Changeset 15451 for branches/Weighted TSNE/3.4/TSNE/TSNEStatic.cs

Timestamp:

11/06/17 15:57:55 (6 years ago)

Author:

bwerth

Message:

#2850 created branch & added WeightedEuclideanDistance

Location:

branches/Weighted TSNE

Files:

• . (added)
• 3.4 (copied) (copied from trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4)
• 3.4/TSNE/TSNEStatic.cs (modified) (22 diffs)

branches/Weighted TSNE/3.4/TSNE/TSNEStatic.cs

@@ -65 +65 @@
   [StorableClass]
   public class TSNEStatic<T> {
-
     [StorableClass]
     public sealed class TSNEState : DeepCloneable {
@@ -170 +169 @@
       [StorableConstructor]
       public TSNEState(bool deserializing) { }
-      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) {
+
+      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, bool randomInit) {
         this.distance = distance;
         this.random = random;
@@ -193 +194 @@
         gains = new double[noDatapoints, newDimensions];
         for (var i = 0; i < noDatapoints; i++)
-          for (var j = 0; j < newDimensions; j++)
-            gains[i, j] = 1.0;
+        for (var j = 0; j < newDimensions; j++)
+          gains[i, j] = 1.0;
 
         p = null;
@@ -212 +213 @@
         var rand = new NormalDistributedRandom(random, 0, 1);
         for (var i = 0; i < noDatapoints; i++)
-          for (var j = 0; j < newDimensions; j++)
-            newData[i, j] = rand.NextDouble() * .0001;
+        for (var j = 0; j < newDimensions; j++)
+          newData[i, j] = rand.NextDouble() * .0001;
+
+        if (data[0] is IReadOnlyList<double> && !randomInit) {
+          for (var i = 0; i < noDatapoints; i++)
+          for (var j = 0; j < newDimensions; j++) {
+            var row = (IReadOnlyList<double>) data[i];
+            newData[i, j] = row[j % row.Count];
+          }
+        }
       }
       #endregion
 
       public double EvaluateError() {
-        return exact ?
-          EvaluateErrorExact(p, newData, noDatapoints, newDimensions) :
-          EvaluateErrorApproximate(rowP, colP, valP, newData, theta);
+        return exact ? EvaluateErrorExact(p, newData, noDatapoints, newDimensions) : EvaluateErrorApproximate(rowP, colP, valP, newData, theta);
       }
 
@@ -226 +233 @@
       private static void CalculateApproximateSimilarities(T[] data, IDistance<T> distance, double perplexity, out int[] rowP, out int[] colP, out double[] valP) {
         // Compute asymmetric pairwise input similarities
-        ComputeGaussianPerplexity(data, distance, out rowP, out colP, out valP, perplexity, (int)(3 * perplexity));
+        ComputeGaussianPerplexity(data, distance, out rowP, out colP, out valP, perplexity, (int) (3 * perplexity));
         // Symmetrize input similarities
         int[] sRowP, symColP;
@@ -290 +297 @@
 
           // Iterate until we found a good perplexity
-          var iter = 0; double sumP = 0;
+          var iter = 0;
+          double sumP = 0;
           while (!found && iter < 200) {
-
             // Compute Gaussian kernel row
             for (var m = 0; m < k; m++) curP[m] = Math.Exp(-beta * distances[m + 1]);
@@ -307 +314 @@
             if (hdiff < tol && -hdiff < tol) {
               found = true;
-            } else {
+            }
+            else {
               if (hdiff > 0) {
                 minBeta = beta;
@@ -314 +322 @@
                 else
                   beta = (beta + maxBeta) / 2.0;
-              } else {
+              }
+              else {
                 maxBeta = beta;
                 if (minBeta.IsAlmost(double.MinValue) || minBeta.IsAlmost(double.MaxValue))
@@ -352 +361 @@
           // Iterate until we found a good perplexity
           var iter = 0;
-          while (!found && iter < 200) {      // 200 iterations as in tSNE implementation by van der Maarten
+          while (!found && iter < 200) { // 200 iterations as in tSNE implementation by van der Maarten
 
