flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF GitHub Bot (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-1807) Stochastic gradient descent optimizer for ML library
Date Thu, 07 May 2015 09:33:00 GMT

    [ https://issues.apache.org/jira/browse/FLINK-1807?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14532329#comment-14532329
] 

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_r29837033
  
    --- Diff: flink-staging/flink-ml/src/main/scala/org/apache/flink/ml/optimization/RegularizationType.scala
---
    @@ -0,0 +1,171 @@
    +/*
    + * 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.{norm => BreezeNorm, max => BreezeMax}
    +
    +
    +
    +// TODO(tvas): Change name to RegularizationPenalty?
    +/** Represents a type of regularization penalty
    +  *
    +  * Regularization penalties are used to restrict the optimization problem to solutions
with
    +  * certain desirable characteristics, such as sparsity for the L1 penalty, or penalizing
large
    +  * weights for the L2 penalty.
    +  *
    +  * The regularization term, $R(w)$ is added to the objective function, $f(w) = L(w)
+ \lambda R(w)$
    +  * where $\lambda$ is the regularization parameter used to tune the amount of regularization
    +  * applied.
    +  */
    +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)
    +  }
    +
    +  /** Adds regularization to the loss value **/
    +  def regLoss(oldLoss: Double, weightVector: FlinkVector, regularizationParameter: Double):
Double
    +
    +}
    +
    +/** Abstract class for regularization penalties that are 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
    +  }
    +
    +  /** Adds regularization gradient to the loss gradient. The gradient is updated in place
**/
    +  def regGradient(
    +      weightVector: FlinkVector,
    +      lossGradient: FlinkVector,
    +      regularizationParameter: Double)
    +}
    +
    +/** Performs no regularization, equivalent to $R(w) = 0$ **/
    +class NoRegularization extends RegularizationType {
    +  /** Adds regularization to the loss value **/
    +  override def regLoss(oldLoss: Double, weightVector: FlinkVector, regularizationParameter:
Double):
    +  Double = {oldLoss}
    --- End diff --
    
    Formatting. We could put the parameters one in each line.


> 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)

Mime
View raw message