Synchronizing seeds at each iteration. This allows newly provisioned workers to "fast-forward" to the latest iteration and have the same random number generator state as all other workers.

Rogan Carr · Rogan Carr · commit 9f9879c0266e · 2018-06-13T12:32:23.000-07:00
As part of this, the dropout parameters were moved into the constructor for `OptimizationAlgorithm` and made `private` and/or `protected`.
diff --git a/src/Microsoft.ML.FastTree/BoostingFastTree.cs b/src/Microsoft.ML.FastTree/BoostingFastTree.cs
@@ -68,13 +68,16 @@ protected override OptimizationAlgorithm ConstructOptimizationAlgorithm(IChannel
             switch (Args.OptimizationAlgorithm)
             {
                 case BoostedTreeArgs.OptimizationAlgorithmType.GradientDescent:
-                    optimizationAlgorithm = new GradientDescent(Ensemble, TrainSet, InitTrainScores, gradientWrapper);
+                    optimizationAlgorithm = new GradientDescent(Ensemble, TrainSet, InitTrainScores, gradientWrapper,
+                        dropoutRate: Args.DropoutRate, dropoutSeed: Args.RngSeed);
                     break;
                 case BoostedTreeArgs.OptimizationAlgorithmType.AcceleratedGradientDescent:
-                    optimizationAlgorithm = new AcceleratedGradientDescent(Ensemble, TrainSet, InitTrainScores, gradientWrapper);
+                    optimizationAlgorithm = new AcceleratedGradientDescent(Ensemble, TrainSet, InitTrainScores, gradientWrapper,
+                        dropoutRate: Args.DropoutRate, dropoutSeed: Args.RngSeed);
                     break;
                 case BoostedTreeArgs.OptimizationAlgorithmType.ConjugateGradientDescent:
-                    optimizationAlgorithm = new ConjugateGradientDescent(Ensemble, TrainSet, InitTrainScores, gradientWrapper);
+                    optimizationAlgorithm = new ConjugateGradientDescent(Ensemble, TrainSet, InitTrainScores, gradientWrapper,
+                        dropoutRate: Args.DropoutRate, dropoutSeed: Args.RngSeed);
                     break;
                 default:
                     throw ch.Except("Unknown optimization algorithm '{0}'", Args.OptimizationAlgorithm);
@@ -83,8 +86,6 @@ protected override OptimizationAlgorithm ConstructOptimizationAlgorithm(IChannel
             optimizationAlgorithm.TreeLearner = ConstructTreeLearner(ch);
             optimizationAlgorithm.ObjectiveFunction = ConstructObjFunc(ch);
             optimizationAlgorithm.Smoothing = Args.Smoothing;
-            optimizationAlgorithm.DropoutRate = Args.DropoutRate;
-            optimizationAlgorithm.DropoutRng = new Random(Args.RngSeed);
             optimizationAlgorithm.PreScoreUpdateEvent += PrintTestGraph;
 
             return optimizationAlgorithm;
diff --git a/src/Microsoft.ML.FastTree/FastTree.cs b/src/Microsoft.ML.FastTree/FastTree.cs
@@ -377,7 +377,7 @@ protected bool[] GetActiveFeatures()
             if (Args.FeatureFraction < 1.0)
             {
                 if (_featureSelectionRandom == null)
-                    _featureSelectionRandom = new Random(Args.FeatureSelectSeed);
+                    _featureSelectionRandom = new Random(Args.FeatureSelectSeed + Ensemble.NumTrees);
 
                 for (int i = 0; i < TrainSet.NumFeatures; ++i)
                 {
@@ -614,6 +614,9 @@ protected virtual void Train(IChannel ch)
                 {
                     using (Timer.Time(TimerEvent.Iteration))
                     {
+                        // Reset Seeds
+                        _featureSelectionRandom = new Random(Args.FeatureSelectSeed + Ensemble.NumTrees);
+
 #if NO_STORE
                         bool[] activeFeatures = GetActiveFeatures();
 #else
diff --git a/src/Microsoft.ML.FastTree/RandomForest.cs b/src/Microsoft.ML.FastTree/RandomForest.cs
@@ -29,9 +29,6 @@ protected override OptimizationAlgorithm ConstructOptimizationAlgorithm(IChannel
             optimizationAlgorithm.TreeLearner = ConstructTreeLearner(ch);
             optimizationAlgorithm.ObjectiveFunction = ConstructObjFunc(ch);
             optimizationAlgorithm.Smoothing = Args.Smoothing;
-            // No notion of dropout for non-boosting applications.
-            optimizationAlgorithm.DropoutRate = 0;
-            optimizationAlgorithm.DropoutRng = null;
             optimizationAlgorithm.PreScoreUpdateEvent += PrintTestGraph;
 
             return optimizationAlgorithm;
diff --git a/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/AcceleratedGradientDescent.cs b/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/AcceleratedGradientDescent.cs
@@ -7,8 +7,9 @@ namespace Microsoft.ML.Runtime.FastTree.Internal
     //Accelerated gradient descent score tracker
     public class AcceleratedGradientDescent : GradientDescent
     {
-        public AcceleratedGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
-            : base(ensemble, trainData, initTrainScores, gradientWrapper)
+        public AcceleratedGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,
+            double dropoutRate = 0, int dropoutSeed = int.MinValue)
+            : base(ensemble, trainData, initTrainScores, gradientWrapper, dropoutRate, dropoutSeed)
         {
             UseFastTrainingScoresUpdate = false;
         }
diff --git a/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/ConjugateGradientDescent.cs b/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/ConjugateGradientDescent.cs
@@ -11,8 +11,9 @@ public class ConjugateGradientDescent : GradientDescent
         private double[] _currentGradient;
         private double[] _currentDk;
 
-        public ConjugateGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
-            : base(ensemble, trainData, initTrainScores, gradientWrapper)
+        public ConjugateGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,
+            double dropoutRate = 0, int dropoutSeed = int.MinValue)
+            : base(ensemble, trainData, initTrainScores, gradientWrapper, dropoutRate, dropoutSeed)
         {
             _currentDk = new double[trainData.NumDocs];
         }
diff --git a/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/GradientDescent.cs b/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/GradientDescent.cs
@@ -21,8 +21,9 @@ public class GradientDescent : OptimizationAlgorithm
         private double[] _droppedScores;
         private double[] _scores;
 
-        public GradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
-            : base(ensemble, trainData, initTrainScores)
+        public GradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,
+            double dropoutRate = 0, int dropoutSeed = int.MinValue)
+            : base(ensemble, trainData, initTrainScores, dropoutRate, dropoutSeed)
         {
             _gradientWrapper = gradientWrapper;
             _treeScores = new List<double[]>();
@@ -36,7 +37,11 @@ protected override ScoreTracker ConstructScoreTracker(string name, Dataset set,
         protected virtual double[] GetGradient(IChannel ch)
         {
             Contracts.AssertValue(ch);
-            if (DropoutRate > 0)
+
+            // Assumes that GetGradient is called at most once per iteration
+            ResetDropoutSeed();
+
+            if (_dropoutRate > 0)
             {
                 if (_droppedScores == null)
                     _droppedScores = new double[TrainingScores.Scores.Length];
@@ -46,16 +51,16 @@ protected virtual double[] GetGradient(IChannel ch)
                     _scores = new double[TrainingScores.Scores.Length];
                 int numberOfTrees = Ensemble.NumTrees;
                 int[] droppedTrees =
-                    Enumerable.Range(0, numberOfTrees).Where(t => (DropoutRng.NextDouble() < DropoutRate)).ToArray();
+                    Enumerable.Range(0, numberOfTrees).Where(t => (_dropoutRng.NextDouble() < _dropoutRate)).ToArray();
                 _numberOfDroppedTrees = droppedTrees.Length;
                 if ((_numberOfDroppedTrees == 0) && (numberOfTrees > 0))
                 {
-                    droppedTrees = new int[] { DropoutRng.Next(numberOfTrees) };
+                    droppedTrees = new int[] { _dropoutRng.Next(numberOfTrees) };
                     // force at least a single tree to be dropped 
                     _numberOfDroppedTrees = droppedTrees.Length;
                 }
                 ch.Trace("dropout: Dropping {0} trees of {1} for rate {2}",
-                    _numberOfDroppedTrees, numberOfTrees, DropoutRate);
+                    _numberOfDroppedTrees, numberOfTrees, _dropoutRate);
                 foreach (int i in droppedTrees)
                 {
                     double[] s = _treeScores[i];
@@ -94,7 +99,7 @@ public override RegressionTree TrainingIteration(IChannel ch, bool[] activeFeatu
         {
             Contracts.CheckValue(ch, nameof(ch));
             // Fit a regression tree to the gradient using least squares.
-            RegressionTree tree = TreeLearner.FitTargets(ch, activeFeatures, AdjustTargetsAndSetWeights(ch));
+            RegressionTree tree = TreeLearner.FitTargets(ch, activeFeatures, AdjustTargetsAndSetWeights(ch), iteration: Iteration);
             if (tree == null)
                 return null; // Could not learn a tree. Exit.
 
@@ -105,7 +110,7 @@ public override RegressionTree TrainingIteration(IChannel ch, bool[] activeFeatu
             {
                 double[] backupScores = null;
                 // when doing dropouts we need to replace the TrainingScores with the scores without the dropped trees 
-                if (DropoutRate > 0)
+                if (_dropoutRate > 0)
                 {
                     backupScores = TrainingScores.Scores;
                     TrainingScores.Scores = _scores;
@@ -117,7 +122,7 @@ public override RegressionTree TrainingIteration(IChannel ch, bool[] activeFeatu
                     (ObjectiveFunction as IStepSearch).AdjustTreeOutputs(ch, tree, TreeLearner.Partitioning, TrainingScores);
                 else
                     throw ch.Except("No AdjustTreeOutputs defined. Objective function should define IStepSearch or AdjustTreeOutputsOverride should be set");
-                if (DropoutRate > 0)
+                if (_dropoutRate > 0)
                 {
                     // Returning the original scores.
                     TrainingScores.Scores = backupScores;
@@ -128,7 +133,7 @@ public override RegressionTree TrainingIteration(IChannel ch, bool[] activeFeatu
                 SmoothTree(tree, Smoothing);
                 UseFastTrainingScoresUpdate = false;
             }
-            if (DropoutRate > 0)
+            if (_dropoutRate > 0)
             {
                 // Don't do shrinkage if you do dropouts.
                 double scaling = (1.0 / (1.0 + _numberOfDroppedTrees));
diff --git a/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/NoOptimizationAlgorithm.cs b/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/NoOptimizationAlgorithm.cs
@@ -12,8 +12,9 @@ public class RandomForestOptimizer : GradientDescent
     {
         private IGradientAdjuster _gradientWrapper;
         // REVIEW: When the FastTree appliation is decoupled with tree learner and boosting logic, this class should be removed.
-        public RandomForestOptimizer(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
-            : base(ensemble, trainData, initTrainScores, gradientWrapper)
+        public RandomForestOptimizer(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,
+            double dropoutRate = 0, int dropoutSeed = int.MinValue)
+            : base(ensemble, trainData, initTrainScores, gradientWrapper, dropoutRate, dropoutSeed)
         {
             _gradientWrapper = gradientWrapper;
         }
@@ -32,7 +33,7 @@ public override RegressionTree TrainingIteration(IChannel ch, bool[] activeFeatu
             double[] targets = GetGradient(ch);
             double[] weightedTargets = _gradientWrapper.AdjustTargetAndSetWeights(targets, ObjectiveFunction, out sampleWeights);
             RegressionTree tree = ((RandomForestLeastSquaresTreeLearner)TreeLearner).FitTargets(ch, activeFeatures, weightedTargets,
-                targets, sampleWeights);
+                targets, sampleWeights, iteration: Iteration);
 
             if (tree != null)
                 Ensemble.AddTree(tree);
diff --git a/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/OptimizationAlgorithm.cs b/src/Microsoft.ML.FastTree/Training/OptimizationAlgorithms/OptimizationAlgorithm.cs
@@ -32,17 +32,22 @@ public abstract class OptimizationAlgorithm
 
         public IStepSearch AdjustTreeOutputsOverride; // if set it overrides IStepSearch possibly implemented by ObejctiveFunctionBase
         public double Smoothing;
-        public double DropoutRate;
-        public Random DropoutRng;
+        protected double _dropoutRate;
+        protected Random _dropoutRng;
+        private int _dropoutSeed;
         public bool UseFastTrainingScoresUpdate;
 
