Github user jkbradley commented on a diff in the pull request:
https://github.com/apache/spark/pull/4677#discussion_r25140445
--- Diff:
mllib/src/test/scala/org/apache/spark/mllib/tree/GradientBoostedTreesSuite.scala
---
@@ -158,6 +158,63 @@ class GradientBoostedTreesSuite extends FunSuite with
MLlibTestSparkContext {
}
}
}
+
+ test("runWithValidation performs better on a validation dataset
(Regression)") {
+ // Set numIterations large enough so that it early stops.
+ val numIterations = 20
+ val trainRdd = sc.parallelize(GradientBoostedTreesSuite.trainData, 2)
+ val validateRdd =
sc.parallelize(GradientBoostedTreesSuite.validateData, 2)
+
+ val treeStrategy = new Strategy(algo = Regression, impurity =
Variance, maxDepth = 2,
+ categoricalFeaturesInfo = Map.empty)
+ Array(SquaredError, AbsoluteError).foreach { error =>
+ val boostingStrategy =
+ new BoostingStrategy(treeStrategy, error, numIterations,
validationTol = 0.0)
+
+ val gbtValidate = new
GradientBoostedTrees(boostingStrategy).runWithValidation(
+ trainRdd, validateRdd)
+ assert(gbtValidate.numTrees != numIterations)
+
+ val gbt = GradientBoostedTrees.train(trainRdd, boostingStrategy)
+ val errorWithoutValidation = error.computeError(gbt, validateRdd)
+ val errorWithValidation = error.computeError(gbtValidate,
validateRdd)
+ assert(errorWithValidation < errorWithoutValidation)
+ }
+ }
+
+ test("runWithValidation performs better on a validation dataset
(Classification)") {
+ // Set numIterations large enough so that it early stops.
+ val numIterations = 20
+ val trainRdd = sc.parallelize(GradientBoostedTreesSuite.trainData, 2)
+ val validateRdd =
sc.parallelize(GradientBoostedTreesSuite.validateData, 2)
+
+ val treeStrategy = new Strategy(algo = Classification, impurity =
Variance, maxDepth = 2,
+ categoricalFeaturesInfo = Map.empty)
+ val boostingStrategy =
+ new BoostingStrategy(treeStrategy, LogLoss, numIterations,
validationTol = 0.0)
+
+ // Test that it stops early.
+ val gbtValidate = new
GradientBoostedTrees(boostingStrategy).runWithValidation(
+ trainRdd, validateRdd)
+ assert(gbtValidate.numTrees != numIterations)
+
+ // Remap labels to {-1, 1}
+ val remappedInput = validateRdd.map(x => new LabeledPoint(2 * x.label
- 1, x.features))
+
+ // The error checked for internally in the GradientBoostedTrees is
based on Regression.
+ // Hence for the validation model, the Classification error need not
be strictly less than
+ // that done with validation.
+ val gbtValidateRegressor = new GradientBoostedTreesModel(
--- End diff --
Hm, I misunderstood this the first time you asked about it. It's weird to
create a regression model and test using LogLoss. I would test on validateRdd,
not on trainRdd. That's really all we need to check. And it should let you
keep the model a Classification model.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
