[SPARK-6345][STREAMING][MLLIB] Fix for training with prediction

This patch fixes a reported bug causing model updates to not properly propagate to model predictions during streaming regression. These minor changes in model declaration fix the problem, and I expanded the tests to include the scenario in which the bug was arising. The two new tests failed prior to the patch and now pass.

cc mengxr

Author: freeman <the.freeman.lab@gmail.com>

Closes #5037 from freeman-lab/train-predict-fix and squashes the following commits:

3af953e [freeman] Expand test coverage to include combined training and prediction
8f84fc8 [freeman] Move model declaration

mllib/src/main/scala/org/apache/spark/mllib/classification/StreamingLogisticRegressionWithSGD.scala

@@ -63,6 +63,8 @@ class StreamingLogisticRegressionWithSGD private[mllib] (
   protected val algorithm = new LogisticRegressionWithSGD(
     stepSize, numIterations, regParam, miniBatchFraction)
 
+  protected var model: Option[LogisticRegressionModel] = None
+
   /** Set the step size for gradient descent. Default: 0.1. */
   def setStepSize(stepSize: Double): this.type = {
     this.algorithm.optimizer.setStepSize(stepSize)

mllib/src/main/scala/org/apache/spark/mllib/regression/StreamingLinearAlgorithm.scala

@@ -60,7 +60,7 @@ abstract class StreamingLinearAlgorithm[
     A <: GeneralizedLinearAlgorithm[M]] extends Logging {
 
   /** The model to be updated and used for prediction. */
-  protected var model: Option[M] = None
+  protected var model: Option[M]
 
   /** The algorithm to use for updating. */
   protected val algorithm: A
@@ -114,7 +114,7 @@ abstract class StreamingLinearAlgorithm[
     if (model.isEmpty) {
       throw new IllegalArgumentException("Model must be initialized before starting prediction.")
     }
-    data.map(model.get.predict)
+    data.map{x => model.get.predict(x)}
   }
 
   /** Java-friendly version of `predictOn`. */
@@ -132,7 +132,7 @@ abstract class StreamingLinearAlgorithm[
     if (model.isEmpty) {
       throw new IllegalArgumentException("Model must be initialized before starting prediction")
     }
-    data.mapValues(model.get.predict)
+    data.mapValues{x => model.get.predict(x)}
   }
 
 

mllib/src/main/scala/org/apache/spark/mllib/regression/StreamingLinearRegressionWithSGD.scala

@@ -59,6 +59,8 @@ class StreamingLinearRegressionWithSGD private[mllib] (
 
   val algorithm = new LinearRegressionWithSGD(stepSize, numIterations, miniBatchFraction)
 
+  protected var model: Option[LinearRegressionModel] = None
+
   /** Set the step size for gradient descent. Default: 0.1. */
   def setStepSize(stepSize: Double): this.type = {
     this.algorithm.optimizer.setStepSize(stepSize)

mllib/src/test/scala/org/apache/spark/mllib/classification/StreamingLogisticRegressionSuite.scala

@@ -132,4 +132,31 @@ class StreamingLogisticRegressionSuite extends FunSuite with TestSuiteBase {
     assert(errors.forall(x => x <= 0.4))
   }
 
+  // Test training combined with prediction
+  test("training and prediction") {
+    // create model initialized with zero weights
+    val model = new StreamingLogisticRegressionWithSGD()
+      .setInitialWeights(Vectors.dense(-0.1))
+      .setStepSize(0.01)
+      .setNumIterations(10)
+
+    // generate sequence of simulated data for testing
+    val numBatches = 10
+    val nPoints = 100
+    val testInput = (0 until numBatches).map { i =>
+      LogisticRegressionSuite.generateLogisticInput(0.0, 5.0, nPoints, 42 * (i + 1))
+    }
+
+    // train and predict
+    val ssc = setupStreams(testInput, (inputDStream: DStream[LabeledPoint]) => {
+      model.trainOn(inputDStream)
+      model.predictOnValues(inputDStream.map(x => (x.label, x.features)))
+    })
+
+    val output: Seq[Seq[(Double, Double)]] = runStreams(ssc, numBatches, numBatches)
+
+    // assert that prediction error improves, ensuring that the updated model is being used
+    val error = output.map(batch => batch.map(p => math.abs(p._1 - p._2)).sum / nPoints).toList
+    assert(error.head > 0.8 & error.last < 0.2)
+  }
 }

mllib/src/test/scala/org/apache/spark/mllib/regression/StreamingLinearRegressionSuite.scala

@@ -139,4 +139,32 @@ class StreamingLinearRegressionSuite extends FunSuite with TestSuiteBase {
     val errors = output.map(batch => batch.map(p => math.abs(p._1 - p._2)).sum / nPoints)
     assert(errors.forall(x => x <= 0.1))
   }
+
+  // Test training combined with prediction
+  test("training and prediction") {
+    // create model initialized with zero weights
+    val model = new StreamingLinearRegressionWithSGD()
+      .setInitialWeights(Vectors.dense(0.0, 0.0))
+      .setStepSize(0.2)
+      .setNumIterations(25)
+
+    // generate sequence of simulated data for testing
+    val numBatches = 10
+    val nPoints = 100
+    val testInput = (0 until numBatches).map { i =>
+      LinearDataGenerator.generateLinearInput(0.0, Array(10.0, 10.0), nPoints, 42 * (i + 1))
+    }
+
+    // train and predict
+    val ssc = setupStreams(testInput, (inputDStream: DStream[LabeledPoint]) => {
+      model.trainOn(inputDStream)
+      model.predictOnValues(inputDStream.map(x => (x.label, x.features)))
+    })
+
+    val output: Seq[Seq[(Double, Double)]] = runStreams(ssc, numBatches, numBatches)
+
+    // assert that prediction error improves, ensuring that the updated model is being used
+    val error = output.map(batch => batch.map(p => math.abs(p._1 - p._2)).sum / nPoints).toList
+    assert((error.head - error.last) > 2)
+  }
 }