source: trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GradientBoostedTrees/GradientBoostedTreesAlgorithmStatic.cs @ 14826

Last change on this file since 14826 was 14826, checked in by gkronber, 6 months ago

#2650: merged the factors branch into trunk

File size: 8.4 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2016 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 readonly uint randSeed;
55      private MersenneTwister random { get; set; }
56
57      // array members (allocate only once)
58      internal double[] pred;
59      internal double[] predTest;
60      internal double[] y;
61      internal int[] activeIdx;
62      internal double[] pseudoRes;
63
64      private readonly IList<IRegressionModel> models;
65      private readonly IList<double> weights;
66
67      public GbmState(IRegressionProblemData problemData, ILossFunction lossFunction, uint randSeed, int maxSize, double r, double m, double nu) {
68        // default settings for MaxSize, Nu and R
69        this.maxSize = maxSize;
70        this.nu = nu;
71        this.r = r;
72        this.m = m;
73
74        this.randSeed = randSeed;
75        random = new MersenneTwister(randSeed);
76        this.problemData = problemData;
77        this.trainingRows = problemData.TrainingIndices.ToArray();
78        this.testRows = problemData.TestIndices.ToArray();
79        this.lossFunction = lossFunction;
80
81        int nRows = trainingRows.Length;
82
83        y = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, trainingRows).ToArray();
84
85        treeBuilder = new RegressionTreeBuilder(problemData, random);
86
87        activeIdx = Enumerable.Range(0, nRows).ToArray();
88
89        var zeros = Enumerable.Repeat(0.0, nRows).ToArray();
90        double f0 = lossFunction.LineSearch(y, zeros, activeIdx, 0, nRows - 1); // initial constant value (mean for squared errors)
91        pred = Enumerable.Repeat(f0, nRows).ToArray();
92        predTest = Enumerable.Repeat(f0, testRows.Length).ToArray();
93        pseudoRes = new double[nRows];
94
95        models = new List<IRegressionModel>();
96        weights = new List<double>();
97        // add constant model
98        models.Add(new ConstantModel(f0, problemData.TargetVariable));
99        weights.Add(1.0);
100      }
101
102      public IRegressionModel GetModel() {
103#pragma warning disable 618
104        var model = new GradientBoostedTreesModel(models, weights);
105#pragma warning restore 618
106        // we don't know the number of iterations here but the number of weights is equal
107        // to the number of iterations + 1 (for the constant model)
108        // wrap the actual model in a surrogate that enables persistence and lazy recalculation of the model if necessary
109        return new GradientBoostedTreesModelSurrogate(problemData, randSeed, lossFunction, weights.Count - 1, maxSize, r, m, nu, model);
110      }
111      public IEnumerable<KeyValuePair<string, double>> GetVariableRelevance() {
112        return treeBuilder.GetVariableRelevance();
113      }
114
115      public double GetTrainLoss() {
116        int nRows = y.Length;
117        return lossFunction.GetLoss(y, pred) / nRows;
118      }
119      public double GetTestLoss() {
120        var yTest = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, testRows);
121        var nRows = testRows.Length;
122        return lossFunction.GetLoss(yTest, predTest) / nRows;
123      }
124
125      internal void AddModel(IRegressionModel m, double weight) {
126        models.Add(m);
127        weights.Add(weight);
128      }
129    }
130
131    // simple interface
132    public static GradientBoostedTreesSolution TrainGbm(IRegressionProblemData problemData, ILossFunction lossFunction, int maxSize, double nu, double r, double m, int maxIterations, uint randSeed = 31415) {
133      Contract.Assert(r > 0);
134      Contract.Assert(r <= 1.0);
135      Contract.Assert(nu > 0);
136      Contract.Assert(nu <= 1.0);
137
138      var state = (GbmState)CreateGbmState(problemData, lossFunction, randSeed, maxSize, r, m, nu);
139
140      for (int iter = 0; iter < maxIterations; iter++) {
141        MakeStep(state);
142      }
143
144      var model = state.GetModel();
145      return new GradientBoostedTreesSolution(model, (IRegressionProblemData)problemData.Clone());
146    }
147
148    // for custom stepping & termination
149    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) {
150      // check input variables. Only double variables are allowed.
151      var invalidInputs =
152        problemData.AllowedInputVariables.Where(name => !problemData.Dataset.VariableHasType<double>(name));
153      if (invalidInputs.Any())
154        throw new NotSupportedException("Gradient tree boosting only supports real-valued variables. Unsupported inputs: " + string.Join(", ", invalidInputs));
155
156      return new GbmState(problemData, lossFunction, randSeed, maxSize, r, m, nu);
157    }
158
159    // use default settings for maxSize, nu, r from state
160    public static void MakeStep(IGbmState state) {
161      var gbmState = state as GbmState;
162      if (gbmState == null) throw new ArgumentException("state");
163
164      MakeStep(gbmState, gbmState.maxSize, gbmState.nu, gbmState.r, gbmState.m);
165    }
166
167    // allow dynamic adaptation of maxSize, nu and r (even though this is not used)
168    public static void MakeStep(IGbmState state, int maxSize, double nu, double r, double m) {
169      var gbmState = state as GbmState;
170      if (gbmState == null) throw new ArgumentException("state");
171
172      var problemData = gbmState.problemData;
173      var lossFunction = gbmState.lossFunction;
174      var yPred = gbmState.pred;
175      var yPredTest = gbmState.predTest;
176      var treeBuilder = gbmState.treeBuilder;
177      var y = gbmState.y;
178      var activeIdx = gbmState.activeIdx;
179      var pseudoRes = gbmState.pseudoRes;
180      var trainingRows = gbmState.trainingRows;
181      var testRows = gbmState.testRows;
182
183      // copy output of gradient function to pre-allocated rim array (pseudo-residual per row and model)
184      int rimIdx = 0;
185      foreach (var g in lossFunction.GetLossGradient(y, yPred)) {
186        pseudoRes[rimIdx++] = g;
187      }
188
189      var tree = treeBuilder.CreateRegressionTreeForGradientBoosting(pseudoRes, yPred, maxSize, activeIdx, lossFunction, r, m);
190
191      int i = 0;
192      foreach (var pred in tree.GetEstimatedValues(problemData.Dataset, trainingRows)) {
193        yPred[i] = yPred[i] + nu * pred;
194        i++;
195      }
196      // update predictions for validation set
197      i = 0;
198      foreach (var pred in tree.GetEstimatedValues(problemData.Dataset, testRows)) {
199        yPredTest[i] = yPredTest[i] + nu * pred;
200        i++;
201      }
202
203      gbmState.AddModel(tree, nu);
204    }
205    #endregion
206  }
207}
Note: See TracBrowser for help on using the repository browser.