use of com.tencent.angel.ml.math2.vector.IntDoubleVector in project angel by Tencent.
the class SimpleBinaryOutNonZAExecutor method apply.
public static Vector apply(IntDoubleVector v1, IntLongVector v2, Binary op) {
IntDoubleVectorStorage newStorage = (IntDoubleVectorStorage) StorageSwitch.apply(v1, v2, op);
if (v1.isDense() && v2.isDense()) {
double[] resValues = newStorage.getValues();
double[] v1Values = v1.getStorage().getValues();
long[] v2Values = v2.getStorage().getValues();
for (int idx = 0; idx < resValues.length; idx++) {
resValues[idx] = op.apply(v1Values[idx], v2Values[idx]);
}
} else if (v1.isDense() && v2.isSparse()) {
double[] resValues = newStorage.getValues();
double[] v1Values = v1.getStorage().getValues();
ObjectIterator<Int2LongMap.Entry> iter = v2.getStorage().entryIterator();
while (iter.hasNext()) {
Int2LongMap.Entry entry = iter.next();
int idx = entry.getIntKey();
resValues[idx] = op.apply(v1Values[idx], entry.getLongValue());
}
} else if (v1.isDense() && v2.isSorted()) {
double[] resValues = newStorage.getValues();
double[] v1Values = v1.getStorage().getValues();
int[] v2Indices = v2.getStorage().getIndices();
long[] v2Values = v2.getStorage().getValues();
int size = v2.size();
for (int i = 0; i < size; i++) {
int idx = v2Indices[i];
resValues[idx] = op.apply(v1Values[idx], v2Values[i]);
}
} else if (v1.isSparse() && v2.isDense()) {
if (op.isKeepStorage()) {
int dim = v1.getDim();
long[] v2Values = v2.getStorage().getValues();
if (v1.size() < Constant.denseLoopThreshold * v1.getDim()) {
for (int i = 0; i < dim; i++) {
newStorage.set(i, op.apply(0, v2Values[i]));
}
ObjectIterator<Int2DoubleMap.Entry> iter = v1.getStorage().entryIterator();
while (iter.hasNext()) {
Int2DoubleMap.Entry entry = iter.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(entry.getDoubleValue(), v2Values[idx]));
}
} else {
for (int i = 0; i < dim; i++) {
if (v1.getStorage().hasKey(i)) {
newStorage.set(i, op.apply(v1.get(i), v2Values[i]));
} else {
newStorage.set(i, op.apply(0, v2Values[i]));
}
}
}
} else {
double[] resValues = newStorage.getValues();
long[] v2Values = v2.getStorage().getValues();
if (v1.size() < Constant.denseLoopThreshold * v1.getDim()) {
for (int i = 0; i < resValues.length; i++) {
resValues[i] = op.apply(0, v2Values[i]);
}
ObjectIterator<Int2DoubleMap.Entry> iter = v1.getStorage().entryIterator();
while (iter.hasNext()) {
Int2DoubleMap.Entry entry = iter.next();
int idx = entry.getIntKey();
resValues[idx] = op.apply(entry.getDoubleValue(), v2Values[idx]);
}
} else {
for (int i = 0; i < resValues.length; i++) {
if (v1.getStorage().hasKey(i)) {
resValues[i] = op.apply(v1.get(i), v2Values[i]);
} else {
resValues[i] = op.apply(0, v2Values[i]);
}
}
}
}
} else if (v1.isSorted() && v2.isDense()) {
if (op.isKeepStorage()) {
int dim = v1.getDim();
int[] resIndices = newStorage.getIndices();
double[] resValues = newStorage.getValues();
long[] v2Values = v2.getStorage().getValues();
int[] v1Indices = v1.getStorage().getIndices();
double[] v1Values = v1.getStorage().getValues();
for (int i = 0; i < dim; i++) {
resIndices[i] = i;
resValues[i] = op.apply(0, v2Values[i]);
}
int size = v1.size();
for (int i = 0; i < size; i++) {
int idx = v1Indices[i];
