You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@spark.apache.org by me...@apache.org on 2016/04/12 20:27:18 UTC
spark git commit: [SPARK-13322][ML] AFTSurvivalRegression supports
feature standardization
Repository: spark
Updated Branches:
refs/heads/master 75e05a5a9 -> 101663f1a
[SPARK-13322][ML] AFTSurvivalRegression supports feature standardization
## What changes were proposed in this pull request?
AFTSurvivalRegression should support feature standardization, it will improve the convergence rate.
Test the convergence rate on the [Ovarian](https://stat.ethz.ch/R-manual/R-devel/library/survival/html/ovarian.html) data which is standard data comes with Survival library in R,
* without standardization(before this PR) -> 74 iterations.
* with standardization(after this PR) -> 38 iterations.
But after this fix, with or without ```standardization``` will converge to the same solution. It means that ```standardization = false``` will run the same code route as ```standardization = true```. Because if the features are not standardized at all, it will result convergency issue when the features have very different scales. This behavior is the same as ML [```LinearRegression``` and ```LogisticRegression```](https://issues.apache.org/jira/browse/SPARK-8522). See more discussion about this topic at #11247.
cc mengxr
## How was this patch tested?
unit test.
Author: Yanbo Liang <yb...@gmail.com>
Closes #11365 from yanboliang/spark-13322.
Project: http://git-wip-us.apache.org/repos/asf/spark/repo
Commit: http://git-wip-us.apache.org/repos/asf/spark/commit/101663f1
Tree: http://git-wip-us.apache.org/repos/asf/spark/tree/101663f1
Diff: http://git-wip-us.apache.org/repos/asf/spark/diff/101663f1
Branch: refs/heads/master
Commit: 101663f1ae222a919fc40510aa4f2bad22d1be6f
Parents: 75e05a5
Author: Yanbo Liang <yb...@gmail.com>
Authored: Tue Apr 12 11:27:16 2016 -0700
Committer: Xiangrui Meng <me...@databricks.com>
Committed: Tue Apr 12 11:27:16 2016 -0700
----------------------------------------------------------------------
.../ml/regression/AFTSurvivalRegression.scala | 105 +++++++++++++------
.../regression/AFTSurvivalRegressionSuite.scala | 22 ++++
2 files changed, 93 insertions(+), 34 deletions(-)
----------------------------------------------------------------------
http://git-wip-us.apache.org/repos/asf/spark/blob/101663f1/mllib/src/main/scala/org/apache/spark/ml/regression/AFTSurvivalRegression.scala
----------------------------------------------------------------------
diff --git a/mllib/src/main/scala/org/apache/spark/ml/regression/AFTSurvivalRegression.scala b/mllib/src/main/scala/org/apache/spark/ml/regression/AFTSurvivalRegression.scala
index afed1f3..89ba6ab 100644
--- a/mllib/src/main/scala/org/apache/spark/ml/regression/AFTSurvivalRegression.scala
+++ b/mllib/src/main/scala/org/apache/spark/ml/regression/AFTSurvivalRegression.scala
@@ -31,6 +31,7 @@ import org.apache.spark.ml.param._
import org.apache.spark.ml.param.shared._
import org.apache.spark.ml.util._
import org.apache.spark.mllib.linalg.{BLAS, Vector, Vectors, VectorUDT}
+import org.apache.spark.mllib.stat.MultivariateOnlineSummarizer
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.{DataFrame, Dataset, Row}
import org.apache.spark.sql.functions._
@@ -198,10 +199,20 @@ class AFTSurvivalRegression @Since("1.6.0") (@Since("1.6.0") override val uid: S
val handlePersistence = dataset.rdd.getStorageLevel == StorageLevel.NONE
if (handlePersistence) instances.persist(StorageLevel.MEMORY_AND_DISK)
- val costFun = new AFTCostFun(instances, $(fitIntercept))
+ val featuresSummarizer = {
+ val seqOp = (c: MultivariateOnlineSummarizer, v: AFTPoint) => c.add(v.features)
+ val combOp = (c1: MultivariateOnlineSummarizer, c2: MultivariateOnlineSummarizer) => {
+ c1.merge(c2)
+ }
+ instances.treeAggregate(new MultivariateOnlineSummarizer)(seqOp, combOp)
+ }
+
+ val featuresStd = featuresSummarizer.variance.toArray.map(math.sqrt)
+
+ val costFun = new AFTCostFun(instances, $(fitIntercept), featuresStd)
val optimizer = new BreezeLBFGS[BDV[Double]]($(maxIter), 10, $(tol))
- val numFeatures = dataset.select($(featuresCol)).take(1)(0).getAs[Vector](0).size
+ val numFeatures = featuresStd.size
/*
The parameters vector has three parts:
the first element: Double, log(sigma), the log of scale parameter
@@ -230,7 +241,13 @@ class AFTSurvivalRegression @Since("1.6.0") (@Since("1.6.0") override val uid: S
if (handlePersistence) instances.unpersist()
- val coefficients = Vectors.dense(parameters.slice(2, parameters.length))
+ val rawCoefficients = parameters.slice(2, parameters.length)
+ var i = 0
+ while (i < numFeatures) {
+ rawCoefficients(i) *= { if (featuresStd(i) != 0.0) 1.0 / featuresStd(i) else 0.0 }
+ i += 1
+ }
+ val coefficients = Vectors.dense(rawCoefficients)
val intercept = parameters(1)
val scale = math.exp(parameters(0))
val model = new AFTSurvivalRegressionModel(uid, coefficients, intercept, scale)
@@ -434,29 +451,36 @@ object AFTSurvivalRegressionModel extends MLReadable[AFTSurvivalRegressionModel]
* @param parameters including three part: The log of scale parameter, the intercept and
* regression coefficients corresponding to the features.
