Free cookie consent management tool by TermsFeed Policy Generator

source: stable/HeuristicLab.Algorithms.DataAnalysis/3.4/TSNE/TSNEStatic.cs @ 16189

Last change on this file since 16189 was 15584, checked in by swagner, 7 years ago

#2640: Updated year of copyrights in license headers on stable

File size: 28.7 KB
Line 
1#region License Information
2/* HeuristicLab
3 * Copyright (C) 2002-2018 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
56using System;
57using System.Collections.Generic;
58using HeuristicLab.Collections;
59using HeuristicLab.Common;
60using HeuristicLab.Core;
61using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
62using HeuristicLab.Random;
63
64namespace HeuristicLab.Algorithms.DataAnalysis {
65  [StorableClass]
66  public class TSNEStatic<T> {
67    [StorableClass]
68    public sealed class TSNEState : DeepCloneable {
69      #region Storables
70      // initialized once
71      [Storable]
72      public IDistance<T> distance;
73      [Storable]
74      public IRandom random;
75      [Storable]
76      public double perplexity;
77      [Storable]
78      public bool exact;
79      [Storable]
80      public int noDatapoints;
81      [Storable]
82      public double finalMomentum;
83      [Storable]
84      public int momSwitchIter;
85      [Storable]
86      public int stopLyingIter;
87      [Storable]
88      public double theta;
89      [Storable]
90      public double eta;
91      [Storable]
92      public int newDimensions;
93
94      // for approximate version: sparse representation of similarity/distance matrix
95      [Storable]
96      public double[] valP; // similarity/distance
97      [Storable]
98      public int[] rowP; // row index
99      [Storable]
100      public int[] colP; // col index
101
102      // for exact version: dense representation of distance/similarity matrix
103      [Storable]
104      public double[,] p;
105
106      // mapped data
107      [Storable]
108      public double[,] newData;
109
110      [Storable]
111      public int iter;
112      [Storable]
113      public double currentMomentum;
114
115      // helper variables (updated in each iteration)
116      [Storable]
117      public double[,] gains;
118      [Storable]
119      public double[,] uY;
120      [Storable]
121      public double[,] dY;
122      #endregion
123
124      #region Constructors & Cloning
125      private TSNEState(TSNEState original, Cloner cloner) : base(original, cloner) {
126        distance = cloner.Clone(original.distance);
127        random = cloner.Clone(original.random);
128        perplexity = original.perplexity;
129        exact = original.exact;
130        noDatapoints = original.noDatapoints;
131        finalMomentum = original.finalMomentum;
132        momSwitchIter = original.momSwitchIter;
133        stopLyingIter = original.stopLyingIter;
134        theta = original.theta;
135        eta = original.eta;
136        newDimensions = original.newDimensions;
137        if (original.valP != null) {
138          valP = new double[original.valP.Length];
139          Array.Copy(original.valP, valP, valP.Length);
140        }
141        if (original.rowP != null) {
142          rowP = new int[original.rowP.Length];
143          Array.Copy(original.rowP, rowP, rowP.Length);
144        }
145        if (original.colP != null) {
146          colP = new int[original.colP.Length];
147          Array.Copy(original.colP, colP, colP.Length);
148        }
149        if (original.p != null) {
150          p = new double[original.p.GetLength(0), original.p.GetLength(1)];
151          Array.Copy(original.p, p, p.Length);
152        }
153        newData = new double[original.newData.GetLength(0), original.newData.GetLength(1)];
154        Array.Copy(original.newData, newData, newData.Length);
155        iter = original.iter;
156        currentMomentum = original.currentMomentum;
157        gains = new double[original.gains.GetLength(0), original.gains.GetLength(1)];
158        Array.Copy(original.gains, gains, gains.Length);
