[
https://issues.apache.org/jira/browse/FLINK-1807?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14532359#comment-14532359
]
ASF GitHub Bot commented on FLINK-1807:
---------------------------------------
Github user thvasilo commented on a diff in the pull request:
https://github.com/apache/flink/pull/613#discussion_r29838305
--- Diff:
flink-staging/flink-ml/src/main/scala/org/apache/flink/ml/optimization/GradientDescent.scala
---
@@ -0,0 +1,254 @@
+/*
+ * 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.common.functions.RichMapFunction
+import org.apache.flink.api.scala._
+import org.apache.flink.configuration.Configuration
+import org.apache.flink.ml.common._
+import org.apache.flink.ml.math._
+import org.apache.flink.ml.optimization.IterativeSolver.{Iterations,
Stepsize}
+import org.apache.flink.ml.optimization.Solver._
+
+/** This [[Solver]] performs Stochastic Gradient Descent optimization
using mini batches
+ *
+ * For each labeled vector in a mini batch the gradient is computed and
added to a partial
+ * gradient. The partial gradients are then summed and divided by the
size of the batches. The
+ * average gradient is then used to updated the weight values, including
regularization.
+ *
+ * At the moment, the whole partition is used for SGD, making it
effectively a batch gradient
+ * descent. Once a sampling operator has been introduced, the algorithm
can be optimized
+ *
+ * @param runParameters The parameters to tune the algorithm. Currently
these include:
+ * [[Solver.LossFunction]] for the loss function to
be used,
+ * [[Solver.RegularizationType]] for the type of
regularization,
+ * [[Solver.RegularizationParameter]] for the
regularization parameter,
+ * [[IterativeSolver.Iterations]] for the maximum
number of iteration,
+ * [[IterativeSolver.Stepsize]] for the learning
rate used.
+ */
+class GradientDescent(runParameters: ParameterMap) extends IterativeSolver
{
+
+ import Solver.WEIGHTVECTOR_BROADCAST
+
+ var parameterMap: ParameterMap = parameters ++ runParameters
+
+ // TODO(tvas): Use once we have proper sampling in place
+// case object MiniBatchFraction extends Parameter[Double] {
+// val defaultValue = Some(1.0)
+// }
+//
+// def setMiniBatchFraction(fraction: Double): GradientDescent = {
+// parameterMap.add(MiniBatchFraction, fraction)
+// this
+// }
+
+ /** Performs one iteration of Stochastic Gradient Descent using mini
batches
+ *
+ * @param data A Dataset of LabeledVector (label, features) pairs
+ * @param currentWeights A Dataset with the current weights to be
optimized as its only element
+ * @return A Dataset containing the weights after one stochastic
gradient descent step
+ */
+ private def SGDStep(data: DataSet[(LabeledVector)], currentWeights:
DataSet[WeightVector]):
+ DataSet[WeightVector] = {
+
+ // TODO: Sample from input to realize proper SGD
+ data.map {
+ new GradientCalculation
+ }.withBroadcastSet(currentWeights, WEIGHTVECTOR_BROADCAST).reduce {
+ (left, right) =>
+ val (leftGradVector, leftLoss, leftCount) = left
+ val (rightGradVector, rightLoss, rightCount) = right
+ // Add the left gradient to the right one
+ BLAS.axpy(1.0, leftGradVector.weights, rightGradVector.weights)
+ val gradients = WeightVector(
+ rightGradVector.weights, leftGradVector.intercept +
rightGradVector.intercept)
+
+ (gradients , leftLoss + rightLoss, leftCount + rightCount)
+ }.map {
+ new WeightsUpdate
+ }.withBroadcastSet(currentWeights, WEIGHTVECTOR_BROADCAST)
+ }
+
+ /** Provides a solution for the given optimization problem
+ *
+ * @param data A Dataset of LabeledVector (label, features) pairs
+ * @param initWeights The initial weights that will be optimized
+ * @return The weights, optimized for the provided data.
+ */
+ override def optimize(data: DataSet[LabeledVector], initWeights:
Option[DataSet[WeightVector]]):
+ DataSet[WeightVector] = {
+ // TODO: Faster way to do this?
+ val dimensionsDS = data.map(_.vector.size).reduce((a, b) => b)
+
+ val numberOfIterations: Int = parameterMap(Iterations)
+
+ val initialWeightsDS: DataSet[WeightVector] = initWeights match {
+ case Some(x) => x
+ case None => createInitialWeightVector(dimensionsDS)
+ }
+
+ // Perform the iterations
+ // TODO: Enable convergence stopping criterion, as in Multiple Linear
regression
+ initialWeightsDS.iterate(numberOfIterations) {
+ weightVector => {
+ SGDStep(data, weightVector)
+ }
+ }
+ }
+
+ /** Mapping function that calculates the weight gradients from the data.
+ *
+ */
+ private class GradientCalculation extends
+ RichMapFunction[LabeledVector, (WeightVector, Double, Int)] {
--- End diff --
Do you mean something like this:
```scala
/** Mapping function that calculates the weight gradients from the data.
*
*/
private class GradientCalculation extends RichMapFunction[
LabeledVector,
(WeightVector, Double, Int)] {
var weightVector: WeightVector = null
...
```
> 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)