* @param fitIntercept Whether to fit an intercept term.
+ * @param featuresStd The standard deviation values of the features.
*/
-private class AFTAggregator(parameters: BDV[Double], fitIntercept: Boolean)
- extends Serializable {
+private class AFTAggregator(
+ parameters: BDV[Double],
+ fitIntercept: Boolean,
+ featuresStd: Array[Double]) extends Serializable {
// the regression coefficients to the covariates
private val coefficients = parameters.slice(2, parameters.length)
- private val intercept = parameters.valueAt(1)
+ private val intercept = parameters(1)
// sigma is the scale parameter of the AFT model
private val sigma = math.exp(parameters(0))
private var totalCnt: Long = 0L
private var lossSum = 0.0
- private var gradientCoefficientSum = BDV.zeros[Double](coefficients.length)
- private var gradientInterceptSum = 0.0
- private var gradientLogSigmaSum = 0.0
+ // Here we optimize loss function over log(sigma), intercept and coefficients
+ private val gradientSumArray = Array.ofDim[Double](parameters.length)
def count: Long = totalCnt
+ def loss: Double = {
+ require(totalCnt > 0.0, s"The number of instances should be " +
+ s"greater than 0.0, but got $totalCnt.")
+ lossSum / totalCnt
+ }
+ def gradient: BDV[Double] = {
+ require(totalCnt > 0.0, s"The number of instances should be " +
+ s"greater than 0.0, but got $totalCnt.")
+ new BDV(gradientSumArray.map(_ / totalCnt.toDouble))
+ }
- def loss: Double = if (totalCnt == 0) 1.0 else lossSum / totalCnt
-
- // Here we optimize loss function over coefficients, intercept and log(sigma)
- def gradient: BDV[Double] = BDV.vertcat(BDV(Array(gradientLogSigmaSum / totalCnt.toDouble)),
- BDV(Array(gradientInterceptSum/totalCnt.toDouble)), gradientCoefficientSum/totalCnt.toDouble)
/**
* Add a new training data to this AFTAggregator, and update the loss and gradient
@@ -466,25 +490,32 @@ private class AFTAggregator(parameters: BDV[Double], fitIntercept: Boolean)
* @return This AFTAggregator object.
*/
def add(data: AFTPoint): this.type = {
-
- val interceptFlag = if (fitIntercept) 1.0 else 0.0
-
- val xi = data.features.toBreeze
+ val xi = data.features
val ti = data.label
val delta = data.censor
- val epsilon = (math.log(ti) - coefficients.dot(xi) - intercept * interceptFlag ) / sigma
- lossSum += math.log(sigma) * delta
- lossSum += (math.exp(epsilon) - delta * epsilon)
+ val margin = {
+ var sum = 0.0
+ xi.foreachActive { (index, value) =>
+ if (featuresStd(index) != 0.0 && value != 0.0) {
+ sum += coefficients(index) * (value / featuresStd(index))
+ }
+ }
+ sum + intercept
+ }
+ val epsilon = (math.log(ti) - margin) / sigma
+
+ lossSum += delta * math.log(sigma) - delta * epsilon + math.exp(epsilon)
- // Sanity check (should never occur):
- assert(!lossSum.isInfinity,
- s"AFTAggregator loss sum is infinity. Error for unknown reason.")
