From issues-return-14312-apmail-flink-issues-archive=flink.apache.org@flink.apache.org Thu May 7 09:45:20 2015 Return-Path: X-Original-To: apmail-flink-issues-archive@minotaur.apache.org Delivered-To: apmail-flink-issues-archive@minotaur.apache.org Received: from mail.apache.org (hermes.apache.org [140.211.11.3]) by minotaur.apache.org (Postfix) with SMTP id 1D14710E6D for ; Thu, 7 May 2015 09:45:20 +0000 (UTC) Received: (qmail 56132 invoked by uid 500); 7 May 2015 09:45:20 -0000 Delivered-To: apmail-flink-issues-archive@flink.apache.org Received: (qmail 56090 invoked by uid 500); 7 May 2015 09:45:20 -0000 Mailing-List: contact issues-help@flink.apache.org; run by ezmlm Precedence: bulk List-Help: List-Unsubscribe: List-Post: List-Id: Reply-To: dev@flink.apache.org Delivered-To: mailing list issues@flink.apache.org Received: (qmail 56081 invoked by uid 99); 7 May 2015 09:45:20 -0000 Received: from Unknown (HELO spamd2-us-west.apache.org) (209.188.14.142) by apache.org (qpsmtpd/0.29) with ESMTP; Thu, 07 May 2015 09:45:20 +0000 Received: from localhost (localhost [127.0.0.1]) by spamd2-us-west.apache.org (ASF Mail Server at spamd2-us-west.apache.org) with ESMTP id 8F34B1A230D for ; Thu, 7 May 2015 09:45:19 +0000 (UTC) X-Virus-Scanned: Debian amavisd-new at spamd2-us-west.apache.org X-Spam-Flag: NO X-Spam-Score: -0.009 X-Spam-Level: X-Spam-Status: No, score=-0.009 tagged_above=-999 required=6.31 tests=[T_RP_MATCHES_RCVD=-0.01, URIBL_BLOCKED=0.001] autolearn=disabled Received: from mx1-us-west.apache.org ([10.40.0.8]) by localhost (spamd2-us-west.apache.org [10.40.0.9]) (amavisd-new, port 10024) with ESMTP id 96HhLlevze7K for ; Thu, 7 May 2015 09:45:04 +0000 (UTC) Received: from mail.apache.org (hermes.apache.org [140.211.11.3]) by mx1-us-west.apache.org (ASF Mail Server at mx1-us-west.apache.org) with SMTP id 1BBC0286DE for ; Thu, 7 May 2015 09:35:56 +0000 (UTC) Received: (qmail 20656 invoked by uid 99); 7 May 2015 09:35:56 -0000 Received: from git1-us-west.apache.org (HELO git1-us-west.apache.org) (140.211.11.23) by apache.org (qpsmtpd/0.29) with ESMTP; Thu, 07 May 2015 09:35:56 +0000 Received: by git1-us-west.apache.org (ASF Mail Server at git1-us-west.apache.org, from userid 33) id DEAD8E4415; Thu, 7 May 2015 09:35:55 +0000 (UTC) From: tillrohrmann To: issues@flink.incubator.apache.org Reply-To: issues@flink.incubator.apache.org References: In-Reply-To: Subject: [GitHub] flink pull request: [WIP] - [FLINK-1807/1889] - Optimization frame... Content-Type: text/plain Message-Id: <20150507093555.DEAD8E4415@git1-us-west.apache.org> Date: Thu, 7 May 2015 09:35:55 +0000 (UTC) Github user tillrohrmann commented on a diff in the pull request: https://github.com/apache/flink/pull/613#discussion_r29837206 --- 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} +} + +/** $L_2$ regularization penalty. + * + * Penalizes large weights, favoring solutions with more small weights rather than few large ones. + * + */ +class L2Regularization extends DiffRegularizationType { + + /** Adds regularization to the loss value **/ + override def regLoss(oldLoss: Double, weightVector: FlinkVector, regParameter: Double) + : Double = { + val brzVector = weightVector.asBreeze + oldLoss + regParameter * (brzVector dot brzVector) / 2 + } + + /** Adds regularization gradient to the loss gradient. The gradient is updated in place **/ + override def regGradient( + weightVector: FlinkVector, + lossGradient: FlinkVector, + regParameter: Double): Unit = { + BLAS.axpy(regParameter, weightVector, lossGradient) + } +} + +/** $L_1$ regularization penalty. + * + * The $L_1$ penalty can be used to drive a number of the solution coefficients to 0, thereby + * producing sparse solutions. + * + */ +class L1Regularization extends RegularizationType { + /** Calculates and applies the regularization amount and the regularization parameter + * + * Implementation was taken from the Apache Spark Mllib library: + * http://git.io/vfZIT + * @param oldWeights The old weights + * @param effectiveStepSize The effective step size for this iteration + * @param regParameter The current regularization parameter + * @return A tuple whose first element is the updated weight FlinkVector and the second is the + * regularization value + */ + override def takeStep( + oldWeights: FlinkVector, + gradient: FlinkVector, + effectiveStepSize: Double, + regParameter: Double) { + BLAS.axpy(-effectiveStepSize, gradient, oldWeights) + val brzWeights = oldWeights.asBreeze.toDenseVector --- End diff -- why converting the oldWeights explicitly to a dense vector. It should be dense since the beginning, or at least we should make sure that it is. --- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastructure@apache.org or file a JIRA ticket with INFRA. ---