[
https://issues.apache.org/jira/browse/MADLIB-1290?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16769573#comment-16769573
]
Domino Valdano commented on MADLIB-1290:
----------------------------------------
I looked into this, and it looks like this _might_ have been intentional
behavior; although I agree that it should probably be changed.
Current behavior: If the parameter is omitted, it divides all independent var
data by 255.0 If it is specified, then you can choose a different value from
255 to divide by. If you don't want it to divide by anything, you have to pass
1.0. NULL is not a valid value to pass.
Upon thinking this through, I'm not sure I like having 255 as the default
value. Intuitively, it seems like if you pass NULL or leave it off, it
shouldn't be dividing by anything, it should just leave the data alone. So I
think I would prefer we just changed the default to 1.0 and accepted NULL as
another way of passing 1.0.
If we keep 255 as the default, but we want to accept NULL, then NULL will have
to mean "divide by 255".
> Minibatch pre-processor for deep learning
> -----------------------------------------
>
> Key: MADLIB-1290
> URL: https://issues.apache.org/jira/browse/MADLIB-1290
> Project: Apache MADlib
> Issue Type: New Feature
> Components: Deep Learning
> Reporter: Frank McQuillan
> Assignee: Nandish Jayaram
> Priority: Major
> Fix For: v1.16
>
>
> The minibatch preprocessor we currently have in MADlib is bloated for DL
> tasks. This feature adds a simplified version of creating buffers, and
> divides each element of the independent array by a normalizing constant
> for standardization (which is 255.0 by default). This is standard practice
> with image data.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
