001/*
002 * Copyright (c) 2021, Oracle and/or its affiliates. All rights reserved.
003 *
004 * Licensed under the Apache License, Version 2.0 (the "License");
005 * you may not use this file except in compliance with the License.
006 * You may obtain a copy of the License at
007 *
008 *     http://www.apache.org/licenses/LICENSE-2.0
009 *
010 * Unless required by applicable law or agreed to in writing, software
011 * distributed under the License is distributed on an "AS IS" BASIS,
012 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express implied.
013 * See the License for the specific language governing permissions and
014 * limitations under the License.
015 */
016
017package org.tribuo.classification.sgd.fm;
018
019import com.oracle.labs.mlrg.olcut.config.Config;
020import org.tribuo.ImmutableFeatureMap;
021import org.tribuo.ImmutableOutputInfo;
022import org.tribuo.classification.Label;
023import org.tribuo.classification.sgd.LabelObjective;
024import org.tribuo.classification.sgd.objectives.LogMulticlass;
025import org.tribuo.common.sgd.AbstractFMTrainer;
026import org.tribuo.common.sgd.FMParameters;
027import org.tribuo.common.sgd.SGDObjective;
028import org.tribuo.math.StochasticGradientOptimiser;
029import org.tribuo.provenance.ModelProvenance;
030
031import java.util.logging.Logger;
032
033/**
034 * A trainer for a classification factorization machine using SGD.
035 * <p>
036 * See:
037 * <pre>
038 * Rendle, S.
039 * Factorization machines.
040 * 2010 IEEE International Conference on Data Mining
041 * </pre>
042 */
043public class FMClassificationTrainer extends AbstractFMTrainer<Label, Integer, FMClassificationModel> {
044    private static final Logger logger = Logger.getLogger(FMClassificationTrainer.class.getName());
045
046    @Config(description = "The classification objective function to use.")
047    private LabelObjective objective = new LogMulticlass();
048
049    /**
050     * Constructs an SGD trainer for a factorization machine.
051     *
052     * @param objective         The objective function to optimise.
053     * @param optimiser         The gradient optimiser to use.
054     * @param epochs            The number of epochs (complete passes through the training data).
055     * @param loggingInterval   Log the loss after this many iterations. If -1 don't log anything.
056     * @param minibatchSize     The size of any minibatches.
057     * @param seed              A seed for the random number generator, used to shuffle the examples before each epoch.
058     * @param factorizedDimSize Size of the factorized feature representation.
059     * @param variance          The variance of the initializer.
060     */
061    public FMClassificationTrainer(LabelObjective objective, StochasticGradientOptimiser optimiser, int epochs,
062                                   int loggingInterval, int minibatchSize, long seed,
063                                   int factorizedDimSize, double variance) {
064        super(optimiser, epochs, loggingInterval, minibatchSize, seed, factorizedDimSize, variance);
065        this.objective = objective;
066    }
067
068    /**
069     * Constructs an SGD trainer for a factorization machine.
070     * <p>
071     * Sets the minibatch size to 1.
072     *
073     * @param objective         The objective function to optimise.
074     * @param optimiser         The gradient optimiser to use.
075     * @param epochs            The number of epochs (complete passes through the training data).
076     * @param loggingInterval   Log the loss after this many iterations. If -1 don't log anything.
077     * @param seed              A seed for the random number generator, used to shuffle the examples before each epoch.
078     * @param factorizedDimSize Size of the factorized feature representation.
079     * @param variance          The variance of the initializer.
080     */
081    public FMClassificationTrainer(LabelObjective objective, StochasticGradientOptimiser optimiser, int epochs,
082                                   int loggingInterval, long seed,
083                                   int factorizedDimSize, double variance) {
084        this(objective, optimiser, epochs, loggingInterval, 1, seed, factorizedDimSize, variance);
085    }
086
087    /**
088     * Constructs an SGD trainer for a factorization machine.
089     * <p>
090     * Sets the minibatch size to 1 and the logging interval to 1000.
091     *
092     * @param objective         The objective function to optimise.
093     * @param optimiser         The gradient optimiser to use.
094     * @param epochs            The number of epochs (complete passes through the training data).
095     * @param seed              A seed for the random number generator, used to shuffle the examples before each epoch.
096     * @param factorizedDimSize Size of the factorized feature representation.
097     * @param variance          The variance of the initializer.
098     */
099    public FMClassificationTrainer(LabelObjective objective, StochasticGradientOptimiser optimiser, int epochs,
100                                   long seed, int factorizedDimSize, double variance) {
101        this(objective, optimiser, epochs, 1000, 1, seed, factorizedDimSize, variance);
102    }
103
104    /**
105     * For olcut.
106     */
107    private FMClassificationTrainer() {
108        super();
109    }
110
111    @Override
112    protected Integer getTarget(ImmutableOutputInfo<Label> outputInfo, Label output) {
113        return outputInfo.getID(output);
114    }
115
116    @Override
117    protected SGDObjective<Integer> getObjective() {
118        return objective;
119    }
120
121    @Override
122    protected FMClassificationModel createModel(String name, ModelProvenance provenance, ImmutableFeatureMap featureMap, ImmutableOutputInfo<Label> outputInfo, FMParameters parameters) {
123        return new FMClassificationModel(name, provenance, featureMap, outputInfo, parameters, objective.getNormalizer(), objective.isProbabilistic());
124    }
125
126    @Override
127    protected String getModelClassName() {
128        return FMClassificationModel.class.getName();
129    }
130
131    @Override
132    public String toString() {
133        return "FMClassificationTrainer(objective=" + objective.toString() + ",optimiser=" + optimiser.toString() +
134                ",epochs=" + epochs + ",minibatchSize=" + minibatchSize + ",seed=" + seed +
135                ",factorizedDimSize=" + factorizedDimSize + ",variance=" + variance +
136                ")";
137    }
138}