1 | #region License Information
|
---|
2 | /* HeuristicLab
|
---|
3 | * Copyright (C) 2002-2016 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 |
|
---|
21 | //Code is based on an implementation from Laurens van der Maaten
|
---|
22 |
|
---|
23 | /*
|
---|
24 | *
|
---|
25 | * Copyright (c) 2014, Laurens van der Maaten (Delft University of Technology)
|
---|
26 | * All rights reserved.
|
---|
27 | *
|
---|
28 | * Redistribution and use in source and binary forms, with or without
|
---|
29 | * modification, are permitted provided that the following conditions are met:
|
---|
30 | * 1. Redistributions of source code must retain the above copyright
|
---|
31 | * notice, this list of conditions and the following disclaimer.
|
---|
32 | * 2. Redistributions in binary form must reproduce the above copyright
|
---|
33 | * notice, this list of conditions and the following disclaimer in the
|
---|
34 | * documentation and/or other materials provided with the distribution.
|
---|
35 | * 3. All advertising materials mentioning features or use of this software
|
---|
36 | * must display the following acknowledgement:
|
---|
37 | * This product includes software developed by the Delft University of Technology.
|
---|
38 | * 4. Neither the name of the Delft University of Technology nor the names of
|
---|
39 | * its contributors may be used to endorse or promote products derived from
|
---|
40 | * this software without specific prior written permission.
|
---|
41 | *
|
---|
42 | * THIS SOFTWARE IS PROVIDED BY LAURENS VAN DER MAATEN ''AS IS'' AND ANY EXPRESS
|
---|
43 | * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
|
---|
44 | * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
|
---|
45 | * EVENT SHALL LAURENS VAN DER MAATEN BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
---|
46 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
---|
47 | * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
|
---|
48 | * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
|
---|
49 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
|
---|
50 | * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY
|
---|
51 | * OF SUCH DAMAGE.
|
---|
52 | *
|
---|
53 | */
|
---|
54 | #endregion
|
---|
55 |
|
---|
56 | using System;
|
---|
57 | using System.Collections.Generic;
|
---|
58 | using HeuristicLab.Collections;
|
---|
59 | using HeuristicLab.Common;
|
---|
60 | using HeuristicLab.Core;
|
---|
61 | using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
|
---|
62 | using HeuristicLab.Random;
|
---|
63 |
|
---|
64 | namespace HeuristicLab.Algorithms.DataAnalysis {
|
---|
65 | [StorableClass]
|
---|
66 | public class TSNEStatic<T> {
|
---|
67 |
|
---|
68 | [StorableClass]
|
---|
69 | public sealed class TSNEState : DeepCloneable {
|
---|
70 | #region Storables
|
---|
71 | // initialized once
|
---|
72 | [Storable]
|
---|
73 | public IDistance<T> distance;
|
---|
74 | [Storable]
|
---|
75 | public IRandom random;
|
---|
76 | [Storable]
|
---|
77 | public double perplexity;
|
---|
78 | [Storable]
|
---|
79 | public bool exact;
|
---|
80 | [Storable]
|
---|
81 | public int noDatapoints;
|
---|
82 | [Storable]
|
---|
83 | public double finalMomentum;
|
---|
84 | [Storable]
|
---|
85 | public int momSwitchIter;
|
---|
86 | [Storable]
|
---|
87 | public int stopLyingIter;
|
---|
88 | [Storable]
|
---|
89 | public double theta;
|
---|
90 | [Storable]
|
---|
91 | public double eta;
|
---|
92 | [Storable]
|
---|
93 | public int newDimensions;
|
---|
94 |
|
---|
95 | // for approximate version: sparse representation of similarity/distance matrix
|
---|
96 | [Storable]
|
---|
97 | public double[] valP; // similarity/distance
|
---|
98 | [Storable]
|
---|
99 | public int[] rowP; // row index
|
---|
100 | [Storable]
|
---|
101 | public int[] colP; // col index
|
---|
102 |
|
---|
103 | // for exact version: dense representation of distance/similarity matrix
|
---|
104 | [Storable]
|
---|
105 | public double[,] p;
|
---|
106 |
|
---|
107 | // mapped data
|
---|
108 | [Storable]
|
---|
109 | public double[,] newData;
|
---|
110 |
|
---|
111 | [Storable]
|
---|
112 | public int iter;
|
---|
113 | [Storable]
|
---|
114 | public double currentMomentum;
|
---|
115 |
|
---|
116 | // helper variables (updated in each iteration)
|
---|
117 | [Storable]
|
---|
118 | public double[,] gains;
|
---|
119 | [Storable]
|
---|
120 | public double[,] uY;
|
---|
121 | [Storable]
|
---|
122 | public double[,] dY;
|
---|
123 | #endregion
|
---|
124 |
|
---|
125 | #region Constructors & Cloning
|
---|
