[SPARK-6496] [MLLIB] GeneralizedLinearAlgorithm.run(input, initialWei…

…ghts) should initialize numFeatures In GeneralizedLinearAlgorithm ```numFeatures``` is default to -1, we need to update it to correct value when we call run() to train a model. ```LogisticRegressionWithLBFGS.run(input)``` works well, but when we call ```LogisticRegressionWithLBFGS.run(input, initialWeights)``` to train multiclass classification model, it will throw exception due to the numFeatures is not updated. In this PR, we just update numFeatures at the beginning of GeneralizedLinearAlgorithm.run(input, initialWeights) and add test case. Author: Yanbo Liang <[email protected]> Closes #5167 from yanboliang/spark-6496 and squashes the following commits: 8131c48 [Yanbo Liang] LogisticRegressionWithLBFGS.run(input, initialWeights) should initialize numFeatures
apache · Mar 25, 2015 · 10c7860 · 10c7860
1 parent 64262ed
commit 10c7860
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 0 deletions.
diff --git a/mllib/src/main/scala/org/apache/spark/mllib/regression/GeneralizedLinearAlgorithm.scala b/mllib/src/main/scala/org/apache/spark/mllib/regression/GeneralizedLinearAlgorithm.scala
@@ -211,6 +211,10 @@ abstract class GeneralizedLinearAlgorithm[M <: GeneralizedLinearModel]
    */
   def run(input: RDD[LabeledPoint], initialWeights: Vector): M = {
 
+    if (numFeatures < 0) {
+      numFeatures = input.map(_.features.size).first()
+    }
+
     if (input.getStorageLevel == StorageLevel.NONE) {
       logWarning("The input data is not directly cached, which may hurt performance if its"
         + " parent RDDs are also uncached.")

diff --git a/mllib/src/test/scala/org/apache/spark/mllib/classification/LogisticRegressionSuite.scala b/mllib/src/test/scala/org/apache/spark/mllib/classification/LogisticRegressionSuite.scala
@@ -425,6 +425,12 @@ class LogisticRegressionSuite extends FunSuite with MLlibTestSparkContext with M
 
     val model = lr.run(testRDD)
 
+    val numFeatures = testRDD.map(_.features.size).first()
+    val initialWeights = Vectors.dense(new Array[Double]((numFeatures + 1) * 2))
+    val model2 = lr.run(testRDD, initialWeights)
+
+    LogisticRegressionSuite.checkModelsEqual(model, model2)
+
     /**
      * The following is the instruction to reproduce the model using R's glmnet package.
      *