Community/training/2019

Background

A desirable side-effect of the NLPL cooperation is community formation, i.e. strengthening interaction and collaboration among Nordic research teams in NLP and advancing a shared level of knowledge and experience in using national e-Infrastructures for large-scale NLP research. Towards these goals, the project will organize some training and outreach events.

For early 2018, NLPL will hold a winter school in conjunction with the NeIC All-Hands Meeting (AHM), which some NLPL team members will attend. The NeIC meeting will be held from January 29 to February 1, 2018, in the Norwegian mountain resort at Skeikampen.

Starting already after lunch on Monday, January 29, NLPL will kick off its winter school in E-Infrastructure and Scientific Computing for Nordic Natural Language Processing Research. The winter school will have a duration of two days, i.e. end with lunch on Wednesday, January 31. We anticipate participation of around 20–25 people, where the majority will come from the NLPL partner sites.

NLPL team members, associates at partner sites (e.g. doctoral and post-doctoral fellows, possibly also some MSc students), and other prospective users of the infrastructure will be invited to attend; NLPL partners(=project team members) can have their cost of participation covered by the project.

Software Environment

The deep learning sessions will include hands-on laboratory sessions following the LxMLS tutorial, although switching over to PyTorch (instead Theano). Stephan Oepen has prepared a Python virtual environment will all LxLMS dependencies installed, including PyTorch with GPU support:

module use -a /projects/nlpl/software/modulefiles/
module add nlpl-lxmls

There is a one-time preparaptory step to obtain the ‘student’ starting package from LXMLS:

git clone https://github.com/LxMLS/lxmls-toolkit.git
cd lxmls-toolkit
git checkout student

Just now at least, there is a minor incompatibility with the standard version of SciPy installed on Abel and a newer version that the LXMLS toolkit assumes. To accomodate for this mismatch, all occurence of scipy.special must be changed to scipy.misc, including in the LXMLS library file lxmls/deep_learning/numpy_models/mlp.py.

Finally, the LXMLS code assumes that its top-level directory is visible for Python import commands; when inside the lxmls-toolkit directory, the following shell command should have the right effect:

export PYTHONPATH=$PYTHONPATH:.