source: trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GradientBoostedTrees/GradientBoostedTreesAlgorithmStatic.cs @ 12710

Last change on this file since 12710 was 12710, checked in by gkronber, 5 years ago

#2261: cached training and test rows in GBT for another speedup of ~1.5 (+renamed test class)

File size: 7.4 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2015 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
4 * and the BEACON Center for the Study of Evolution in Action.
5 *
6 * This file is part of HeuristicLab.
7 *
8 * HeuristicLab is free software: you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation, either version 3 of the License, or
11 * (at your option) any later version.
12 *
13 * HeuristicLab is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
20 */
21#endregion
22
23using System;
24using System.Collections.Generic;
25using System.Diagnostics.Contracts;
26using System.Linq;
27using HeuristicLab.Problems.DataAnalysis;
28using HeuristicLab.Random;
29
30namespace HeuristicLab.Algorithms.DataAnalysis {
31  public static class GradientBoostedTreesAlgorithmStatic {
32    #region static API
33
34    public interface IGbmState {
35      IRegressionModel GetModel();
36      double GetTrainLoss();
37      double GetTestLoss();
38      IEnumerable<KeyValuePair<string, double>> GetVariableRelevance();
39    }
40
41    // created through factory method
42    // GbmState details are private API users can only use methods from IGbmState
43    private class GbmState : IGbmState {
44      internal IRegressionProblemData problemData { get; private set; }
45      internal ILossFunction lossFunction { get; private set; }
46      internal int maxSize { get; private set; }
47      internal double nu { get; private set; }
48      internal double r { get; private set; }
49      internal double m { get; private set; }
50      internal int[] trainingRows { get; private set; }
51      internal int[] testRows { get; private set; }
52      internal RegressionTreeBuilder treeBuilder { get; private set; }
53
54      private MersenneTwister random { get; set; }
55
56      // array members (allocate only once)
57      internal double[] pred;
58      internal double[] predTest;
59      internal double[] y;
60      internal int[] activeIdx;
61      internal double[] pseudoRes;
62
63      private readonly IList<IRegressionModel> models;
64      private readonly IList<double> weights;
65
66      public GbmState(IRegressionProblemData problemData, ILossFunction lossFunction, uint randSeed, int maxSize, double r, double m, double nu) {
67        // default settings for MaxSize, Nu and R
68        this.maxSize = maxSize;
69        this.nu = nu;
70        this.r = r;
71        this.m = m;
72
73        random = new MersenneTwister(randSeed);
74        this.problemData = problemData;
75        this.trainingRows = problemData.TrainingIndices.ToArray();
76        this.testRows = problemData.TestIndices.ToArray();
77        this.lossFunction = lossFunction;
78
79        int nRows = trainingRows.Length;
80
81        y = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, trainingRows).ToArray();
82
83        treeBuilder = new RegressionTreeBuilder(problemData, random);
84
85        activeIdx = Enumerable.Range(0, nRows).ToArray();
86
87        var zeros = Enumerable.Repeat(0.0, nRows).ToArray();
88        double f0 = lossFunction.LineSearch(y, zeros, activeIdx, 0, nRows - 1); // initial constant value (mean for squared errors)
89        pred = Enumerable.Repeat(f0, nRows).ToArray();
90        predTest = Enumerable.Repeat(f0, testRows.Length).ToArray();
91        pseudoRes = new double[nRows];
92
93        models = new List<IRegressionModel>();
94        weights = new List<double>();
95        // add constant model
96        models.Add(new ConstantRegressionModel(f0));
97        weights.Add(1.0);
98      }
99
100      public IRegressionModel GetModel() {
101        return new GradientBoostedTreesModel(models, weights);
102      }
103      public IEnumerable<KeyValuePair<string, double>> GetVariableRelevance() {
104        return treeBuilder.GetVariableRelevance();
105      }
106
107      public double GetTrainLoss() {
108        int nRows = y.Length;
109        return lossFunction.GetLoss(y, pred) / nRows;
110      }
111      public double GetTestLoss() {
112        var yTest = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, testRows);
113        var nRows = testRows.Length;
114        return lossFunction.GetLoss(yTest, predTest) / nRows;
115      }
116
117      internal void AddModel(IRegressionModel m, double weight) {
118        models.Add(m);
119        weights.Add(weight);
120      }
121    }
122
123    // simple interface
124    public static IRegressionSolution TrainGbm(IRegressionProblemData problemData, ILossFunction lossFunction, int maxSize, double nu, double r, double m, int maxIterations, uint randSeed = 31415) {
125      Contract.Assert(r > 0);
126      Contract.Assert(r <= 1.0);
127      Contract.Assert(nu > 0);
128      Contract.Assert(nu <= 1.0);
129
130      var state = (GbmState)CreateGbmState(problemData, lossFunction, randSeed, maxSize, r, m, nu);
131
132      for (int iter = 0; iter < maxIterations; iter++) {
133        MakeStep(state);
134      }
135
136      var model = state.GetModel();
137      return new RegressionSolution(model, (IRegressionProblemData)problemData.Clone());
138    }
139
140    // for custom stepping & termination
141    public static IGbmState CreateGbmState(IRegressionProblemData problemData, ILossFunction lossFunction, uint randSeed, int maxSize = 3, double r = 0.66, double m = 0.5, double nu = 0.01) {
142      return new GbmState(problemData, lossFunction, randSeed, maxSize, r, m, nu);
143    }
144
145    // use default settings for maxSize, nu, r from state
146    public static void MakeStep(IGbmState state) {
147      var gbmState = state as GbmState;
148      if (gbmState == null) throw new ArgumentException("state");
149
150      MakeStep(gbmState, gbmState.maxSize, gbmState.nu, gbmState.r, gbmState.m);
151    }
152
153    // allow dynamic adaptation of maxSize, nu and r (even though this is not used)
154    public static void MakeStep(IGbmState state, int maxSize, double nu, double r, double m) {
155      var gbmState = state as GbmState;
156      if (gbmState == null) throw new ArgumentException("state");
157
158      var problemData = gbmState.problemData;
159      var lossFunction = gbmState.lossFunction;
160      var yPred = gbmState.pred;
161      var yPredTest = gbmState.predTest;
162      var treeBuilder = gbmState.treeBuilder;
163      var y = gbmState.y;
164      var activeIdx = gbmState.activeIdx;
165      var pseudoRes = gbmState.pseudoRes;
166      var trainingRows = gbmState.trainingRows;
167      var testRows = gbmState.testRows;
168
169      // copy output of gradient function to pre-allocated rim array (pseudo-residual per row and model)
170      int rimIdx = 0;
171      foreach (var g in lossFunction.GetLossGradient(y, yPred)) {
172        pseudoRes[rimIdx++] = g;
173      }
174
175      var tree = treeBuilder.CreateRegressionTreeForGradientBoosting(pseudoRes, yPred, maxSize, activeIdx, lossFunction, r, m);
176
177      int i = 0;
178      foreach (var pred in tree.GetEstimatedValues(problemData.Dataset, trainingRows)) {
179        yPred[i] = yPred[i] + nu * pred;
180        i++;
181      }
182      // update predictions for validation set
183      i = 0;
184      foreach (var pred in tree.GetEstimatedValues(problemData.Dataset, testRows)) {
185        yPredTest[i] = yPredTest[i] + nu * pred;
186        i++;
187      }
188
189      gbmState.AddModel(tree, nu);
190    }
191    #endregion
192  }
193}
Note: See TracBrowser for help on using the repository browser.