159        uY = new double[original.uY.GetLength(0), original.uY.GetLength(1)];
160        Array.Copy(original.uY, uY, uY.Length);
161        dY = new double[original.dY.GetLength(0), original.dY.GetLength(1)];
162        Array.Copy(original.dY, dY, dY.Length);
163      }
164
165      public override IDeepCloneable Clone(Cloner cloner) {
166        return new TSNEState(this, cloner);
167      }
168
169      [StorableConstructor]
170      public TSNEState(bool deserializing) { }
171
172      public TSNEState(IReadOnlyList<T> data, IDistance<T> distance, IRandom random, int newDimensions, double perplexity,
173        double theta, int stopLyingIter, int momSwitchIter, double momentum, double finalMomentum, double eta, bool randomInit) {
174        this.distance = distance;
175        this.random = random;
176        this.newDimensions = newDimensions;
177        this.perplexity = perplexity;
178        this.theta = theta;
179        this.stopLyingIter = stopLyingIter;
180        this.momSwitchIter = momSwitchIter;
181        currentMomentum = momentum;
182        this.finalMomentum = finalMomentum;
183        this.eta = eta;
184
185        // initialize
186        noDatapoints = data.Count;
187        if (noDatapoints - 1 < 3 * perplexity)
188          throw new ArgumentException("Perplexity too large for the number of data points!");
189
190        exact = Math.Abs(theta) < double.Epsilon;
191        newData = new double[noDatapoints, newDimensions];
192        dY = new double[noDatapoints, newDimensions];
193        uY = new double[noDatapoints, newDimensions];
194        gains = new double[noDatapoints, newDimensions];
195        for (var i = 0; i < noDatapoints; i++)
196        for (var j = 0; j < newDimensions; j++)
197          gains[i, j] = 1.0;
198
199        p = null;
200        rowP = null;
201        colP = null;
202        valP = null;
203
204        //Calculate Similarities
205        if (exact) p = CalculateExactSimilarites(data, distance, perplexity);
206        else CalculateApproximateSimilarities(data, distance, perplexity, out rowP, out colP, out valP);
207
208        // Lie about the P-values (factor is 4 in the MATLAB implementation)
209        if (exact) for (var i = 0; i < noDatapoints; i++) for (var j = 0; j < noDatapoints; j++) p[i, j] *= 12.0;
210        else for (var i = 0; i < rowP[noDatapoints]; i++) valP[i] *= 12.0;
211
212        // Initialize solution (randomly)
213        var rand = new NormalDistributedRandom(random, 0, 1);
214        for (var i = 0; i < noDatapoints; i++)
215        for (var j = 0; j < newDimensions; j++)
216          newData[i, j] = rand.NextDouble() * .0001;
217
218        if (!(data[0] is IReadOnlyList<double>) || randomInit) return;
219        for (var i = 0; i < noDatapoints; i++)
220        for (var j = 0; j < newDimensions; j++) {
221          var row = (IReadOnlyList<double>) data[i];
222          newData[i, j] = row[j % row.Count];
223        }
224      }
225      #endregion
226
227      public double EvaluateError() {
228        return exact ? EvaluateErrorExact(p, newData, noDatapoints, newDimensions) : EvaluateErrorApproximate(rowP, colP, valP, newData, theta);
229      }
230
231      #region Helpers
232      private static void CalculateApproximateSimilarities(IReadOnlyList<T> data, IDistance<T> distance, double perplexity, out int[] rowP, out int[] colP, out double[] valP) {
233        // Compute asymmetric pairwise input similarities
234        ComputeGaussianPerplexity(data, distance, out rowP, out colP, out valP, perplexity, (int) (3 * perplexity));
235        // Symmetrize input similarities
236        int[] sRowP, symColP;
237        double[] sValP;
238        SymmetrizeMatrix(rowP, colP, valP, out sRowP, out symColP, out sValP);
239        rowP = sRowP;
240        colP = symColP;
241        valP = sValP;
242        var sumP = .0;
243        for (var i = 0; i < rowP[data.Count]; i++) sumP += valP[i];
244        for (var i = 0; i < rowP[data.Count]; i++) valP[i] /= sumP;
