SKIP 2 — scikit-image mission statement¶
Juan Nunez-Iglesias <email@example.com>
Stéfan van der Walt <firstname.lastname@example.org>
scikit-image should adopt the document below as its mission statement. This statement will feature prominently in the scikit-image home page and readme, as well as the contributor and core developer guides. Decisions about the API and the future of the library would be referenced against this document. (See SKIP 1 — scikit-image governance and decision-making.)
In July 2018, I (Juan) published a blog post that broadly outlined what I would want from a roadmap for scikit-image 1, but requested comments from the community before it would be finalized. I consider that we have collected comments for long enough and can move forward with formal adoption. Most comments were positive, so below I’ll just summarize the “negative” comments under “rejected ideas”.
(Or: What problem does this proposal solve?)
Over the past few years, scikit-image has been slightly “adrift”, with new and old contributors coming in, adding what small bits they need at the time, and disappearing again for a while. This is fine and we want to encourage more of it, but it also lacks direction. Additionally, without a concerted roadmap to concentrate effort, many of these contributions just fall by the wayside, as it is difficult for new contributors to live up to our stringent (and largely unwritten) standards of code.
scikit-image aims to be the reference library for scientific image analysis in Python. We accomplish this by:
being easy to use and install. We are careful in taking on new dependencies, and sometimes cull existing ones, or make them optional. All functions in our API have thorough docstrings clarifying expected inputs and outputs.
providing a consistent API. Conceptually identical arguments have the same name and position in a function signature.
ensuring correctness. Test coverage is close to 100% and code is reviewed by at least two core developers before being included in the library.
caring for users’ data. We have a functional 2 API and don’t modify input arrays unless explicitly directed to do so.
promoting education in image processing, with extensive pedagogical documentation.
We are inclusive. We continue to welcome and mentor newcomers who are making their first contribution.
We are community-driven. Decisions about the API and features are driven by our users’ requirements, not by the whims of the core team. (See SKIP 1 — scikit-image governance and decision-making.)
We serve scientific applications primarily, over “consumer” image editing in the vein of Photoshop or GIMP. This often means prioritizing n-dimensional data support, and rejecting implementations of “flashy” filters that have little scientific value.
We value simple, readable implementations over getting every last ounce of performance. Readable code that is easy to understand, for newcomers and maintainers alike, makes it easier to contribute new code as well as prevent bugs. This means that we will prefer a 20% slowdown if it reduces lines of code two-fold, for example.
We value education and documentation. All functions should have NumPy-style docstrings 3, preferably with examples, as well as gallery examples that showcase how that function is used in a scientific application. Core developers take an active role in finishing documentation examples.
Much in the same way that the Zen of Python 5 and PEP8 guide style and implementation details in most Python code, this guide is meant to guide decisions about the future of scikit-image, be it in terms of code style, whether to accept new functionality, or whether to take on new dependencies, among other things.
To find out more about the history of this document, please read the following:
The use of NumPy arrays was the most supported of the statement’s components, together with the points about inclusivity, mentorship, and documentation. We had +1s from Mark Harfouche, Royi Avital, and Greg Lee, among others.
This SKIP formalizes what had been the unwritten culture of scikit-image, so it does not raise any backward compatibility concerns.
Two topics in the original discussion were ultimately rejected, detailed below:
In my original post, I suggested that scikit-image should have some form of metadata handling before 1.0. Among others, Mark Harfouche, Curtis Rueden, and Dan Allan all advised that (a) maybe scikit-image doesn’t need to handle metadata, and can instead focus on being a robust lower-level library that another like XArray can use to include metadata handling, and (b) anyway, metadata support can be added later without breaking the 1.0 API. I think these are very good points and furthermore metadata handling is super hard and I don’t mind keeping this off our plate for the moment.
See References below.
CC0 1.0 Universal (CC0 1.0) Public Domain Dedication, https://creativecommons.org/publicdomain/zero/1.0/