126 | private TSNEState(TSNEState original, Cloner cloner) : base(original, cloner) {
|
---|
127 | distance = cloner.Clone(original.distance);
|
---|
128 | random = cloner.Clone(original.random);
|
---|
129 | perplexity = original.perplexity;
|
---|
130 | exact = original.exact;
|
---|
131 | noDatapoints = original.noDatapoints;
|
---|
132 | finalMomentum = original.finalMomentum;
|
---|
133 | momSwitchIter = original.momSwitchIter;
|
---|
134 | stopLyingIter = original.stopLyingIter;
|
---|
135 | theta = original.theta;
|
---|
136 | eta = original.eta;
|
---|
137 | newDimensions = original.newDimensions;
|
---|
138 | if (original.valP != null) {
|
---|
139 | valP = new double[original.valP.Length];
|
---|
140 | Array.Copy(original.valP, valP, valP.Length);
|
---|
141 | }
|
---|
142 | if (original.rowP != null) {
|
---|
143 | rowP = new int[original.rowP.Length];
|
---|
144 | Array.Copy(original.rowP, rowP, rowP.Length);
|
---|
145 | }
|
---|
146 | if (original.colP != null) {
|
---|
147 | colP = new int[original.colP.Length];
|
---|
148 | Array.Copy(original.colP, colP, colP.Length);
|
---|
149 | }
|
---|
150 | if (original.p != null) {
|
---|
151 | p = new double[original.p.GetLength(0), original.p.GetLength(1)];
|
---|
152 | Array.Copy(original.p, p, p.Length);
|
---|
153 | }
|
---|
154 | newData = new double[original.newData.GetLength(0), original.newData.GetLength(1)];
|
---|
155 | Array.Copy(original.newData, newData, newData.Length);
|
---|
156 | iter = original.iter;
|
---|
157 | currentMomentum = original.currentMomentum;
|
---|
158 | gains = new double[original.gains.GetLength(0), original.gains.GetLength(1)];
|
---|
159 | Array.Copy(original.gains, gains, gains.Length);
|
---|
160 | uY = new double[original.uY.GetLength(0), original.uY.GetLength(1)];
|
---|
161 | Array.Copy(original.uY, uY, uY.Length);
|
---|
162 | dY = new double[original.dY.GetLength(0), original.dY.GetLength(1)];
|
---|
163 | Array.Copy(original.dY, dY, dY.Length);
|
---|
164 | }
|
---|
165 |
|
---|
166 | public override IDeepCloneable Clone(Cloner cloner) {
|
---|
167 | return new TSNEState(this, cloner);
|
---|
168 | }
|
---|
169 |
|
---|
170 | [StorableConstructor]
|
---|
171 | public TSNEState(bool deserializing) { }
|
---|
172 | public TSNEState(T[] data, IDistance<T> distance, IRandom random, int newDimensions, double perplexity, double theta, int stopLyingIter, int momSwitchIter, double momentum, double finalMomentum, double eta) {
|
---|
173 | this.distance = distance;
|
---|
174 | this.random = random;
|
---|
175 | this.newDimensions = newDimensions;
|
---|
176 | this.perplexity = perplexity;
|
---|
177 | this.theta = theta;
|
---|
178 | this.stopLyingIter = stopLyingIter;
|
---|
179 | this.momSwitchIter = momSwitchIter;
|
---|
180 | currentMomentum = momentum;
|
---|
181 | this.finalMomentum = finalMomentum;
|
---|
182 | this.eta = eta;
|
---|
183 |
|
---|
184 | // initialize
|
---|
185 | noDatapoints = data.Length;
|
---|
186 | if (noDatapoints - 1 < 3 * perplexity)
|
---|
187 | throw new ArgumentException("Perplexity too large for the number of data points!");
|
---|
188 |
|
---|
189 | exact = Math.Abs(theta) < double.Epsilon;
|
---|
190 | newData = new double[noDatapoints, newDimensions];
|
---|
191 | dY = new double[noDatapoints, newDimensions];
|
---|
192 | uY = new double[noDatapoints, newDimensions];
|
---|
193 | gains = new double[noDatapoints, newDimensions];
|
---|
194 | for (var i = 0; i < noDatapoints; i++)
|
---|
195 | for (var j = 0; j < newDimensions; j++)
|
---|
196 | gains[i, j] = 1.0;
|
---|
197 |
|
---|
198 | p = null;
|
---|
199 | rowP = null;
|
---|
200 | colP = null;
|
---|
201 | valP = null;
|
---|
202 |
|
---|
203 | //Calculate Similarities
|
---|
204 | if (exact) p = CalculateExactSimilarites(data, distance, perplexity);
|
---|
205 | else CalculateApproximateSimilarities(data, distance, perplexity, out rowP, out colP, out valP);
|
---|
206 |
|
---|
207 | // Lie about the P-values (factor is 4 in the MATLAB implementation)
|
---|
208 | if (exact) for (var i = 0; i < noDatapoints; i++) for (var j = 0; j < noDatapoints; j++) p[i, j] *= 12.0;
|
---|
209 | else for (var i = 0; i < rowP[noDatapoints]; i++) valP[i] *= 12.0;
|
---|
210 |
|
---|
211 | // Initialize solution (randomly)
|
---|
212 | var rand = new NormalDistributedRandom(random, 0, 1);
|
---|
213 | for (var i = 0; i < noDatapoints; i++)
|
---|
214 | for (var j = 0; j < newDimensions; j++)
|
---|
215 | newData[i, j] = rand.NextDouble() * .0001;
|
---|
216 | }
|
---|
217 | #endregion
|
---|
218 |
|
---|
219 | public double EvaluateError() {
|
---|
220 | return exact ?