245      }
246      private static double[,] CalculateExactSimilarites(IReadOnlyList<T> data, IDistance<T> distance, double perplexity) {
247        // Compute similarities
248        var p = new double[data.Count, data.Count];
249        ComputeGaussianPerplexity(data, distance, p, perplexity);
250        // Symmetrize input similarities
251        for (var n = 0; n < data.Count; n++) {
252          for (var m = n + 1; m < data.Count; m++) {
253            p[n, m] += p[m, n];
254            p[m, n] = p[n, m];
255          }
256        }
257        var sumP = .0;
258        for (var i = 0; i < data.Count; i++) {
259          for (var j = 0; j < data.Count; j++) {
260            sumP += p[i, j];
261          }
262        }
263        for (var i = 0; i < data.Count; i++) {
264          for (var j = 0; j < data.Count; j++) {
265            p[i, j] /= sumP;
266          }
267        }
268        return p;
269      }
270      private static void ComputeGaussianPerplexity(IReadOnlyList<T> x, IDistance<T> distance, out int[] rowP, out int[] colP, out double[] valP, double perplexity, int k) {
271        if (perplexity > k) throw new ArgumentException("Perplexity should be lower than k!");
272
273        var n = x.Count;
274        // Allocate the memory we need
275        rowP = new int[n + 1];
276        colP = new int[n * k];
277        valP = new double[n * k];
278        var curP = new double[n - 1];
279        rowP[0] = 0;
280        for (var i = 0; i < n; i++) rowP[i + 1] = rowP[i] + k;
281
282        var objX = new List<IndexedItem<T>>();
283        for (var i = 0; i < n; i++) objX.Add(new IndexedItem<T>(i, x[i]));
284
285        // Build ball tree on data set
286        var tree = new VantagePointTree<IndexedItem<T>>(new IndexedItemDistance<T>(distance), objX);
287
288        // Loop over all points to find nearest neighbors
289        for (var i = 0; i < n; i++) {
290          IList<IndexedItem<T>> indices;
291          IList<double> distances;
292
293          // Find nearest neighbors
294          tree.Search(objX[i], k + 1, out indices, out distances);
295
296          // Initialize some variables for binary search
297          var found = false;
298          var beta = 1.0;
299          var minBeta = double.MinValue;
300          var maxBeta = double.MaxValue;
301          const double tol = 1e-5;
302
303          // Iterate until we found a good perplexity
304          var iter = 0;
305          double sumP = 0;
306          while (!found && iter < 200) {
307            // Compute Gaussian kernel row
308            for (var m = 0; m < k; m++) curP[m] = Math.Exp(-beta * distances[m + 1]);
309
310            // Compute entropy of current row
311            sumP = double.Epsilon;
312            for (var m = 0; m < k; m++) sumP += curP[m];
313            var h = .0;
314            for (var m = 0; m < k; m++) h += beta * (distances[m + 1] * curP[m]);
315            h = h / sumP + Math.Log(sumP);
316
317            // Evaluate whether the entropy is within the tolerance level
318            var hdiff = h - Math.Log(perplexity);
319            if (hdiff < tol && -hdiff < tol) {
320              found = true;
321            }
322            else {
323              if (hdiff > 0) {
324                minBeta = beta;
325                if (maxBeta.IsAlmost(double.MaxValue) || maxBeta.IsAlmost(double.MinValue))
326                  beta *= 2.0;
327                else
328                  beta = (beta + maxBeta) / 2.0;
329              }
330              else {
331                maxBeta = beta;
332                if (minBeta.IsAlmost(double.MinValue) || minBeta.IsAlmost(double.MaxValue))
333                  beta /= 2.0;
334                else
335                  beta = (beta + minBeta) / 2.0;
336              }
337            }
338
339            // Update iteration counter
340            iter++;
341          }
342
343          // Row-normalize current row of P and store in matrix
344          for (var m = 0; m < k; m++) curP[m] /= sumP;
