Code Quality in the knittingpattern Python Library

In our Google Summer of Code project a part of our work is to bring knitting to the digital age. We is Kirstin Heidler and Nicco Kunzmann. Our knittingpattern library aims at being the exchange and conversion format between different types of knit work representations: hand knitting instructions, machine commands for different machines and SVG schemata.

Cafe instructions
The generated schema from the knittingpattern library.
The original pattern schema Cafe.








The image above was generated by this Python code:

import knittingpattern, webbrowser
example = knittingpattern.load_from().example("Cafe.json")".svg"))

So far about the context. Now about the Quality tools we use:


Continuous integration

We use Travis CI [FOSSASIA] to upload packages of a specific git tag  automatically. The Travis build runs under Python 3.3 to 3.5. It first builds the package and then installs it with its dependencies. To upload tags automatically, one can configure Travis, preferably with the command line interface, to save username and password for the Python Package Index (Pypi).[TravisDocs] Our process of releasing a new version is the following:

  1. Increase the version in the knitting pattern library and create a new pull request for it.
  2. Merge the pull request after the tests passed.
  3. Pull and create a new release with a git tag using tag_and_deploy

Travis then builds the new tag and uploads it to Pypi.

With this we have a basic quality assurance. Pull-requests need to run all tests before they can be merge. Travis can be configured to automatically reject a request with errors.

Documentation Driven Development

As mentioned in a blog post, documentation-driven development was something worth to check out. In our case that means writing the documentation first, then the tests and then the code.

Writing the documentation first means thinking in the space of the mental model you have for the code. It defines the interfaces you would be happy to use. A lot of edge cases can be thought of at this point.

When writing the tests, they are often split up and do not represent the flow of thought any more that you had when thinking about your wishes. Tests can be seen as the glue between the code and the documentation. As it is with writing code to pass the tests, in the conversation between the tests and the documentation I find out some things I have forgotten.

When writing the code in a test-driven way, another conversation starts. I call implementing the tests conversation because the tests talk to the code that it should be different and the code tells the tests their inconsistencies like misspellings and bloated interfaces.

With writing documentation first, we have the chance to have two conversations about our code, in spoken language and in code. I like it when the code hears my wishes, so I prefer to talk a bit more.

Testing the Documentation

Our documentation is hosted on Read the Docs. It should have these properties:

  1. Every module is documented.
  2. Everything that is public is documented.
  3. The documentation is syntactically correct.

These are qualities that can be tested, so they are tested. The code can not be deployed if it does not meet these standards. We use Sphinx for building the docs. That makes it possible to tests these properties in this way:

  1. For every module there exists a .rst file which automatically documents the module with autodoc.
  2. A Sphinx build outputs a list of objects that should be covered by documentation but are not.
  3. Sphinx outputs warnings throughout the build.

testing out documentation allows us to have it in higher quality. Many more tests could be imagined, but the basic ones already help.

Code Coverage

It is possible to test your code coverage and see how well we do using It gives us the files we need to work on when we want to improve the quality of the package.


Landscape is also free for open source projects. It can give hints about where to improve next. Also it is possible to fail pull requests if the quality decreases. It shows code duplication and can run pylint. Currently, most of the style problems arise from undocumented tests.


When starting with the more strict quality assurance, the question arose if that would only slow us down. Now, we have learned to write properly styled pep8 code and begin to automatically do what pylint demands. High test-coverage allows us to change the underlying functionality without changing the interface and without fear we may break something irrecoverably. I feel like having a burden taken from me with all those free tools for open-source software that spare my time to set quality assurance up.

Future Work

In the future we like to also create a user interface. It is hard, sometimes, to test these. So, we plan not to put it into the package but build it on the package.

Knitting Pattern Conversion


We can convert knitting patterns to svg (middle) which proves the concept but is still a different from the original (right)

Our goal is to create a knit-work exchange format. This includes the conversion to a scematic view of the knittting pattern as svg – to make it endlessly scalable and allow conversions to png, pdf and paper.

This week we ended the prototype of the SVG conversion. The positions are a bit off and instructions are placed above eachother. Most of the work is done.

We are also able to load and save knitting patterns a png files.

(1)a34e6d2c-372d-11e6-9bbd-71c846ead7f9 (2)f6a6bf82-372e-11e6-9467-8bab0e07c099


We loaded them (1), converted them to a knitting pattern and then saved them again as png (2). This way we path our way towards using the ayab software and actually knitting the pattern. Also we can convert the knitting pattern to an svg consisting all of knit instructions (3). Here is the code for it in version 0.0.8.