-        public OptimizationAlgorithm(Ensemble ensemble, Dataset trainData, double[] initTrainScores)
+        public OptimizationAlgorithm(Ensemble ensemble, Dataset trainData, double[] initTrainScores,
+            double dropoutRate = 0, int dropoutSeed = int.MinValue)
         {
             Ensemble = ensemble;
             TrainingScores = ConstructScoreTracker("train", trainData, initTrainScores);
             TrackedScores = new List<ScoreTracker>();
             TrackedScores.Add(TrainingScores);
-            DropoutRng = new Random();
+            if (dropoutSeed != int.MinValue)
+                _dropoutRng = new Random(dropoutSeed);
+            _dropoutSeed = dropoutSeed;
+            _dropoutRate = dropoutRate;
             UseFastTrainingScoresUpdate = true;
         }
 
@@ -117,5 +122,12 @@ public virtual void FinalizeLearning(int bestIteration)
                 TrackedScores.Clear();  //Invalidate all precomputed scores as they are not valid anymore //slow method of score computation will be used instead
             }
         }
+
+        protected void ResetDropoutSeed()
+        {
+            _dropoutRng = new Random(_dropoutSeed + Iteration);
+        }
+
+        protected int Iteration => Ensemble.NumTrees;
     }
 }
diff --git a/src/Microsoft.ML.FastTree/Training/TreeLearners/FastForestLeastSquaresTreeLearner.cs b/src/Microsoft.ML.FastTree/Training/TreeLearners/FastForestLeastSquaresTreeLearner.cs
@@ -28,9 +28,11 @@ protected override RegressionTree NewTree()
             return new QuantileRegressionTree(NumLeaves);
         }
 
-        public RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, Double[] weightedtargets, Double[] targets, Double[] weights)
+        public RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, Double[] weightedtargets, Double[] targets, Double[] weights, int iteration = 0)
         {
-            var tree = (QuantileRegressionTree)FitTargets(ch, activeFeatures, weightedtargets);
+            ResetRandomNumberGenerator(iteration);
+
+            var tree = (QuantileRegressionTree)FitTargets(ch, activeFeatures, weightedtargets, iteration: iteration);
             if (tree != null && _quantileEnabled)
             {
                 Double[] distributionWeights = null;
diff --git a/src/Microsoft.ML.FastTree/Training/TreeLearners/LeastSquaresRegressionTreeLearner.cs b/src/Microsoft.ML.FastTree/Training/TreeLearners/LeastSquaresRegressionTreeLearner.cs
@@ -62,7 +62,8 @@ public class LeastSquaresRegressionTreeLearner : TreeLearner
         // how many times each feature has been split on, for diversity penalty
         protected readonly int[] FeatureUseCount;
 
-        protected readonly Random Rand;
+        protected Random Rand;
+        private readonly int _randomSeed;
 
         protected readonly double SplitFraction;
         protected readonly bool FilterZeros;
@@ -163,6 +164,7 @@ public LeastSquaresRegressionTreeLearner(Dataset trainData, int numLeaves, int m
 
