source: branches/2883_GBTModelStorage/HeuristicLab.Algorithms.DataAnalysis/3.4/GradientBoostedTrees/GradientBoostedTreesAlgorithmStatic.cs @ 15678

Last change on this file since 15678 was 15678, checked in by fholzing, 3 years ago

#2883: Implemented third option for complete storage

File size: 8.5 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2018 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(bool shallow = true);
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(bool shallow = true) {
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        if (shallow) {
110          return new GradientBoostedTreesModelSurrogate(problemData, randSeed, lossFunction, weights.Count - 1, maxSize, r, m, nu, model);
111        } else {
112          return model;
113        }
114      }
115      public IEnumerable<KeyValuePair<string, double>> GetVariableRelevance() {
116        return treeBuilder.GetVariableRelevance();
117      }
118
119      public double GetTrainLoss() {
120        int nRows = y.Length;
121        return lossFunction.GetLoss(y, pred) / nRows;
122      }
123      public double GetTestLoss() {
124        var yTest = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, testRows);
125        var nRows = testRows.Length;
126        return lossFunction.GetLoss(yTest, predTest) / nRows;
127      }
128
129      internal void AddModel(IRegressionModel m, double weight) {
130        models.Add(m);
131        weights.Add(weight);
132      }
133    }
134
135    // simple interface
136    public static GradientBoostedTreesSolution TrainGbm(IRegressionProblemData problemData, ILossFunction lossFunction, int maxSize, double nu, double r, double m, int maxIterations, uint randSeed = 31415) {
137      Contract.Assert(r > 0);
138      Contract.Assert(r <= 1.0);
139      Contract.Assert(nu > 0);
140      Contract.Assert(nu <= 1.0);
141
142      var state = (GbmState)CreateGbmState(problemData, lossFunction, randSeed, maxSize, r, m, nu);
143
144      for (int iter = 0; iter < maxIterations; iter++) {
145        MakeStep(state);
146      }
147
148      var model = state.GetModel();
149      return new GradientBoostedTreesSolution(model, (IRegressionProblemData)problemData.Clone());
150    }
151
152    // for custom stepping & termination
153    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) {
154      // check input variables. Only double variables are allowed.
155      var invalidInputs =
156        problemData.AllowedInputVariables.Where(name => !problemData.Dataset.VariableHasType<double>(name));
157      if (invalidInputs.Any())
158        throw new NotSupportedException("Gradient tree boosting only supports real-valued variables. Unsupported inputs: " + string.Join(", ", invalidInputs));
159
160      return new GbmState(problemData, lossFunction, randSeed, maxSize, r, m, nu);
161    }
162
163    // use default settings for maxSize, nu, r from state
164    public static void MakeStep(IGbmState state) {
165      var gbmState = state as GbmState;
166      if (gbmState == null) throw new ArgumentException("state");
167
168      MakeStep(gbmState, gbmState.maxSize, gbmState.nu, gbmState.r, gbmState.m);
169    }
170
171    // allow dynamic adaptation of maxSize, nu and r (even though this is not used)
172    public static void MakeStep(IGbmState state, int maxSize, double nu, double r, double m) {
173      var gbmState = state as GbmState;
174      if (gbmState == null) throw new ArgumentException("state");
175
176      var problemData = gbmState.problemData;
177      var lossFunction = gbmState.lossFunction;
178      var yPred = gbmState.pred;
179      var yPredTest = gbmState.predTest;
180      var treeBuilder = gbmState.treeBuilder;
181      var y = gbmState.y;
182      var activeIdx = gbmState.activeIdx;
183      var pseudoRes = gbmState.pseudoRes;
184      var trainingRows = gbmState.trainingRows;
185      var testRows = gbmState.testRows;
186
187      // copy output of gradient function to pre-allocated rim array (pseudo-residual per row and model)
188      int rimIdx = 0;
189      foreach (var g in lossFunction.GetLossGradient(y, yPred)) {
190        pseudoRes[rimIdx++] = g;
191      }
192
193      var tree = treeBuilder.CreateRegressionTreeForGradientBoosting(pseudoRes, yPred, maxSize, activeIdx, lossFunction, r, m);
194
195      int i = 0;
196      foreach (var pred in tree.GetEstimatedValues(problemData.Dataset, trainingRows)) {
197        yPred[i] = yPred[i] + nu * pred;
198        i++;
199      }
200      // update predictions for validation set
201      i = 0;
202      foreach (var pred in tree.GetEstimatedValues(problemData.Dataset, testRows)) {
203        yPredTest[i] = yPredTest[i] + nu * pred;
204        i++;
205      }
206
207      gbmState.AddModel(tree, nu);
208    }
209    #endregion
210  }
211}
Note: See TracBrowser for help on using the repository browser.