com.expleague.ml.methods.greedyRegion.GreedyRegion Maven / Gradle / Ivy
package com.expleague.ml.methods.greedyRegion;
import com.expleague.commons.math.vectors.Vec;
import com.expleague.commons.util.Holder;
import com.expleague.ml.data.set.VecDataSet;
import com.expleague.ml.BFGrid;
import com.expleague.ml.impl.BinaryFeatureImpl;
import com.expleague.ml.loss.WeightedLoss;
import com.expleague.commons.util.ArrayTools;
import com.expleague.ml.data.set.DataSet;
import com.expleague.ml.loss.L2;
import com.expleague.ml.methods.VecOptimization;
import com.expleague.ml.models.Region;
import java.util.ArrayList;
import java.util.List;
import java.util.Random;
import java.util.concurrent.ArrayBlockingQueue;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ThreadPoolExecutor;
import java.util.concurrent.TimeUnit;
/**
* User: solar
* Date: 15.11.12
* Time: 15:19
*/
public class GreedyRegion extends VecOptimization.Stub> {
public static final int NN_NEIGHBORHOOD = 1000;
private final Random rng;
private final BFGrid grid;
byte[][] binarization;
int[] nn;
private final double betta = 0.00001;
public GreedyRegion(final Random rng, final BFGrid grid) {
this.rng = rng;
this.grid = grid;
}
private void prepareNN(final DataSet ds) {
final int total = ds.length();
binarization = new byte[total][];
for (int i = 0; i < ds.length(); i++) {
final byte[] folds = binarization[i] = new byte[((VecDataSet) ds).xdim()];
grid.binarizeTo(((VecDataSet) ds).data().row(i), folds);
}
nn = new int[total * NN_NEIGHBORHOOD];
final int[] l1dist = new int[total];
for (int i = 0; i < total; i++) {
final byte[] folds = binarization[i];
final int[] order = ArrayTools.sequence(0, total);
{
for (int t = 0; t < binarization.length; t++) {
final byte[] currentFolds = binarization[t];
int l1 = 0;
for (int f = 0; f < folds.length; f++) {
final int diff = folds[f] - currentFolds[f];
l1 += diff > 0 ? diff : -diff;
}
l1dist[t] = l1;
}
ArrayTools.parallelSort(l1dist, order);
for (int t = 0; t < NN_NEIGHBORHOOD; t++) {
nn[i * NN_NEIGHBORHOOD + t] = order[t];
}
}
}
}
public static final int POOL_SIZE = Runtime.getRuntime().availableProcessors();
ThreadPoolExecutor exec = new ThreadPoolExecutor(POOL_SIZE, POOL_SIZE, 1, TimeUnit.SECONDS, new ArrayBlockingQueue(100));
@Override
public synchronized Region fit(final VecDataSet learn, final WeightedLoss loss) {
prepareNN(learn);
final Holder answer = new Holder(null);
final CountDownLatch latch = new CountDownLatch(POOL_SIZE);
for (int i = 0; i < Runtime.getRuntime().availableProcessors(); i++) {
exec.execute(new Runnable() {
@Override
public void run() {
final Region model = fitInner(learn, loss);
synchronized (answer) {
if (answer.getValue() == null || answer.getValue().score() > model.score())
answer.setValue(model);
}
latch.countDown();
}
});
}
try {
latch.await();
} catch (InterruptedException e) {
// skip
}
return answer.getValue();
}
public Region fitInner(final VecDataSet learn, final WeightedLoss loss) {
final int pointIdx = choosePointAtRandomNN(learn, loss.base());
final byte[] folds = binarization[pointIdx];
final int total = learn.length();
final int[] order = ArrayTools.sequence(0, total);
final int[] l1dist = new int[total];
{
for (int i = 0; i < binarization.length; i++) {
final byte[] currentFolds = binarization[i];
int l1 = 0;
for (int f = 0; f < folds.length; f++) {
final int diff = folds[f] - currentFolds[f];
l1 += diff > 0 ? diff : -diff;
}
l1dist[i] = l1;
}
ArrayTools.parallelSort(l1dist, order);
}
final List conditions = new ArrayList<>(grid.size());