345          for (var m = 0; m < k; m++) {
346            colP[rowP[i] + m] = indices[m + 1].Index;
347            valP[rowP[i] + m] = curP[m];
348          }
349        }
350      }
351      private static void ComputeGaussianPerplexity(IReadOnlyList<T> x, IDistance<T> distance, double[,] p, double perplexity) {
352        // Compute the distance matrix
353        var dd = ComputeDistances(x, distance);
354
355        var n = x.Count;
356        // Compute the Gaussian kernel row by row
357        for (var i = 0; i < n; i++) {
358          // Initialize some variables
359          var found = false;
360          var beta = 1.0;
361          var minBeta = double.MinValue;
362          var maxBeta = double.MaxValue;
363          const double tol = 1e-5;
364          double sumP = 0;
365
366          // Iterate until we found a good perplexity
367          var iter = 0;
368          while (!found && iter < 200) { // 200 iterations as in tSNE implementation by van der Maarten
369
370            // Compute Gaussian kernel row
371            for (var m = 0; m < n; m++) p[i, m] = Math.Exp(-beta * dd[i][m]);
372            p[i, i] = double.Epsilon;
373
374            // Compute entropy of current row
375            sumP = double.Epsilon;
376            for (var m = 0; m < n; m++) sumP += p[i, m];
377            var h = 0.0;
378            for (var m = 0; m < n; m++) h += beta * (dd[i][m] * p[i, m]);
379            h = h / sumP + Math.Log(sumP);
380
381            // Evaluate whether the entropy is within the tolerance level
382            var hdiff = h - Math.Log(perplexity);
383            if (hdiff < tol && -hdiff < tol) {
384              found = true;
385            }
386            else {
387              if (hdiff > 0) {
388                minBeta = beta;
389                if (maxBeta.IsAlmost(double.MaxValue) || maxBeta.IsAlmost(double.MinValue))
390                  beta *= 2.0;
391                else
392                  beta = (beta + maxBeta) / 2.0;
393              }
394              else {
395                maxBeta = beta;
396                if (minBeta.IsAlmost(double.MinValue) || minBeta.IsAlmost(double.MaxValue))
397                  beta /= 2.0;
398                else
399                  beta = (beta + minBeta) / 2.0;
400              }
401            }
402
403            // Update iteration counter
404            iter++;
405          }
406
407          // Row normalize P
408          for (var m = 0; m < n; m++) p[i, m] /= sumP;
409        }
410      }
411      private static double[][] ComputeDistances(IReadOnlyList<T> x, IDistance<T> distance) {
412        var res = new double[x.Count][];
413        for (var r = 0; r < x.Count; r++) {
414          var rowV = new double[x.Count];
415          // all distances must be symmetric
416          for (var c = 0; c < r; c++) {
417            rowV[c] = res[c][r];
418          }
419          rowV[r] = 0.0; // distance to self is zero for all distances
420          for (var c = r + 1; c < x.Count; c++) {
421            rowV[c] = distance.Get(x[r], x[c]);
422          }
423          res[r] = rowV;
424        }
425        return res;
426        // return x.Select(m => x.Select(n => distance.Get(m, n)).ToArray()).ToArray();
427      }
428      private static double EvaluateErrorExact(double[,] p, double[,] y, int n, int d) {
429        // Compute the squared Euclidean distance matrix
430        var dd = new double[n, n];
431        var q = new double[n, n];
432        ComputeSquaredEuclideanDistance(y, n, d, dd);
433
434        // Compute Q-matrix and normalization sum
435        var sumQ = double.Epsilon;
436        for (var n1 = 0; n1 < n; n1++) {
437          for (var m = 0; m < n; m++) {
438            if (n1 != m) {
439              q[n1, m] = 1 / (1 + dd[n1, m]);
440              sumQ += q[n1, m];
441            }
442            else q[n1, m] = double.Epsilon;
443          }
444        }
445        for (var i = 0; i < n; i++) for (var j = 0; j < n; j++) q[i, j] /= sumQ;
