/**
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.mahout.math.hadoop.stochasticsvd;
import java.io.IOException;
import java.util.List;
import java.util.Map;
import java.util.Random;
import com.google.common.collect.Lists;
import com.google.common.collect.Maps;
import org.apache.mahout.common.RandomUtils;
import org.apache.mahout.common.iterator.CopyConstructorIterator;
import org.apache.mahout.math.DenseMatrix;
import org.apache.mahout.math.DenseVector;
import org.apache.mahout.math.SingularValueDecomposition;
import org.apache.mahout.math.Vector;
import org.apache.mahout.math.function.Functions;
import org.apache.mahout.math.hadoop.stochasticsvd.qr.GivensThinSolver;
import org.apache.mahout.math.ssvd.EigenSolverWrapper;
/**
* SSVD protoptype: non-MR concept verification for Givens QR & SSVD basic
* algorithms
*/
public class SSVDPrototype {
private final Omega omega;
private final int kp; // k+p
private GivensThinSolver qSolver;
private final double[] yRow;
private int cnt;
private int blckCnt;
private final int r;
private final List<UpperTriangular> rBlocks = Lists.newArrayList();
private final List<double[][]> qtBlocks = Lists.newArrayList();
private final List<double[]> yLookahead;
public SSVDPrototype(long seed, int kp, int r) {
this.kp = kp;
omega = new Omega(seed, kp / 2, kp - kp / 2);
yRow = new double[kp];
// m_yRowV = new DenseVector(m_yRow,true);
this.r = r;
yLookahead = Lists.newArrayList();
}
void firstPass(Vector aRow) {
omega.computeYRow(aRow, yRow);
yLookahead.add(yRow.clone()); // bad for GC but it's just a prototype,
// hey. in real thing we'll rotate usage
// of y buff
while (yLookahead.size() > kp) {
if (qSolver == null) {
qSolver = new GivensThinSolver(r, kp);
}
qSolver.appendRow(yLookahead.remove(0));
if (qSolver.isFull()) {
UpperTriangular r = qSolver.getRTilde();
double[][] qt = qSolver.getThinQtTilde();
qSolver = null;
qtBlocks.add(qt);
rBlocks.add(r);
}
}
cnt++;
}
void finishFirstPass() {
if (qSolver == null && yLookahead.isEmpty()) {
return;
}
if (qSolver == null) {
qSolver = new GivensThinSolver(yLookahead.size(), kp);
}
// grow q solver up if necessary
qSolver.adjust(qSolver.getCnt() + yLookahead.size());
while (!yLookahead.isEmpty()) {
qSolver.appendRow(yLookahead.remove(0));
if (qSolver.isFull()) {
UpperTriangular r = qSolver.getRTilde();
double[][] qt = qSolver.getThinQtTilde();
qSolver = null;
qtBlocks.add(qt);
rBlocks.add(r);
}
}
// simulate reducers -- produce qHats
for (int i = 0; i < rBlocks.size(); i++) {
qtBlocks.set(i,
GivensThinSolver.computeQtHat(qtBlocks.get(i), i,
new CopyConstructorIterator<UpperTriangular>(rBlocks.listIterator())));
}
cnt = 0;
blckCnt = 0;
}
void secondPass(Vector aRow, PartialRowEmitter btEmitter)
throws IOException {
int n = aRow.size();
double[][] qtHat = qtBlocks.get(blckCnt);
int r = qtHat[0].length;
int qRowBlckIndex = r - cnt - 1; // <-- reverse order since we fed A in
// reverse
double[] qRow = new double[kp];
for (int i = 0; i < kp; i++) {
qRow[i] = qtHat[i][qRowBlckIndex];
}
Vector qRowV = new DenseVector(qRow, true);
if (++cnt == r) {
blckCnt++;
cnt = 0;
}
for (int i = 0; i < n; i++) {
btEmitter.emitRow(i, qRowV.times(aRow.getQuick(i)));
}
}
public static void testThinQr(int dims, int kp) throws IOException {
DenseMatrix mx = new DenseMatrix(dims << 2, dims);
Random rnd = RandomUtils.getRandom();
for (int i = 0; i < mx.rowSize(); i++) {
for (int j = 0; j < mx.columnSize(); j++) {
mx.set(i, j, rnd.nextDouble() * 1000);
}
}
mx.setQuick(0, 0, 1);
mx.setQuick(0, 1, 2);
mx.setQuick(0, 2, 3);
mx.setQuick(1, 0, 4);
mx.setQuick(1, 1, 5);
mx.setQuick(1, 2, 6);
mx.setQuick(2, 0, 7);
mx.setQuick(2, 1, 8);
mx.setQuick(2, 2, 9);
SingularValueDecomposition svd2 = new SingularValueDecomposition(mx);
double[] svaluesControl = svd2.getSingularValues();
for (int i = 0; i < kp; i++) {
System.out.printf("%.3e ", svaluesControl[i]);
}
System.out.println();
int m = mx.rowSize(); /* ,n=mx.columnSize(); */
long seed = RandomUtils.getRandom().nextLong();
final Map<Integer, Vector> btRows = Maps.newHashMap();
PartialRowEmitter btEmitter = new PartialRowEmitter() {
@Override
public void emitRow(int rowNum, Vector row) {
Vector btRow = btRows.get(rowNum);
if (btRow != null) {
btRow.assign(row, Functions.PLUS);
}
btRows.put(rowNum, btRow == null ? new DenseVector(row) : btRow);
}
};
SSVDPrototype mapperSimulation = new SSVDPrototype(seed, kp, 3000);
for (int i = 0; i < m; i++) {
mapperSimulation.firstPass(mx.viewRow(i));
}
mapperSimulation.finishFirstPass();
for (int i = 0; i < m; i++) {
mapperSimulation.secondPass(mx.viewRow(i), btEmitter);
}
// LocalSSVDTest.assertOrthonormality(mapperSimulation.m_qt.transpose(),
// false,1e-10);
// reconstruct bbt
final Map<Integer, Vector> bbt = Maps.newHashMap();
PartialRowEmitter bbtEmitter = new PartialRowEmitter() {
@Override
public void emitRow(int rowNum, Vector row) {
Vector bbtRow = bbt.get(rowNum);
if (bbtRow != null) {
bbtRow.assign(row, Functions.PLUS);
}
bbt.put(rowNum, bbtRow == null ? new DenseVector(row) : bbtRow);
}
};
for (Map.Entry<Integer, Vector> btRowEntry : btRows.entrySet()) {
Vector btRow = btRowEntry.getValue();
assert btRow.size() == kp;
for (int i = 0; i < kp; i++) {
bbtEmitter.emitRow(i, btRow.times(btRow.getQuick(i)));
}
}
double[][] bbtValues = new double[kp][];
for (int i = 0; i < kp; i++) {
bbtValues[i] = new double[kp];
Vector bbtRow = bbt.get(i);
for (int j = 0; j < kp; j++) {
bbtValues[i][j] = bbtRow.getQuick(j);
}
}
EigenSolverWrapper eigenWrapper = new EigenSolverWrapper(bbtValues);
double[] eigenva2 = eigenWrapper.getEigenValues();
double[] svalues = new double[kp];
for (int i = 0; i < kp; i++) {
svalues[i] = Math.sqrt(eigenva2[i]); // sqrt?
}
for (int i = 0; i < kp; i++) {
System.out.printf("%.3e ", svalues[i]);
}
System.out.println();
}
public static void testBlockQrWithSSVD(int dims, int kp, int r, long rndSeed) throws IOException {
DenseMatrix mx = new DenseMatrix(dims << 2, dims);
Random rnd = RandomUtils.getRandom();
for (int i = 0; i < mx.rowSize(); i++) {
for (int j = 0; j < mx.columnSize(); j++) {
mx.set(i, j, (rnd.nextDouble() - 0.5) * 1000);
}
}
mx.setQuick(0, 0, 1);
mx.setQuick(0, 1, 2);
mx.setQuick(0, 2, 3);
mx.setQuick(1, 0, 4);
mx.setQuick(1, 1, 5);
mx.setQuick(1, 2, 6);
mx.setQuick(2, 0, 7);
mx.setQuick(2, 1, 8);
mx.setQuick(2, 2, 9);
SingularValueDecomposition svd2 = new SingularValueDecomposition(mx);
double[] svaluesControl = svd2.getSingularValues();
for (int i = 0; i < kp; i++) {
System.out.printf("%e ", svaluesControl[i]);
}
System.out.println();
int m = mx.rowSize(); /* ,n=mx.columnSize(); */
final Map<Integer, Vector> btRows = Maps.newHashMap();
PartialRowEmitter btEmitter = new PartialRowEmitter() {
@Override
public void emitRow(int rowNum, Vector row) {
Vector btRow = btRows.get(rowNum);
if (btRow != null) {
btRow.assign(row, Functions.PLUS);
}
btRows.put(rowNum, btRow == null ? new DenseVector(row) : btRow);
}
};
SSVDPrototype mapperSimulation = new SSVDPrototype(rndSeed, kp, r);
for (int i = 0; i < m; i++) {
mapperSimulation.firstPass(mx.viewRow(i));
}
mapperSimulation.finishFirstPass();
for (int i = 0; i < m; i++) {
mapperSimulation.secondPass(mx.viewRow(i), btEmitter);
}
// LocalSSVDTest.assertOrthonormality(mapperSimulation.m_qt.transpose(),
// false,1e-10);
// reconstruct bbt
final Map<Integer, Vector> bbt = Maps.newHashMap();
PartialRowEmitter bbtEmitter = new PartialRowEmitter() {
@Override
public void emitRow(int rowNum, Vector row) {
Vector bbtRow = bbt.get(rowNum);
if (bbtRow != null) {
bbtRow.assign(row, Functions.PLUS);
}
bbt.put(rowNum, bbtRow == null ? new DenseVector(row) : bbtRow);
}
};
for (Map.Entry<Integer, Vector> btRowEntry : btRows.entrySet()) {
Vector btRow = btRowEntry.getValue();
assert btRow.size() == kp;
for (int i = 0; i < kp; i++) {
bbtEmitter.emitRow(i, btRow.times(btRow.getQuick(i)));
}
}
double[][] bbtValues = new double[kp][];
for (int i = 0; i < kp; i++) {
bbtValues[i] = new double[kp];
Vector bbtRow = bbt.get(i);
for (int j = 0; j < kp; j++) {
bbtValues[i][j] = bbtRow.getQuick(j);
}
}
EigenSolverWrapper eigenWrapper = new EigenSolverWrapper(bbtValues);
double[] eigenva2 = eigenWrapper.getEigenValues();
double[] svalues = new double[kp];
for (int i = 0; i < kp; i++) {
svalues[i] = Math.sqrt(eigenva2[i]); // sqrt?
}
for (int i = 0; i < kp; i++) {
System.out.printf("%e ", svalues[i]);
}
System.out.println();
}
public static void main(String[] args) throws Exception {
// testThinQr();
long seed = RandomUtils.getRandom().nextLong();
testBlockQrWithSSVD(200, 200, 800, seed);
testBlockQrWithSSVD(200, 20, 800, seed);
testBlockQrWithSSVD(200, 20, 850, seed); // test trimming
testBlockQrWithSSVD(200, 20, 90, seed);
testBlockQrWithSSVD(200, 20, 99, seed);
}
}