[
https://issues.apache.org/jira/browse/MADLIB-1037?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15667937#comment-15667937
]
Xiaocheng Tang commented on MADLIB-1037:
----------------------------------------
The first PR is created here: https://github.com/apache/incubator-madlib/pull/75
This PR implements multi-class support vector machine and a new training
mechanism using buffered mini-batch stochastic gradient descent (hence
"in-memory" computations across epochs). Preliminary experiments demonstrate
potential on greatly improving training speed and accuracy in MADlib. For
example, training on mnist datasets (10 classes) for 1s on HAWQ 2.0 half-rack
DCA and the model is able to achieve over 90% accuracy on holdout data. The
training and testing set in this case consist of 46900 and 23100 images,
respectively.
> Multi-class SVM with mini-batching
> -----------------------------------
>
> Key: MADLIB-1037
> URL: https://issues.apache.org/jira/browse/MADLIB-1037
> Project: Apache MADlib
> Issue Type: New Feature
> Components: Module: Support Vector Machines
> Reporter: Frank McQuillan
> Assignee: Xiaocheng Tang
> Fix For: v1.10
>
>
> Prototype of structure svm trained with mini-batch
> Acceptance
> 1) train on mnist and achieve acceptable accuracy
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