resValues[idx] = op.apply(v1Values[i], v2Values[idx]);
}
} else {
double[] resValues = newStorage.getValues();
long[] v2Values = v2.getStorage().getValues();
if (v1.size() < Constant.denseLoopThreshold * v1.getDim()) {
int[] v1Indices = v1.getStorage().getIndices();
double[] v1Values = v1.getStorage().getValues();
for (int i = 0; i < resValues.length; i++) {
resValues[i] = op.apply(0, v2Values[i]);
}
int size = v1.size();
for (int i = 0; i < size; i++) {
int idx = v1Indices[i];
resValues[idx] = op.apply(v1Values[i], v2Values[idx]);
}
} else {
IntDoubleVectorStorage v1Storage = v1.getStorage();
for (int i = 0; i < resValues.length; i++) {
if (v1Storage.hasKey(i)) {
resValues[i] = op.apply(v1.get(i), v2Values[i]);
} else {
resValues[i] = op.apply(0, v2Values[i]);
}
}
}
}
} else if (v1.isSparse() && v2.isSparse()) {
int v1Size = v1.size();
int v2Size = v2.size();
if (v1Size >= v2Size * Constant.sparseThreshold && (v1Size + v2Size) * Constant.intersectionCoeff <= Constant.sparseDenseStorageThreshold * v1.dim()) {
// we gauss the indices of v2 maybe is a subset of v1, or overlap is very large
ObjectIterator<Int2LongMap.Entry> iter = v2.getStorage().entryIterator();
while (iter.hasNext()) {
Int2LongMap.Entry entry = iter.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(v1.get(idx), entry.getLongValue()));
}
} else if ((v1Size + v2Size) * Constant.intersectionCoeff >= Constant.sparseDenseStorageThreshold * v1.dim()) {
// we gauss dense storage is more efficient
ObjectIterator<Int2DoubleMap.Entry> iter1 = v1.getStorage().entryIterator();
while (iter1.hasNext()) {
Int2DoubleMap.Entry entry = iter1.next();
int idx = entry.getIntKey();
newStorage.set(idx, entry.getDoubleValue());
}
ObjectIterator<Int2LongMap.Entry> iter2 = v2.getStorage().entryIterator();
while (iter2.hasNext()) {
Int2LongMap.Entry entry = iter2.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(v1.get(idx), entry.getLongValue()));
}
} else {
// to avoid multi-rehash
int capacity = 1 << (32 - Integer.numberOfLeadingZeros((int) (v1.size() / 0.75)));
if (v1.size() + v2.size() <= 1.5 * capacity) {
// no rehashor one onle rehash is required, nothing to optimization
ObjectIterator<Int2LongMap.Entry> iter = v2.getStorage().entryIterator();
while (iter.hasNext()) {
Int2LongMap.Entry entry = iter.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(v1.get(idx), entry.getLongValue()));
}
} else {
// multi-rehash
ObjectIterator<Int2DoubleMap.Entry> iter1 = v1.getStorage().entryIterator();
while (iter1.hasNext()) {
Int2DoubleMap.Entry entry = iter1.next();
int idx = entry.getIntKey();
newStorage.set(idx, entry.getDoubleValue());
}
ObjectIterator<Int2LongMap.Entry> iter2 = v2.getStorage().entryIterator();
while (iter2.hasNext()) {
Int2LongMap.Entry entry = iter2.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(v1.get(idx), entry.getLongValue()));
}
}
}
} else if (v1.isSparse() && v2.isSorted()) {
int v1Size = v1.size();
int v2Size = v2.size();
if (v1Size >= v2Size * Constant.sparseThreshold && (v1Size + v2Size) * Constant.intersectionCoeff <= Constant.sparseDenseStorageThreshold * v1.dim()) {
// we gauss the indices of v2 maybe is a subset of v1, or overlap is very large
int[] v2Indices = v2.getStorage().getIndices();
