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

[Thomas Röblitz via RT]

Hei Stephan,

DyNet 2.0 seems to be installed already.

module load dynet/2.0
dynet_python
>>> import dynet
>>> dynet.__version__
2.0

TensorFlow 1.3, but as explained to Aaron, it is terrible to install :) Of course, can be done, but it will take time, and given previous experience it is a rocky road to get it working. What's the benefit compared to the installed version?

Both DyNet and TensorFlow are installed in containers. If you need them together, the best option is probably to create a new container combining both.

I have no idea how expensive the bullet is.

Cheers

Thomas

On 2017-11-21 14:34:49, oe wrote:
> dear colleagues,
> 
> our NeIC-funded NLPL project is preparing to put various project-
> specific software stacks on Abel (and, where possible, also the
> Finnish Taito system), and we are back to dependencies on the
> TensorFlow and DyNet machine learning frameworks.  both seem in
> principle general enough to be maintained globally, i.e. by USIT
> rather than NLPL.  could we kindly ask for the global TensorFlow and
> DyNet installations on Abel to be updated to current versions, i.e.
> 1.3 and 2.0, respectively?  as before, the hours required for these
> installations can be billed to the NLPL service agreement.
> 
> a related question regarding containerization: i believe trond ended
> up installing at least one of these two libraries as a singularity
> container, which made it simpler to configure an installation
> environment as required by the libraries.  however, in NLPL we remain
> a bit unsure about what containerization does to interoperability?
> specifically, we need to be able to use both TensorFlow and DyNet
> together, i.e. from one Python process.  would that be possible if
> either library resides in its own container?
> 
> in sum, i wonder whether we should not try to bite the bullet and
> rather try ‘native’ (non-container) installations of these basic
> libraries?
> 
> with thanks in advance, oe