[NLPL Task Force (A)] Keras and DyNet

[Stephan Oepen]

hi ali,

i am copying the NLPL infrastructure task force, who we should try to
keep up-to-date on all technical matters.

> 1- We need Keras (based on Tensorflow with GPU support for Python-3.x) for our experiments. I could not find the module on the cluster. Is it possible to install it as one of the NLPL module or in some other way?
>
> 2- The new version of UUParser is based on Python-3.x and as you (may) already know the parser uses DyNet internally. It seems that the current installation of the DyNet is based on Python-2.7. Is it possible to have DyNet based on Python-3.x.

it almost sounds as if you are only using the relatively old
installations of DyNet and TensorFlow that were installed system-wide
(by USIT, the administrators) early on in the NLPL project, and which
somewhat unfortunately are packaged each in their own container.

so, just to make sure you are actually aware of the more recent
NLPL-maintained collection of modules:

http://wiki.nlpl.eu/index.php/Infrastructure/software/catalogue
http://wiki.nlpl.eu/index.php/Infrastructure/software/python

to get an up-to-date inventory of available NLPL modules, please try:

$ module use -a /projects/nlpl/software/modulefiles
$ module avail 2>&1 | grep nlpl-

keras on tensorflow with gpu support should work in the following modules:

$ module purge; module load nlpl-tensorflow/1.13/3.7
$ python3 -m pip freeze

unfortunately, the DyNet situation is more mixed: there is a current
version available, but we have never succeeded in enabling gpu support
for it.  truth be told, we have stopped trying recently, because gpu
capacity on Abel has been so severely limited anyway.  so, the
following should give you DyNet in Python 3.7, but (last i tried at
least), this will not work on gpus:

$ module purge; module load nlpl-dynet/2.1/3.7
$ python3 -c "import dynet as dy; print(dy.__version__);"
[dynet] random seed: 1294263076
[dynet] allocating memory: 512MB
[dynet] memory allocation done.
2.0.3

please do not hesitate to follow up with more questions!  but in
general, we are preparing to migrate to the Abel successor system
(called Saga, running a slightly more modern operating system) and
re-build the NLPL software stack there ...

best wishes, oe