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] [Updated] (SYSTEMML-1681) Add a unit test for 2D convolution comparing against another system
Date Wed, 05 Jul 2017 17:24:00 GMT

     [ https://issues.apache.org/jira/browse/SYSTEMML-1681?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Mike Dusenberry updated SYSTEMML-1681:
--------------------------------------
    Description: 
We should add a comparison against TensorFlow in {{nn/test/test.dml}}, like we have for other
layers, and check that the same values are produced on CPU & GPU for both TensorFlow and
SystemML.  There is already a {{conv2d}} function in {{nn/test/test.dml}} that compares the
output of three different 2D convolution layers that we have (built-in, DML im2col, DML basic).
 We should add a new section to this {{conv2d}} function (similar to how the {{max_pool2d}}
test function has multiple test sections) to compare the built-in conv2d layer against the
output from TensorFlow.  We should use the same {{X}}, {{W}}, and {{b}} initialization values
as used in {{conv2d_transpose}}.


> Add a unit test for 2D convolution comparing against another system
> -------------------------------------------------------------------
>
>                 Key: SYSTEMML-1681
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-1681
>             Project: SystemML
>          Issue Type: New Feature
>            Reporter: Mike Dusenberry
>            Assignee: Mike Dusenberry
>
> We should add a comparison against TensorFlow in {{nn/test/test.dml}}, like we have for
other layers, and check that the same values are produced on CPU & GPU for both TensorFlow
and SystemML.  There is already a {{conv2d}} function in {{nn/test/test.dml}} that compares
the output of three different 2D convolution layers that we have (built-in, DML im2col, DML
basic).  We should add a new section to this {{conv2d}} function (similar to how the {{max_pool2d}}
test function has multiple test sections) to compare the built-in conv2d layer against the
output from TensorFlow.  We should use the same {{X}}, {{W}}, and {{b}} initialization values
as used in {{conv2d_transpose}}.



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

Mime
View raw message