long[] v2Values = v2.getStorage().getValues();
for (int i = 0; i < v2.size(); i++) {
int idx = v2Indices[i];
newStorage.set(idx, op.apply(v1.get(idx), v2Values[i]));
}
} else if ((v1Size + v2Size) * Constant.intersectionCoeff >= Constant.sparseDenseStorageThreshold * v1.dim()) {
ObjectIterator<Int2DoubleMap.Entry> iter1 = v1.getStorage().entryIterator();
while (iter1.hasNext()) {
Int2DoubleMap.Entry entry = iter1.next();
int idx = entry.getIntKey();
newStorage.set(idx, entry.getDoubleValue());
}
int[] v2Indices = v2.getStorage().getIndices();
long[] v2Values = v2.getStorage().getValues();
int size = v2.size();
for (int i = 0; i < size; i++) {
int idx = v2Indices[i];
newStorage.set(idx, op.apply(v1.get(idx), v2Values[i]));
}
} else {
// to avoid multi-rehash
int capacity = 1 << (32 - Integer.numberOfLeadingZeros((int) (v1.size() / 0.75)));
if (v1.size() + v2.size() <= 1.5 * capacity) {
int[] v2Indices = v2.getStorage().getIndices();
long[] v2Values = v2.getStorage().getValues();
for (int i = 0; i < v2.size(); i++) {
int idx = v2Indices[i];
newStorage.set(idx, op.apply(v1.get(idx), v2Values[i]));
}
} else {
ObjectIterator<Int2DoubleMap.Entry> iter1 = v1.getStorage().entryIterator();
while (iter1.hasNext()) {
Int2DoubleMap.Entry entry = iter1.next();
int idx = entry.getIntKey();
newStorage.set(idx, entry.getDoubleValue());
}
int[] v2Indices = v2.getStorage().getIndices();
long[] v2Values = v2.getStorage().getValues();
int size = v2.size();
for (int i = 0; i < size; i++) {
int idx = v2Indices[i];
newStorage.set(idx, op.apply(v1.get(idx), v2Values[i]));
}
}
}
} else if (v1.isSorted() && v2.isSparse()) {
int v1Size = v1.size();
int v2Size = v2.size();
if ((v1Size + v2Size) * Constant.intersectionCoeff >= Constant.sortedDenseStorageThreshold * v1.dim()) {
if (op.isKeepStorage()) {
int[] v1Indices = v1.getStorage().getIndices();
int[] idxiter = v2.getStorage().indexIterator().toIntArray();
int[] indices = new int[(int) (v1Size + v2Size)];
System.arraycopy(v1Indices, 0, indices, 0, (int) v1.size());
System.arraycopy(idxiter, 0, indices, (int) v1.size(), (int) v2.size());
IntAVLTreeSet avl = new IntAVLTreeSet(indices);
IntBidirectionalIterator iter = avl.iterator();
double[] values = new double[indices.length];
int i = 0;
while (iter.hasNext()) {
int idx = iter.nextInt();
indices[i] = idx;
values[i] = op.apply(v1.get(idx), v2.get(idx));
i++;
}
while (i < indices.length) {
indices[i] = 0;
i++;
}
newStorage = new IntDoubleSortedVectorStorage(v1.getDim(), (int) avl.size(), indices, values);
} else {
int[] v1Indices = v1.getStorage().getIndices();
double[] v1Values = v1.getStorage().getValues();
int size = v1.size();
for (int i = 0; i < size; i++) {
int idx = v1Indices[i];
newStorage.set(idx, v1Values[i]);
}
ObjectIterator<Int2LongMap.Entry> iter = v2.getStorage().entryIterator();
while (iter.hasNext()) {
Int2LongMap.Entry entry = iter.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(newStorage.get(idx), entry.getLongValue()));
}
}
} else {
if (op.isKeepStorage()) {
int[] v1Indices = v1.getStorage().getIndices();
int[] idxiter = v2.getStorage().indexIterator().toIntArray();
int[] indices = new int[(int) (v1Size + v2Size)];
System.arraycopy(v1Indices, 0, indices, 0, (int) v1.size());
