hivemall.regression.AROWRegressionUDTF Maven / Gradle / Ivy
The newest version!
/*
* 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 hivemall.regression;
import hivemall.annotations.Cite;
import hivemall.model.FeatureValue;
import hivemall.model.IWeightValue;
import hivemall.model.PredictionResult;
import hivemall.model.WeightValue.WeightValueWithCovar;
import hivemall.optimizer.LossFunctions;
import hivemall.utils.stats.OnlineVariance;
import javax.annotation.Nonnull;
import org.apache.commons.cli.CommandLine;
import org.apache.commons.cli.Options;
import org.apache.hadoop.hive.ql.exec.Description;
import org.apache.hadoop.hive.ql.exec.UDFArgumentException;
import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.StructObjectInspector;
// @formatter:off
@Description(name = "train_arow_regr",
value = "_FUNC_(array features, float target [, constant string options])"
+ " - a standard AROW (Adaptive Reguralization of Weight Vectors) regressor "
+ "that uses `y - w^Tx` for the loss function.",
extended = "SELECT \n" +
" feature,\n" +
" argmin_kld(weight, covar) as weight\n" +
"FROM (\n" +
" SELECT \n" +
" train_arow_regr(features,label) as (feature,weight,covar)\n" +
" FROM \n" +
" training_data\n" +
" ) t \n" +
"GROUP BY feature")
// @formatter:on
@Cite(description = "K. Crammer, A. Kulesza, and M. Dredze, \"Adaptive Regularization of Weight Vectors\", In Proc. NIPS, 2009.",
url = "https://papers.nips.cc/paper/3848-adaptive-regularization-of-weight-vectors.pdf")
public class AROWRegressionUDTF extends RegressionBaseUDTF {
/** Regularization parameter r */
protected float r;
@Override
public StructObjectInspector initialize(ObjectInspector[] argOIs) throws UDFArgumentException {
final int numArgs = argOIs.length;
if (numArgs != 2 && numArgs != 3) {
throw new UDFArgumentException(
"_FUNC_ takes arguments: List features, float target [, constant string options]");
}
return super.initialize(argOIs);
}
@Override
protected boolean useCovariance() {
return true;
}
@Override
protected Options getOptions() {
Options opts = super.getOptions();
opts.addOption("r", "regularization", true,
"Regularization parameter for some r > 0 [default 0.1]");
return opts;
}
@Override
protected CommandLine processOptions(ObjectInspector[] argOIs) throws UDFArgumentException {
final CommandLine cl = super.processOptions(argOIs);
float r = 0.1f;
if (cl != null) {
String r_str = cl.getOptionValue("r");
if (r_str != null) {
r = Float.parseFloat(r_str);
if (!(r > 0)) {
throw new UDFArgumentException(
"Regularization parameter must be greater than 0: " + r_str);
}
}
}
this.r = r;
return cl;
}
@Override
protected void train(@Nonnull final FeatureValue[] features, float target) {
PredictionResult margin = calcScoreAndVariance(features);
float predicted = margin.getScore();
float loss = loss(target, predicted);
float var = margin.getVariance();
float beta = 1.f / (var + r);
update(features, loss, beta);
}
/**
* @return target - predicted
*/
protected float loss(float target, float predicted) {
return target - predicted; // y - w^Tx
}
@Override
protected void update(@Nonnull final FeatureValue[] features, final float coeff,
final float beta) {
for (FeatureValue f : features) {
if (f == null) {
continue;
}
Object k = f.getFeature();
float v = f.getValueAsFloat();
IWeightValue old_w = model.get(k);
IWeightValue new_w = getNewWeight(old_w, v, coeff, beta);
model.set(k, new_w);
}
}
private static IWeightValue getNewWeight(final IWeightValue old, final float x,
final float coeff, final float beta) {
final float old_w;
final float old_cov;
if (old == null) {
old_w = 0.f;
old_cov = 1.f;
} else {
old_w = old.get();
old_cov = old.getCovariance();
}
float cov_x = old_cov * x;
float new_w = old_w + coeff * cov_x * beta;
float new_cov = old_cov - (beta * cov_x * cov_x);
return new WeightValueWithCovar(new_w, new_cov);
}
// @formatter:off
@Description(name = "train_arowe_regr",
value = "_FUNC_(array features, float target [, constant string options])"
+ " - a refined version of AROW (Adaptive Reguralization of Weight Vectors) regressor "
+ "that usages epsilon-insensitive hinge loss `|w^t - y| - epsilon` for the loss function",
extended = "SELECT \n" +
" feature,\n" +
" argmin_kld(weight, covar) as weight\n" +
"FROM (\n" +
" SELECT \n" +
" train_arowe_regr(features,label) as (feature,weight,covar)\n" +
" FROM \n" +
" training_data\n" +
" ) t \n" +
"GROUP BY feature")
// @formatter:on
public static class AROWe extends AROWRegressionUDTF {
/** Sensitivity to prediction mistakes */
protected float epsilon;
@Override
protected Options getOptions() {
Options opts = super.getOptions();
opts.addOption("e", "epsilon", true,
"Sensitivity to prediction mistakes [default 0.1]");
return opts;
}
@Override
protected CommandLine processOptions(ObjectInspector[] argOIs) throws UDFArgumentException {
CommandLine cl = super.processOptions(argOIs);
float epsilon = 0.1f;
if (cl != null) {
String opt_epsilon = cl.getOptionValue("epsilon");
if (opt_epsilon != null) {
epsilon = Float.parseFloat(opt_epsilon);
}
}
this.epsilon = epsilon;
return cl;
}
@Override
protected void train(@Nonnull final FeatureValue[] features, float target) {
preTrain(target);
PredictionResult margin = calcScoreAndVariance(features);
float predicted = margin.getScore();
float loss = loss(target, predicted);
if (loss > 0.f) {
float coeff = (target - predicted) > 0.f ? loss : -loss;
float var = margin.getVariance();
float beta = 1.f / (var + r);
update(features, coeff, beta);
}
}
protected void preTrain(float target) {}
/**
* |w^t - y| - epsilon
*/
protected float loss(float target, float predicted) {
return LossFunctions.epsilonInsensitiveLoss(predicted, target, epsilon);
}
}
// @formatter:off
@Description(name = "train_arowe2_regr",
value = "_FUNC_(array features, float target [, constant string options])"
+ " - a refined version of AROW (Adaptive Reguralization of Weight Vectors) regressor "
+ "that usages adaptive epsilon-insensitive hinge loss `|w^t - y| - epsilon * stddev` for the loss function",
extended = "SELECT \n" +
" feature,\n" +
" argmin_kld(weight, covar) as weight\n" +
"FROM (\n" +
" SELECT \n" +
" train_arowe2_regr(features,label) as (feature,weight,covar)\n" +
" FROM \n" +
" training_data\n" +
" ) t \n" +
"GROUP BY feature")
// @formatter:on
public static class AROWe2 extends AROWe {
private OnlineVariance targetStdDev;
@Override
public StructObjectInspector initialize(ObjectInspector[] argOIs)
throws UDFArgumentException {
this.targetStdDev = new OnlineVariance();
return super.initialize(argOIs);
}
@Override
protected void preTrain(float target) {
targetStdDev.handle(target);
}
@Override
protected float loss(float target, float predicted) {
float stddev = (float) targetStdDev.stddev();
float e = epsilon * stddev;
return LossFunctions.epsilonInsensitiveLoss(predicted, target, e);
}
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy