[
https://issues.apache.org/jira/browse/SYSTEMML-2080?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16338772#comment-16338772
]
Matthias Boehm commented on SYSTEMML-2080:
------------------------------------------
with a correctness fix for sparse relu-backward CP operations, we get the
expected output again:
{code}
Iter:100, training loss:0.1781972225785432, training accuracy:93.75
Iter:200, training loss:0.19475235762456966, training accuracy:93.75
Iter:300, training loss:0.04969600094002809, training accuracy:98.4375
Iter:400, training loss:0.059094595267147325, training accuracy:98.4375
Iter:500, training loss:0.09571390248083674, training accuracy:96.875
Iter:500, validation loss:328.0176739861353, validation
accuracy:96.21829710144928
Iter:600, training loss:0.017400919930351314, training accuracy:100.0
Iter:700, training loss:0.013629615706048195, training accuracy:100.0
{code}
> Lenet over mnist shows accuracy issues on CPU backend
> -----------------------------------------------------
>
> Key: SYSTEMML-2080
> URL: https://issues.apache.org/jira/browse/SYSTEMML-2080
> Project: SystemML
> Issue Type: Bug
> Reporter: Matthias Boehm
> Priority: Major
>
> {code}
> ter:100, training loss:2.301559079510293, training accuracy:12.5
> Iter:200, training loss:2.3023935631583994, training accuracy:21.875
> Iter:300, training loss:2.3033211472430044, training accuracy:6.25
> Iter:400, training loss:2.3015396319042303, training accuracy:10.9375
> Iter:500, training loss:2.3030694920810917, training accuracy:15.625
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