System.arraycopy(idxiter, 0, indices, (int) v1.size(), (int) v2.size());
IntAVLTreeSet avl = new IntAVLTreeSet(indices);
IntBidirectionalIterator iter = avl.iterator();
double[] values = new double[indices.length];
int i = 0;
while (iter.hasNext()) {
int idx = iter.nextInt();
indices[i] = idx;
values[i] = op.apply(v1.get(idx), v2.get(idx));
i++;
}
while (i < indices.length) {
indices[i] = 0;
i++;
}
newStorage = new IntDoubleSortedVectorStorage(v1.getDim(), (int) avl.size(), indices, values);
} else {
int[] v1Indices = v1.getStorage().getIndices();
double[] v1Values = v1.getStorage().getValues();
int size = v1.size();
for (int i = 0; i < size; i++) {
int idx = v1Indices[i];
newStorage.set(idx, v1Values[i]);
}
ObjectIterator<Int2LongMap.Entry> iter = v2.getStorage().entryIterator();
while (iter.hasNext()) {
Int2LongMap.Entry entry = iter.next();
int idx = entry.getIntKey();
newStorage.set(idx, op.apply(newStorage.get(idx), entry.getLongValue()));
}
}
}
} else if (v1.isSorted() && v2.isSorted()) {
int v1Pointor = 0;
int v2Pointor = 0;
int size1 = v1.size();
int size2 = v2.size();
int[] v1Indices = v1.getStorage().getIndices();
double[] v1Values = v1.getStorage().getValues();
int[] v2Indices = v2.getStorage().getIndices();
long[] v2Values = v2.getStorage().getValues();
if ((size1 + size2) * Constant.intersectionCoeff >= Constant.sortedDenseStorageThreshold * v1.dim()) {
if (op.isKeepStorage()) {
// sorted
int[] resIndices = newStorage.getIndices();
double[] resValues = newStorage.getValues();
int global = 0;
while (v1Pointor < size1 && v2Pointor < size2) {
if (v1Indices[v1Pointor] == v2Indices[v2Pointor]) {
resIndices[global] = v1Indices[v1Pointor];
resValues[global] = op.apply(v1Values[v1Pointor], v2Values[v2Pointor]);
global++;
v1Pointor++;
v2Pointor++;
} else if (v1Indices[v1Pointor] < v2Indices[v2Pointor]) {
resIndices[global] = v1Indices[v1Pointor];
resValues[global] = v1Values[v1Pointor];
global++;
v1Pointor++;
} else {
// v1Indices[v1Pointor] > v2Indices[v2Pointor]
resIndices[global] = v2Indices[v2Pointor];
resValues[global] = op.apply(0, v2Values[v2Pointor]);
global++;
v2Pointor++;
}
}
} else {
// dense
while (v1Pointor < size1 || v2Pointor < size2) {
if ((v1Pointor < size1 && v2Pointor < size2) && v1Indices[v1Pointor] == v2Indices[v2Pointor]) {
newStorage.set(v1Indices[v1Pointor], op.apply(v1Values[v1Pointor], v2Values[v2Pointor]));
v1Pointor++;
v2Pointor++;
} else if ((v1Pointor < size1 && v2Pointor < size2) && v1Indices[v1Pointor] < v2Indices[v2Pointor] || (v1Pointor < size1 && v2Pointor >= size2)) {
newStorage.set(v1Indices[v1Pointor], v1Values[v1Pointor]);
v1Pointor++;
} else if (((v1Pointor < size1 && v2Pointor < size2) && v1Indices[v1Pointor] >= v2Indices[v2Pointor]) || (v1Pointor >= size1 && v2Pointor < size2)) {
newStorage.set(v2Indices[v2Pointor], op.apply(0, v2Values[v2Pointor]));
v2Pointor++;
}
}
}
} else {
if (op.isKeepStorage()) {
int[] resIndices = newStorage.getIndices();
double[] resValues = newStorage.getValues();
int globalPointor = 0;
while (v1Pointor < size1 && v2Pointor < size2) {
if (v1Indices[v1Pointor] == v2Indices[v2Pointor]) {
resIndices[globalPointor] = v1Indices[v1Pointor];
