1 | #region License Information
|
---|
2 | /* HeuristicLab
|
---|
3 | * Copyright (C) Heuristic and Evolutionary Algorithms Laboratory (HEAL)
|
---|
4 | *
|
---|
5 | * This file is part of HeuristicLab.
|
---|
6 | *
|
---|
7 | * HeuristicLab is free software: you can redistribute it and/or modify
|
---|
8 | * it under the terms of the GNU General Public License as published by
|
---|
9 | * the Free Software Foundation, either version 3 of the License, or
|
---|
10 | * (at your option) any later version.
|
---|
11 | *
|
---|
12 | * HeuristicLab is distributed in the hope that it will be useful,
|
---|
13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
15 | * GNU General Public License for more details.
|
---|
16 | *
|
---|
17 | * You should have received a copy of the GNU General Public License
|
---|
18 | * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
|
---|
19 | */
|
---|
20 | #endregion
|
---|
21 |
|
---|
22 | using System;
|
---|
23 | using System.Collections.Generic;
|
---|
24 | using System.Linq;
|
---|
25 | using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
|
---|
26 | using Tensorflow;
|
---|
27 | using static Tensorflow.Binding;
|
---|
28 | using DoubleVector = MathNet.Numerics.LinearAlgebra.Vector<double>;
|
---|
29 |
|
---|
30 | namespace HeuristicLab.Problems.DataAnalysis.Symbolic {
|
---|
31 | public class TreeToTensorConverter {
|
---|
32 |
|
---|
33 | #region helper class
|
---|
34 | public class DataForVariable {
|
---|
35 | public readonly string variableName;
|
---|
36 | public readonly string variableValue; // for factor vars
|
---|
37 |
|
---|
38 | public DataForVariable(string varName, string varValue) {
|
---|
39 | this.variableName = varName;
|
---|
40 | this.variableValue = varValue;
|
---|
41 | }
|
---|
42 |
|
---|
43 | public override bool Equals(object obj) {
|
---|
44 | var other = obj as DataForVariable;
|
---|
45 | if (other == null) return false;
|
---|
46 | return other.variableName.Equals(this.variableName) &&
|
---|
47 | other.variableValue.Equals(this.variableValue);
|
---|
48 | }
|
---|
49 |
|
---|
50 | public override int GetHashCode() {
|
---|
51 | return variableName.GetHashCode() ^ variableValue.GetHashCode();
|
---|
52 | }
|
---|
53 | }
|
---|
54 | #endregion
|
---|
55 |
|
---|
56 | public static bool TryConvert(ISymbolicExpressionTree tree, int numRows, int? vectorLength,
|
---|
57 | bool makeVariableWeightsVariable, bool addLinearScalingTerms,
|
---|
58 | out Tensor graph, out Dictionary<Tensor, string> parameters, out List<Tensor> variables
|
---|
59 | /*, out double[] initialConstants*/) {
|
---|
60 |
|
---|
61 | try {
|
---|
62 | var converter = new TreeToTensorConverter(numRows, vectorLength, makeVariableWeightsVariable, addLinearScalingTerms);
|
---|
63 | graph = converter.ConvertNode(tree.Root.GetSubtree(0));
|
---|
64 |
|
---|
65 | //var parametersEntries = converter.parameters.ToList(); // guarantee same order for keys and values
|
---|
66 | parameters = converter.parameters; // parametersEntries.Select(kvp => kvp.Value).ToList();
|
---|
67 | variables = converter.variables;
|
---|
68 | //initialConstants = converter.initialConstants.ToArray();
|
---|
69 | return true;
|
---|
70 | } catch (NotSupportedException) {
|
---|
71 | graph = null;
|
---|
72 | parameters = null;
|
---|
73 | variables = null;
|
---|
74 | //initialConstants = null;
|
---|
75 | return false;
|
---|
76 | }
|
---|
77 | }
|
---|
78 |
|
---|
79 | private readonly int numRows;
|