446
447        // Sum t-SNE error
448        var c = .0;
449        for (var i = 0; i < n; i++)
450        for (var j = 0; j < n; j++) {
451          c += p[i, j] * Math.Log((p[i, j] + float.Epsilon) / (q[i, j] + float.Epsilon));
452        }
453        return c;
454      }
455      private static double EvaluateErrorApproximate(IReadOnlyList<int> rowP, IReadOnlyList<int> colP, IReadOnlyList<double> valP, double[,] y, double theta) {
456        // Get estimate of normalization term
457        var n = y.GetLength(0);
458        var d = y.GetLength(1);
459        var tree = new SpacePartitioningTree(y);
460        var buff = new double[d];
461        var sumQ = 0.0;
462        for (var i = 0; i < n; i++) tree.ComputeNonEdgeForces(i, theta, buff, ref sumQ);
463
464        // Loop over all edges to compute t-SNE error
465        var c = .0;
466        for (var k = 0; k < n; k++) {
467          for (var i = rowP[k]; i < rowP[k + 1]; i++) {
468            var q = .0;
469            for (var j = 0; j < d; j++) buff[j] = y[k, j];
470            for (var j = 0; j < d; j++) buff[j] -= y[colP[i], j];
471            for (var j = 0; j < d; j++) q += buff[j] * buff[j];
472            q = (1.0 / (1.0 + q)) / sumQ;
473            c += valP[i] * Math.Log((valP[i] + float.Epsilon) / (q + float.Epsilon));
474          }
475        }
476        return c;
477      }
478      private static void SymmetrizeMatrix(IReadOnlyList<int> rowP, IReadOnlyList<int> colP, IReadOnlyList<double> valP, out int[] symRowP, out int[] symColP, out double[] symValP) {
479        // Count number of elements and row counts of symmetric matrix
480        var n = rowP.Count - 1;
481        var rowCounts = new int[n];
482        for (var j = 0; j < n; j++) {
483          for (var i = rowP[j]; i < rowP[j + 1]; i++) {
484            // Check whether element (col_P[i], n) is present
485            var present = false;
486            for (var m = rowP[colP[i]]; m < rowP[colP[i] + 1]; m++) {
487              if (colP[m] == j) present = true;
488            }
489            if (present) rowCounts[j]++;
490            else {
491              rowCounts[j]++;
492              rowCounts[colP[i]]++;
493            }
494          }
495        }
496        var noElem = 0;
497        for (var i = 0; i < n; i++) noElem += rowCounts[i];
498
499        // Allocate memory for symmetrized matrix
500        symRowP = new int[n + 1];
501        symColP = new int[noElem];
502        symValP = new double[noElem];
503
504        // Construct new row indices for symmetric matrix
505        symRowP[0] = 0;
506        for (var i = 0; i < n; i++) symRowP[i + 1] = symRowP[i] + rowCounts[i];
507
508        // Fill the result matrix
509        var offset = new int[n];
510        for (var j = 0; j < n; j++) {
511          for (var i = rowP[j]; i < rowP[j + 1]; i++) { // considering element(n, colP[i])
512
513            // Check whether element (col_P[i], n) is present
514            var present = false;
515            for (var m = rowP[colP[i]]; m < rowP[colP[i] + 1]; m++) {
516              if (colP[m] != j) continue;
517              present = true;
518              if (j > colP[i]) continue; // make sure we do not add elements twice
519              symColP[symRowP[j] + offset[j]] = colP[i];
520              symColP[symRowP[colP[i]] + offset[colP[i]]] = j;
521              symValP[symRowP[j] + offset[j]] = valP[i] + valP[m];
522              symValP[symRowP[colP[i]] + offset[colP[i]]] = valP[i] + valP[m];
523            }
524
525            // If (colP[i], n) is not present, there is no addition involved
526            if (!present) {
527              symColP[symRowP[j] + offset[j]] = colP[i];
528              symColP[symRowP[colP[i]] + offset[colP[i]]] = j;
529              symValP[symRowP[j] + offset[j]] = valP[i];
530              symValP[symRowP[colP[i]] + offset[colP[i]]] = valP[i];
531            }
532
533            // Update offsets
534            if (present && (j > colP[i])) continue;
535            offset[j]++;
