[NLPL Task Force (A)] [Parsing] [rt.uio.no #2747687] newer versions of TensorFlow and DyNet

[aaron.smith@lingfil.uu.se via RT]

Hi! The earlier communication arose as I had a small problem getting the currently installed version of TensorFlow to run, and I wondered at the same time about a possible upgrade, at which point Thomas mentioned how difficult it is and suggested I go through official NLPL channels if necessary.


We were hoping to be able to get our segmenter software running with the older version of TensorFlow, but after some experiments yesterday that now looks more difficult. Certain functions have been renamed and redefined between versions so we decided that it would in fact be a good time for us to request an upgrade. We are aware that version 1.4 is available, the reason we requested 1.3 is that the software is built and tested with that version.


We are able to run our parser with the currently installed version of DyNet, so a first version of the parsing laboratory, without the segmenter, could be released in the meantime if the update to TensorFlow is a longer term project.


Best,


Aaron

________________________________
From: parsing-bounces at emmtee.net <parsing-bounces at emmtee.net> on behalf of Joakim Nivre <joakim.nivre at lingfil.uu.se>
Sent: 22 November 2017 09:09:31
To: Thomas Röblitz via RT
Cc: parsing at nlpl.eu
Subject: Re: [Parsing] [rt.uio.no #2747687] newer versions of TensorFlow and DyNet

I think this is a good illustration of the challenges involved in building an HPC NLP lab. Many of the software packages that we need for our research are under development themselves and therefore may not conform to standard conventions for installation, etc. I completely realize that this may cause a lot of headache, but the more we can overcome these challenges the more useful the resulting laboratory will be for the community.

Joakim

> On 22 Nov 2017, at 08:42, Thomas Röblitz via RT <hpc-drift at usit.uio.no> wrote:
>
> I just told him that TensorFlow is terrible to install, and I'll extend this to DyNet. I gave it a quick shot yesterday, but it just doesn't work.
>
> These packages seem to be targeted for Ubuntu or Windows. They also provide source code, but, at least for DyNet, their installation instructions lack even basic information such as required libraries, compilers, etc. All this needs to be figured out, tried and this costs time.
>
> BTW, it seems TensorFlow 1.4 is out.
>
> Thomas
>
> On 2017-11-21 23:00:25, oe wrote:
>>> TensorFlow 1.3, but as explained to Aaron, it is terrible to install :)
>>
>> could you forward that earlier thread to ‘infrastructure at nlpl.eu’?
>> this is a task force comprising bjørn, joerg, martin, and myself.  we
>> would like to keep ourselves up-to-date on all technical communication
>> with the infrastructure providers regarding NLPL.
>>
>> i will let aaron answer regarding the benefits of moving to the
>> current version of TensorFlow.
>>
>> with thanks in advance, oe
>>
>