systemml-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Mike Dusenberry (JIRA)" <j...@apache.org>
Subject [jira] [Assigned] (SYSTEMML-1819) Create Keras2DML: Keras frontend to SystemML
Date Mon, 31 Jul 2017 21:25:01 GMT

     [ https://issues.apache.org/jira/browse/SYSTEMML-1819?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Mike Dusenberry reassigned SYSTEMML-1819:
-----------------------------------------

    Assignee: Mike Dusenberry  (was: Anooj Patel)

> Create Keras2DML: Keras frontend to SystemML
> --------------------------------------------
>
>                 Key: SYSTEMML-1819
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-1819
>             Project: SystemML
>          Issue Type: New Feature
>            Reporter: Mike Dusenberry
>            Assignee: Mike Dusenberry
>
> This task covers the creation of a "Keras2DML" frontend for SystemML, built upon the
[Caffe2DML | http://apache.github.io/systemml/beginners-guide-caffe2dml] infrastructure, that
will allow users to define (and even train) models in Keras and then import them into SystemML
for distributed training and prediction.  As an initial set of thoughts, the input could be
either (1) a Keras {{Model}} object, or (2) a saved Keras model hdf5 file, and the output
of training could be either (1) a Keras {{Model}} object, (2) a saved Keras model hdf5 file,
or (3) a SystemML model.
> This would be a step towards a full-blown, official backend for Keras.  The main goal
here would be to allow users to be able to transparently make use of distributed training,
without having to learn the details of SystemML.
> Initial steps:
> 1. Learn Keras
> 2. Learn Caffe2DML: [http://apache.github.io/systemml/beginners-guide-caffe2dml]  Basically,
Caffe2DML lets users import Caffe models (architecture and trained weights if available) into
SystemML and train/predict on Spark with a scikit-learn compatible API without the user having
to learn SystemML.  The main benefit is distributed training without the user needing to think
about it much.  A bunch of the infrastructure is in place that I think Keras2DML would use.
> 3. Import a simple Keras model definition with a single Dense layer, and focus on hooking
up the new Keras2DML class to the existing infrastructure.
> 4. Add reading of trained weights from Keras for the simple model, and hook up to existing
infrastructure.
> 5. Expand out to increasingly complex models, aiming to be able to import all of the
pretrained models from Keras, starting with VGG16 & ResNet50. [https://keras.io/applications/]



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message