             // Compute Gaussian kernel row
@@ -369 +378 @@
             if (hdiff < tol && -hdiff < tol) {
               found = true;
-            } else {
+            }
+            else {
               if (hdiff > 0) {
                 minBeta = beta;
@@ -376 +386 @@
                 else
                   beta = (beta + maxBeta) / 2.0;
-              } else {
+              }
+              else {
                 maxBeta = beta;
                 if (minBeta.IsAlmost(double.MinValue) || minBeta.IsAlmost(double.MaxValue))
@@ -425 +436 @@
               q[n1, m] = 1 / (1 + dd[n1, m]);
               sumQ += q[n1, m];
-            } else q[n1, m] = double.Epsilon;
+            }
+            else q[n1, m] = double.Epsilon;
           }
         }
@@ -433 +445 @@
         var c = .0;
         for (var i = 0; i < n; i++)
-          for (var j = 0; j < n; j++) {
-            c += p[i, j] * Math.Log((p[i, j] + float.Epsilon) / (q[i, j] + float.Epsilon));
-          }
+        for (var j = 0; j < n; j++) {
+          c += p[i, j] * Math.Log((p[i, j] + float.Epsilon) / (q[i, j] + float.Epsilon));
+        }
         return c;
       }
@@ -463 +475 @@
       }
       private static void SymmetrizeMatrix(IReadOnlyList<int> rowP, IReadOnlyList<int> colP, IReadOnlyList<double> valP, out int[] symRowP, out int[] symColP, out double[] symValP) {
-
         // Count number of elements and row counts of symmetric matrix
         var n = rowP.Count - 1;
@@ -469 +480 @@
         for (var j = 0; j < n; j++) {
           for (var i = rowP[j]; i < rowP[j + 1]; i++) {
-
             // Check whether element (col_P[i], n) is present
             var present = false;
@@ -497 +507 @@
         var offset = new int[n];
         for (var j = 0; j < n; j++) {
-          for (var i = rowP[j]; i < rowP[j + 1]; i++) {                                  // considering element(n, colP[i])
+          for (var i = rowP[j]; i < rowP[j + 1]; i++) { // considering element(n, colP[i])
 
             // Check whether element (col_P[i], n) is present
@@ -552 +562 @@
       int stopLyingIter = 0, int momSwitchIter = 0, double momentum = .5,
       double finalMomentum = .8, double eta = 10.0
-      ) {
+    ) {
       var state = CreateState(data, distance, random, newDimensions, perplexity,
         theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta);
@@ -565 +575 @@
       int newDimensions = 2, double perplexity = 25, double theta = 0,
       int stopLyingIter = 0, int momSwitchIter = 0, double momentum = .5,
-      double finalMomentum = .8, double eta = 10.0
-      ) {
-      return new TSNEState(data, distance, random, newDimensions, perplexity, theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta);
+      double finalMomentum = .8, double eta = 10.0, bool randomInit = true
+    ) {
+      return new TSNEState(data, distance, random, newDimensions, perplexity, theta, stopLyingIter, momSwitchIter, momentum, finalMomentum, eta, randomInit);
     }
 
@@ -580 +590 @@
         for (var j = 0; j < state.newDimensions; j++) {
           state.gains[i, j] = Math.Sign(state.dY[i, j]) != Math.Sign(state.uY[i, j])
-            ? state.gains[i, j] + .2  // +0.2 nd *0.8 are used in two separate implementations of tSNE -> seems to be correct
+            ? state.gains[i, j] + .2 // +0.2 nd *0.8 are used in two separate implementations of tSNE -> seems to be correct
             : state.gains[i, j] * .8;
 
@@ -590 +600 @@
       // Perform gradient update (with momentum and gains)
       for (var i = 0; i < state.noDatapoints; i++)
-        for (var j = 0; j < state.newDimensions; j++)
-          state.uY[i, j] = state.currentMomentum * state.uY[i, j] - state.eta * state.gains[i, j] * state.dY[i, j];
+      for (var j = 0; j < state.newDimensions; j++)
+        state.uY[i, j] = state.currentMomentum * state.uY[i, j] - state.eta * state.gains[i, j] * state.dY[i, j];
 
       for (var i = 0; i < state.noDatapoints; i++)
-        for (var j = 0; j < state.newDimensions; j++)
-          state.newData[i, j] = state.newData[i, j] + state.uY[i, j];
+      for (var j = 0; j < state.newDimensions; j++)
+        state.newData[i, j] = state.newData[i, j] + state.uY[i, j];
 
       // Make solution zero-mean
@@ -604 +614 @@
         if (state.exact)
           for (var i = 0; i < state.noDatapoints; i++)
-            for (var j = 0; j < state.noDatapoints; j++)
-              state.p[i, j] /= 12.0;
+          for (var j = 0; j < state.noDatapoints; j++)
+            state.p[i, j] /= 12.0;
         else
           for (var i = 0; i < state.rowP[state.noDatapoints]; i++)
@@ -634 +644 @@
       // Compute final t-SNE gradient
       for (var i = 0; i < n; i++)
-        for (var j = 0; j < d; j++) {
-          dC[i, j] = posF[i, j] - negF[i, j] / sumQ;
-        }
+      for (var j = 0; j < d; j++) {
+        dC[i, j] = posF[i, j] - negF[i, j] / sumQ;
+      }
     }