[jira] [Commented] (SPARK-1359) SGD implementation is not efficient
[ https://issues.apache.org/jira/browse/SPARK-1359?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16013998#comment-16013998 ] Nick Pentreath commented on SPARK-1359: --- Do we care much about this now, since {{mllib}}'s SGD is in maintenance mode and currently {{ml}} supports only L-BFGS (or IRLS)? Shall we close this off? > SGD implementation is not efficient > --- > > Key: SPARK-1359 > URL: https://issues.apache.org/jira/browse/SPARK-1359 > Project: Spark > Issue Type: Improvement > Components: MLlib >Affects Versions: 0.9.0, 1.0.0 >Reporter: Xiangrui Meng > > The SGD implementation samples a mini-batch to compute the stochastic > gradient. This is not efficient because examples are provided via an iterator > interface. We have to scan all of them to obtain a sample. -- This message was sent by Atlassian JIRA (v6.3.15#6346) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-1359) SGD implementation is not efficient
[ https://issues.apache.org/jira/browse/SPARK-1359?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15981794#comment-15981794 ] yu peng commented on SPARK-1359: i think by randomly shuffle partitions and do gradient Descent by toLocalIterator is better and without compromising accuracy > SGD implementation is not efficient > --- > > Key: SPARK-1359 > URL: https://issues.apache.org/jira/browse/SPARK-1359 > Project: Spark > Issue Type: Improvement > Components: MLlib >Affects Versions: 0.9.0, 1.0.0 >Reporter: Xiangrui Meng > > The SGD implementation samples a mini-batch to compute the stochastic > gradient. This is not efficient because examples are provided via an iterator > interface. We have to scan all of them to obtain a sample. -- This message was sent by Atlassian JIRA (v6.3.15#6346) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-1359) SGD implementation is not efficient
[ https://issues.apache.org/jira/browse/SPARK-1359?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15219190#comment-15219190 ] Yu Ishikawa commented on SPARK-1359: [~mbaddar] Since the current ann in mllib depends on `GradientDescent`, we should modify the efficienty. How do we evaluate new implementation against the current implementation? And What are better tasks to evaluate it? - Metrics 1. Convergence Effieiency 2. Compute Cost 3. Compute Time 4. Other - Task 1. Logistic Regression and Linear Regression with random generated data 2. Logistic Regression and Linear Regression with any Kaggle data 3. Other I make an implementation of Parallelized Stochastic Gradient Descent. https://github.com/yu-iskw/spark-parallelized-sgd > SGD implementation is not efficient > --- > > Key: SPARK-1359 > URL: https://issues.apache.org/jira/browse/SPARK-1359 > Project: Spark > Issue Type: Improvement > Components: MLlib >Affects Versions: 0.9.0, 1.0.0 >Reporter: Xiangrui Meng > > The SGD implementation samples a mini-batch to compute the stochastic > gradient. This is not efficient because examples are provided via an iterator > interface. We have to scan all of them to obtain a sample. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-1359) SGD implementation is not efficient
[ https://issues.apache.org/jira/browse/SPARK-1359?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15197062#comment-15197062 ] Mohamed Baddar commented on SPARK-1359: --- [~mengxr] If this issue is still of interest and nobody is working on it , I can start implementation. > SGD implementation is not efficient > --- > > Key: SPARK-1359 > URL: https://issues.apache.org/jira/browse/SPARK-1359 > Project: Spark > Issue Type: Improvement > Components: MLlib >Affects Versions: 0.9.0, 1.0.0 >Reporter: Xiangrui Meng > > The SGD implementation samples a mini-batch to compute the stochastic > gradient. This is not efficient because examples are provided via an iterator > interface. We have to scan all of them to obtain a sample. -- This message was sent by Atlassian JIRA (v6.3.4#6332) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org