[
https://issues.apache.org/jira/browse/SPARK-22126?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16312264#comment-16312264
]
Bago Amirbekian commented on SPARK-22126:
-----------------------------------------
[~bryanc] thanks for taking the time to put together the PR and share thoughts.
I like the idea of being able to preserve the existing APIs and not needing to
add a new fitMultiple API but I'm concerned the existing APIs aren't quite
flexible enough.
One of the use cases that motivated the {{ fitMultiple }} API was optimizing
the Pipeline Estimator. The Pipeline Estimator seems like in important one to
optimize because I believe it's required in order for CrossValidator to be able
to exploit optimized implementations of the {{ fit }}/{{ fitMultiple }} methods
of Pipeline stages.
The way one would optimize the Pipeline Estimator is to group the paramMaps
into a tree structure where each level represents a stage with a param that can
take multiple values. One would then traverse the tree in depth first order.
Notice that in this case the params need not be estimator params, but could
actually be transformer params as well since we can avoid applying expensive
transformers multiple times.
With this approach all the params for a pipeline estimator after the top level
of the tree are "optimizable" so simply being group on optimizable params isn't
sufficient, we need to actually order the paramMaps to match the depth first
traversal of the stages tree.
I'm still thinking through all this in my head so let me know if any of it is
off base or not clear, but I think the advantage of the {{ fitMultiple }}
approach gives us full flexibility in order to these kinds of optimizations.
> Fix model-specific optimization support for ML tuning
> -----------------------------------------------------
>
> Key: SPARK-22126
> URL: https://issues.apache.org/jira/browse/SPARK-22126
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Affects Versions: 2.3.0
> Reporter: Weichen Xu
>
> Fix model-specific optimization support for ML tuning. This is discussed in
> SPARK-19357
> more discussion is here
> https://gist.github.com/MrBago/f501b9e7712dc6a67dc9fea24e309bf0
> Anyone who's following might want to scan the design doc (in the links
> above), the latest api proposal is:
> {code}
> def fitMultiple(
> dataset: Dataset[_],
> paramMaps: Array[ParamMap]
> ): java.util.Iterator[scala.Tuple2[java.lang.Integer, Model]]
> {code}
> Old discussion:
> I copy discussion from gist to here:
> I propose to design API as:
> {code}
> def fitCallables(dataset: Dataset[_], paramMaps: Array[ParamMap]):
> Array[Callable[Map[Int, M]]]
> {code}
> Let me use an example to explain the API:
> {quote}
> It could be possible to still use the current parallelism and still allow
> for model-specific optimizations. For example, if we doing cross validation
> and have a param map with regParam = (0.1, 0.3) and maxIter = (5, 10). Lets
> say that the cross validator could know that maxIter is optimized for the
> model being evaluated (e.g. a new method in Estimator that return such
> params). It would then be straightforward for the cross validator to remove
> maxIter from the param map that will be parallelized over and use it to
> create 2 arrays of paramMaps: ((regParam=0.1, maxIter=5), (regParam=0.1,
> maxIter=10)) and ((regParam=0.3, maxIter=5), (regParam=0.3, maxIter=10)).
> {quote}
> In this example, we can see that, models computed from ((regParam=0.1,
> maxIter=5), (regParam=0.1, maxIter=10)) can only be computed in one thread
> code, models computed from ((regParam=0.3, maxIter=5), (regParam=0.3,
> maxIter=10)) in another thread. In this example, there're 4 paramMaps, but
> we can at most generate two threads to compute the models for them.
> The API above allow "callable.call()" to return multiple models, and return
> type is {code}Map[Int, M]{code}, key is integer, used to mark the paramMap
> index for corresponding model. Use the example above, there're 4 paramMaps,
> but only return 2 callable objects, one callable object for ((regParam=0.1,
> maxIter=5), (regParam=0.1, maxIter=10)), another one for ((regParam=0.3,
> maxIter=5), (regParam=0.3, maxIter=10)).
> and the default "fitCallables/fit with paramMaps" can be implemented as
> following:
> {code}
> def fitCallables(dataset: Dataset[_], paramMaps: Array[ParamMap]):
> Array[Callable[Map[Int, M]]] = {
> paramMaps.zipWithIndex.map { case (paramMap: ParamMap, index: Int) =>
> new Callable[Map[Int, M]] {
> override def call(): Map[Int, M] = Map(index -> fit(dataset, paramMap))
> }
> }
> }
> def fit(dataset: Dataset[_], paramMaps: Array[ParamMap]): Seq[M] = {
> fitCallables(dataset, paramMaps).map { _.call().toSeq }
> .flatMap(_).sortBy(_._1).map(_._2)
> }
> {code}
> If use the API I proposed above, the code in
> [CrossValidation|https://github.com/apache/spark/blob/master/mllib/src/main/scala/org/apache/spark/ml/tuning/CrossValidator.scala#L149-L159]
> can be changed to:
> {code}
> val trainingDataset = sparkSession.createDataFrame(training,
> schema).cache()
> val validationDataset = sparkSession.createDataFrame(validation,
> schema).cache()
> // Fit models in a Future for training in parallel
> val modelMapFutures = fitCallables(trainingDataset, paramMaps).map {
> callable =>
> Future[Map[Int, Model[_]]] {
> val modelMap = callable.call()
> if (collectSubModelsParam) {
> ...
> }
> modelMap
> } (executionContext)
> }
> // Unpersist training data only when all models have trained
> Future.sequence[Model[_], Iterable](modelMapFutures)(implicitly,
> executionContext)
> .onComplete { _ => trainingDataset.unpersist() } (executionContext)
> // Evaluate models in a Future that will calulate a metric and allow
> model to be cleaned up
> val foldMetricMapFutures = modelMapFutures.map { modelMapFuture =>
> modelMapFuture.map { modelMap =>
> modelMap.map { case (index: Int, model: Model[_]) =>
> val metric = eval.evaluate(model.transform(validationDataset,
> paramMaps(index)))
> (index, metric)
> }
> } (executionContext)
> }
> // Wait for metrics to be calculated before unpersisting validation
> dataset
> val foldMetrics = foldMetricMapFutures.map(ThreadUtils.awaitResult(_,
> Duration.Inf))
> .map(_.toSeq).sortBy(_._1).map(_._2)
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