---|
80 | private readonly int? vectorLength;
|
---|
81 | private readonly bool makeVariableWeightsVariable;
|
---|
82 | private readonly bool addLinearScalingTerms;
|
---|
83 |
|
---|
84 | //private readonly List<double> initialConstants = new List<double>();
|
---|
85 | private readonly Dictionary<Tensor, string> parameters = new Dictionary<Tensor, string>();
|
---|
86 | private readonly List<Tensor> variables = new List<Tensor>();
|
---|
87 |
|
---|
88 | private TreeToTensorConverter(int numRows, int? vectorLength, bool makeVariableWeightsVariable, bool addLinearScalingTerms) {
|
---|
89 | this.numRows = numRows;
|
---|
90 | this.vectorLength = vectorLength;
|
---|
91 | this.makeVariableWeightsVariable = makeVariableWeightsVariable;
|
---|
92 | this.addLinearScalingTerms = addLinearScalingTerms;
|
---|
93 | }
|
---|
94 |
|
---|
95 |
|
---|
96 | private Tensor ConvertNode(ISymbolicExpressionTreeNode node) {
|
---|
97 | if (node.Symbol is Constant) {
|
---|
98 | var value = ((ConstantTreeNode)node).Value;
|
---|
99 | //initialConstants.Add(value);
|
---|
100 | var var = tf.Variable(value, name: $"c_{variables.Count}", dtype: tf.float64);
|
---|
101 | variables.Add(var);
|
---|
102 | return var;
|
---|
103 | }
|
---|
104 |
|
---|
105 | if (node.Symbol is Variable/* || node.Symbol is BinaryFactorVariable*/) {
|
---|
106 | var varNode = node as VariableTreeNodeBase;
|
---|
107 | //var factorVarNode = node as BinaryFactorVariableTreeNode;
|
---|
108 | // factor variable values are only 0 or 1 and set in x accordingly
|
---|
109 | //var varValue = factorVarNode != null ? factorVarNode.VariableValue : string.Empty;
|
---|
110 | //var par = FindOrCreateParameter(parameters, varNode.VariableName, varValue);
|
---|
111 | var shape = vectorLength.HasValue
|
---|
112 | ? new TensorShape(numRows, vectorLength.Value)
|
---|
113 | : new TensorShape(numRows);
|
---|
114 | var par = tf.placeholder(tf.float64, shape: shape, name: varNode.VariableName);
|
---|
115 | parameters.Add(par, varNode.VariableName);
|
---|
116 |
|
---|
117 | if (makeVariableWeightsVariable) {
|
---|
118 | //initialConstants.Add(varNode.Weight);
|
---|
119 | var w = tf.Variable(varNode.Weight, name: $"w_{varNode.VariableName}_{variables.Count}", dtype: tf.float64);
|
---|
120 | variables.Add(w);
|
---|
121 | return w * par;
|
---|
122 | } else {
|
---|
123 | return varNode.Weight * par;
|
---|
124 | }
|
---|
125 | }
|
---|
126 |
|
---|
127 | if (node.Symbol is FactorVariable) {
|
---|
128 | var factorVarNode = node as FactorVariableTreeNode;
|
---|
129 | var products = new List<Tensor>();
|
---|
130 | foreach (var variableValue in factorVarNode.Symbol.GetVariableValues(factorVarNode.VariableName)) {
|
---|
131 | //var par = FindOrCreateParameter(parameters, factorVarNode.VariableName, variableValue);
|
---|
132 | var par = tf.placeholder(tf.float64, shape: new TensorShape(numRows), name: factorVarNode.VariableName);
|
---|
133 | parameters.Add(par, factorVarNode.VariableName);
|
---|
134 |
|
---|
135 | var value = factorVarNode.GetValue(variableValue);
|
---|
136 | //initialConstants.Add(value);
|
---|
137 | var wVar = tf.Variable(value, name: $"f_{factorVarNode.VariableName}_{variables.Count}");
|
---|
138 | variables.Add(wVar);
|
---|
139 |
|
---|
140 | products.add(wVar * par);
|
---|
141 | }
|
---|
142 |
|
---|
143 | return products.Aggregate((a, b) => a + b);
|
---|
144 | }
|
---|
145 |
|
---|
146 | if (node.Symbol is Addition) {
|
---|