|
---|
221 | EvaluateErrorExact(p, newData, noDatapoints, newDimensions) :
|
---|
222 | EvaluateErrorApproximate(rowP, colP, valP, newData, theta);
|
---|
223 | }
|
---|
224 |
|
---|
225 | #region Helpers
|
---|
226 | private static void CalculateApproximateSimilarities(T[] data, IDistance<T> distance, double perplexity, out int[] rowP, out int[] colP, out double[] valP) {
|
---|
227 | // Compute asymmetric pairwise input similarities
|
---|
228 | ComputeGaussianPerplexity(data, distance, out rowP, out colP, out valP, perplexity, (int)(3 * perplexity));
|
---|
229 | // Symmetrize input similarities
|
---|
230 | int[] sRowP, symColP;
|
---|
231 | double[] sValP;
|
---|
232 | SymmetrizeMatrix(rowP, colP, valP, out sRowP, out symColP, out sValP);
|
---|
233 | rowP = sRowP;
|
---|
234 | colP = symColP;
|
---|
235 | valP = sValP;
|
---|
236 | var sumP = .0;
|
---|
237 | for (var i = 0; i < rowP[data.Length]; i++) sumP += valP[i];
|
---|
238 | for (var i = 0; i < rowP[data.Length]; i++) valP[i] /= sumP;
|
---|
239 | }
|
---|
240 |
|
---|
241 | private static double[,] CalculateExactSimilarites(T[] data, IDistance<T> distance, double perplexity) {
|
---|
242 | // Compute similarities
|
---|
243 | var p = new double[data.Length, data.Length];
|
---|
244 | ComputeGaussianPerplexity(data, distance, p, perplexity);
|
---|
245 | // Symmetrize input similarities
|
---|
246 | for (var n = 0; n < data.Length; n++) {
|
---|
247 | for (var m = n + 1; m < data.Length; m++) {
|
---|
248 | p[n, m] += p[m, n];
|
---|
249 | p[m, n] = p[n, m];
|
---|
250 | }
|
---|
251 | }
|
---|
252 | var sumP = .0;
|
---|
253 | for (var i = 0; i < data.Length; i++) for (var j = 0; j < data.Length; j++) sumP += p[i, j];
|
---|
254 | for (var i = 0; i < data.Length; i++) for (var j = 0; j < data.Length; j++) p[i, j] /= sumP;
|
---|
255 | return p;
|
---|
256 | }
|
---|
257 |
|
---|
258 | private static void ComputeGaussianPerplexity(IReadOnlyList<T> x, IDistance<T> distance, out int[] rowP, out int[] colP, out double[] valP, double perplexity, int k) {
|
---|
259 | if (perplexity > k) throw new ArgumentException("Perplexity should be lower than k!");
|
---|
260 |
|
---|
261 | var n = x.Count;
|
---|
262 | // Allocate the memory we need
|
---|
263 | rowP = new int[n + 1];
|
---|
264 | colP = new int[n * k];
|
---|
265 | valP = new double[n * k];
|
---|
266 | var curP = new double[n - 1];
|
---|
267 | rowP[0] = 0;
|
---|
268 | for (var i = 0; i < n; i++) rowP[i + 1] = rowP[i] + k;
|
---|
269 |
|
---|
270 | var objX = new List<IndexedItem<T>>();
|
---|
271 | for (var i = 0; i < n; i++) objX.Add(new IndexedItem<T>(i, x[i]));
|
---|
272 |
|
---|
273 | // Build ball tree on data set
|
---|
274 | var tree = new VantagePointTree<IndexedItem<T>>(new IndexedItemDistance<T>(distance), objX);
|
---|
275 |
|
---|
276 | // Loop over all points to find nearest neighbors
|
---|
277 | for (var i = 0; i < n; i++) {
|
---|
278 | IList<IndexedItem<T>> indices;
|
---|
279 | IList<double> distances;
|
---|
280 |
|
---|
281 | // Find nearest neighbors
|
---|
282 | tree.Search(objX[i], k + 1, out indices, out distances);
|
---|
283 |
|
---|
284 | // Initialize some variables for binary search
|
---|
285 | var found = false;
|
---|
286 | var beta = 1.0;
|
---|
287 | var minBeta = double.MinValue;
|
---|
288 | var maxBeta = double.MaxValue;
|
---|
289 | const double tol = 1e-5;
|
---|
290 |
|
---|
291 | // Iterate until we found a good perplexity
|
---|
292 | var iter = 0; double sumP = 0;
|
---|
293 | while (!found && iter < 200) {
|
---|
294 |
|
---|
295 | // Compute Gaussian kernel row
|
---|
296 | for (var m = 0; m < k; m++) curP[m] = Math.Exp(-beta * distances[m + 1]);
|
---|
