Remove toDense.

Author: viirya

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

@@ -21,8 +21,6 @@ import java.lang.{Iterable => JIterable}
 
 import scala.collection.JavaConverters._
 
-import breeze.linalg.{Axis, DenseMatrix => BDM, DenseVector => BDV, argmax => brzArgmax, sum => brzSum}
-import breeze.numerics.{exp => brzExp, log => brzLog}
 import org.json4s.JsonDSL._
 import org.json4s.jackson.JsonMethods._
 
@@ -50,8 +48,8 @@ class NaiveBayesModel private[mllib] (
     val modelType: String)
   extends ClassificationModel with Serializable with Saveable {
 
-  val piVector = Vectors.dense(pi).asInstanceOf[DenseVector]
-  val thetaMatrix = new DenseMatrix(labels.size, theta(0).size, theta.flatten, true)
+  private val piVector = new DenseVector(pi)
+  private val thetaMatrix = new DenseMatrix(labels.size, theta(0).size, theta.flatten, true)
 
   private[mllib] def this(labels: Array[Double], pi: Array[Double], theta: Array[Array[Double]]) =
     this(labels, pi, theta, "Multinomial")
@@ -66,15 +64,15 @@ class NaiveBayesModel private[mllib] (
   // Bernoulli scoring requires log(condprob) if 1, log(1-condprob) if 0.
   // This precomputes log(1.0 - exp(theta)) and its sum which are used for the linear algebra
   // application of this condition (in predict function).
-  private val (thetaMinusnegTheta, brzNegThetaSum) = modelType match {
+  private val (thetaMinusNegTheta, negThetaSum) = modelType match {
     case "Multinomial" => (None, None)
     case "Bernoulli" =>
       val negTheta = thetaMatrix.map(value => math.log(1.0 - math.exp(value)))
-      val ones = Vectors.dense(Array.fill(thetaMatrix.numCols){1.0}).asInstanceOf[DenseVector]
-      val thetaMinusnegTheta = thetaMatrix.map { value =>
+      val ones = new DenseVector(Array.fill(thetaMatrix.numCols){1.0})
+      val thetaMinusNegTheta = thetaMatrix.map { value =>
         value - math.log(1.0 - math.exp(value))
       }
-      (Option(thetaMinusnegTheta), Option(negTheta.multiply(ones)))
+      (Option(thetaMinusNegTheta), Option(negTheta.multiply(ones)))
     case _ =>
       // This should never happen.
       throw new UnknownError(s"NaiveBayesModel was created with an unknown ModelType: $modelType")
@@ -91,7 +89,7 @@ class NaiveBayesModel private[mllib] (
   override def predict(testData: Vector): Double = {
     modelType match {
       case "Multinomial" =>
-        val prob = thetaMatrix.multiply(testData.toDense)
+        val prob = thetaMatrix.multiply(testData)
         BLAS.axpy(1.0, piVector, prob)
         labels(prob.argmax)
       case "Bernoulli" =>
@@ -101,9 +99,9 @@ class NaiveBayesModel private[mllib] (
               s"Bernoulli Naive Bayes requires 0 or 1 feature values but found $testData.")
           }
         }
-        val prob = thetaMinusnegTheta.get.multiply(testData.toDense)
+        val prob = thetaMinusNegTheta.get.multiply(testData)
         BLAS.axpy(1.0, piVector, prob)
-        BLAS.axpy(1.0, brzNegThetaSum.get, prob)
+        BLAS.axpy(1.0, negThetaSum.get, prob)
         labels(prob.argmax)
       case _ =>
         // This should never happen.

mllib/src/main/scala/org/apache/spark/mllib/linalg/BLAS.scala

@@ -463,7 +463,7 @@ private[spark] object BLAS extends Serializable with Logging {
   def gemv(
       alpha: Double,
       A: Matrix,
-      x: DenseVector,
+      x: Vector,
       beta: Double,
       y: DenseVector): Unit = {
     require(A.numCols == x.size,
@@ -473,13 +473,16 @@ private[spark] object BLAS extends Serializable with Logging {
     if (alpha == 0.0) {
       logDebug("gemv: alpha is equal to 0. Returning y.")
     } else {
-      A match {
-        case sparse: SparseMatrix =>
-          gemv(alpha, sparse, x, beta, y)
-        case dense: DenseMatrix =>
-          gemv(alpha, dense, x, beta, y)
+      (A, x) match {
+        case (sparse: SparseMatrix, dx: DenseVector) =>
+          gemv(alpha, sparse, dx, beta, y)
+        case (dense: DenseMatrix, dx: DenseVector) =>
+          gemv(alpha, dense, dx, beta, y)
+        case (dense: DenseMatrix, sx: SparseVector) =>
+          gemv(alpha, dense, sx, beta, y)
         case _ =>
-          throw new IllegalArgumentException(s"gemv doesn't support matrix type ${A.getClass}.")
+          throw new IllegalArgumentException(s"gemv doesn't support running on matrix type " +
+            s"${A.getClass} and vector type ${x.getClass}.")
       }
     }
   }
@@ -500,6 +503,55 @@ private[spark] object BLAS extends Serializable with Logging {
     nativeBLAS.dgemv(tStrA, mA, nA, alpha, A.values, mA, x.values, 1, beta,
       y.values, 1)
   }
+ 
+  /**
+   * y := alpha * A * x + beta * y
+   * For `DenseMatrix` A and SparseVector x.
+   */
+  private def gemv(
+      alpha: Double,
+      A: DenseMatrix,
+      x: SparseVector,
+      beta: Double,
+      y: DenseVector): Unit =  {
+    val mA: Int = A.numRows
+    val nA: Int = A.numCols
+
+    val Avals = A.values
+    var colCounterForA = 0
+
+    var xIndices = x.indices
+    var xNnz = xIndices.size
+    var xValues = x.values
+
+    scal(beta, y)
+
+    if (!A.isTransposed) {
+      var rowCounterForA = 0
+      while (rowCounterForA < mA) {
+        var sum = 0.0
+        var k = 0
+        while (k < xNnz) {
+          sum += xValues(k) * Avals(xIndices(k) * mA + rowCounterForA)
+          k += 1
+        }
+        y.values(rowCounterForA) += sum * alpha
+        rowCounterForA += 1
+      }
+    } else {
+      var rowCounterForA = 0
+      while (rowCounterForA < mA) {
+        var sum = 0.0
+        var k = 0
+        while (k < xNnz) {
+          sum += xValues(k) * Avals(xIndices(k) + rowCounterForA * nA)
+          k += 1
+        }
+        y.values(rowCounterForA) += sum * alpha
+        rowCounterForA += 1
+      }
+    }
+  }
 