536            if (colP[i] != j) offset[colP[i]]++;
537          }
538        }
539
540        for (var i = 0; i < noElem; i++) symValP[i] /= 2.0;
541      }
542      #endregion
543    }
544
545    /// <summary>
546    /// Static interface to tSNE
547    /// </summary>
548    /// <param name="data"></param>
549    /// <param name="distance">The distance function used to differentiate similar from non-similar points, e.g. Euclidean distance.</param>
550    /// <param name="random">Random number generator</param>
551    /// <param name="newDimensions">Dimensionality of projected space (usually 2 for easy visual analysis).</param>
552    /// <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>
553    /// <param name="iterations">Maximum number of iterations for gradient descent.</param>
554    /// <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>
555    /// <param name="stopLyingIter">Number of iterations after which p is no longer approximated.</param>
556    /// <param name="momSwitchIter">Number of iterations after which the momentum in the gradient descent is switched.</param>
557    /// <param name="momentum">The initial momentum in the gradient descent.</param>
558    /// <param name="finalMomentum">The final momentum in gradient descent (after momentum switch).</param>
559    /// <param name="eta">Gradient descent learning rate.</param>
560    /// <returns></returns>
561    public static double[,] Run(T[] data, IDistance<T> distance, IRandom random,
562      int newDimensions = 2, double perplexity = 25, int iterations = 1000,
563      double theta = 0, int stopLyingIter = 0, int momSwitchIter = 0, double momentum = .5,
564      double finalMomentum = .8, double eta = 10.0
565    ) {
566      var state = CreateState(data, distance, random, newDimensions, perplexity,
567        theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta);
568
569      for (var i = 0; i < iterations - 1; i++) {
570        Iterate(state);
571      }
572      return Iterate(state);
573    }
574
575    public static TSNEState CreateState(T[] data, IDistance<T> distance, IRandom random,
576      int newDimensions = 2, double perplexity = 25, double theta = 0,
577      int stopLyingIter = 0, int momSwitchIter = 0, double momentum = .5,
578      double finalMomentum = .8, double eta = 10.0, bool randomInit = true
579    ) {
580      return new TSNEState(data, distance, random, newDimensions, perplexity, theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta, randomInit);
581    }
582
583    public static double[,] Iterate(TSNEState state) {
584      if (state.exact)
585        ComputeExactGradient(state.p, state.newData, state.noDatapoints, state.newDimensions, state.dY);
586      else
587        ComputeApproximateGradient(state.rowP, state.colP, state.valP, state.newData, state.noDatapoints, state.newDimensions, state.dY, state.theta);
588
589      // Update gains
590      for (var i = 0; i < state.noDatapoints; i++) {
591        for (var j = 0; j < state.newDimensions; j++) {
592          state.gains[i, j] = Math.Sign(state.dY[i, j]) != Math.Sign(state.uY[i, j])
593            ? state.gains[i, j] + .2 // +0.2 nd *0.8 are used in two separate implementations of tSNE -> seems to be correct
594            : state.gains[i, j] * .8;
595          if (state.gains[i, j] < .01) state.gains[i, j] = .01;
596        }
597      }
598
599      // Perform gradient update (with momentum and gains)
600      for (var i = 0; i < state.noDatapoints; i++)
601      for (var j = 0; j < state.newDimensions; j++)
602        state.uY[i, j] = state.currentMomentum * state.uY[i, j] - state.eta * state.gains[i, j] * state.dY[i, j];
603
604      for (var i = 0; i < state.noDatapoints; i++)
605      for (var j = 0; j < state.newDimensions; j++)
606        state.newData[i, j] = state.newData[i, j] + state.uY[i, j];
607
608      // Make solution zero-mean
609      ZeroMean(state.newData);
610
611      // Stop lying about the P-values after a while, and switch momentum