297 |
|
---|
298 | // Compute entropy of current row
|
---|
299 | sumP = double.Epsilon;
|
---|
300 | for (var m = 0; m < k; m++) sumP += curP[m];
|
---|
301 | var h = .0;
|
---|
302 | for (var m = 0; m < k; m++) h += beta * (distances[m + 1] * curP[m]);
|
---|
303 | h = h / sumP + Math.Log(sumP);
|
---|
304 |
|
---|
305 | // Evaluate whether the entropy is within the tolerance level
|
---|
306 | var hdiff = h - Math.Log(perplexity);
|
---|
307 | if (hdiff < tol && -hdiff < tol) {
|
---|
308 | found = true;
|
---|
309 | } else {
|
---|
310 | if (hdiff > 0) {
|
---|
311 | minBeta = beta;
|
---|
312 | if (maxBeta.IsAlmost(double.MaxValue) || maxBeta.IsAlmost(double.MinValue))
|
---|
313 | beta *= 2.0;
|
---|
314 | else
|
---|
315 | beta = (beta + maxBeta) / 2.0;
|
---|
316 | } else {
|
---|
317 | maxBeta = beta;
|
---|
318 | if (minBeta.IsAlmost(double.MinValue) || minBeta.IsAlmost(double.MaxValue))
|
---|
319 | beta /= 2.0;
|
---|
320 | else
|
---|
321 | beta = (beta + minBeta) / 2.0;
|
---|
322 | }
|
---|
323 | }
|
---|
324 |
|
---|
325 | // Update iteration counter
|
---|
326 | iter++;
|
---|
327 | }
|
---|
328 |
|
---|
329 | // Row-normalize current row of P and store in matrix
|
---|
330 | for (var m = 0; m < k; m++) curP[m] /= sumP;
|
---|
331 | for (var m = 0; m < k; m++) {
|
---|
332 | colP[rowP[i] + m] = indices[m + 1].Index;
|
---|
333 | valP[rowP[i] + m] = curP[m];
|
---|
334 | }
|
---|
335 | }
|
---|
336 | }
|
---|
337 | private static void ComputeGaussianPerplexity(T[] x, IDistance<T> distance, double[,] p, double perplexity) {
|
---|
338 | // Compute the distance matrix
|
---|
339 | var dd = ComputeDistances(x, distance);
|
---|
340 |
|
---|
341 | var n = x.Length;
|
---|
342 | // Compute the Gaussian kernel row by row
|
---|
343 | for (var i = 0; i < n; i++) {
|
---|
344 | // Initialize some variables
|
---|
345 | var found = false;
|
---|
346 | var beta = 1.0;
|
---|
347 | var minBeta = double.MinValue;
|
---|
348 | var maxBeta = double.MaxValue;
|
---|
349 | const double tol = 1e-5;
|
---|
350 | double sumP = 0;
|
---|
351 |
|
---|
352 | // Iterate until we found a good perplexity
|
---|
353 | var iter = 0;
|
---|
354 | while (!found && iter < 200) { // 200 iterations as in tSNE implementation by van der Maarten
|
---|
355 |
|
---|
356 | // Compute Gaussian kernel row
|
---|
357 | for (var m = 0; m < n; m++) p[i, m] = Math.Exp(-beta * dd[i][m]);
|
---|
358 | p[i, i] = double.Epsilon;
|
---|
359 |
|
---|
360 | // Compute entropy of current row
|
---|
361 | sumP = double.Epsilon;
|
---|
362 | for (var m = 0; m < n; m++) sumP += p[i, m];
|
---|
363 | var h = 0.0;
|
---|
364 | for (var m = 0; m < n; m++) h += beta * (dd[i][m] * p[i, m]);
|
---|
365 | h = h / sumP + Math.Log(sumP);
|
---|
366 |
|
---|
367 | // Evaluate whether the entropy is within the tolerance level
|
---|
368 | var hdiff = h - Math.Log(perplexity);
|
---|
369 | if (hdiff < tol && -hdiff < tol) {
|
---|
370 | found = true;
|
---|
371 | } else {
|
---|
372 | if (hdiff > 0) {
|
---|
373 | minBeta = beta;
|
---|
374 | if (maxBeta.IsAlmost(double.MaxValue) || maxBeta.IsAlmost(double.MinValue))
|
---|
375 | beta *= 2.0;
|
---|
376 | else
|
---|
377 | beta = (beta + maxBeta) / 2.0;
|
---|
378 | } else {
|
---|
379 | maxBeta = beta;
|
---|
380 | if (minBeta.IsAlmost(double.MinValue) || minBeta.IsAlmost(double.MaxValue))
|
---|
381 | beta /= 2.0;
|
---|
382 | else
|
---|
383 | beta = (beta + minBeta) / 2.0;
|
---|
384 | }
|
---|
385 | }
|
---|
386 |
|
---|
387 | // Update iteration counter
|
---|
388 | iter++;
|
---|
389 | }
|
---|
390 |
|
---|
391 | // Row normalize P
|
---|
392 | for (var m = 0; m < n; m++) p[i, m] /= sumP;
|
---|
393 | }
|
---|
