sparkling-water-ml:test produces:
H2O logs
DistributedException from michals-mbp/192.168.1.65:54321, caused by java.lang.AssertionError: Logloss is only defined for values in 0...1, but got 1.5250916823676084
at water.MRTask.getResult(MRTask.java:478)
at water.MRTask.getResult(MRTask.java:486)
{noformat]
For now the test was disabled.
Interesting thing is that tests are passing in travis:
are we running this test as part of travis tests?
what is main difference between travis and jenkins tests? Spark version?
in this branch https://travis-ci.org/h2oai/sparkling-water/builds/197547170, where I removed ignore("") from the test, it seems to be failing randomly for some scala/jdk combinations. Tried it locally a lot of times and couldn't reproduce so it's either I'm unlucky or I need to try with a different scala/java version.
The weird part is, it's not failing on assertions or anywhere in the test itself but on:
Is there any way to get this file from travis: file:///home/travis/build/h2oai/sparkling-water/ml/build/reports/tests/test/index.html ? They come in handy sometimes.
The issue here is that the result of SVM are not probabilities but distances from the separating hyperplane. We need to map those values to propabilities. This issue was intermittent because the numbers on which we were creating the SVMModel were generating randomly