[
https://issues.apache.org/jira/browse/SINGA-298?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15897993#comment-15897993
]
Sai Narsi Reddy Donthi Reddy commented on SINGA-298:
----------------------------------------------------
Hey [~hiwayne],
Sorry for not replying soon enough. I had a lot of work and did not get a
chance to test the program till yesterday.
And YES it runs with out any problem.
However, when I train a alexnet without batchnorm, within first batch the loss
if going to *nan*.
So I tested with +squared errors loss+ and +no softmax activation+.
I started to get loss ~= 0.35 and acc = 14%, how much ever train the model.
> Check failed: input.device()->lang() == kOpencl
> ------------------------------------------------
>
> Key: SINGA-298
> URL: https://issues.apache.org/jira/browse/SINGA-298
> Project: Singa
> Issue Type: Bug
> Components: PySINGA
> Environment: Ubuntu 16.04, AMD rx480, intel i7k 6-gen
> Reporter: Sai Narsi Reddy Donthi Reddy
> Labels: newbie
>
> Hello I came across the framework when I was searching for a deep learning
> framework with opencl support.
> I cloned the GITHUB master branch version and downloaded as shown on the
> install page.
> After sucessful install, I tried to run cifar-10 example with opencl instead
> of CUDA, I am receiving following errors:
> WARNING: Logging before InitGoogleLogging() is written to STDERR
> F0127 17:07:02.757550 27607 opencl_convolution.cc:30] Check failed:
> input.device()->lang() == kOpencl (0 vs. 2)
> *** Check failure stack trace: ***
> Aborted (core dumped)
> I am new to this and I am learning as I go.
> Sorry if I posted here mistakenly.
> Thanks,
> Narsi
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
