Nandish Jayaram commented on MADLIB-1200:

A suggestion regarding the buffer_size:

We could consider making it a factor of the page size rather than the 1GB 
limit. By page size I mean the default page size that the underlying database 
uses while fetching data from the disk (4 MB?). This might result in packing 
fewer rows together, but that may be a good thing from user's perspective. If 
we pack around 10000 rows into one, running a select query on that table, even 
with limit 1 takes a long time to load. With 1GB based factoring, we might be 
able to pack hundreds of thousands of rows, which makes it harder for the user 
to view it I guess.

> Pre-processing helper function for mini-batching 
> -------------------------------------------------
>                 Key: MADLIB-1200
>                 URL: https://issues.apache.org/jira/browse/MADLIB-1200
>             Project: Apache MADlib
>          Issue Type: New Feature
>          Components: Module: Utilities
>            Reporter: Frank McQuillan
>            Priority: Major
>             Fix For: v1.14
> Related to
>  https://issues.apache.org/jira/browse/MADLIB-1037
>  https://issues.apache.org/jira/browse/MADLIB-1048
> Story
> {{As a}}
>  data scientist
>  {{I want to}}
>  pre-process input files for use with mini-batching
>  {{so that}}
>  the optimization part of MLP, SVM, etc. runs faster when I do multiple runs, 
> perhaps because I am tuning parameters (i.e., pre-processing is an occasional 
> operation that I don't want to re-do every time that I train a model)
> Interface
> {code:java}
> minibatch_preprocessor (
> source_table,                     -- Name of the table containing the input 
> data.
> output_table,                      -- Name of the table suitable for 
> mini-batching.
> dependent_varname,         -- Name of the dependent variable column. 
> independent_varname,      -- Expression list to evaluate for the independent 
> variables.
> buffer_size,                        --  buffer_size? Default should be to 
> pack as much as possible in the 1GB limit imposed by postgres/gpdb.
> )
> {code}
> The main purpose of the function is to prepare the training data for 
> minibatching algorithms. This will be achieved in 2 stages
>  # Based on the batch size, group all the dependent and independent variables 
> in a single tuple representative of the batch.
>  # If the independent variables are boolean or text, perform one hot 
> encoding.  N/A for integer and floats. Note that if the integer vars are 
> actually categorical, they must be case to ::TEXT so that they get encoded.  
> Notes
> 1) Random shuffle needed for mini-batch.
>  2) Naive approach may be OK to start, not worth big investment to make run 
> 10% or 20% faster.
> Acceptance
> Summary 
>   1) Convert from standard to special format for mini-batching
>   2) Standardize by default for now but the user cannot opt out of it. We may 
> decide to add a flag later.
>   3) Some scale testing OK (does not need to be comprehensive)
>   4) Document as a helper function user docs
>   5) Always ignore nulls in dependent variable
>   6) IC

This message was sent by Atlassian JIRA

Reply via email to