[ https://issues.apache.org/jira/browse/SPARK-30661?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17478437#comment-17478437 ]
zhengruifeng commented on SPARK-30661: -------------------------------------- recently, I spend some time on testing blockify kmeans and apply GEMM in finding the closest cluster. In short: 1, for sparse datasets, blockifying kmeans still cause regression in most cases; (existing impl with triangle-inequality can skip some distance computation, but scala-based sparse BLAS will always compute all distances) 2, for dense datasets and small k, blockifying kmeans (without native BLAS) is competitive; with native BLAS, it should be significantly faster than existing impl. So I plan to add a new parameter {{solver}} by making KMeans extending HasSolver, and support both two training impls, so that end users can switch to the blockify version. How do you think about it? [~srowen] [~WeichenXu123] [~mengxr] [~huaxingao] > KMeans blockify input vectors > ----------------------------- > > Key: SPARK-30661 > URL: https://issues.apache.org/jira/browse/SPARK-30661 > Project: Spark > Issue Type: Sub-task > Components: ML, PySpark > Affects Versions: 3.0.0 > Reporter: zhengruifeng > Assignee: zhengruifeng > Priority: Minor > -- This message was sent by Atlassian Jira (v8.20.1#820001) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org