>>> import knittingpattern
>>> from knittingpattern.convert.image_to_knittingpattern import *
>>> convert_image_to_knitting_pattern.path("head-band.png").temporary_path(".json")
>>> k = knittingpattern.load_from_path("head-band.json")
>>> k.to_svg(10).temporary_path(".svg")

Here you can see a proof of concept video:


A look into the knitting pattern format

We are currently working on a format that allows to exchange instructions for knitting independent of how it is going to be knit: by a machine like Brother, Pfaff, or by hand.

For this to be possible we need the format to be as general as possible and have no ties to a specific form of knitting. At some point the transition to a format specifically designed for a certain machine is necessary. However, we believe that it is possible to have a format so general that instructions for all types of machines and for hand knitting could be generated from it.

We have decided to use JSON as the language to describe the format, because it is machine readable and human readable at the same time.
The structure of our format follows the rows in knitting.

Our first thought was about creating a format that would describe how meshes are connected and how the thread travels. This would allow for great flexibility and it should be possible to represent everything like this. However, we have decided against it, because we think this format would be quite complicated (different orientations and twists of meshes possible, different threads for multiple colors…) and would get quite big very quickly, because of all the different properties for each mesh and because of all the meshes. Furthermore, and this point is probably more important, knitters do not think this way. If we had a format like that, it would not be easy to understand what was happening for human beings.

Whenever I knit by hand, I never think about how all the meshes are connected by this single thread I am using. I always think about which operations I am performing when knitting in each row. Instructions for creating patterns in knitting are also written this way. They give the knitter a set of knitting instructions to do and possibly repeat. We have concluded, that most knitters think in knitting operations performed rather than connections between meshes.


Knitting instructions from Garnstudio’s Café.

Therefore we have decided to base our format on knitting instructions/operations. The most common instructions probably are: knit, purl, cast on, bind off, knit two together, yarn over. Of course for increases and decreases there are many different operations which work in a similar way but have slight differences (e.g. skp, k2tog).
Since in knitting many things are possible and it is unlikely that we ever manage to create a complete list of all the possible operations you can perform in knitting we have decided to have a very open format, that allows the definition of new instructions.

Instructions are also defined in JSON format. Here is an example, the “k2tog” instruction:

        "type" : "k2tog",
        "title" : {
            "en-en" : "Knit 2 Together"
        "number of consumed meshes" : 2,
        "description" : {
            "wikipedia" : {
                "en-en" : ""
            "text" : {
                "en-en" : "Knit two stitches together, as if they were one stitch."


Knitting patterns consist of multiple rows, which consist of multiple instructions. Furthermore we want to define the connections between rows. This is important, so we can express gaps or slits which are multiple rows long. For example when knitting pants the two legs will be separate. They will be knit separately and their combined width will be increased in comparison to the width of the hip.
Here is an example for a pattern which specifies a cast on in the first row, then a row where all stitches are knit, then the last row is bound off.

  "type" : "knitting pattern",
  "version" : "0.1",
  "comment" : {
    "content" : "cast on and bind off",
    "type" : "markdown"
  "patterns" : [
      "id" : "knit",
      "name" : "cobo",
      "rows" : [
          "id" : 1,
          "instructions" : [
            {"id": "1.0", "type": "co"},
            {"id": "1.1", "type": "co"},
            {"id": "1.2", "type": "co"},
            {"id": "1.3", "type": "co"}
          "id" : 2,
          "instructions" : [
            {"id": "2.0"},
            {"id": "2.1"},
            {"id": "2.2"},
            {"id": "2.3"}
          "id" : 3,
          "instructions" : [
            {"id": "3.0", "type": "bo"},
            {"id": "3.1", "type": "bo"},
            {"id": "3.2", "type": "bo"},
            {"id": "3.3", "type": "bo"}
      "connections" : [
          "from" : {
            "id" : 1
          "to" : {
            "id" : 2
          "from" : {
            "id" : 2
          "to" : {
            "id" : 3

Connections are defined “from” one row “to” another.  The ids identify the rows. The optional attribute start defines the mesh where the connection starts. If start is not defined, the first mesh of the row is assumed. When indexing the list of meshes in a row the first index is 1. The optional attribute “meshes” describes how many meshes will be connected, starting from the mesh defined in “start”.

The resulting parsed Python object structure looks like this:

row model

The Python object structure for working with the parsed knitting pattern.

Each row has a list of instructions. Each instruction produces a number of meshes and consumes a number of meshes. These meshes are also the meshes that are consumed/produced by the rows.