getDim(); i++) {
                        resPart.set(i, op.apply(0, v2Values[i + base]));
                    }
                    ObjectIterator<Int2FloatMap.Entry> iter = part.getStorage().entryIterator();
                    while (iter.hasNext()) {
                        Int2FloatMap.Entry entry = iter.next();
                        int idx = entry.getIntKey();
                        resPart.set(idx, op.apply(entry.getFloatValue(), v2Values[idx + base]));
                    }
                } else {
                    for (int i = 0; i < newValues.length; i++) {
                        if (part.getStorage().hasKey(i)) {
                            resPart.set(i, op.apply(part.get(i), v2Values[i + base]));
                        } else {
                            resPart.set(i, op.apply(0, v2Values[i + base]));
                        }
                    }
                }
            } else {
                // sorted
                if (op.isKeepStorage()) {
                    int dim = part.getDim();
                    int[] resIndices = resPart.getIndices();
                    float[] resValues = resPart.getValues();
                    int[] partIndices = part.getStorage().getIndices();
                    float[] partValues = part.getStorage().getValues();
                    for (int i = 0; i < dim; i++) {
                        resIndices[i] = i;
                        resValues[i] = op.apply(0, v2Values[i + base]);
                    }
                    int size = part.size();
                    for (int i = 0; i < size; i++) {
                        int idx = partIndices[i];
                        resValues[idx] = op.apply(partValues[i], v2Values[idx + base]);
                    }
                } else {
                    if (part.size() < Constant.denseLoopThreshold * part.getDim()) {
                        int[] partIndices = part.getStorage().getIndices();
                        float[] partValues = part.getStorage().getValues();
                        for (int i = 0; i < part.getDim(); i++) {
                            newValues[i] = op.apply(0, v2Values[i + base]);
                        }
                        int size = part.size();
                        for (int i = 0; i < size; i++) {
                            int idx = partIndices[i];
                            newValues[idx] = op.apply(partValues[i], v2Values[idx + base]);
                        }
                    } else {
                        IntFloatVectorStorage partStorage = part.getStorage();
                        for (int i = 0; i < newValues.length; i++) {
                            if (partStorage.hasKey(i)) {
                                newValues[i] = op.apply(partStorage.get(i), v2Values[i + base]);
                            } else {
                                newValues[i] = op.apply(0, v2Values[i + base]);
                            }
                        }
                    }
                }
            }
            base += part.getDim();
            k++;
        }
    } else {
        if (v2.isSparse()) {
            if (!op.isKeepStorage()) {
                for (int i = 0; i < parts.length; i++) {
                    if (parts[i].getStorage() instanceof IntFloatSortedVectorStorage) {
                        resParts[i] = new IntFloatSparseVectorStorage(parts[i].getDim(), parts[i].getStorage().getIndices(), parts[i].getStorage().getValues());
                    }
                }
            }
            int subDim = (v1.getDim() + v1.getNumPartitions() - 1) / v1.getNumPartitions();
            ObjectIterator<Int2LongMap.Entry> iter = v2.getStorage().entryIterator();
            while (iter.hasNext()) {
                Int2LongMap.Entry entry = iter.next();
                int gidx = entry.getIntKey();
                int pidx = (int) (gidx / subDim);
                int subidx = gidx % subDim;
                ((IntFloatVectorStorage) resParts[pidx]).set(subidx, op.apply(parts[pidx].get(subidx), entry.getLongValue()));
            }
        } else {
            // sorted
            if (!op.isKeepStorage()) {
                for (int i = 0; i < parts.length; i++) {
                    if (parts[i].getStorage() instanceof IntFloatSortedVectorStorage) {
                        resParts[i] = new IntFloatSparseVectorStorage(parts[i].getDim(), parts[i].getStorage().getIndices(), parts[i].getStorage().getValues());
                    }
                }
            }
            int subDim = (v1.getDim() + v1.getNumPartitions() - 1) / v1.getNumPartitions();
            int[] v2Indices = v2.getStorage().getIndices();
            long[] v2Values = v2.getStorage().getValues();
            for (int i = 0; i < v2Indices.length; i++) {
                int gidx = v2Indices[i];
                int pidx = (int) (gidx / subDim);
                int subidx = gidx % subDim;
                ((IntFloatVectorStorage) resParts[pidx]).set(subidx, op.apply(parts[pidx].get(subidx), v2Values[i]));
            }
        }
    }
    IntFloatVector[] res = new IntFloatVector[parts.length];
    int i = 0;
    for (IntFloatVector part : parts) {
        res[i] = new IntFloatVector(part.getMatrixId(), part.getRowId(), part.getClock(), part.getDim(), (IntFloatVectorStorage) resParts[i]);
        i++;
    }
    return new CompIntFloatVector(v1.getMatrixId(), v1.getRowId(), v1.getClock(), v1.getDim(), res, v1.getSubDim());
}

Example 85 with IntFloatVector

use of com.tencent.angel.ml.math2.vector.IntFloatVector in project angel by Tencent.

the class MixedBinaryOutNonZAExecutor method apply.

private static Vector apply(CompIntFloatVector v1, IntDummyVector v2, Binary op) {
    IntFloatVector[] parts = v1.getPartitions();
    Storage[] resParts = StorageSwitch.applyComp(v1, v2, op);
    if (!op.isKeepStorage()) {
        for (int i = 0; i < parts.length; i++) {
            if (parts[i].getStorage() instanceof IntFloatSortedVectorStorage) {
                resParts[i] = new IntFloatSparseVectorStorage(parts[i].getDim(), parts[i].getStorage().getIndices(), parts[i].getStorage().getValues());
            }
        }
    }
    int subDim = (v1.getDim() + v1.getNumPartitions() - 1) / v1.getNumPartitions();
    int[] v2Indices = v2.getIndices();
    for (int i = 0; i < v2Indices.length; i++) {
        int gidx = v2Indices[i];
        int pidx = (int) (gidx / subDim);
        int subidx = gidx % subDim;
        ((IntFloatVectorStorage) resParts[pidx]).set(subidx, op.apply(parts[pidx].get(subidx), 1));
    }
    IntFloatVector[] res = new IntFloatVector[parts.length];
    int i = 0;
    for (IntFloatVector part : parts) {
        res[i] = new IntFloatVector(part.getMatrixId(), part.getRowId(), part.getClock(), part.getDim(), (IntFloatVectorStorage) resParts[i]);
        i++;
    }
    return new CompIntFloatVector(v1.getMatrixId(), v1.getRowId(), v1.getClock(), v1.getDim(), res, v1.getSubDim());
}