612      if (state.iter == state.stopLyingIter) {
613        if (state.exact)
614          for (var i = 0; i < state.noDatapoints; i++)
615          for (var j = 0; j < state.noDatapoints; j++)
616            state.p[i, j] /= 12.0;
617        else
618          for (var i = 0; i < state.rowP[state.noDatapoints]; i++)
619            state.valP[i] /= 12.0;
620      }
621
622      if (state.iter == state.momSwitchIter)
623        state.currentMomentum = state.finalMomentum;
624
625      state.iter++;
626      return state.newData;
627    }
628
629    #region Helpers
630    private static void ComputeApproximateGradient(int[] rowP, int[] colP, double[] valP, double[,] y, int n, int d, double[,] dC, double theta) {
631      var tree = new SpacePartitioningTree(y);
632      var sumQ = 0.0;
633      var posF = new double[n, d];
634      var negF = new double[n, d];
635      SpacePartitioningTree.ComputeEdgeForces(rowP, colP, valP, n, posF, y, d);
636      var row = new double[d];
637      for (var n1 = 0; n1 < n; n1++) {
638        Array.Clear(row, 0, row.Length);
639        tree.ComputeNonEdgeForces(n1, theta, row, ref sumQ);
640        Buffer.BlockCopy(row, 0, negF, (sizeof(double) * n1 * d), d * sizeof(double));
641      }
642
643      // Compute final t-SNE gradient
644      for (var i = 0; i < n; i++)
645      for (var j = 0; j < d; j++) {
646        dC[i, j] = posF[i, j] - negF[i, j] / sumQ;
647      }
648    }
649
650    private static void ComputeExactGradient(double[,] p, double[,] y, int n, int d, double[,] dC) {
651      // Make sure the current gradient contains zeros
652      for (var i = 0; i < n; i++) for (var j = 0; j < d; j++) dC[i, j] = 0.0;
653
654      // Compute the squared Euclidean distance matrix
655      var dd = new double[n, n];
656      ComputeSquaredEuclideanDistance(y, n, d, dd);
657
658      // Compute Q-matrix and normalization sum
659      var q = new double[n, n];
660      var sumQ = .0;
661      for (var n1 = 0; n1 < n; n1++) {
662        for (var m = 0; m < n; m++) {
663          if (n1 == m) continue;
664          q[n1, m] = 1 / (1 + dd[n1, m]);
665          sumQ += q[n1, m];
666        }
667      }
668
669      // Perform the computation of the gradient
670      for (var n1 = 0; n1 < n; n1++) {
671        for (var m = 0; m < n; m++) {
672          if (n1 == m) continue;
673          var mult = (p[n1, m] - q[n1, m] / sumQ) * q[n1, m];
674          for (var d1 = 0; d1 < d; d1++) {
675            dC[n1, d1] += (y[n1, d1] - y[m, d1]) * mult;
676          }
677        }
678      }
679    }
680
681    private static void ComputeSquaredEuclideanDistance(double[,] x, int n, int d, double[,] dd) {
682      var dataSums = new double[n];
683      for (var i = 0; i < n; i++) {
684        for (var j = 0; j < d; j++) {
685          dataSums[i] += x[i, j] * x[i, j];
686        }
687      }
688      for (var i = 0; i < n; i++) {
689        for (var m = 0; m < n; m++) {
690          dd[i, m] = dataSums[i] + dataSums[m];
691        }
692      }
693      for (var i = 0; i < n; i++) {
694        dd[i, i] = 0.0;
695        for (var m = i + 1; m < n; m++) {
696          dd[i, m] = 0.0;
697          for (var j = 0; j < d; j++) {
698            dd[i, m] += (x[i, j] - x[m, j]) * (x[i, j] - x[m, j]);
699          }
700          dd[m, i] = dd[i, m];
701        }
702      }
703    }
704
705    private static void ZeroMean(double[,] x) {
706      // Compute data mean
707      var n = x.GetLength(0);
708      var d = x.GetLength(1);
709      var mean = new double[d];
710      for (var i = 0; i < n; i++) {
711        for (var j = 0; j < d; j++) {
712          mean[j] += x[i, j];
713        }
714      }
715      for (var i = 0; i < d; i++) {
716        mean[i] /= n;
717      }
718      // Subtract data mean
719      for (var i = 0; i < n; i++) {
720        for (var j = 0; j < d; j++) {
721          x[i, j] -= mean[j];
722        }
723      }
724    }
725    #endregion
726  }
727}
Note: See TracBrowser for help on using the repository browser.