CSE 40657/60657
Homeworks

Programming environment

The “official” programming environment for homework assignments will be:

• Python 3.5
• NumPy 1.10
• Autograd 1.1

Examples, skeleton code, and solutions will use the above; however, if you prefer to use something else, you are free to.

On student{00-03}.cse.nd.edu, the Python installed at /afs/nd.edu/coursefa.16/cse/cse40657.01/bin/python has all of the above installed.

Python

You probably have Python installed already; if you haven't switched to Python 3 yet, now's the time!

1. If you need to install Python, the simplest way to get it is probably Miniconda.
2. If you want to keep packages for this course separated from your system-wide packages, create a virtualenv:

   virtualenv nlp
   source nlp/bin/activate      # sh/ksh/bash/zsh
   #source nlp/bin/activate.csh # csh/tcsh
   

3. If you need a refresher on Python, see the tutorial. Or if you're used to Python 2, here is a list of differences.

Other learning toolkits

We're using Autograd because it's simple and clean. However, it's not a deep learning toolkit per se, and certainly not industrial-strength. If you'd like to try an alternative, here are few pointers:

• Some toolkits have you define a graph of basic operations (e.g., Theano, TensorFlow) and some have you define a tree of neural network layers (e.g., Torch). The latter are not general-purpose enough for this course.
• In some graph-based toolkits (Theano and TensorFlow), creating a graph is rather expensive, so you create a graph once and reuse it over and over. But in NLP, the graphs are, as a rule, different sizes for each sentence. So you have two choices: make one graph for each size (which some people find inelegant), or code a loop into the graph (which some people find difficult). You can use these toolkits, but it will be more work.
• That leaves the following toolkits that I'm aware of:
  • CNN/PyCNN: from CMU
  • Chainer: from Preferred Networks, a Japan-based data analytics company
  • Penne: Go Irish!