[
https://issues.apache.org/jira/browse/SPARK-14053?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15205514#comment-15205514
]
DB Tsai edited comment on SPARK-14053 at 3/22/16 1:13 AM:
----------------------------------------------------------
This makes sense for me. We just need to document it properly. Also, the
current code for comparing double is symmetric. We can do
```If (abs(y) > eps / t && abs(x) > eps / t) test abs(y - x) < t *
math.min(absX, absY)
else test abs(y - x) < eps
```
was (Author: dbtsai):
This makes sense for me. We just need to document it properly. Also, the
current code for comparing double is symmetric. We can do
If (abs(y) > eps / t && abs(x) > eps / t) test abs(y - x) < t * math.min(absX,
absY)
else test abs(y - x) < eps
> Merge absTol and relTol into one in MLlib tests
> -----------------------------------------------
>
> Key: SPARK-14053
> URL: https://issues.apache.org/jira/browse/SPARK-14053
> Project: Spark
> Issue Type: Improvement
> Components: ML, MLlib, Tests
> Affects Versions: 2.0.0
> Reporter: Xiangrui Meng
> Assignee: Xiangrui Meng
>
> We have absTol and relTol in MLlib tests to compare values with possible
> numerical differences. However, in most cases we should just use relTol. Many
> absTol are not used properly. See
> https://github.com/apache/spark/search?q=absTol. One corner case relTol
> doesn't handle is when the target value is 0. We can make the following
> change to relTol to solve the issue. Consider `x ~== y relTol t`.
> 1. If abs( y ) > eps / t, test abs(y - x) / abs( y ) < t,
> 2. else test abs(y - x) < eps
> where eps is a reasonably small value, e.g., 1e-14. Note that the transition
> is smooth at abs( y ) = eps / t.
> cc [~dbtsai]
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