resValues[globalPointor] = op.apply(v1Values[v1Pointor], v2Values[v2Pointor]);
v1Pointor++;
v2Pointor++;
globalPointor++;
} else if (v1Indices[v1Pointor] < v2Indices[v2Pointor]) {
resIndices[globalPointor] = v1Indices[v1Pointor];
resValues[globalPointor] = v1Values[v1Pointor];
v1Pointor++;
globalPointor++;
} else {
// v1Indices[v1Pointor] > v2Indices[v2Pointor]
resIndices[globalPointor] = v2Indices[v2Pointor];
resValues[globalPointor] = op.apply(0, v2Values[v2Pointor]);
v2Pointor++;
globalPointor++;
}
}
} else {
while (v1Pointor < size1 || v2Pointor < size2) {
if ((v1Pointor < size1 && v2Pointor < size2) && v1Indices[v1Pointor] == v2Indices[v2Pointor]) {
newStorage.set(v1Indices[v1Pointor], op.apply(v1Values[v1Pointor], v2Values[v2Pointor]));
v1Pointor++;
v2Pointor++;
} else if ((v1Pointor < size1 && v2Pointor < size2) && v1Indices[v1Pointor] < v2Indices[v2Pointor] || (v1Pointor < size1 && v2Pointor >= size2)) {
newStorage.set(v1Indices[v1Pointor], v1Values[v1Pointor]);
v1Pointor++;
} else if (((v1Pointor < size1 && v2Pointor < size2) && v1Indices[v1Pointor] >= v2Indices[v2Pointor]) || (v1Pointor >= size1 && v2Pointor < size2)) {
newStorage.set(v2Indices[v2Pointor], op.apply(0, v2Values[v2Pointor]));
v2Pointor++;
}
}
}
}
} else {
throw new AngelException("The operation is not support!");
}
return new IntDoubleVector(v1.getMatrixId(), v1.getRowId(), v1.getClock(), v1.getDim(), newStorage);
}
use of com.tencent.angel.ml.math2.vector.IntDoubleVector in project angel by Tencent.
the class SimpleUnaryExecutor method apply.
private static Vector apply(IntDoubleVector v1, Unary op) {
IntDoubleVector res;
if (op.isOrigin() || v1.isDense()) {
if (!op.isInplace()) {
res = v1.copy();
} else {
res = v1;
}
if (v1.isDense()) {
double[] values = res.getStorage().getValues();
for (int i = 0; i < values.length; i++) {
values[i] = op.apply(values[i]);
}
} else if (v1.isSparse()) {
ObjectIterator<Int2DoubleMap.Entry> iter = res.getStorage().entryIterator();
while (iter.hasNext()) {
Int2DoubleMap.Entry entry = iter.next();
entry.setValue(op.apply(entry.getDoubleValue()));
}
} else if (v1.isSorted()) {
double[] values = res.getStorage().getValues();
for (int i = 0; i < v1.size(); i++) {
values[i] = op.apply(values[i]);
}
} else {
throw new AngelException("The operation is not support!");
}
} else {
IntDoubleVectorStorage newstorage = v1.getStorage().emptyDense();
IntDoubleVectorStorage storage = v1.getStorage();
double[] values = newstorage.getValues();
double tmp = op.apply((double) 0);
int dim = v1.getDim();
for (int i = 0; i < dim; i++) {
values[i] = tmp;
}
if (v1.isSparse()) {
ObjectIterator<Int2DoubleMap.Entry> iter = storage.entryIterator();
while (iter.hasNext()) {
Int2DoubleMap.Entry entry = iter.next();
values[entry.getIntKey()] = op.apply(entry.getDoubleValue());
}
} else {
// sort
int[] idxs = storage.getIndices();
double[] v1Values = storage.getValues();
for (int k = 0; k < idxs.length; k++) {
values[idxs[k]] = op.apply(v1Values[k]);
}
}
if (op.isInplace()) {
v1.setStorage(newstorage);
res = v1;
} else {
res = new IntDoubleVector(v1.getMatrixId(), v1.getRowId(), v1.getClock(), v1.getDim(), newstorage);
}
}
return res;
}
use of com.tencent.angel.ml.math2.vector.IntDoubleVector in project angel by Tencent.
