Contributing to Kur

Development Setup

  1. Check out the code:

    git clone
    cd kur
  2. (Optional, but recommended) Set up a virtualenv.

    There are lots of ways to do this. The easiest way if you don’t know how is to use virtualenv. You do this once:

    virtualenv -p /usr/bin/python3.5 venv

    This will create a new folder in the repo root called venv (it is in the .gitignore, so don’t worry about it polluting anything).

    Now every time you are ready to work on Kur, activate the environment:

    source venv/bin/activate

    This puts you in an isolated Python environment, with its own packages. If you install packages while the virtual environment is activatd, they will only be installed within the virtual environment, and the system packages will be left untouched.

    To leave the virtual environment, deactivate it:

  3. Install an editable version of Kur.

    pip install -e .

    This will install Kur (within the virtual environment only, if one is active), but any changes you make to the Kur source code will be immediately “seen” by programs that use Kur (rather than having to remove/reinstall).


    This is very similar to the functionality provided by python develop, but the unit testing framework that Kur uses (pytest) is slightly more annoying to run, as it won’t “see” the main Kur package installed. If you really insist on using python develop instead, then instead of running py.test, you need to run PYTHONPATH=.:$PYTHONPATH py.test or python -m pytest tests/ instead.

  4. Install the unit-testing packages and pylint.

    pip install tox pytest pytest-xdist pylint

Running the Unit Tests

Kur uses pytest as its unit-testing framework, and tox for running the unit tests in a number of different, isolated environments (i.e., against different versions of Python, each in their own virutal environment).

Running the Unit Tests with tox

To run the entire unit-testing suite for all versions of Python, you can simply do this:



Kur does not need to be installed to run the unit tests through tox. This means that if you installed Kur in a virtual environment, you do not need to activate the virtual environment before running the unit tests (although there is no harm in running tox from within the virtual environment, too).

To run the unit-test suite through tox for a particular Python version (for example, Python 3.5):

tox -e py35

You can enumerate all defined tox environments using tox -l.

Running the Unit Tests with pytest

tox already uses pytest behind the scenes to run the unit tests. But if you want to run the tests manually, you can do so. They will only test against the current Python environment.

python -m pytest --boxed tests/


Unlike running the unit tests through tox, if you want to call pytest directly like this, you will need Kur installed (or your virtual environment activated).


Like we mentioned earlier, pytest is a little naïve about its Python path. If you installed Kur into a virtual environment, you’ll need to tell pytest where it is (even if the environment is already activated). If your virtual environment is called venv in the repository root, you can do (be sure to change your Python version as appropriate):

PYTHONPATH=venv/lib/python3.5/site-packages:$PYTHONPATH pytest


What’s up with the --boxed option? It’s an option for the pytest-xdist plugin which runs each test in its own subprocess. This is important when testing backends like Keras, which do not seem to allow easy swapping of Theano/TensorFlow backends on-the-fly. Thus, when a test does import keras, the Keras backend will get “stuck” for that process.

Style Guide

We loosely adhere to the PEP 8 style guide. The most notable exception is that our code is indented with tabs instead of spaces. Why? Although Python suggests using spaces for indentation, spaces can be awkward to use: they do not convey semantic information and they make it difficult for people to adjust the indentation appearance to fit their preferences (on the other hand, editors can usually be customized to display the tab character as any number of spaces). Maybe this will change someday with enough public outcry. For now, tabs rule.

We have a Pylint configuration file so that you, too, can use the linter to check code quality. To do this, make sure pylint is installed (if it is in a virtual environment, make sure the environment is activated) and then:

pylint kur

Please make sure all linting issues are addressed before submitting a pull request.

We do not lint our tests directory, because they break lots of rules due to the magic of pytest (e.g., through fixtures and files).

Bug Reporting

Bugs should be reported as issues on GitHub. Please provide this information to help us get things fixed!

  • If you encountered a bug using the Python API:

    • Please actually think about the problem yourself a little, and tell us what you’ve tried to do to avoid the problem.
    • Please describe what you expected the code to do.
    • Please provide a minimal working example (the smallest program that reproduces your error) in Python.
    • Please include debug-level output: kur -vv ...
  • If you encountered a bug using the specification file and command-line API:

    • Please provide a minimum working example in YAML.
    • Please provide the command-line invocation(s) used.
    • Please tell us what you expected to happen.
    • Please include debug-level output: kur -vv ...

In both cases, if you bug needs a data source to reproduce, you should:

  • Check if the example data suppliers can be used to recreate your problem. This is definitely the most convenient way to check your problem, since we don’t need to download and understand your data.
  • If the examples don’t cut it, see if you can include an example Numpy array that produces the problem, either hard-coded or via some little Python snippet that creates the array (e.g., with numpy.random).
  • As a very last resort, you can try submitting small datasets (with as few elements in them as possible to reproduce the problem). But doing this will very likely deter us from addressing your issue, because it is more frustrating having to deal with dataset problems than actual Kur problems.