394 | }
|
---|
395 |
|
---|
396 | private static double[][] ComputeDistances(T[] x, IDistance<T> distance) {
|
---|
397 | var res = new double[x.Length][];
|
---|
398 | for (var r = 0; r < x.Length; r++) {
|
---|
399 | var rowV = new double[x.Length];
|
---|
400 | // all distances must be symmetric
|
---|
401 | for (var c = 0; c < r; c++) {
|
---|
402 | rowV[c] = res[c][r];
|
---|
403 | }
|
---|
404 | rowV[r] = 0.0; // distance to self is zero for all distances
|
---|
405 | for (var c = r + 1; c < x.Length; c++) {
|
---|
406 | rowV[c] = distance.Get(x[r], x[c]);
|
---|
407 | }
|
---|
408 | res[r] = rowV;
|
---|
409 | }
|
---|
410 | return res;
|
---|
411 | // return x.Select(m => x.Select(n => distance.Get(m, n)).ToArray()).ToArray();
|
---|
412 | }
|
---|
413 |
|
---|
414 | private static double EvaluateErrorExact(double[,] p, double[,] y, int n, int d) {
|
---|
415 | // Compute the squared Euclidean distance matrix
|
---|
416 | var dd = new double[n, n];
|
---|
417 | var q = new double[n, n];
|
---|
418 | ComputeSquaredEuclideanDistance(y, n, d, dd);
|
---|
419 |
|
---|
420 | // Compute Q-matrix and normalization sum
|
---|
421 | var sumQ = double.Epsilon;
|
---|
422 | for (var n1 = 0; n1 < n; n1++) {
|
---|
423 | for (var m = 0; m < n; m++) {
|
---|
424 | if (n1 != m) {
|
---|
425 | q[n1, m] = 1 / (1 + dd[n1, m]);
|
---|
426 | sumQ += q[n1, m];
|
---|
427 | } else q[n1, m] = double.Epsilon;
|
---|
428 | }
|
---|
429 | }
|
---|
430 | for (var i = 0; i < n; i++) for (var j = 0; j < n; j++) q[i, j] /= sumQ;
|
---|
431 |
|
---|
432 | // Sum t-SNE error
|
---|
433 | var c = .0;
|
---|
434 | for (var i = 0; i < n; i++)
|
---|
435 | for (var j = 0; j < n; j++) {
|
---|
436 | c += p[i, j] * Math.Log((p[i, j] + float.Epsilon) / (q[i, j] + float.Epsilon));
|
---|
437 | }
|
---|
438 | return c;
|
---|
439 | }
|
---|
440 |
|
---|
441 | private static double EvaluateErrorApproximate(IReadOnlyList<int> rowP, IReadOnlyList<int> colP, IReadOnlyList<double> valP, double[,] y, double theta) {
|
---|
442 | // Get estimate of normalization term
|
---|
443 | var n = y.GetLength(0);
|
---|
444 | var d = y.GetLength(1);
|
---|
445 | var tree = new SpacePartitioningTree(y);
|
---|
446 | var buff = new double[d];
|
---|
447 | var sumQ = 0.0;
|
---|
448 | for (var i = 0; i < n; i++) tree.ComputeNonEdgeForces(i, theta, buff, ref sumQ);
|
---|
449 |
|
---|
450 | // Loop over all edges to compute t-SNE error
|
---|
451 | var c = .0;
|
---|
452 | for (var k = 0; k < n; k++) {
|
---|
453 | for (var i = rowP[k]; i < rowP[k + 1]; i++) {
|
---|
454 | var q = .0;
|
---|
455 | for (var j = 0; j < d; j++) buff[j] = y[k, j];
|
---|
456 | for (var j = 0; j < d; j++) buff[j] -= y[colP[i], j];
|
---|
457 | for (var j = 0; j < d; j++) q += buff[j] * buff[j];
|
---|
458 | q = (1.0 / (1.0 + q)) / sumQ;
|
---|
459 | c += valP[i] * Math.Log((valP[i] + float.Epsilon) / (q + float.Epsilon));
|
---|
460 | }
|
---|
461 | }
|
---|
462 | return c;
|
---|
463 | }
|
---|
464 | private static void SymmetrizeMatrix(IReadOnlyList<int> rowP, IReadOnlyList<int> colP, IReadOnlyList<double> valP, out int[] symRowP, out int[] symColP, out double[] symValP) {
|
---|
465 |
|
---|
466 | // Count number of elements and row counts of symmetric matrix
|
---|
467 | var n = rowP.Count - 1;
|
---|
468 | var rowCounts = new int[n];
|
---|
469 | for (var j = 0; j < n; j++) {
|
---|
470 | for (var i = rowP[j]; i < rowP[j + 1]; i++) {
|
---|
471 |
|
---|
472 | // Check whether element (col_P[i], n) is present
|
---|
473 | var present = false;
|
---|
474 | for (var m = rowP[colP[i]]; m < rowP[colP[i] + 1]; m++) {
|
---|
475 | if (colP[m] == j) present = true;
|
---|
476 | }
|
---|
