[jira] [Updated] (SYSTEMML-1659) New rewrite eliminate unnecessary aggregates

["Matthias Boehm (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/SYSTEMML-1659?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Matthias Boehm updated SYSTEMML-1659:
-------------------------------------
    Description: 
There are some example scripts which use unnecessary aggregations for convenience of a consistent specification - hence, we should automatically eliminate these unnecessary aggregates, via a simple static rewrite. For example, in below script, we should remove the unnecessary rowSums, but the same applies to min-min, max-max, sum_sq-sum.

{code}
 # nn lib - cross_entropy_loss::forward (returns only loss)
 N = nrow(y)
 eps = 1e-10  # numerical stability to avoid log(0)
 losses = rowSums(-y * log(pred+eps))
 loss = sum(losses) / N
{code}

  was:
There are some example scripts which use unnecessary aggregations for convenience of a consistent specification - hence, we should automatically eliminate these unnecessary aggregates, via a simple static rewrite.

{code}
 # nn lib - cross_entropy_loss::forward (returns only loss)
 N = nrow(y)
 eps = 1e-10  # numerical stability to avoid log(0)
 losses = rowSums(-y * log(pred+eps))
 loss = sum(losses) / N
{code}


> New rewrite eliminate unnecessary aggregates
> --------------------------------------------
>
>                 Key: SYSTEMML-1659
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-1659
>             Project: SystemML
>          Issue Type: Sub-task
>          Components: Compiler
>            Reporter: Matthias Boehm
>             Fix For: SystemML 1.0
>
>
> There are some example scripts which use unnecessary aggregations for convenience of a consistent specification - hence, we should automatically eliminate these unnecessary aggregates, via a simple static rewrite. For example, in below script, we should remove the unnecessary rowSums, but the same applies to min-min, max-max, sum_sq-sum.
> {code}
>  # nn lib - cross_entropy_loss::forward (returns only loss)
>  N = nrow(y)
>  eps = 1e-10  # numerical stability to avoid log(0)
>  losses = rowSums(-y * log(pred+eps))
>  loss = sum(losses) / N
> {code}



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)