[SPARK-4526][MLLIB]GradientDescent get a wrong gradient value according to the gradient formula.

This is caused by the miniBatchSize parameter.The number of `RDD.sample` returns is not fixed.
cc mengxr

Author: GuoQiang Li <[email protected]>

Closes #3399 from witgo/GradientDescent and squashes the following commits:

13cb228 [GuoQiang Li] review commit
668ab66 [GuoQiang Li] Double to Long
b6aa11a [GuoQiang Li] Check miniBatchSize is greater than 0
0b5c3e3 [GuoQiang Li] Minor fix
12e7424 [GuoQiang Li] GradientDescent get a wrong gradient value according to the gradient formula, which is caused by the miniBatchSize parameter.

Author: witgo

mllib/src/main/scala/org/apache/spark/mllib/optimization/GradientDescent.scala

@@ -160,14 +160,15 @@ object GradientDescent extends Logging {
     val stochasticLossHistory = new ArrayBuffer[Double](numIterations)
 
     val numExamples = data.count()
-    val miniBatchSize = numExamples * miniBatchFraction
 
     // if no data, return initial weights to avoid NaNs
     if (numExamples == 0) {
-
-      logInfo("GradientDescent.runMiniBatchSGD returning initial weights, no data found")
+      logWarning("GradientDescent.runMiniBatchSGD returning initial weights, no data found")
       return (initialWeights, stochasticLossHistory.toArray)
+    }
 
+    if (numExamples * miniBatchFraction < 1) {
+      logWarning("The miniBatchFraction is too small")
     }
 
     // Initialize weights as a column vector
@@ -185,25 +186,31 @@ object GradientDescent extends Logging {
       val bcWeights = data.context.broadcast(weights)
       // Sample a subset (fraction miniBatchFraction) of the total data
       // compute and sum up the subgradients on this subset (this is one map-reduce)
-      val (gradientSum, lossSum) = data.sample(false, miniBatchFraction, 42 + i)
-        .treeAggregate((BDV.zeros[Double](n), 0.0))(
-          seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, features)) =>
-            val l = gradient.compute(features, label, bcWeights.value, Vectors.fromBreeze(grad))
-            (grad, loss + l)
+      val (gradientSum, lossSum, miniBatchSize) = data.sample(false, miniBatchFraction, 42 + i)
+        .treeAggregate((BDV.zeros[Double](n), 0.0, 0L))(
+          seqOp = (c, v) => {
+            // c: (grad, loss, count), v: (label, features)
+            val l = gradient.compute(v._2, v._1, bcWeights.value, Vectors.fromBreeze(c._1))
+            (c._1, c._2 + l, c._3 + 1)
           },
-          combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
-            (grad1 += grad2, loss1 + loss2)
+          combOp = (c1, c2) => {
+            // c: (grad, loss, count)
+            (c1._1 += c2._1, c1._2 + c2._2, c1._3 + c2._3)
           })
 
-      /**
-       * NOTE(Xinghao): lossSum is computed using the weights from the previous iteration
-       * and regVal is the regularization value computed in the previous iteration as well.
-       */
-      stochasticLossHistory.append(lossSum / miniBatchSize + regVal)
-      val update = updater.compute(
-        weights, Vectors.fromBreeze(gradientSum / miniBatchSize), stepSize, i, regParam)
-      weights = update._1
-      regVal = update._2
+      if (miniBatchSize > 0) {
+        /**
+         * NOTE(Xinghao): lossSum is computed using the weights from the previous iteration
+         * and regVal is the regularization value computed in the previous iteration as well.
+         */
+        stochasticLossHistory.append(lossSum / miniBatchSize + regVal)
+        val update = updater.compute(
+          weights, Vectors.fromBreeze(gradientSum / miniBatchSize.toDouble), stepSize, i, regParam)
+        weights = update._1
+        regVal = update._2
+      } else {
+        logWarning(s"Iteration ($i/$numIterations). The size of sampled batch is zero")
+      }
     }
 
     logInfo("GradientDescent.runMiniBatchSGD finished. Last 10 stochastic losses %s".format(