477 | if (present) rowCounts[j]++;
|
---|
478 | else {
|
---|
479 | rowCounts[j]++;
|
---|
480 | rowCounts[colP[i]]++;
|
---|
481 | }
|
---|
482 | }
|
---|
483 | }
|
---|
484 | var noElem = 0;
|
---|
485 | for (var i = 0; i < n; i++) noElem += rowCounts[i];
|
---|
486 |
|
---|
487 | // Allocate memory for symmetrized matrix
|
---|
488 | symRowP = new int[n + 1];
|
---|
489 | symColP = new int[noElem];
|
---|
490 | symValP = new double[noElem];
|
---|
491 |
|
---|
492 | // Construct new row indices for symmetric matrix
|
---|
493 | symRowP[0] = 0;
|
---|
494 | for (var i = 0; i < n; i++) symRowP[i + 1] = symRowP[i] + rowCounts[i];
|
---|
495 |
|
---|
496 | // Fill the result matrix
|
---|
497 | var offset = new int[n];
|
---|
498 | for (var j = 0; j < n; j++) {
|
---|
499 | for (var i = rowP[j]; i < rowP[j + 1]; i++) { // considering element(n, colP[i])
|
---|
500 |
|
---|
501 | // Check whether element (col_P[i], n) is present
|
---|
502 | var present = false;
|
---|
503 | for (var m = rowP[colP[i]]; m < rowP[colP[i] + 1]; m++) {
|
---|
504 | if (colP[m] != j) continue;
|
---|
505 | present = true;
|
---|
506 | if (j > colP[i]) continue; // make sure we do not add elements twice
|
---|
507 | symColP[symRowP[j] + offset[j]] = colP[i];
|
---|
508 | symColP[symRowP[colP[i]] + offset[colP[i]]] = j;
|
---|
509 | symValP[symRowP[j] + offset[j]] = valP[i] + valP[m];
|
---|
510 | symValP[symRowP[colP[i]] + offset[colP[i]]] = valP[i] + valP[m];
|
---|
511 | }
|
---|
512 |
|
---|
513 | // If (colP[i], n) is not present, there is no addition involved
|
---|
514 | if (!present) {
|
---|
515 | symColP[symRowP[j] + offset[j]] = colP[i];
|
---|
516 | symColP[symRowP[colP[i]] + offset[colP[i]]] = j;
|
---|
517 | symValP[symRowP[j] + offset[j]] = valP[i];
|
---|
518 | symValP[symRowP[colP[i]] + offset[colP[i]]] = valP[i];
|
---|
519 | }
|
---|
520 |
|
---|
521 | // Update offsets
|
---|
522 | if (present && (j > colP[i])) continue;
|
---|
523 | offset[j]++;
|
---|
524 | if (colP[i] != j) offset[colP[i]]++;
|
---|
525 | }
|
---|
526 | }
|
---|
527 |
|
---|
528 | for (var i = 0; i < noElem; i++) symValP[i] /= 2.0;
|
---|
529 | }
|
---|
530 | #endregion
|
---|
531 | }
|
---|
532 |
|
---|
533 | /// <summary>
|
---|
534 | /// Static interface to tSNE
|
---|
535 | /// </summary>
|
---|
536 | /// <param name="data"></param>
|
---|
537 | /// <param name="distance">The distance function used to differentiate similar from non-similar points, e.g. Euclidean distance.</param>
|
---|
538 | /// <param name="random">Random number generator</param>
|
---|
539 | /// <param name="newDimensions">Dimensionality of projected space (usually 2 for easy visual analysis).</param>
|
---|
540 | /// <param name="perplexity">Perplexity parameter of tSNE. Comparable to k in a k-nearest neighbour algorithm. Recommended value is floor(number of points /3) or lower</param>
|
---|
541 | /// <param name="iterations">Maximum number of iterations for gradient descent.</param>
|
---|
542 | /// <param name="theta">Value describing how much appoximated gradients my differ from exact gradients. Set to 0 for exact calculation and in [0,1] otherwise. CAUTION: exact calculation of forces requires building a non-sparse N*N matrix where N is the number of data points. This may exceed memory limitations.</param>
|
---|
543 | /// <param name="stopLyingIter">Number of iterations after which p is no longer approximated.</param>
|
---|
544 | /// <param name="momSwitchIter">Number of iterations after which the momentum in the gradient descent is switched.</param>
|
---|
545 | /// <param name="momentum">The initial momentum in the gradient descent.</param>
|
---|
