ReForeSt

API Docs

How to learn a random forest classification model with ReForeSt

In order to use ReForeSt it is required to add the ReForeSt and the Apache Spark jar files in your classpath.

import reforest.rf.{RFProperty, RFRunner}

// Create the ReForeSt configuration.

val property = RFParameterBuilder.apply

  .addParameter(RFParameterType.Dataset, "data/test10k-labels")

  .addParameter(RFParameterType.NumFeatures, 794)

  .addParameter(RFParameterType.NumClasses, 10)

  .addParameter(RFParameterType.NumTrees, 100)

  .addParameter(RFParameterType.Depth, 10)

  .addParameter(RFParameterType.BinNumber, 32)

  .addParameter(RFParameterType.SparkMaster, "local[4]")

  .addParameter(RFParameterType.SparkCoresMax, 4)

  .addParameter(RFParameterType.SparkPartition, 4 * 4)

  .addParameter(RFParameterType.SparkExecutorMemory, "4096m")

  .addParameter(RFParameterType.SparkExecutorInstances, 1)

  .build

val sc = CCUtil.getSparkContext(property)

// Create the Random Forest classifier.

val timeStart = System.currentTimeMillis()

val rfRunner = ReForeStTrainerBuilder.apply(property).build(sc)

// Train a Random Forest model.

val model = rfRunner.trainClassifier()

val timeEnd = System.currentTimeMillis()

// Evaluate model on test instances and compute test error

val labelAndPreds = rfRunner.getDataLoader.getTestingData.map { point =>

  val prediction = model.predict(point.features)

  (point.label, prediction)

val testErr = labelAndPreds.filter(r => r._1 != r._2).count.toDouble / rfRunner.getDataLoader.getTestingData.count()

println("Accuracy: "+(1 - testErr))

println("Time: " + (timeEnd - timeStart))

rfRunner.sparkStop()

How to perform Model Selection

Since version v1.2, ReForeSt support Model Selection.

In order to enable ReForeSt model selection it is required to activate the functionality and to provide a list of values for each of the parameters. In the following is presented how to configure ReForeSt to enable the model selection.

val property = RFParameterBuilder.apply

  .addParameter(RFParameterType.Dataset, "data/test10k-labels")

  .addParameter(RFParameterType.NumFeatures, 794)

  .addParameter(RFParameterType.NumClasses, 10)

  .addParameter(RFParameterType.ModelSelectionEpsilon, 0.001)

  .addParameter(RFParameterType.ModelSelectionEpsilonRemove, 0.02)

  .addParameter(RFParameterType.NumTrees, Array(50, 80, 100))

  .addParameter(RFParameterType.Depth, Array(5, 10))

  .addParameter(RFParameterType.FeatureMultiplierPerNode, Array(1d, 2d))

  .addParameter(RFParameterType.BinNumber, Array(8, 16, 32))

  .addParameter(RFParameterType.SparkMaster, "local[4]")

  .addParameter(RFParameterType.SparkCoresMax, 4)

  .addParameter(RFParameterType.SparkPartition, 4 * 4)

  .addParameter(RFParameterType.SparkExecutorMemory, "4096m")

  .addParameter(RFParameterType.SparkExecutorInstances, 1)

  .addParameter(RFParameterType.ModelSelection, true)

  .build

How to perform Random Rotations

The Random Forest Rotation is a new ensemble method that extends the classic Random Forest formulation. The original definition can be found at: http://www.jmlr.org/papers/volume17/blaser16a/blaser16a.pdf

Since version v.1.1, ReForeSt support Random Forest Rotation. In order to activate the support for the Random Rotations it is required to activate its support as following.

val property = RFParameterBuilder.apply

  .addParameter(RFParameterType.Dataset, "data/sample-covtype.libsvm")

  .addParameter(RFParameterType.NumFeatures, 54)

  .addParameter(RFParameterType.NumClasses, 2)

  .addParameter(RFParameterType.NumTrees, 100)

  .addParameter(RFParameterType.Depth, 10)

  .addParameter(RFParameterType.BinNumber, 32)

  .addParameter(RFParameterType.SparkMaster, "local[4]")

  .addParameter(RFParameterType.SparkCoresMax, 4)

  .addParameter(RFParameterType.SparkPartition, 4 * 4)

  .addParameter(RFParameterType.SparkExecutorMemory, "4096m")

  .addParameter(RFParameterType.SparkExecutorInstances, 1)

  .addParameter(RFParameterType.Rotation, true)

  .addParameter(RFParameterType.NumRotations, 4)

  .build

ReForeSt has been developed at Smartlab - DIBRIS - University of Genoa

Google Sites

Report abuse