the class MatrixUtils method rbCompDense2Blas.
public static BlasDoubleMatrix rbCompDense2Blas(RBCompIntDoubleMatrix mat) {
assert mat != null;
int dim = (int) mat.getDim();
int subDim = mat.getSubDim();
CompIntDoubleVector[] rows = mat.getRows();
double[] data = new double[rows.length * dim];
int rowId = 0;
for (CompIntDoubleVector row : rows) {
IntDoubleVector[] partitions = row.getPartitions();
int partId = 0;
for (IntDoubleVector part : partitions) {
assert part.isDense();
double[] src = part.getStorage().getValues();
System.arraycopy(src, 0, data, rowId * dim + partId * subDim, src.length);
partId += 1;
}
rowId += 1;
}
return MFactory.denseDoubleMatrix(rows.length, dim, data);
}
use of com.tencent.angel.ml.math2.vector.IntDoubleVector in project angel by Tencent.
the class RangeRouterUtils method splitIntDoubleVector.
public static KeyValuePart[] splitIntDoubleVector(MatrixMeta matrixMeta, IntDoubleVector vector) {
IntDoubleVectorStorage storage = vector.getStorage();
if (storage.isSparse()) {
// Get keys and values
IntDoubleSparseVectorStorage sparseStorage = (IntDoubleSparseVectorStorage) storage;
int[] keys = sparseStorage.getIndices();
double[] values = sparseStorage.getValues();
return split(matrixMeta, vector.getRowId(), keys, values, false);
} else if (storage.isDense()) {
// Get values
IntDoubleDenseVectorStorage denseStorage = (IntDoubleDenseVectorStorage) storage;
double[] values = denseStorage.getValues();
return split(matrixMeta, vector.getRowId(), values);
} else {
// Key and value array pair
IntDoubleSortedVectorStorage sortStorage = (IntDoubleSortedVectorStorage) storage;
int[] keys = sortStorage.getIndices();
double[] values = sortStorage.getValues();
return split(matrixMeta, vector.getRowId(), keys, values, true);
}
}
use of com.tencent.angel.ml.math2.vector.IntDoubleVector in project angel by Tencent.
the class GBDTController method updateNodeGradStats.
// update node's grad stats on PS
// called during splitting in GradHistHelper, update the grad stats of children nodes after finding the best split
// the root node's stats is updated by leader worker
public void updateNodeGradStats(int nid, GradStats gradStats) throws Exception {
LOG.debug(String.format("Update gradStats of node[%d]: sumGrad[%f], sumHess[%f]", nid, gradStats.sumGrad, gradStats.sumHess));
// 1. create the update
IntDoubleVector vec = new IntDoubleVector(2 * this.activeNode.length, new IntDoubleDenseVectorStorage(2 * this.activeNode.length));
vec.set(nid, gradStats.sumGrad);
vec.set(nid + this.activeNode.length, gradStats.sumHess);
// 2. push the update to PS
PSModel nodeGradStats = this.model.getPSModel(this.param.nodeGradStatsName);
nodeGradStats.increment(this.currentTree, vec);
}
Aggregations