             FeatureUseCount = new int[TrainData.NumFeatures];
             Rand = new Random(randomSeed);
+            _randomSeed = randomSeed;
             SplitFraction = splitFraction;
             FilterZeros = filterZeros;
             BsrMaxTreeOutput = bsrMaxTreeOutput;
@@ -213,7 +215,7 @@ protected virtual void MakeDummyRootSplit(RegressionTree tree, double rootTarget
         /// Learns a new tree for the current outputs
         /// </summary>
         /// <returns>A regression tree</returns>
-        public sealed override RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, double[] targets)
+        public sealed override RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, double[] targets, int iteration = 0)
         {
             int maxLeaves = base.NumLeaves;
             using (Timer.Time(TimerEvent.TreeLearnerGetTree))
@@ -224,7 +226,7 @@ public sealed override RegressionTree FitTargets(IChannel ch, bool[] activeFeatu
                 tree.ActiveFeatures = (bool[])activeFeatures.Clone();
 
                 // clear memory
-                Initialize(activeFeatures);
+                Initialize(activeFeatures, iteration);
 
                 // find the best split of the root node.
                 FindBestSplitOfRoot(targets);
@@ -318,14 +320,25 @@ protected virtual void PerformSplit(RegressionTree tree, int bestLeaf, double[]
         /// <summary>
         /// Clears data structures
         /// </summary>
-        private void Initialize(bool[] activeFeatures)
+        private void Initialize(bool[] activeFeatures, int iteration)
         {
+            // Synchronize the Random Number Generator
+            ResetRandomNumberGenerator(iteration);
             _parallelTraining.InitIteration(ref activeFeatures);
             ActiveFeatures = activeFeatures;
             HistogramArrayPool.Reset();
             Partitioning.Initialize();
         }
 
+        /// <summary>
+        /// Reset the random number generator to synchronize at iteration boundaries
+        /// </summary>
+        /// <param name="iteration">The iteration</param>
+        protected void ResetRandomNumberGenerator(int iteration)
+        {
+            Rand = new Random(_randomSeed + iteration);
+        }
+
         protected bool HasWeights => TrainData?.SampleWeights != null;
 
         protected double[] GetTargetWeights()
diff --git a/src/Microsoft.ML.FastTree/Training/TreeLearners/TreeLearner.cs b/src/Microsoft.ML.FastTree/Training/TreeLearners/TreeLearner.cs
@@ -22,7 +22,7 @@ protected TreeLearner(Dataset trainData, int numLeaves)
 
         public static string TargetWeightsDatasetName { get { return "TargetWeightsDataset"; } }
 
-        public abstract RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, double[] targets);
+        public abstract RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, double[] targets, int iteration = 0);
 
         /// <summary>
         /// Get size of reserved memory for the tree learner.

