[jira] [Comment Edited] (SPARK-2372) Grouped Optimization/Learning
[ https://issues.apache.org/jira/browse/SPARK-2372?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14094668#comment-14094668 ] Kyle Ellrott edited comment on SPARK-2372 at 8/13/14 6:06 AM: -- GroupedBinaryClassificationMetrics has been added to the pull request connected to this issue. GroupedBinaryClassificationMetrics is an re-write of the BinaryClassificationMetrics methods, but it work on a RDD[KEY,(Double,Double)] structure (rather then the RDD[(Double,Double)] that BinaryClassificationMetrics takes), where KEY is a generic that will be the type of the key used to identified the data set. A unit test is included do validate these function work in the same way as the BinaryClassificationMetrics implementations. https://github.com/kellrott/spark/commit/dcabb2f6a39c0940afc39e809a50601f46e50162 was (Author: kellrott): GroupedBinaryClassificationMetrics has been added to the pull request connected to this issue. GroupedBinaryClassificationMetrics is an re-write of the BinaryClassificationMetrics methods, but it work on a RDD[KEY,(Double,Double)] structure (rather then the RDD[(Double,Double)] that BinaryClassificationMetrics takes), where KEY is a generic that will be the type of the key used to identified the data set. Now methods return Map[KEY,Double], with a different score for each data set, rather then a single 'Double' A unit test is included do validate these function work in the same way as the BinaryClassificationMetrics implementations. https://github.com/kellrott/spark/commit/dcabb2f6a39c0940afc39e809a50601f46e50162 Grouped Optimization/Learning - Key: SPARK-2372 URL: https://issues.apache.org/jira/browse/SPARK-2372 Project: Spark Issue Type: New Feature Components: MLlib Affects Versions: 1.0.1, 1.1.0, 1.0.2 Reporter: Kyle Ellrott The purpose of this patch is the enable MLLib to better handle scenarios where the user would want to do learning on multiple feature/label sets at the same time. Rather then schedule each learning task separately, this patch lets the user create a single RDD with an Int key to represent the 'group' sets of entries belong to. This patch establishing the GroupedOptimizer trait, for which GroupedGradientDescent has been implemented. This systems differs from the original Optimizer trait in that the original optimize method accepted RDD[(Int, Vector)] the new GroupedOptimizer accepts RDD[(Int, (Double, Vector))]. The difference is that the GroupedOptimizer uses a 'group' ID key in the RDD to multiplex multiple optimization operations in the same RDD. This patch also establishes the GroupedGeneralizedLinearAlgorithm trait, for which the 'run' method has had the RDD[LabeledPoint] input replaced with RDD[(Int,LabeledPoint)]. This patch also provides a unit test and utility to take the results of MLUtils.kFold and turn it into a single grouped RDD, ready for simultaneous learning. https://github.com/apache/spark/pull/1292 -- This message was sent by Atlassian JIRA (v6.2#6252) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-1486) Support multi-model training in MLlib
[ https://issues.apache.org/jira/browse/SPARK-1486?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14094673#comment-14094673 ] Kyle Ellrott commented on SPARK-1486: - It would be helpful to get some feedback if the work being done for SPARK-2372 would help with this issue. Support multi-model training in MLlib - Key: SPARK-1486 URL: https://issues.apache.org/jira/browse/SPARK-1486 Project: Spark Issue Type: Improvement Components: MLlib Reporter: Xiangrui Meng Assignee: Xiangrui Meng Priority: Critical It is rare in practice to train just one model with a given set of parameters. Usually, this is done by training multiple models with different sets of parameters and then select the best based on their performance on the validation set. MLlib should provide native support for multi-model training/scoring. It requires decoupling of concepts like problem, formulation, algorithm, parameter set, and model, which are missing in MLlib now. MLI implements similar concepts, which we can borrow. There are different approaches for multi-model training: 0) Keep one copy of the data, and train models one after another (or maybe in parallel, depending on the scheduler). 1) Keep one copy of the data, and train multiple models at the same time (similar to `runs` in KMeans). 2) Make multiple copies of the data (still stored distributively), and use more cores to distribute the work. 3) Collect the data, make the entire dataset available on workers, and train one or more models on each worker. Users should be able to choose which execution mode they want to use. Note that 3) could cover many use cases in practice when the training data is not huge, e.g., 1GB. This task will be divided into sub-tasks and this JIRA is created to discuss the design and track the overall progress. -- This message was sent by Atlassian JIRA (v6.2#6252) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-2372) Grouped Optimization/Learning
