[NLPL Task Force (A)] Update to TensorFlow on Abel

[Aaron Smith]

Thanks Stephan.


We're perhaps not tech-savvy enough to understand exactly how these singularity containers work, but I think the general opinion here is that it's clunky to have to call separate versions of Python to run DyNet and TensorFlow. It also creates problems for example when we want to load other modules at the same time. We haven't yet tried having DyNet and TensorFlow loaded at the same time, but that's something that we would ideally need for the complete parsing pipeline, and I suspect at least that might be complicated in the current setup.


Aaron

________________________________
From: Stephan Oepen <oe at ifi.uio.no>
Sent: 21 November 2017 11:05:49
To: Aaron Smith
Cc: infrastructure
Subject: Re: [NLPL Task Force (A)] Update to TensorFlow on Abel

hi aaron,

yes, certainly, we will request that from the Abel maintainers (a
group called USIT) and copy you on the communication.

from last i recall, they had resorted to installing DyNet and
TensorFlow inside singularity containers, which likely has both
advantages and downsides for NLPL usage of these modules.  do you have
any feedback on that choice already?

cheers, oe


On Tue, Nov 21, 2017 at 11:02 AM, Aaron Smith <aaron.smith at lingfil.uu.se> wrote:
> Hi,
>
>
> We are having issues in Uppsala getting some of our software to run on Abel
> due to the version of TensorFlow available there (currently 1.0.1). I
> understand that it's a bit complicated to install new versions, but it's
> somewhat critical to enable us to fully install our parsing software suite.
> Would it be possible to mandate an update to TensorFlow 1.3?
>
>
> Best,
>
>
> Aaron
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.nlpl.eu/archives/infrastructure/attachments/20171121/b7e2419f/attachment.htm>