[
https://issues.apache.org/jira/browse/OPENNLP-722?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14175062#comment-14175062
]
Vinh Khuc commented on OPENNLP-722:
-----------------------------------
Looks like this issue comes from changes in Java 8's floating point arithmetic.
If in this line
https://github.com/apache/opennlp/blob/trunk/opennlp-tools/src/main/java/opennlp/tools/ml/perceptron/PerceptronTrainer.java#L264
stepsize is changed from double to float, the test will pass although the
training accuracy is different.
stepsize only involves in the expression "stepsize*values[ei][ci]" during
estimating params. I notice that values[ei][ci] is actually float.
Still haven't figured out why.
> PerceptronPrepAttachTest fails only on Java 8
> ---------------------------------------------
>
> Key: OPENNLP-722
> URL: https://issues.apache.org/jira/browse/OPENNLP-722
> Project: OpenNLP
> Issue Type: Bug
> Components: Build, Packaging and Test
> Affects Versions: 1.6.0
> Reporter: Joern Kottmann
> Assignee: Joern Kottmann
> Priority: Minor
> Fix For: 1.6.0
>
>
> The test
> PerceptronPrepAttachTest.testPerceptronOnPrepAttachDataWithStepSizeDecrease
> fails if executed on Java 8.
> It would be really nice to track down the cause of that.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
