First test for a multithreaded svm learner

mikerabat · Jan 26, 2022 · 9f2dff5 · 9f2dff5
1 parent 1cfc94e
commit 9f2dff5
Showing 1 changed file with 35 additions and 9 deletions.
diff --git a/SimpleClassifier/SVM.pas b/SimpleClassifier/SVM.pas
@@ -99,6 +99,7 @@   TSVMLearner = class(TCustomWeightedLearner)
 
     faStar : IMatrix;
     fBias : IMatrix;
+    fMtxClass : TMatrixClass;
 
     fVectIdx : TIntegerDynArray;
 
@@ -117,18 +118,23 @@   TSVMLearner = class(TCustomWeightedLearner)
     // todo: add support for weighted learning
     function DoLearn(const weights : Array of double) : TCustomClassifier; override;
   public
+    class var Threaded : boolean;
+
     property Margin : double read GetMargin;  // just a property
 
     procedure AfterConstruction; override;
 
     procedure SetProps(const Props : TSVMProps);
 
     class function CanLearnClassifier(Classifier : TCustomClassifierClass) : boolean; override;
+
+    constructor Create;
+    destructor Destroy; override;
   end;
 
 implementation
 
-uses BaseMatrixExamples, MatrixConst, MathUtilFunc, math;
+uses BaseMatrixExamples, MatrixConst, MathUtilFunc, math, ThreadedMatrix;
 
 procedure TanhFunc(var value : double);
 begin
@@ -162,15 +168,14 @@ function TSVMLearner.TrainSet(doAugment : boolean): IMatrix;
      if doAugment then
         augsize := 1;
 
-     Result := TDoubleMatrix.Create(DataSet.Count, DataSet.Example[0].FeatureVec.FeatureVecLen + augsize, 1);
+     Result := fMtxClass.MatrixClass.Create(DataSet.Count, DataSet.Example[0].FeatureVec.FeatureVecLen + augsize, 1);
 
      if DataSet is TMatrixLearnerExampleList
      then
          Result.AssignSubMatrix(TMatrixLearnerExampleList(DataSet).Matrix, 0, 0)
      else
      begin
           // Create a matrix of the feature vectors -> this classifier only understands matrices
-          Result := TDoubleMatrix.Create(DataSet.Count, DataSet[0].FeatureVec.FeatureVecLen);
           for y := 0 to DataSet[0].FeatureVec.FeatureVecLen - 1 do
               for x := 0 to DataSet.Count - 1 do
                   Result[x, y] := DataSet[x].FeatureVec[y];
@@ -183,6 +188,20 @@ class function TSVMLearner.CanLearnClassifier(
      Result := Classifier = TSVMClassifier;
 end;
 
+constructor TSVMLearner.Create;
+begin
+     inherited;
+
+     fMtxClass := TMatrixClass.Create;
+end;
+
+destructor TSVMLearner.Destroy;
+begin
+     fMtxClass.Free;
+
+     inherited;
+end;
+
 function TSVMLearner.DoLearn(const weights : Array of double): TCustomClassifier;
 var augExmpl : IMatrix;
     y : IMatrix;
@@ -198,6 +217,12 @@ function TSVMLearner.DoLearn(const weights : Array of double): TCustomClassifier
 
     //fDists : TDoubleDynArray;
 begin
+     if Threaded
+     then
+         fMtxClass.MatrixClass := TThreadedMatrix
+     else
+         fMtxClass.MatrixClass := TDoubleMatrix;
+
      numCl := IndexOfClasses(idx, cl);
      if numCl <> 2 then
         raise ESVMClassException.Create('Error only 2 class problems are supported');
@@ -230,7 +255,7 @@ function TSVMLearner.DoLearn(const weights : Array of double): TCustomClassifier
                   scaleFact[0, counter] := 1;
           end;
 
-          ones := TDoubleMatrix.Create(scaleFact.Width, scaleFact.Height, 1);
+          ones := fMtxClass.MatrixClass.Create(scaleFact.Width, scaleFact.Height, 1);
           scaleFact := ones.ElementWiseDiv(scaleFact);
 
           // now scale the examples!
@@ -269,7 +294,7 @@ function TSVMLearner.DoLearn(const weights : Array of double): TCustomClassifier
      end;
 
      faStar.UseFullMatrix;
-     if Length(fVectIdx) < faStar.Height then
+     if (Length(fVectIdx) < faStar.Height) or Threaded then
      begin
           faStar.SetSubMatrix(0, 0, 1, Length(fVectIdx));
           tmp := TDoubleMatrix.Create;
@@ -437,7 +462,7 @@ procedure TSVMLearner.LearnLeastSquares(y: IMatrix);
      gIdx := nil;    
 
      // create augmented matrix A
-     A := TDoubleMatrix.Create(DataSet.Count + 1, DataSet.Count + 1);
+     A := fMtxClass.MatrixClass.Create(DataSet.Count + 1, DataSet.Count + 1);
      A.SetSubMatrix(1, 0, Length(grIndex), 1);
      A.SetRow(0, grIndex);
      A.SetSubMatrix(0, 1, 1, Length(grIndex));
@@ -479,7 +504,7 @@ function TSVMLearner.PolyKernelData(augTrainSet : IMatrix): IMatrix;
      if fProps.kernelType = svmPolyInhomogen then
         dotproduct.AddInplace(1);
 
-     Result := TDoubleMatrix.Create;
+     Result := fMtxClass.MatrixClass.Create;
      Result.Assign(dotproduct);
 
      for x := 1 to fProps.order - 1 do
@@ -491,7 +516,7 @@ function TSVMLearner.SigmoidKernelData(augTrainSet: IMatrix): IMatrix;
     tTrainSet : IMatrix;
     col : IMatrix;
 begin
-     Result := TDoubleMatrix.Create(augTrainSet.Width, augTrainSet.Width);
+     Result := fMtxClass.MatrixClass.Create(augTrainSet.Width, augTrainSet.Width);
      tTrainSet := augTrainSet.Transpose;
 
      for x := 0 to Result.Height - 1 do
@@ -514,7 +539,7 @@ function TSVMLearner.GaussKernelData(augTrainSet: IMatrix): IMatrix;
     xi, sum : IMatrix;
     val : double;
 begin
-     Result := TDoubleMatrix.Create(augTrainSet.Width, augTrainSet.Width);
+     Result := fMtxClass.MatrixClass.Create(augTrainSet.Width, augTrainSet.Width);
 
      // y(:,i)    = exp(-sum((train_patterns-train_patterns(:,i)*ones(1,Nf)).^2)'/(2*ker_param^2));
      // according to wikipedia k(xi, xj) = exp(-s * ||xi - xj]]^2)   s = 1/(2*sigma^2);
@@ -1008,5 +1033,6 @@ procedure TSVMLearner.SigmoidTraining(cl: TSVMClassifier; var confA, confB: doub
 
 initialization
    RegisterMathIO(TSVMClassifier);
+   TSVMLearner.Threaded := False;
 
 end.