Github user dbtsai commented on a diff in the pull request:
https://github.com/apache/spark/pull/13796#discussion_r74984115
--- Diff:
mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala
---
@@ -952,13 +963,160 @@ private class LogisticAggregator(
val bcFeaturesStd: Broadcast[Array[Double]],
private val numFeatures: Int,
numClasses: Int,
- fitIntercept: Boolean) extends Serializable {
+ fitIntercept: Boolean,
+ multinomial: Boolean) extends Serializable with Logging {
+
+ private val numFeaturesPlusIntercept = if (fitIntercept) numFeatures + 1
else numFeatures
+ private val coefficientSize = bcCoefficients.value.size
+ if (multinomial) {
+ require(numClasses == coefficientSize / numFeaturesPlusIntercept,
s"The number of " +
+ s"coefficients should be ${numClasses * numFeaturesPlusIntercept}
but was $coefficientSize")
+ } else {
+ require(coefficientSize == numFeaturesPlusIntercept, s"Expected
$numFeaturesPlusIntercept " +
+ s"coefficients but got $coefficientSize")
+ require(numClasses <= 2, s"Binary logistic aggregator requires
numClasses in {1, 2}" +
+ s" but found $numClasses.")
+ }
private var weightSum = 0.0
private var lossSum = 0.0
- private val gradientSumArray =
- Array.ofDim[Double](if (fitIntercept) numFeatures + 1 else numFeatures)
+ private val totalCoefficientLength = {
+ val cols = if (fitIntercept) numFeatures + 1 else numFeatures
+ val rows = if (multinomial) numClasses else 1
+ rows * cols
+ }
+
+ private val gradientSumArray =
Array.ofDim[Double](totalCoefficientLength)
+
+ if (multinomial && numClasses < 2) {
+ logInfo(s"Multinomial logistic regression for binary classification
yields separate " +
+ s"coefficients for positive and negative classes. When no
regularization is applied, the" +
+ s"result will be effectively the same as binary logistic regression.
When regularization" +
+ s"is applied, multinomial loss will produce a result different from
binary loss.")
+ }
+
+ /** Update gradient and loss using binary loss function. */
+ private def binaryUpdateInPlace(
+ features: Vector,
+ weight: Double,
+ label: Double,
+ coefficients: Array[Double],
+ gradient: Array[Double],
+ featuresStd: Array[Double],
+ numFeaturesPlusIntercept: Int): Unit = {
+ val margin = - {
+ var sum = 0.0
+ features.foreachActive { (index, value) =>
+ if (featuresStd(index) != 0.0 && value != 0.0) {
+ sum += coefficients(index) * value / featuresStd(index)
+ }
+ }
+ sum + {
+ if (fitIntercept) coefficients(numFeaturesPlusIntercept - 1) else
0.0
+ }
+ }
+
+ val multiplier = weight * (1.0 / (1.0 + math.exp(margin)) - label)
+
+ features.foreachActive { (index, value) =>
+ if (featuresStd(index) != 0.0 && value != 0.0) {
+ gradient(index) += multiplier * value / featuresStd(index)
+ }
+ }
+
+ if (fitIntercept) {
+ gradient(numFeaturesPlusIntercept - 1) += multiplier
+ }
+
+ if (label > 0) {
+ // The following is equivalent to log(1 + exp(margin)) but more
numerically stable.
+ lossSum += weight * MLUtils.log1pExp(margin)
+ } else {
+ lossSum += weight * (MLUtils.log1pExp(margin) - margin)
+ }
+ }
+
+ /** Update gradient and loss using multinomial (softmax) loss function.
*/
+ private def multinomialUpdateInPlace(
+ features: Vector,
+ weight: Double,
+ label: Double,
+ coefficients: Array[Double],
+ gradient: Array[Double],
+ featuresStd: Array[Double],
+ numFeaturesPlusIntercept: Int): Unit = {
+ // TODO: use level 2 BLAS operations
+ /*
+ Note: this can still be used when numClasses = 2 for binary
+ logistic regression without pivoting.
+ */
+
+ // marginOfLabel is margins(label) in the formula
+ var marginOfLabel = 0.0
+ var maxMargin = Double.NegativeInfinity
+
+ val margins = Array.tabulate(numClasses) { i =>
+ var margin = 0.0
+ features.foreachActive { (index, value) =>
--- End diff --
I quickly did a benchmark. For high classes problem (10+ classes) and many
zeros in features, looping the # of numClasses inside `foreachActive` improves
the performance dramatically.
---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