+ val multiplier = (delta - math.exp(epsilon)) / sigma
- val deltaMinusExpEps = delta - math.exp(epsilon)
- gradientCoefficientSum += xi * deltaMinusExpEps / sigma
- gradientInterceptSum += interceptFlag * deltaMinusExpEps / sigma
- gradientLogSigmaSum += delta + deltaMinusExpEps * epsilon
+ gradientSumArray(0) += delta + multiplier * sigma * epsilon
+ gradientSumArray(1) += { if (fitIntercept) multiplier else 0.0 }
+ xi.foreachActive { (index, value) =>
+ if (featuresStd(index) != 0.0 && value != 0.0) {
+ gradientSumArray(index + 2) += multiplier * (value / featuresStd(index))
+ }
+ }
totalCnt += 1
this
@@ -503,9 +534,12 @@ private class AFTAggregator(parameters: BDV[Double], fitIntercept: Boolean)
totalCnt += other.totalCnt
lossSum += other.lossSum
- gradientCoefficientSum += other.gradientCoefficientSum
- gradientInterceptSum += other.gradientInterceptSum
- gradientLogSigmaSum += other.gradientLogSigmaSum
+ var i = 0
+ val len = this.gradientSumArray.length
+ while (i < len) {
+ this.gradientSumArray(i) += other.gradientSumArray(i)
+ i += 1
+ }
}
this
}
@@ -516,12 +550,15 @@ private class AFTAggregator(parameters: BDV[Double], fitIntercept: Boolean)
* It returns the loss and gradient at a particular point (parameters).
* It's used in Breeze's convex optimization routines.
*/
-private class AFTCostFun(data: RDD[AFTPoint], fitIntercept: Boolean)
- extends DiffFunction[BDV[Double]] {
+private class AFTCostFun(
+ data: RDD[AFTPoint],
+ fitIntercept: Boolean,
+ featuresStd: Array[Double]) extends DiffFunction[BDV[Double]] {
override def calculate(parameters: BDV[Double]): (Double, BDV[Double]) = {
- val aftAggregator = data.treeAggregate(new AFTAggregator(parameters, fitIntercept))(
+ val aftAggregator = data.treeAggregate(
+ new AFTAggregator(parameters, fitIntercept, featuresStd))(
seqOp = (c, v) => (c, v) match {
case (aggregator, instance) => aggregator.add(instance)
},
http://git-wip-us.apache.org/repos/asf/spark/blob/101663f1/mllib/src/test/scala/org/apache/spark/ml/regression/AFTSurvivalRegressionSuite.scala
----------------------------------------------------------------------
diff --git a/mllib/src/test/scala/org/apache/spark/ml/regression/AFTSurvivalRegressionSuite.scala b/mllib/src/test/scala/org/apache/spark/ml/regression/AFTSurvivalRegressionSuite.scala
index f4844cc..76891ad 100644
--- a/mllib/src/test/scala/org/apache/spark/ml/regression/AFTSurvivalRegressionSuite.scala
+++ b/mllib/src/test/scala/org/apache/spark/ml/regression/AFTSurvivalRegressionSuite.scala
@@ -33,6 +33,7 @@ class AFTSurvivalRegressionSuite
@transient var datasetUnivariate: DataFrame = _
@transient var datasetMultivariate: DataFrame = _
+ @transient var datasetUnivariateScaled: DataFrame = _
override def beforeAll(): Unit = {
super.beforeAll()
@@ -42,6 +43,11 @@ class AFTSurvivalRegressionSuite
datasetMultivariate = sqlContext.createDataFrame(
sc.parallelize(generateAFTInput(
2, Array(0.9, -1.3), Array(0.7, 1.2), 1000, 42, 1.5, 2.5, 2.0)))
+ datasetUnivariateScaled = sqlContext.createDataFrame(
+ sc.parallelize(generateAFTInput(
+ 1, Array(5.5), Array(0.8), 1000, 42, 1.0, 2.0, 2.0)).map { x =>
+ AFTPoint(Vectors.dense(x.features(0) * 1.0E3), x.label, x.censor)
+ })
}
/**
@@ -356,6 +362,22 @@ class AFTSurvivalRegressionSuite
}
}
+ test("numerical stability of standardization") {
+ val trainer = new AFTSurvivalRegression()
+ val model1 = trainer.fit(datasetUnivariate)
+ val model2 = trainer.fit(datasetUnivariateScaled)
+
+ /**
+ * During training we standardize the dataset first, so no matter how we multiple
+ * a scaling factor into the dataset, the convergence rate should be the same,
+ * and the coefficients should equal to the original coefficients multiple by
+ * the scaling factor. It will have no effect on the intercept and scale.
+ */
+ assert(model1.coefficients(0) ~== model2.coefficients(0) * 1.0E3 absTol 0.01)
+ assert(model1.intercept ~== model2.intercept absTol 0.01)
+ assert(model1.scale ~== model2.scale absTol 0.01)
+ }
+
test("read/write") {
def checkModelData(
model: AFTSurvivalRegressionModel,
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@spark.apache.org
For additional commands, e-mail: commits-help@spark.apache.org