[
https://issues.apache.org/jira/browse/FLINK-1807?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14526606#comment-14526606
]
ASF GitHub Bot commented on FLINK-1807:
---------------------------------------
Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/613#discussion_r29582777
--- Diff:
flink-staging/flink-ml/src/main/scala/org/apache/flink/ml/optimization/RegularizationType.scala
---
@@ -0,0 +1,143 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.flink.ml.optimization
+
+import org.apache.flink.api.scala._
+import org.apache.flink.ml.math.{Vector => FlinkVector, BLAS}
+import org.apache.flink.ml.math.Breeze._
+
+import breeze.numerics._
+import breeze.linalg.max
+
+
+
+// TODO(tvas): Change name to RegularizationPenalty?
+/** Represents a type of regularization penalty
+ *
+ */
+abstract class RegularizationType extends Serializable{
+
+ /** Updates the weights by taking a step according to the gradient and
regularization applied
+ *
+ * @param oldWeights The weights to be updated
+ * @param gradient The gradient according to which we will update the
weights
+ * @param effectiveStepSize The effective step size for this iteration
+ * @param regParameter The regularization parameter to be applied in
the case of L1
+ * regularization
+ */
+ def takeStep(
+ oldWeights: FlinkVector,
+ gradient: FlinkVector,
+ effectiveStepSize: Double,
+ regParameter: Double) {
+ BLAS.axpy(-effectiveStepSize, gradient, oldWeights)
+ }
+
+}
+
+/** A regularization penalty that is differentiable
+ *
+ */
+abstract class DiffRegularizationType extends RegularizationType {
+
+ /** Compute the regularized gradient loss for the given data.
+ * The provided cumGradient is updated in place.
+ *
+ * @param weightVector The current weight vector
+ * @param lossGradient The vector to which the gradient will be added
to, in place.
+ * @return The regularized loss. The gradient is updated in place.
+ */
+ def regularizedLossAndGradient(
+ loss: Double,
+ weightVector: FlinkVector,
+ lossGradient: FlinkVector,
+ regularizationParameter: Double) : Double ={
+ val adjustedLoss = regLoss(loss, weightVector, regularizationParameter)
+ regGradient(weightVector, lossGradient, regularizationParameter)
+
+ adjustedLoss
+ }
+
+ /** Calculates the regularized loss **/
+ def regLoss(oldLoss: Double, weightVector: FlinkVector,
regularizationParameter: Double): Double
+
+ /** Calculates the regularized gradient **/
--- End diff --
Comment can be confusing, since we're only adding the gradient of the
regularization term here. Maybe: Adds gradient of regularization term to loss
gradient.
Would be good to mention in the comments that the `weightVector` is updated
in place.
> Stochastic gradient descent optimizer for ML library
> ----------------------------------------------------
>
> Key: FLINK-1807
> URL: https://issues.apache.org/jira/browse/FLINK-1807
> Project: Flink
> Issue Type: Improvement
> Components: Machine Learning Library
> Reporter: Till Rohrmann
> Assignee: Theodore Vasiloudis
> Labels: ML
>
> Stochastic gradient descent (SGD) is a widely used optimization technique in
> different ML algorithms. Thus, it would be helpful to provide a generalized
> SGD implementation which can be instantiated with the respective gradient
> computation. Such a building block would make the development of future
> algorithms easier.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
