Changeset 8455 for trunk/sources/HeuristicLab.Algorithms.DataAnalysis

Timestamp:

08/09/12 09:33:23 (12 years ago)

Author:

gkronber

Message:

#1902 changed calculation of gradients for covariance functions to reduce allocations of arrays

Location:

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess

Files:

• CovarianceLinear.cs (modified) (2 diffs)
• CovariancePeriodic.cs (modified) (1 diff)
• CovarianceProd.cs (modified) (2 diffs)
• CovarianceSEard.cs (modified) (1 diff)
• CovarianceSEiso.cs (modified) (1 diff)
• CovarianceSum.cs (modified) (1 diff)
• GaussianProcessModel.cs (modified) (1 diff)
• ICovarianceFunction.cs (modified) (1 diff)

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/CovarianceLinear.cs

@@ -29 +29 @@
   [Item(Name = "CovarianceLinear", Description = "Linear covariance function for Gaussian processes.")]
   public class CovarianceLinear : Item, ICovarianceFunction {
-    private static readonly double[] emptyArray = new double[0];
-
     [Storable]
     private double[,] x;
@@ -89 +87 @@
     }
 
-    public double[] GetGradient(int i, int j) {
-      return emptyArray;
+    public double GetGradient(int i, int j, int k) {
+      throw new NotSupportedException("CovarianceLinear does not have hyperparameters.");
     }
 

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/CovariancePeriodic.cs

@@ -117 +117 @@
     }
 
-    public double[] GetGradient(int i, int j) {
+    public double GetGradient(int i, int j, int k) {
+      double v = Math.PI * sd[i, j] / p;
+      switch (k) {
+        case 0: {
+            double newK = Math.Sin(v) / l;
+            newK = newK * newK;
+            return 4 * sf2 * Math.Exp(-2 * newK) * newK;
+          }
+        case 1: {
+            double r = Math.Sin(v) / l;
+            return 4 * sf2 / l * Math.Exp(-2 * r * r) * r * Math.Cos(v) * v;
+          }
+        case 2: {
+            double newK = Math.Sin(v) / l;
+            newK = newK * newK;
+            return 2 * sf2 * Math.Exp(-2 * newK);
 
-      var res = new double[3];
-      double k = sd[i, j];
-      k = Math.PI * k / p;
-      {
-        double newK = Math.Sin(k) / l;
-        newK = newK * newK;
-        res[0] = 4 * sf2 * Math.Exp(-2 * newK) * newK;
+          }
+        default: {
+            throw new ArgumentException("CovariancePeriodic only has three hyperparameters.", "k");
+          }
       }
-      {
-        double r = Math.Sin(k) / l;
-        res[1] = 4 * sf2 / l * Math.Exp(-2 * r * r) * r * Math.Cos(k) * k;
-      }
-      {
-        double newK = Math.Sin(k) / l;
-        newK = newK * newK;
-        res[2] = 2 * sf2 * Math.Exp(-2 * newK);
-      }
-
-      return res;
     }
 

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/CovarianceProd.cs

@@ -86 +86 @@
     }
 
-    public double[] GetGradient(int i, int j) {
-      return Enumerable.Range(0, GetNumberOfParameters(numberOfVariables)).Select(k => GetGradient(i, j, k)).ToArray();
-    }
     public double GetGradient(int i, int j, int k) {
       // map from parameter index to factor
@@ -97 +94 @@
         var f = factors[ii];
         if (ii == vi[k]) {
-          res *= f.GetGradient(i, j)[jj];
+          res *= f.GetGradient(i, j, jj);
         } else {
           res *= f.GetCovariance(i, j);

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/CovarianceSEard.cs

@@ -99 +99 @@
     }
 
-    public double[] GetGradient(int i, int j) {
-      var res = new double[l.Length + 1];
-      for (int k = 0; k < l.Length; k++) {
+    public double GetGradient(int i, int j, int k) {
+      if (k < l.Length) {
         double sqrDist = Util.SqrDist(x[i, k] / l[k], xt[j, k] / l[k]);
-
-        res[k] = sf2 * Math.Exp(-sd[i, j] / 2.0) * sqrDist;
+        return sf2 * Math.Exp(-sd[i, j] / 2.0) * sqrDist;
+      } else if (k == l.Length) {
+        return 2.0 * sf2 * Math.Exp(-sd[i, j] / 2.0);
+      } else {
+        throw new ArgumentException("CovarianceSEard has dimension+1 hyperparameters.", "k");
       }
-      res[res.Length - 1] = 2.0 * sf2 * Math.Exp(-sd[i, j] / 2.0);
-      return res;
     }
 

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/CovarianceSEiso.cs

@@ -101 +101 @@
     }
 
-    public double[] GetGradient(int i, int j) {
-      var res = new double[2];
-      res[0] = sf2 * Math.Exp(-sd[i, j] / 2.0) * sd[i, j];
-      res[1] = 2.0 * sf2 * Math.Exp(-sd[i, j] / 2.0);
-      return res;
+    public double GetGradient(int i, int j, int k) {
+      switch (k) {
+        case 0: return sf2 * Math.Exp(-sd[i, j] / 2.0) * sd[i, j];
+        case 1: return 2.0 * sf2 * Math.Exp(-sd[i, j] / 2.0);
+        default: throw new ArgumentException("CovarianceSEiso has two hyperparameters", "k");
+      }
     }
 

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/CovarianceSum.cs

@@ -86 +86 @@
     }
 
-    public double[] GetGradient(int i, int j) {
-      return terms.Select(t => t.GetGradient(i, j)).SelectMany(seq => seq).ToArray();
+    public double GetGradient(int i, int j, int k) {
+      int ii = 0;
+      while (k > terms[ii].GetNumberOfParameters(numberOfVariables)) {
+        k -= terms[ii].GetNumberOfParameters(numberOfVariables);
+      }
+      return terms[ii].GetGradient(i, j, k);
     }
   }

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/GaussianProcessModel.cs

@@ -188 +188 @@
         for (int i = 0; i < n; i++) {
           for (int j = 0; j < n; j++) {
-            var covDeriv = covarianceFunction.GetGradient(i, j);
             for (int k = 0; k < covGradients.Length; k++) {
-              covGradients[k] += q[i, j] * covDeriv[k];
+              covGradients[k] += q[i, j] * covarianceFunction.GetGradient(i, j, k);
             }
           }

trunk/sources/HeuristicLab.Algorithms.DataAnalysis/3.4/GaussianProcess/ICovarianceFunction.cs

@@ -30 +30 @@
 
     double GetCovariance(int i, int j);
-    double[] GetGradient(int i, int j);
+    double GetGradient(int i, int j, int k);
   }
 }