spark-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Bryan Cutler (JIRA)" <>
Subject [jira] [Commented] (SPARK-22126) Fix model-specific optimization support for ML tuning
Date Tue, 02 Jan 2018 23:09:00 GMT


Bryan Cutler commented on SPARK-22126:

Thanks for taking a look [~josephkb]!  I believe it's possible to still use the current fit()
API that returns a {{Seq[Model[_]]}} and avoid materializing all models in memory at once.
 If an estimator has model-specific optimizations and is creating multiple models, it could
return a lazy sequence (such as a SeqVew or Stream).  Then if the CrossValidator just converts
the sequence of Models to an iterator, it will only hold a reference to the model currently
being evaluated and previous models can be GC'd.  Does that sound like it would be worth a
shot here?

As for the issue of parallelism in model-specific optimizations, it's true there might be
some benefit in the Estimator being able to allow the CrossValidator to handle the parallelism
under certain cases.  But until there are more examples of this to look at, it's hard to know
if making a new API for it is worth that benefit.  I saw a reference to a Keras model in another
JIRA, is that an example that could have a model-specific optimization while still allowing
the CrossValidator to parallelize over it?

> Fix model-specific optimization support for ML tuning
> -----------------------------------------------------
>                 Key: SPARK-22126
>                 URL:
>             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
> 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,
> {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 "" 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]]] = {
> { 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 { }
>      .flatMap(_).sortBy(_._1).map(_._2)
> }
> {code}
> If use the API I proposed above, the code in [CrossValidation|]
> 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 =
>             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 = { modelMapFuture =>
> { modelMap =>
>  { 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 =, Duration.Inf))
>           .map(_.toSeq).sortBy(_._1).map(_._2)
> {code}

This message was sent by Atlassian JIRA

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message