Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/692#discussion_r30780043
--- Diff:
flink-staging/flink-ml/src/main/scala/org/apache/flink/ml/optimization/GradientDescent.scala
---
@@ -76,86 +77,163 @@ class GradientDescent(runParameters: ParameterMap)
extends IterativeSolver {
}.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
+ * @param initialWeights 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)
+ initialWeights: Option[DataSet[WeightVector]]): DataSet[WeightVector]
= {
+ val numberOfIterations: Int = parameters(Iterations)
+ // TODO(tvas): This looks out of place, why don't we get back an
Option from
+ // parameters(ConvergenceThreshold)?
+ val convergenceThresholdOption = parameters.get(ConvergenceThreshold)
// Initialize weights
- val initialWeightsDS: DataSet[WeightVector] = initWeights match {
- // Ensure provided weight vector is a DenseVector
- case Some(wvDS) => {
- wvDS.map{wv => {
- val denseWeights = wv.weights match {
- case dv: DenseVector => dv
- case sv: SparseVector => sv.toDenseVector
+ val initialWeightsDS: DataSet[WeightVector] =
createInitialWeightsDS(initialWeights, data)
+
+ // Perform the iterations
+ val optimizedWeights = convergenceThresholdOption match {
+ // No convergence criterion
+ case None =>
+ initialWeightsDS.iterate(numberOfIterations) {
+ weightVectorDS => {
+ SGDStep(data, weightVectorDS)
}
- WeightVector(denseWeights, wv.intercept)
}
-
+ case Some(convergence) =>
+ /** Calculates the regularized loss, from the data and given
weights **/
--- End diff --
One line comments should use `//`
---
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 infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
