use in-place gradient computation

andrewor14 · Mar 27, 2014 · 44733e1 · 44733e1
1 parent e981396
commit 44733e1
Show file tree

Hide file tree

Showing 2 changed files with 60 additions and 4 deletions.
diff --git a/mllib/src/main/scala/org/apache/spark/mllib/optimization/Gradient.scala b/mllib/src/main/scala/org/apache/spark/mllib/optimization/Gradient.scala
@@ -17,6 +17,8 @@
 
 package org.apache.spark.mllib.optimization
 
+import breeze.linalg.{axpy => brzAxpy}
+
 import org.apache.spark.mllib.linalg.{Vectors, Vector}
 
 /**
@@ -33,6 +35,19 @@ abstract class Gradient extends Serializable {
    * @return (gradient: Vector, loss: Double)
    */
   def compute(data: Vector, label: Double, weights: Vector): (Vector, Double)
+
+  /**
+   * Compute the gradient and loss given the features of a single data point, add the gradient to a provided vector to
+   * avoid creating new objects, and return loss.
+   *
+   * @param data features for one data point
+   * @param label label for this data point
+   * @param weights weights/coefficients corresponding to features
+   * @param gradientAddTo gradient will be added to this vector
+   *
+   * @return (gradient: Vector, loss: Double)
+   */
+  def compute(data: Vector, label: Double, weights: Vector, gradientAddTo: Vector): Double
 }
 
 /**
@@ -55,6 +70,21 @@ class LogisticGradient extends Gradient {
 
     (Vectors.fromBreeze(gradient), loss)
   }
+
+  override def compute(data: Vector, label: Double, weights: Vector, gradientAddTo: Vector): Double = {
+    val brzData = data.toBreeze
+    val brzWeights = weights.toBreeze
+    val margin: Double = -1.0 * brzWeights.dot(brzData)
+    val gradientMultiplier = (1.0 / (1.0 + math.exp(margin))) - label
+
+    brzAxpy(gradientMultiplier, brzData, gradientAddTo.toBreeze)
+
+    if (label > 0) {
+      math.log(1 + math.exp(margin))
+    } else {
+      math.log(1 + math.exp(margin)) - margin
+    }
+  }
 }
 
 /**
@@ -73,6 +103,16 @@ class LeastSquaresGradient extends Gradient {
 
     (Vectors.fromBreeze(gradient), loss)
   }
+
+  override def compute(data: Vector, label: Double, weights: Vector, gradientAddTo: Vector): Double = {
+    val brzData = data.toBreeze
+    val brzWeights = weights.toBreeze
+    val diff = brzWeights.dot(brzData) - label
+
+    brzAxpy(2.0 * diff, brzData, gradientAddTo.toBreeze)
+
+    diff * diff
+  }
 }
 
 /**
@@ -96,4 +136,21 @@ class HingeGradient extends Gradient {
       (Vectors.dense(new Array[Double](weights.size)), 0.0)
     }
   }
+
+  override def compute(data: Vector, label: Double, weights: Vector, gradientAddTo: Vector): Double = {
+    val brzData = data.toBreeze
+    val brzWeights = weights.toBreeze
+    val dotProduct = brzWeights.dot(brzData)
+
+    // Our loss function with {0, 1} labels is max(0, 1 - (2y – 1) (f_w(x)))
+    // Therefore the gradient is -(2y - 1)*x
+    val labelScaled = 2 * label - 1.0
+
+    if (1.0 > labelScaled * dotProduct) {
+      brzAxpy(-labelScaled, brzData, gradientAddTo.toBreeze)
+      1.0 - labelScaled * dotProduct
+    } else {
+      0.0
+    }
+  }
 }
diff --git a/mllib/src/main/scala/org/apache/spark/mllib/optimization/GradientDescent.scala b/mllib/src/main/scala/org/apache/spark/mllib/optimization/GradientDescent.scala
@@ -162,13 +162,12 @@ object GradientDescent extends Logging {
       val (gradientSum, lossSum) = data.sample(false, miniBatchFraction, 42 + i)
         .aggregate((BDV.zeros[Double](weights.size), 0.0))(
           seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, features)) =>
-            val (g, l) = gradient.compute(features, label, weights)
-            (grad += g.toBreeze, loss + l)
+            val l = gradient.compute(features, label, weights, Vectors.fromBreeze(grad))
+            (grad, loss + l)
           },
           combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
             (grad1 += grad2, loss1 + loss2)
-          }
-        )
+          })
 
       /**
        * NOTE(Xinghao): lossSum is computed using the weights from the previous iteration