Original file line number	Diff line number	Diff line change
`@@ -377,7 +377,7 @@ protected bool[] GetActiveFeatures()`
`377`	`377`	`if (Args.FeatureFraction < 1.0)`
`378`	`378`	`{`
`379`	`379`	`if (_featureSelectionRandom == null)`
`380`		`- _featureSelectionRandom = new Random(Args.FeatureSelectSeed);`
	`380`	`+ _featureSelectionRandom = new Random(Args.FeatureSelectSeed + Ensemble.NumTrees);`
`381`	`381`
`382`	`382`	`for (int i = 0; i < TrainSet.NumFeatures; ++i)`
`383`	`383`	`{`
`@@ -614,6 +614,9 @@ protected virtual void Train(IChannel ch)`
`614`	`614`	`{`
`615`	`615`	`using (Timer.Time(TimerEvent.Iteration))`
`616`	`616`	`{`
	`617`	`+ // Reset Seeds`
	`618`	`+ _featureSelectionRandom = new Random(Args.FeatureSelectSeed + Ensemble.NumTrees);`
	`619`	`+`
`617`	`620`	`#if NO_STORE`
`618`	`621`	`bool[] activeFeatures = GetActiveFeatures();`
`619`	`622`	`#else`
Original file line number	Diff line number	Diff line change
`@@ -7,8 +7,9 @@ namespace Microsoft.ML.Runtime.FastTree.Internal`
`7`	`7`	`//Accelerated gradient descent score tracker`
`8`	`8`	`public class AcceleratedGradientDescent : GradientDescent`
`9`	`9`	`{`
`10`		`- public AcceleratedGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)`
`11`		`- : base(ensemble, trainData, initTrainScores, gradientWrapper)`
	`10`	`+ public AcceleratedGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,`
	`11`	`+ double dropoutRate = 0, int dropoutSeed = int.MinValue)`
	`12`	`+ : base(ensemble, trainData, initTrainScores, gradientWrapper, dropoutRate, dropoutSeed)`
`12`	`13`	`{`
`13`	`14`	`UseFastTrainingScoresUpdate = false;`
`14`	`15`	`}`
Original file line number	Diff line number	Diff line change
`@@ -11,8 +11,9 @@ public class ConjugateGradientDescent : GradientDescent`
`11`	`11`	`private double[] _currentGradient;`
`12`	`12`	`private double[] _currentDk;`
`13`	`13`
`14`		`- public ConjugateGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)`
`15`		`- : base(ensemble, trainData, initTrainScores, gradientWrapper)`
	`14`	`+ public ConjugateGradientDescent(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,`
	`15`	`+ double dropoutRate = 0, int dropoutSeed = int.MinValue)`
	`16`	`+ : base(ensemble, trainData, initTrainScores, gradientWrapper, dropoutRate, dropoutSeed)`
`16`	`17`	`{`
`17`	`18`	`_currentDk = new double[trainData.NumDocs];`
`18`	`19`	`}`
Original file line number	Diff line number	Diff line change
`@@ -12,8 +12,9 @@ public class RandomForestOptimizer : GradientDescent`
`12`	`12`	`{`
`13`	`13`	`private IGradientAdjuster _gradientWrapper;`
`14`	`14`	`// REVIEW: When the FastTree appliation is decoupled with tree learner and boosting logic, this class should be removed.`
`15`		`- public RandomForestOptimizer(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)`
`16`		`- : base(ensemble, trainData, initTrainScores, gradientWrapper)`
	`15`	`+ public RandomForestOptimizer(Ensemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper,`
	`16`	`+ double dropoutRate = 0, int dropoutSeed = int.MinValue)`
	`17`	`+ : base(ensemble, trainData, initTrainScores, gradientWrapper, dropoutRate, dropoutSeed)`
`17`	`18`	`{`
`18`	`19`	`_gradientWrapper = gradientWrapper;`
`19`	`20`	`}`
`@@ -32,7 +33,7 @@ public override RegressionTree TrainingIteration(IChannel ch, bool[] activeFeatu`
`32`	`33`	`double[] targets = GetGradient(ch);`
`33`	`34`	`double[] weightedTargets = _gradientWrapper.AdjustTargetAndSetWeights(targets, ObjectiveFunction, out sampleWeights);`
`34`	`35`	`RegressionTree tree = ((RandomForestLeastSquaresTreeLearner)TreeLearner).FitTargets(ch, activeFeatures, weightedTargets,`
`35`		`- targets, sampleWeights);`
	`36`	`+ targets, sampleWeights, iteration: Iteration);`
`36`	`37`
`37`	`38`	`if (tree != null)`
`38`	`39`	`Ensemble.AddTree(tree);`
Original file line number	Diff line number	Diff line change
`@@ -28,9 +28,11 @@ protected override RegressionTree NewTree()`
`28`	`28`	`return new QuantileRegressionTree(NumLeaves);`
`29`	`29`	`}`
`30`	`30`
`31`		`- public RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, Double[] weightedtargets, Double[] targets, Double[] weights)`
	`31`	`+ public RegressionTree FitTargets(IChannel ch, bool[] activeFeatures, Double[] weightedtargets, Double[] targets, Double[] weights, int iteration = 0)`
`32`	`32`	`{`
`33`		`- var tree = (QuantileRegressionTree)FitTargets(ch, activeFeatures, weightedtargets);`
	`33`	`+ ResetRandomNumberGenerator(iteration);`
	`34`	`+`
	`35`	`+ var tree = (QuantileRegressionTree)FitTargets(ch, activeFeatures, weightedtargets, iteration: iteration);`
`34`	`36`	`if (tree != null && _quantileEnabled)`
`35`	`37`	`{`
`36`	`38`	`Double[] distributionWeights = null;`