147 | var terms = new List<Tensor>();
|
---|
148 | foreach (var subTree in node.Subtrees) {
|
---|
149 | terms.Add(ConvertNode(subTree));
|
---|
150 | }
|
---|
151 |
|
---|
152 | return terms.Aggregate((a, b) => a + b);
|
---|
153 | }
|
---|
154 |
|
---|
155 | if (node.Symbol is Subtraction) {
|
---|
156 | var terms = new List<Tensor>();
|
---|
157 | for (int i = 0; i < node.SubtreeCount; i++) {
|
---|
158 | var t = ConvertNode(node.GetSubtree(i));
|
---|
159 | if (i > 0) t = -t;
|
---|
160 | terms.Add(t);
|
---|
161 | }
|
---|
162 |
|
---|
163 | if (terms.Count == 1) return -terms[0];
|
---|
164 | else return terms.Aggregate((a, b) => a + b);
|
---|
165 | }
|
---|
166 |
|
---|
167 | if (node.Symbol is Multiplication) {
|
---|
168 | var terms = new List<Tensor>();
|
---|
169 | foreach (var subTree in node.Subtrees) {
|
---|
170 | terms.Add(ConvertNode(subTree));
|
---|
171 | }
|
---|
172 |
|
---|
173 | if (terms.Count == 1) return terms[0];
|
---|
174 | else return terms.Aggregate((a, b) => a * b);
|
---|
175 | }
|
---|
176 |
|
---|
177 | if (node.Symbol is Division) {
|
---|
178 | var terms = new List<Tensor>();
|
---|
179 | foreach (var subTree in node.Subtrees) {
|
---|
180 | terms.Add(ConvertNode(subTree));
|
---|
181 | }
|
---|
182 |
|
---|
183 | if (terms.Count == 1) return 1.0 / terms[0];
|
---|
184 | else return terms.Aggregate((a, b) => a * (1.0 / b));
|
---|
185 | }
|
---|
186 |
|
---|
187 | if (node.Symbol is Absolute) {
|
---|
188 | var x1 = ConvertNode(node.GetSubtree(0));
|
---|
189 | return tf.abs(x1);
|
---|
190 | }
|
---|
191 |
|
---|
192 | if (node.Symbol is AnalyticQuotient) {
|
---|
193 | var x1 = ConvertNode(node.GetSubtree(0));
|
---|
194 | var x2 = ConvertNode(node.GetSubtree(1));
|
---|
195 | return x1 / tf.pow(1 + x2 * x2, 0.5);
|
---|
196 | }
|
---|
197 |
|
---|
198 | if (node.Symbol is Logarithm) {
|
---|
199 | return math_ops.log(
|
---|
200 | ConvertNode(node.GetSubtree(0)));
|
---|
201 | }
|
---|
202 |
|
---|
203 | if (node.Symbol is Exponential) {
|
---|
204 | return math_ops.pow(
|
---|
205 | Math.E,
|
---|
206 | ConvertNode(node.GetSubtree(0)));
|
---|
207 | }
|
---|
208 |
|
---|
209 | if (node.Symbol is Square) {
|
---|
210 | return tf.square(
|
---|
211 | ConvertNode(node.GetSubtree(0)));
|
---|
212 | }
|
---|
213 |
|
---|
214 | if (node.Symbol is SquareRoot) {
|
---|
215 | return math_ops.sqrt(
|
---|
216 | ConvertNode(node.GetSubtree(0)));
|
---|
217 | }
|
---|
218 |
|
---|
219 | if (node.Symbol is Cube) {
|
---|
220 | return math_ops.pow(
|
---|
221 | ConvertNode(node.GetSubtree(0)), 3.0);
|
---|
222 | }
|
---|
223 |
|
---|
224 | if (node.Symbol is CubeRoot) {
|
---|
225 | return math_ops.pow(
|
---|
226 | ConvertNode(node.GetSubtree(0)), 1.0 / 3.0);
|
---|
227 | // TODO
|
---|
228 | // f: x < 0 ? -Math.Pow(-x, 1.0 / 3) : Math.Pow(x, 1.0 / 3),
|
---|
229 | // g: { var cbrt_x = x < 0 ? -Math.Pow(-x, 1.0 / 3) : Math.Pow(x, 1.0 / 3); return 1.0 / (3 * cbrt_x * cbrt_x); }
|
---|
230 | }
|
---|
231 |
|
---|
232 | if (node.Symbol is Sine) {
|
---|
233 | return tf.sin(
|
---|
234 | ConvertNode(node.GetSubtree(0)));
|
---|
235 | }
|
---|
236 |
|
---|
237 | if (node.Symbol is Cosine) {
|
---|
238 | return tf.cos(
|
---|
239 | ConvertNode(node.GetSubtree(0)));
|
---|
240 | }
|
---|
241 |
|
---|
242 | if (node.Symbol is Tangent) {
|
---|
243 | return tf.tan(
|
---|
244 | ConvertNode(node.GetSubtree(0)));
|
---|
245 | }
|
---|
246 |
|
---|
247 | if (node.Symbol is Mean) {
|
---|
248 | return tf.reduce_mean(