   /**
    * y := alpha * A * x + beta * y

mllib/src/main/scala/org/apache/spark/mllib/linalg/Matrices.scala

@@ -78,7 +78,7 @@ sealed trait Matrix extends Serializable {
   }
 
   /** Convenience method for `Matrix`-`DenseVector` multiplication. */
-  def multiply(y: DenseVector): DenseVector = {
+  def multiply(y: Vector): DenseVector = {
     val output = new DenseVector(new Array[Double](numRows))
     BLAS.gemv(1.0, this, y, 0.0, output)
     output

mllib/src/test/scala/org/apache/spark/mllib/linalg/BLASSuite.scala

@@ -257,11 +257,12 @@ class BLASSuite extends FunSuite {
       new DenseMatrix(4, 3, Array(0.0, 1.0, 0.0, 0.0, 2.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 3.0))
     val sA = new SparseMatrix(4, 3, Array(0, 1, 3, 4), Array(1, 0, 2, 3), Array(1.0, 2.0, 1.0, 3.0))
 
-    val x = new DenseVector(Array(1.0, 2.0, 3.0))
+    val dx = new DenseVector(Array(1.0, 2.0, 3.0))
+    val sx = dx.toSparse
     val expected = new DenseVector(Array(4.0, 1.0, 2.0, 9.0))
 
-    assert(dA.multiply(x) ~== expected absTol 1e-15)
-    assert(sA.multiply(x) ~== expected absTol 1e-15)
+    assert(dA.multiply(dx) ~== expected absTol 1e-15)
+    assert(sA.multiply(dx) ~== expected absTol 1e-15)
 
     val y1 = new DenseVector(Array(1.0, 3.0, 1.0, 0.0))
     val y2 = y1.copy
@@ -270,17 +271,26 @@ class BLASSuite extends FunSuite {
     val expected2 = new DenseVector(Array(6.0, 7.0, 4.0, 9.0))
     val expected3 = new DenseVector(Array(10.0, 8.0, 6.0, 18.0))
 
-    gemv(1.0, dA, x, 2.0, y1)
-    gemv(1.0, sA, x, 2.0, y2)
-    gemv(2.0, dA, x, 2.0, y3)
-    gemv(2.0, sA, x, 2.0, y4)
+    gemv(1.0, dA, dx, 2.0, y1)
+    gemv(1.0, sA, dx, 2.0, y2)
+    gemv(2.0, dA, dx, 2.0, y3)
+    gemv(2.0, sA, dx, 2.0, y4)
     assert(y1 ~== expected2 absTol 1e-15)
     assert(y2 ~== expected2 absTol 1e-15)
     assert(y3 ~== expected3 absTol 1e-15)
     assert(y4 ~== expected3 absTol 1e-15)
+
+    val y1_copy = new DenseVector(Array(1.0, 3.0, 1.0, 0.0))
+    val y3_copy = y1_copy.copy
+
+    gemv(1.0, dA, sx, 2.0, y1_copy)
+    gemv(2.0, dA, sx, 2.0, y3_copy)
+    assert(y1_copy ~== expected2 absTol 1e-15)
+    assert(y3_copy ~== expected3 absTol 1e-15)
+ 
     withClue("columns of A don't match the rows of B") {
       intercept[Exception] {
-        gemv(1.0, dA.transpose, x, 2.0, y1)
+        gemv(1.0, dA.transpose, dx, 2.0, y1)
       }
     }
     val dAT =
@@ -291,7 +301,7 @@ class BLASSuite extends FunSuite {
     val dATT = dAT.transpose
     val sATT = sAT.transpose
 
-    assert(dATT.multiply(x) ~== expected absTol 1e-15)
-    assert(sATT.multiply(x) ~== expected absTol 1e-15)
+    assert(dATT.multiply(dx) ~== expected absTol 1e-15)
+    assert(sATT.multiply(dx) ~== expected absTol 1e-15)
   }
 }