[jira] [Commented] (MADLIB-1172) Logistic regression produces empty output table but no error message on Greenplum

[Himanshu Pandey (JIRA)]

    [ 
https://issues.apache.org/jira/browse/MADLIB-1172?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16478113#comment-16478113
 ] 

Himanshu Pandey commented on MADLIB-1172:
-----------------------------------------

Hi [~fmcquillan], 

This looks like a data issue. Following warning is captured in segment logs : 

{code}
2018-05-16 18:51:58.499974 
UTC,"gpadmin","gpadmin",p2130,th2056992640,"127.0.0.1","55898",2018-05-16 
18:51:58 
UTC,5644,con141,cmd40,seg1,slice3,dx2066,x5644,sx1,"WARNING","01000","Over- or 
underflow in Newton step, while updating coefficients.Input data is likely of 
poor numerical condition.",,,,,,"

{code}


The above warning is inside the logregr_irls_step_final function of 
logistic.cpp file: 

{code}
 if (!state.X_transp_AX.is_finite() || !state.X_transp_Az.is_finite()){
        //throw NoSolutionFoundException(
        //    "Over- or underflow in intermediate calulation. Input data is "
        //    "likely of poor numerical condition.");
        warning("Over- or underflow in intermediate calulation. Input data is "
              "likely of poor numerical condition.");
        state.status = TERMINATED;
        return state;
    }
{code}



> Logistic regression produces empty output table but no error message on 
> Greenplum
> ---------------------------------------------------------------------------------
>
>                 Key: MADLIB-1172
>                 URL: https://issues.apache.org/jira/browse/MADLIB-1172
>             Project: Apache MADlib
>          Issue Type: Bug
>          Components: Module: Logistic Regression
>            Reporter: Frank McQuillan
>            Assignee: Himanshu Pandey
>            Priority: Minor
>             Fix For: v1.15
>
>         Attachments: Logistic-regression-empty-output.ipynb, 
> load-data-sep.sql, load-data-singular.sql, load-data.sql
>
>
> Separated and singular data sets may produce and empty model table on 
> Greenplum 4.3.x.  On Postgres 9.6 the same example works OK. 
> See the attache jupyter notebook and data sets for details.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)