546 | /// <param name="finalMomentum">The final momentum in gradient descent (after momentum switch).</param>
|
---|
547 | /// <param name="eta">Gradient descent learning rate.</param>
|
---|
548 | /// <returns></returns>
|
---|
549 | public static double[,] Run(T[] data, IDistance<T> distance, IRandom random,
|
---|
550 | int newDimensions = 2, double perplexity = 25, int iterations = 1000,
|
---|
551 | double theta = 0,
|
---|
552 | int stopLyingIter = 0, int momSwitchIter = 0, double momentum = .5,
|
---|
553 | double finalMomentum = .8, double eta = 10.0
|
---|
554 | ) {
|
---|
555 | var state = CreateState(data, distance, random, newDimensions, perplexity,
|
---|
556 | theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta);
|
---|
557 |
|
---|
558 | for (var i = 0; i < iterations - 1; i++) {
|
---|
559 | Iterate(state);
|
---|
560 | }
|
---|
561 | return Iterate(state);
|
---|
562 | }
|
---|
563 |
|
---|
564 | public static TSNEState CreateState(T[] data, IDistance<T> distance, IRandom random,
|
---|
565 | int newDimensions = 2, double perplexity = 25, double theta = 0,
|
---|
566 | int stopLyingIter = 0, int momSwitchIter = 0, double momentum = .5,
|
---|
567 | double finalMomentum = .8, double eta = 10.0
|
---|
568 | ) {
|
---|
569 | return new TSNEState(data, distance, random, newDimensions, perplexity, theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta);
|
---|
570 | }
|
---|
571 |
|
---|
572 | public static double[,] Iterate(TSNEState state) {
|
---|
573 | if (state.exact)
|
---|
574 | ComputeExactGradient(state.p, state.newData, state.noDatapoints, state.newDimensions, state.dY);
|
---|
575 | else
|
---|
576 | ComputeApproximateGradient(state.rowP, state.colP, state.valP, state.newData, state.noDatapoints, state.newDimensions, state.dY, state.theta);
|
---|
577 |
|
---|
578 | // Update gains
|
---|
579 | for (var i = 0; i < state.noDatapoints; i++) {
|
---|
580 | for (var j = 0; j < state.newDimensions; j++) {
|
---|
581 | state.gains[i, j] = Math.Sign(state.dY[i, j]) != Math.Sign(state.uY[i, j])
|
---|
582 | ? state.gains[i, j] + .2 // +0.2 nd *0.8 are used in two separate implementations of tSNE -> seems to be correct
|
---|
583 | : state.gains[i, j] * .8;
|
---|
584 |
|
---|
585 | if (state.gains[i, j] < .01) state.gains[i, j] = .01;
|
---|
586 | }
|
---|
587 | }
|
---|
588 |
|
---|
589 |
|
---|
590 | // Perform gradient update (with momentum and gains)
|
---|
591 | for (var i = 0; i < state.noDatapoints; i++)
|
---|
592 | for (var j = 0; j < state.newDimensions; j++)
|
---|
593 | state.uY[i, j] = state.currentMomentum * state.uY[i, j] - state.eta * state.gains[i, j] * state.dY[i, j];
|
---|
594 |
|
---|
595 | for (var i = 0; i < state.noDatapoints; i++)
|
---|
596 | for (var j = 0; j < state.newDimensions; j++)
|
---|
597 | state.newData[i, j] = state.newData[i, j] + state.uY[i, j];
|
---|
598 |
|
---|
599 | // Make solution zero-mean
|
---|
600 | ZeroMean(state.newData);
|
---|
601 |
|
---|
602 | // Stop lying about the P-values after a while, and switch momentum
|
---|
603 | if (state.iter == state.stopLyingIter) {
|
---|
604 | if (state.exact)
|
---|
605 | for (var i = 0; i < state.noDatapoints; i++)
|
---|
606 | for (var j = 0; j < state.noDatapoints; j++)
|
---|
607 | state.p[i, j] /= 12.0;
|
---|
608 | else
|
---|
609 | for (var i = 0; i < state.rowP[state.noDatapoints]; i++)
|
---|
610 | state.valP[i] /= 12.0;
|
---|
611 | }
|
---|
612 |
|
---|
613 | if (state.iter == state.momSwitchIter)
|
---|
614 | state.currentMomentum = state.finalMomentum;
|
---|
615 |
|
---|
616 | state.iter++;
|
---|
617 | return state.newData;
|
---|
618 | }
|
---|
619 |
|
---|
620 | #region Helpers
|
---|
