Merge pull request #11 from Ru-Xiang/master

add elastic net and L1 bug fix
LIBBLE · Mar 30, 2017 · 73cdc97 · 73cdc97
2 parents bb526af + 033905a
commit 73cdc97
Showing 1 changed file with 75 additions and 1 deletion.
diff --git a/src/main/scala/generalizedLinear/Updater.scala b/src/main/scala/generalizedLinear/Updater.scala
@@ -130,7 +130,7 @@ class L1Updater extends Updater {
 
   override def update(data: RDD[(Double, Vector)], weights: Vector, mu: Vector, lossfunc: LossFunc, stepSize: Double, factor: Double, regParam: Double): Vector = {
 
-    val preFact = 1.0 - stepSize * (regParam + factor)
+    val preFact = 1.0 - stepSize *  factor
 
     val upFact = -stepSize / preFact
 
@@ -246,3 +246,77 @@ class L2Updater extends Updater {
   }
 }
 
+class elasticNetUpdater(val alpha:Double) extends Updater{
+  val beta= 1-alpha
+
+
+  override def update(data: RDD[(Double, Vector)], weights: Vector, mu: Vector, lossfunc: LossFunc, stepSize: Double, factor: Double, regParam: Double): Vector = {
+
+    val preFact = 1.0 - stepSize * (regParam*alpha + factor)
+    val upFact = -stepSize / preFact
+    mu.plusax(-factor, weights)
+    val w_0 = data.sparkContext.broadcast(weights)
+    val fix = data.sparkContext.broadcast(mu)
+    val partsNum = data.partitions.length
+    val chkSize = findChkSize(preFact)
+
+    val l1fact=stepSize * regParam*beta
+
+    data.mapPartitions(iter => {
+      val omiga = new WeightsVector(w_0.value.copy, fix.value)
+      val indexedSeq = iter.toIndexedSeq
+      val pNum = indexedSeq.size
+
+      val rand = new Random(partsNum * pNum)
+
+      val flags = new Array[Int](omiga.size)
+      util.Arrays.fill(flags, 0)
+
+      for (j <- 1 to pNum) {
+        val e = indexedSeq(rand.nextInt(pNum))
+        val f1 = lossfunc.deltaF(e._2, e._1, omiga)
+        f1 -= lossfunc.deltaF(e._2, e._1, w_0.value)
+        //          val delta = f1 x e._2
+        //          delta += mu
+        if (j % chkSize == 0)
+          omiga.merge()
+
+        val oValues = omiga.partA.toArray
+        e._2.foreachActive { (i, v) =>
+          val wi = omiga.apply(i)
+          oValues(i) = (math.signum(wi) * max(0.0, abs(wi) - (j - 1 - flags(i)) * l1fact) - omiga.fac_b * omiga.partB(i)) / omiga.fac_a
+          flags(i) = j - 1
+        }
+
+        omiga.partA.plusax(upFact / omiga.fac_a, f1 x e._2)
+        omiga.fac_a *= preFact
+        omiga.fac_b *= preFact
+        omiga.fac_b -= stepSize
+
+      }
+      Iterator(omiga.toDense())
+
+    }, true).treeAggregate(new DenseVector(weights.size))(seqOp = (c, w) => {
+      c += w
+    }, combOp = (c1, c2) => {
+      c1 += c2
+    }) /= (partsNum)
+
+
+
+  }
+  /**
+    * In this method, we give the cost of the regularizer.
+    *
+    * @param weight
+    * @param regParam
+    * @return regCost
+    */
+  override def getRegVal(weight: Vector, regParam: Double): Double = {
+    val norm1= weight.norm1()
+    val norm2=weight.norm2()
+    regParam*(0.5*alpha*norm2*norm2+beta*norm1)
+
+  }
+}
+