|
---|
249 | ConvertNode(node.GetSubtree(0)),
|
---|
250 | axis: new[] { 1 });
|
---|
251 | }
|
---|
252 |
|
---|
253 | //if (node.Symbol is StandardDeviation) {
|
---|
254 | // return tf.reduce_std(
|
---|
255 | // ConvertNode(node.GetSubtree(0)),
|
---|
256 | // axis: new [] { 1 }
|
---|
257 | // );
|
---|
258 | //}
|
---|
259 |
|
---|
260 | if (node.Symbol is Sum) {
|
---|
261 | return tf.reduce_sum(
|
---|
262 | ConvertNode(node.GetSubtree(0)),
|
---|
263 | axis: new[] { 1 });
|
---|
264 | }
|
---|
265 |
|
---|
266 | if (node.Symbol is StartSymbol) {
|
---|
267 | if (addLinearScalingTerms) {
|
---|
268 | // scaling variables α, β are given at the beginning of the parameter vector
|
---|
269 | var alpha = tf.Variable(1.0, name: $"alpha_{1.0}", dtype: tf.float64);
|
---|
270 | var beta = tf.Variable(0.0, name: $"beta_{0.0}", dtype: tf.float64);
|
---|
271 | variables.Add(alpha);
|
---|
272 | variables.Add(beta);
|
---|
273 | var t = ConvertNode(node.GetSubtree(0));
|
---|
274 | return t * alpha + beta;
|
---|
275 | } else return ConvertNode(node.GetSubtree(0));
|
---|
276 | }
|
---|
277 |
|
---|
278 | throw new NotSupportedException($"Node symbol {node.Symbol} is not supported.");
|
---|
279 | }
|
---|
280 |
|
---|
281 | // for each factor variable value we need a parameter which represents a binary indicator for that variable & value combination
|
---|
282 | // each binary indicator is only necessary once. So we only create a parameter if this combination is not yet available
|
---|
283 | private static Tensor FindOrCreateParameter(Dictionary<DataForVariable, Tensor> parameters, string varName, string varValue = "") {
|
---|
284 | var data = new DataForVariable(varName, varValue);
|
---|
285 |
|
---|
286 | if (!parameters.TryGetValue(data, out var par)) {
|
---|
287 | // not found -> create new parameter and entries in names and values lists
|
---|
288 | par = tf.placeholder(tf.float64, shape: new TensorShape(-1), name: varName);
|
---|
289 | parameters.Add(data, par);
|
---|
290 | }
|
---|
291 | return par;
|
---|
292 | }
|
---|
293 |
|
---|
294 | public static bool IsCompatible(ISymbolicExpressionTree tree) {
|
---|
295 | var containsUnknownSymbol = (
|
---|
296 | from n in tree.Root.GetSubtree(0).IterateNodesPrefix()
|
---|
297 | where
|
---|
298 | !(n.Symbol is Variable) &&
|
---|
299 | //!(n.Symbol is BinaryFactorVariable) &&
|
---|
300 | //!(n.Symbol is FactorVariable) &&
|
---|
301 | !(n.Symbol is Constant) &&
|
---|
302 | !(n.Symbol is Addition) &&
|
---|
303 | !(n.Symbol is Subtraction) &&
|
---|
304 | !(n.Symbol is Multiplication) &&
|
---|
305 | !(n.Symbol is Division) &&
|
---|
306 | !(n.Symbol is Logarithm) &&
|
---|
307 | !(n.Symbol is Exponential) &&
|
---|
308 | !(n.Symbol is SquareRoot) &&
|
---|
309 | !(n.Symbol is Square) &&
|
---|
310 | !(n.Symbol is Sine) &&
|
---|
311 | !(n.Symbol is Cosine) &&
|
---|
312 | !(n.Symbol is Tangent) &&
|
---|
313 | !(n.Symbol is HyperbolicTangent) &&
|
---|
314 | !(n.Symbol is Erf) &&
|
---|
315 | !(n.Symbol is Norm) &&
|
---|
316 | !(n.Symbol is StartSymbol) &&
|
---|
317 | !(n.Symbol is Absolute) &&
|
---|
318 | !(n.Symbol is AnalyticQuotient) &&
|
---|
319 | !(n.Symbol is Cube) &&
|
---|
320 | !(n.Symbol is CubeRoot) &&
|
---|
321 | !(n.Symbol is Mean) &&
|
---|
322 | //!(n.Symbol is StandardDeviation) &&
|
---|
323 | !(n.Symbol is Sum)
|
---|
324 | select n).Any();
|
---|
325 | return !containsUnknownSymbol;
|
---|
326 | }
|
---|
327 | }
|
---|
328 | }
|
---|