621 | private static void ComputeApproximateGradient(int[] rowP, int[] colP, double[] valP, double[,] y, int n, int d, double[,] dC, double theta) {
|
---|
622 | var tree = new SpacePartitioningTree(y);
|
---|
623 | var sumQ = 0.0;
|
---|
624 | var posF = new double[n, d];
|
---|
625 | var negF = new double[n, d];
|
---|
626 | SpacePartitioningTree.ComputeEdgeForces(rowP, colP, valP, n, posF, y, d);
|
---|
627 | var row = new double[d];
|
---|
628 | for (var n1 = 0; n1 < n; n1++) {
|
---|
629 | Array.Clear(row, 0, row.Length);
|
---|
630 | tree.ComputeNonEdgeForces(n1, theta, row, ref sumQ);
|
---|
631 | Buffer.BlockCopy(row, 0, negF, (sizeof(double) * n1 * d), d * sizeof(double));
|
---|
632 | }
|
---|
633 |
|
---|
634 | // Compute final t-SNE gradient
|
---|
635 | for (var i = 0; i < n; i++)
|
---|
636 | for (var j = 0; j < d; j++) {
|
---|
637 | dC[i, j] = posF[i, j] - negF[i, j] / sumQ;
|
---|
638 | }
|
---|
639 | }
|
---|
640 |
|
---|
641 | private static void ComputeExactGradient(double[,] p, double[,] y, int n, int d, double[,] dC) {
|
---|
642 | // Make sure the current gradient contains zeros
|
---|
643 | for (var i = 0; i < n; i++) for (var j = 0; j < d; j++) dC[i, j] = 0.0;
|
---|
644 |
|
---|
645 | // Compute the squared Euclidean distance matrix
|
---|
646 | var dd = new double[n, n];
|
---|
647 | ComputeSquaredEuclideanDistance(y, n, d, dd);
|
---|
648 |
|
---|
649 | // Compute Q-matrix and normalization sum
|
---|
650 | var q = new double[n, n];
|
---|
651 | var sumQ = .0;
|
---|
652 | for (var n1 = 0; n1 < n; n1++) {
|
---|
653 | for (var m = 0; m < n; m++) {
|
---|
654 | if (n1 == m) continue;
|
---|
655 | q[n1, m] = 1 / (1 + dd[n1, m]);
|
---|
656 | sumQ += q[n1, m];
|
---|
657 | }
|
---|
658 | }
|
---|
659 |
|
---|
660 | // Perform the computation of the gradient
|
---|
661 | for (var n1 = 0; n1 < n; n1++) {
|
---|
662 | for (var m = 0; m < n; m++) {
|
---|
663 | if (n1 == m) continue;
|
---|
664 | var mult = (p[n1, m] - q[n1, m] / sumQ) * q[n1, m];
|
---|
665 | for (var d1 = 0; d1 < d; d1++) {
|
---|
666 | dC[n1, d1] += (y[n1, d1] - y[m, d1]) * mult;
|
---|
667 | }
|
---|
668 | }
|
---|
669 | }
|
---|
670 | }
|
---|
671 |
|
---|
672 | private static void ComputeSquaredEuclideanDistance(double[,] x, int n, int d, double[,] dd) {
|
---|
673 | var dataSums = new double[n];
|
---|
674 | for (var i = 0; i < n; i++) {
|
---|
675 | for (var j = 0; j < d; j++) {
|
---|
676 | dataSums[i] += x[i, j] * x[i, j];
|
---|
677 | }
|
---|
678 | }
|
---|
679 | for (var i = 0; i < n; i++) {
|
---|
680 | for (var m = 0; m < n; m++) {
|
---|
681 | dd[i, m] = dataSums[i] + dataSums[m];
|
---|
682 | }
|
---|
683 | }
|
---|
684 | for (var i = 0; i < n; i++) {
|
---|
685 | dd[i, i] = 0.0;
|
---|
686 | for (var m = i + 1; m < n; m++) {
|
---|
687 | dd[i, m] = 0.0;
|
---|
688 | for (var j = 0; j < d; j++) {
|
---|
689 | dd[i, m] += (x[i, j] - x[m, j]) * (x[i, j] - x[m, j]);
|
---|
690 | }
|
---|
691 | dd[m, i] = dd[i, m];
|
---|
692 | }
|
---|
693 | }
|
---|
694 | }
|
---|
695 |
|
---|
696 | private static void ZeroMean(double[,] x) {
|
---|
697 | // Compute data mean
|
---|
698 | var n = x.GetLength(0);
|
---|
699 | var d = x.GetLength(1);
|
---|
700 | var mean = new double[d];
|
---|
701 | for (var i = 0; i < n; i++) {
|
---|
702 | for (var j = 0; j < d; j++) {
|
---|
703 | mean[j] += x[i, j];
|
---|
704 | }
|
---|
705 | }
|
---|
706 | for (var i = 0; i < d; i++) {
|
---|
707 | mean[i] /= n;
|
---|
708 | }
|
---|
709 | // Subtract data mean
|
---|
710 | for (var i = 0; i < n; i++) {
|
---|
711 | for (var j = 0; j < d; j++) {
|
---|
712 | x[i, j] -= mean[j];
|
---|
713 | }
|
---|
714 | }
|
---|
715 | }
|
---|
716 | #endregion
|
---|
717 | }
|
---|
718 | }
|
---|