[
https://issues.apache.org/jira/browse/MAHOUT-1736?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14599877#comment-14599877
]
ASF GitHub Bot commented on MAHOUT-1736:
----------------------------------------
Github user dlyubimov commented on the pull request:
https://github.com/apache/mahout/pull/143#issuecomment-114967473
@avati question.
so all in-core operations work for H20Matrix in the backend too? for
example, if allreduceBlock defines reduce as x1 rbind x2 (stacking
concatenation) that would work ?
i don't think we repeat all in-core tests for H20Matrix
mostly because we assumed the backends would use the same classes in the
back from mahout-math.
We confirm functionality, indirectly, by running current distributed
algorithms, but they don't cover everything that in-core tests in
mahout-math-scala cover for in-core matrices.
> H20 implementation for allreduceBlock
> -------------------------------------
>
> Key: MAHOUT-1736
> URL: https://issues.apache.org/jira/browse/MAHOUT-1736
> Project: Mahout
> Issue Type: Improvement
> Affects Versions: 0.10.1
> Reporter: Dmitriy Lyubimov
> Assignee: Anand Avati
> Fix For: 0.10.2
>
>
> allreduce aggregator on matrix blocks:
> {code}
> override def allreduceBlock[K: ClassTag](drm: CheckpointedDrm[K], bmf:
> BlockMapFunc2[K], rf: BlockReduceFunc)
> : Matrix = ???
> {code}
> This is pretty important for custom aggregated statistics
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
