[jira] [Created] (MADLIB-1400) Modify warm start logic for DL to handle case of missing weight

["Frank McQuillan (Jira)" <ji...@apache.org>]

Frank McQuillan created MADLIB-1400:
---------------------------------------

             Summary: Modify warm start logic for DL to handle case of missing weight
                 Key: MADLIB-1400
                 URL: https://issues.apache.org/jira/browse/MADLIB-1400
             Project: Apache MADlib
          Issue Type: Improvement
          Components: Deep Learning
            Reporter: Frank McQuillan
             Fix For: v1.17
         Attachments: 20191219_163748.jpg

I was trying to implement an autoML algorithm on top of the new multi-model fit and ran into an issue with warm start.  I would suggest a slight change in logic:

Currently if there are not existing models+weights in the model table for every single MST key in the MST table, we error out when warm start = TRUE.

I suggest if there is not an entry in the model table+weights for an MST key in the MST table, then randomly initialize the weights (or whatever the default is in Keras) and do not error out.  Of course, if there are existing models+weights in the model table for an MST key in the MST table, then use them (which is currently what we do). 

Again this all applies to warm start = TRUE only.

Oh, if warm start = TRUE but the model table does not exist at all, we should error out like we do today.

Logic for warm start = FALSE as implemented seems OK to me.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)