[ https://issues.apache.org/jira/browse/SPARK-2372?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14094668#comment-14094668 ] Kyle Ellrott commented on SPARK-2372: - GroupedBinaryClassificationMetrics has been added to the pull request connected to this issue. GroupedBinaryClassificationMetrics is an re-write of the BinaryClassificationMetrics methods, but it work on a RDD[KEY,(Double,Double)] structure (rather then the RDD[(Double,Double)] that BinaryClassificationMetrics takes), where KEY is a generic that will be the type of the key used to identified the data set. Now methods return Map[KEY,Double], with a different score for each data set, rather then a single 'Double' A unit test is included do validate these function work in the same way as the BinaryClassificationMetrics implementations. https://github.com/kellrott/spark/commit/dcabb2f6a39c0940afc39e809a50601f46e50162 Grouped Optimization/Learning - Key: SPARK-2372 URL: https://issues.apache.org/jira/browse/SPARK-2372 Project: Spark Issue Type: New Feature Components: MLlib Affects Versions: 1.0.1, 1.1.0, 1.0.2 Reporter: Kyle Ellrott The purpose of this patch is the enable MLLib to better handle scenarios where the user would want to do learning on multiple feature/label sets at the same time. Rather then schedule each learning task separately, this patch lets the user create a single RDD with an Int key to represent the 'group' sets of entries belong to. This patch establishing the GroupedOptimizer trait, for which GroupedGradientDescent has been implemented. This systems differs from the original Optimizer trait in that the original optimize method accepted RDD[(Int, Vector)] the new GroupedOptimizer accepts RDD[(Int, (Double, Vector))]. The difference is that the GroupedOptimizer uses a 'group' ID key in the RDD to multiplex multiple optimization operations in the same RDD. This patch also establishes the GroupedGeneralizedLinearAlgorithm trait, for which the 'run' method has had the RDD[LabeledPoint] input replaced with RDD[(Int,LabeledPoint)]. This patch also provides a unit test and utility to take the results of MLUtils.kFold and turn it into a single grouped RDD, ready for simultaneous learning. https://github.com/apache/spark/pull/1292 -- This message was sent by Atlassian JIRA (v6.2#6252) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Created] (SPARK-2372) Grouped Optimization/Learning
Kyle Ellrott created SPARK-2372: --- Summary: Grouped Optimization/Learning Key: SPARK-2372 URL: https://issues.apache.org/jira/browse/SPARK-2372 Project: Spark Issue Type: New Feature Components: MLlib Affects Versions: 1.0.1, 1.1.0, 1.0.2 Reporter: Kyle Ellrott The purpose of this patch is the enable MLLib to better handle scenarios where the user would want to do learning on multiple feature/label sets at the same time. Rather then schedule each learning task separately, this patch lets the user create a single RDD with an Int key to represent the 'group' sets of entries belong to. This patch establishing the GroupedOptimizer trait, for which GroupedGradientDescent has been implemented. This systems differs from the original Optimizer trait in that the original optimize method accepted RDD[(Int, Vector)] the new GroupedOptimizer accepts RDD[(Int, (Double, Vector))]. The difference is that the GroupedOptimizer uses a 'group' ID key in the RDD to multiplex multiple optimization operations in the same RDD. This patch also establishes the GroupedGeneralizedLinearAlgorithm trait, for which the 'run' method has had the RDD[LabeledPoint] input replaced with RDD[(Int,LabeledPoint)]. This patch also provides a unit test and utility to take the results of MLUtils.kFold and turn it into a single grouped RDD, ready for simultaneous learning. https://github.com/apache/spark/pull/1292 -- This message was sent by Atlassian JIRA (v6.2#6252)
[jira] [Commented] (SPARK-1486) Support multi-model training in MLlib
[ https://issues.apache.org/jira/browse/SPARK-1486?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14052783#comment-14052783 ] Kyle Ellrott commented on SPARK-1486: - In the case where you are using the same training/optimization procedure, you can group together the sets of samples using a key. There is an example of this in the pull request linked to SPARK-2372. This could be a very effective way to deal with simultaneously training all of the different folds from a kFolds split. Support multi-model training in MLlib - Key: SPARK-1486 URL: https://issues.apache.org/jira/browse/SPARK-1486 Project: Spark Issue Type: Improvement Components: MLlib Reporter: Xiangrui Meng Assignee: Xiangrui Meng Priority: Critical Fix For: 1.1.0 It is rare in practice to train just one model with a given set of parameters. Usually, this is done by training multiple models with different sets of parameters and then select the best based on their performance on the validation set. MLlib should provide native support for multi-model training/scoring. It requires decoupling of concepts like problem, formulation, algorithm, parameter set, and model, which are missing in MLlib now. MLI implements similar concepts, which we can borrow. There are different approaches for multi-model training: 0) Keep one copy of the data, and train models one after another (or maybe in parallel, depending on the scheduler). 1) Keep one copy of the data, and train multiple models at the same time (similar to `runs` in KMeans). 2) Make multiple copies of the data (still stored distributively), and use more cores to distribute the work. 3) Collect the data, make the entire dataset available on workers, and train one or more models on each worker. Users should be able to choose which execution mode they want to use. Note that 3) could cover many use cases in practice when the training data is not huge, e.g., 1GB. This task will be divided into sub-tasks and this JIRA is created to discuss the design and track the overall progress. -- This message was sent by Atlassian JIRA (v6.2#6252)