for (int bf = 0; bf < grid.size(); bf++) {
final BinaryCond bc = new BinaryCond();
bc.bf = grid.bf(bf);
bc.mask = bc.bf.value(folds);
// if (rng.nextDouble() > 100. / grid.size())
conditions.add(bc);
}
List best = null;
double bestMean = 0.;
double bestScore = Double.MAX_VALUE;
int bestCount = 0;
final Vec target = loss.base().target;
while (!conditions.isEmpty()) {
final int currentConditionsCount = conditions.size();
final double[] csum = new double[currentConditionsCount];
final double[] csum2 = new double[currentConditionsCount];
final int[] ccount = new int[currentConditionsCount];
double sum = 0;
double sum2 = 0;
int count = 0;
for (int t = 0; t < order.length; t++) {
if (l1dist[t] > grid.size() - currentConditionsCount - 1)
break;
int matches = currentConditionsCount;
int lastUnmatch = 0;
final int index = order[t];
final byte[] currentFolds = binarization[index];
for (int i = 0; i < currentConditionsCount && matches > currentConditionsCount - 2; i++) {
final BinaryCond next = conditions.get(i);
if (!next.yes(currentFolds)) {
matches--;
lastUnmatch = i;
}
}
if (matches == currentConditionsCount) {
final double y = target.get(index);
sum += y;
sum2 += y * y;
count++;
}
else if (matches == currentConditionsCount - 1) { // the only binary feature has not matched
final double y = target.get(index);
csum[lastUnmatch] += y;
csum2[lastUnmatch] += y * y;
ccount[lastUnmatch]++;
}
}
{ // best region update
final double score = score(total, count, sum, sum2, currentConditionsCount);
if (score < bestScore) {
best = new ArrayList<>(conditions);
bestScore = score;
bestMean = sum/count;
bestCount = count;
}
}
{ // choose what condition should be dropped
int worst = (int)(currentConditionsCount * rng.nextDouble());
double score = score(total, count + ccount[worst], sum + csum[worst], sum2 + csum2[worst], currentConditionsCount - 1);
for (int i = 0; i < currentConditionsCount; i++) {
final double cscore = score(total, count + ccount[i], sum + csum[i], sum2 + csum2[i], currentConditionsCount - 1);
if (cscore < score) {
worst = i;
score = cscore;
}
}
conditions.remove(worst);
}
}
final List features = new ArrayList<>();
final boolean[] mask = new boolean[best.size()];
for (int i = 0; i < best.size(); i++) {
features.add(best.get(i).bf);
mask[i] = best.get(i).mask;
}
return new Region(features, mask, bestMean, bestCount, bestScore);
}
private double score(final int total, final int count, final double sum, final double sum2, final int ccount) {
final double err = -sum * sum / count;
return err * (1. - 2 * (Math.log(2)/ Math.log(count + 1.) + (total > count ? Math.log(2)/ Math.log(total - count + 1.) : 0))) + betta * ccount;
}
private int choosePointAtRandomNN(final VecDataSet learn, final L2 target) {
double total = 0.;
final double[] weights = new double[learn.length()];
double max = 0;
for (int i = 0; i < weights.length; i++) {
double sum = 0;
for (int t = 0; t < NN_NEIGHBORHOOD; t++) {
sum += target.get(nn[i * NN_NEIGHBORHOOD + t]);
}
weights[i] = sum * sum;
total += weights[i];
if (max < weights[i]) {
max = weights[i];
}
}
// return maxIndex;
double rnd = total * rng.nextDouble();
for (int i = 0; i < weights.length; i++) {
if (rnd <= weights[i])
return i;
rnd -= weights[i];
}
return 0;
}
private static class BinaryCond {
BFGrid.Feature bf;
boolean mask;
public boolean yes(final byte[] folds) {
return bf.value(folds) == mask;
}
@Override
public String toString() {
return " " + bf.findex() + (mask ? ">=" : "<") + bf.condition();
}
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy