Back to Scikit Image

SKIP 4 — Transitioning to scikit-image 2.0

doc/source/skips/4-transition-to-v2.rst

0.26.019.5 KB
Original Source

.. _skip_4_transition_v2:

========================================== SKIP 4 — Transitioning to scikit-image 2.0

:Author: Juan Nunez-Iglesias [email protected] :Author: Lars Grüter :Author: Stéfan van der Walt :Author: Matthew Brett :Author: Marianne Corvellec :Status: Draft :Type: Standards Track :Created: 2025-08-16 :Resolved: <null> :Resolution: <null> :Version effective: None

Abstract

scikit-image is working towards a 2.0 release. This :ref:is an opportunity <skip_3_transition_v1> to clean up the API, and make long-outstanding backward incompatible changes. Some of these changes involve changing return values without changing function signatures, which cannot be done without adding an otherwise useless keyword argument (such as new_return_style=True) whose default value would need to change over several releases. That procedure would still result in backward incompatible changes, but made over a longer time period.

Despite being in beta and in a 0.x series of releases, scikit-image is used extremely broadly, and any backwards incompatible changes are likely to be disruptive. Given the rejection of :ref:SKIP-3 <skip_3_transition_v1>, this document proposes an alternative pathway towards cleaning and modernizing our API. The new pathway involves the following steps:

  • Introduce a new namespace skimage2 that will be included in the package alongside the existing skimage namespace during a transition period.
  • The new API will be implemented in skimage2 and will initially be marked as unstable and experimental. The old API in skimage will continue working.
  • When the new API in skimage2 is complete, the old namespace skimage will be successively deprecated and eventually removed.

See the :ref:skip4_implementation section for a more detailed description of the changes and steps involved.

Motivation and Scope

.. note:: This is largely duplicated from :ref:SKIP-3 <skip_3_transition_v1>.

scikit-image has grown organically over the past 12+ years, with functionality being added by a broad community of contributors from different backgrounds. This has resulted in various parts of the API being inconsistent: for example, skimage.transform.warp inverts the order of coordinates, so that a translation of (45, 32) actually moves the values in a NumPy array by 32 along the 0th axis, and 45 along the 1st, but only in 2D. In 3D, a translation of (45, 32, 77) moves the values in each axis by the number in the corresponding position.

Additionally, as our user base has grown, it has become apparent that certain early API choices turned out to be more confusing than helpful. For example, scikit-image will automatically convert images to various data types, rescaling them in the process. A uint8 image in the range [0, 255] will automatically be converted to a float64 image in [0, 1]. This might initially seem reasonable, but, for consistency, uint16 images in [0, 65535] are rescaled to [0, 1] floats, and uint16 images with 12-bit range in [0, 4095], which are common in microscopy, are rescaled to [0, 0.0625]. These silent conversions have resulted in much user confusion.

Changing this convention would require adding a preserve_range= keyword argument to almost all scikit-image functions, whose default value would change from False to True over 4 versions. Eventually, the change would be backwards-incompatible, no matter how gentle we made the deprecation curve.

Other major functions, such as skimage.measure.regionprops, could use an API tweak, for example by returning a dictionary mapping labels to properties, rather than a list.

Given the accumulation of potential API changes that have turned out to be too burdensome and noisy to fix with a standard deprecation cycle, principally because they involve changing function outputs for the same inputs, it makes sense to make all those changes in a transition to version 2.0.

Although semantic versioning [6]_ technically allows API changes with major version bumps, we must acknowledge that (1) an enormous number of projects depend on scikit-image and would thus be affected by backwards incompatible changes, and (2) it is not yet common practice in the scientific Python community to put upper version bounds on dependencies, so it is very unlikely that anyone used scikit-image<1.* or scikit-image<2.* in their dependency list. This implies that releasing a version 2.0 of scikit-image with breaking API changes would disrupt a large number of users. Additionally, such wide-sweeping changes would invalidate a large number of StackOverflow and other user guides. Finally, releasing a new version with a large number of changes prevents users from gradually migrating to the new API: an old code base must be migrated wholesale because it is impossible to depend on both versions of the API. This would represent an enormous barrier of entry for many users.

Given the above, this SKIP proposes that we release a new package where we can apply everything we have learned from over a decade of development, without disrupting our existing user base.

Detailed description

It is beyond the scope of this document to list all of the proposed API changes for skimage2, many of which have yet to be decided upon. Indeed, the scope and ambition of the 2.0 transition could grow if this SKIP is accepted. This SKIP instead proposes a mechanism for managing the transition without breaking users' code. A meta-issue tracking the proposed changes can be found on GitHub, scikit-image/scikit-image#5439 [7]_. Some examples are briefly included below for illustrative purposes:

  • Stop rescaling input arrays when the dtype must be coerced to float.
  • Stop swapping coordinate axis order in different contexts, such as drawing or warping.
  • Allow automatic return of non-NumPy types, so long as they are coercible to NumPy with numpy.asarray.
  • Harmonizing similar parameters in different functions to have the same name; for example, we currently have random_seed, random_state, seed, or sample_seed in different functions, all to mean the same thing.
  • Changing measure.regionprops to return a dictionary instead of a list.
  • Combine functions that have the same purpose, such as watershed, slic, or felzenschwalb, into a common namespace. This would make it easier for new users to find out which functions they should try out for a specific task. It would also help the community grow around common APIs, where now scikit-image APIs are essentially unique for each function.

More examples can be found in "API changes for skimage2" on our Wiki <https://github.com/scikit-image/scikit-image/wiki/API-changes-for-skimage2>_.

To make this transition with a minimum amount of user disruption, this SKIP proposes releasing a new namespace, skimage2, that would provide the new API, but only if users explicitly opt in. Additionally, by releasing a new namespace, users could use both APIs at the same time, allowing users to migrate their code progressively.

pandas released 1.0.0 in January 2020, including many backwards-incompatible API changes [3]. SciPy released version 1.0 in 2017, but, given its stage of maturity and position at the base of the scientific Python ecosystem, opted not to make major breaking changes [4]. However, SciPy has adopted a policy of adding upper-bounds on dependencies [5]_, acknowledging that the ecosystem as a whole makes backwards incompatible changes on a 2 version deprecation cycle.

Several libraries have successfully migrated their user community to a new namespace with a version number on it, such as OpenCV (imported as cv2) and BeautifulSoup (imported as bs4), Jinja (jinja2) and psycopg (currently imported as psycopg2). Further afield, R's ggplot is used as ggplot2.

.. _skip4_implementation:

Implementation

As a first execution step of this SKIP, scikit-image 1.0 will be released, celebrating the maturity of the project.

First phase: Building skimage2 ................................

Afterward, a new empty skimage2 namespace will be created in our repository alongside the skimage namespace. It will be marked as experimental – importing it will warn that content in skimage2 is still unstable. This namespace should be included in releases on PyPI or elsewhere early to facilitate testing downstream.

With the new namespace available, we will start building the new API inside it. This process orients itself around the following principles:

Only one implementation If possible, only one implementation should exist, and one API should be a simple wrapper around the other. The implementation can live in either namespace depending on what is more opportune – preferably it should live in skimage2.

Independent test suite Each API should have its own independent test suite.

Small API difference Keep the differences between the old and new API small to make the eventual transition easier for users. Prefer conventional deprecations in the skimage namespace if possible.

Backwards compatible It should be possible to achieve the old behavior of the skimage API by some call or set of calls with the skimage2 API. There may be some situations where we have to break this general rule, but an argument should be made for the relevant change that breaks this rule. In those cases, we will try to provide helper functions instead.

Migration guide We will record the pathway for migrating from the old to the new API in detail in a migration guide. Each API difference will be documented.

.. _sk2-local-warning:

Local deprecation warnings Specific deprecation warnings will be added to the old API. These warnings will be silent (subclass of PendingDeprecationWarning_) and not be shown to users by default.

For example, skimage.data.binary_blobs may emit a :class:.PendingSkimage2Change warning that advises users to use skimage2.data.binary_blobs instead and how to adapt to a new signature.

During this phase, new (additional) features can still be introduced into the old skimage namespace, not only in the new one.

.. _PendingDeprecationWarning: https://docs.python.org/3/library/exceptions.html#PendingDeprecationWarning

Second phase: Transitioning to skimage2 .........................................

Once we consider the API in skimage2 complete and stable, we will publish it in a full release versioned 2.0.0. Starting with that version, importing skimage2 is encouraged and won't raise warnings.

Instead, we will mark the API in skimage as deprecated with a single top-level warning that is raised on import. This warning will encourage users to transition to skimage2. It should link to the migration guide and should explain how to enable :ref:more specific warnings <sk2-local-warning>.

Not earlier than 1 year after the release of 2.0.0, we will begin to successively remove parts of the deprecated API from skimage. Implementations and internal code that still live in skimage will be moved to skimage2. This process can and should be spread over multiple releases.

Before completely removing parts of the API, relevant :ref:warnings from the first phase <sk2-local-warning> should be made visible to users. They should be visible for 2 releases before API is actually removed (or whatever our existing deprecation policy <dep_pol>_ recommends). This helps users who want to transition slowly.

Once the skimage namespace is empty, it will be removed.

.. _dep_pol: https://scikit-image.org/docs/dev/development/contribute.html#deprecation-cycle

Code translation helper .......................

While unclear that we will have time to do so, we would like to explore the viability of building a code translation tool, to help users automate the transition to skimage2. This should alleviate the cost and work involved for switching – especially in cases that can be easily automated.

Still, this tool might not support more ambiguous or complex updates of our API, or all the complex ways in which users might use our library. Supporting these cases might be impossible or might require prohibitive development effort. Therefore, users and downstream libraries must always have other means of completing the transition manually, e.g., with the help of conventional deprecation warnings.

If this tool is successfully implemented, it will be included at the start of the second phase as an entry point_ alongside skimage2.

.. _entry point: https://packaging.python.org/en/latest/specifications/entry-points/

Backward compatibility

This proposal breaks backward compatibility in numerous places in the library. However, it does so in a new namespace, so that this proposal does not raise backward compatibility concerns for our users. That said, the authors will attempt to limit the number of backward incompatible changes to those likely to substantially improve the overall user experience. It is anticipated that porting skimage code to skimage2 will be a straightforward process and we will publish a user guide for making the transition by the time of the skimage2 release. Users will be notified about these resources - among other things - by a warning in scikit-image 1.1.

Alternatives

Releasing the new API in the same package using semantic versioning ...................................................................

This is :ref:SKIP-3 <skip_3_transition_v1>, which was rejected after discussion with the community.

Continuous deprecation over multiple versions .............................................

This transition could occur gradually over many versions. For example, for functions automatically converting and rescaling float inputs, we could add a preserve_range keyword argument that would initially default to False, but the default value of False would be deprecated, with users getting a warning to switch to True. After the switch, we could (optionally) deprecate the argument, arriving, after a further two releases, at the same place: scikit-image no longer rescales data automatically, there are no unnecessary keyword arguments lingering all over the API.

Of course, this kind of operation would have to be done simultaneously over all of the above proposed changes.

Ultimately, the core team felt that this approach generates more work for both the scikit-image developers and the developers of downstream libraries, for dubious benefit: ultimately, later versions of scikit-image will still be incompatible with prior versions, although over a longer time scale.

A new package name ..................

Since the import name is changing, it would be possible to also change the package name from scikit-image to skimage2 for example. This was proposed in a previous version of this SKIP. It shares many of the same strengths as the current proposal – chiefly – the new skimage2 namespace. This option also requires informing users about the new package. Similarly to the suggestion here, we could raise a warning when the old package is imported. It could advise users to install the new package.

However, managing and releasing two packages from the same repository is problematic. At the same time, introducing a new repository would eventually leave behind issues and pull requests and would also make it prohibitively difficult to implement one API as a wrapper around the other. Therefore, this SKIP recommends keeping the scikit-image package name.

Not making the proposed API changes ...................................

Another possibility is to reject backwards incompatible API changes outright, except in extreme cases. The core team feels that this is essentially equivalent to pinning the library at 0.19.

"scikit-image2" as the new package name .......................................

The authors acknowledge that the new names should be chosen with care to keep the disruption to scikit-image's user base and community as small as possible. However, to protect users without upper version constraints from accidentally upgrading to the new API, the package name scikit-image must be changed. Changing the import name skimage is similarly advantageous because it allows using both APIs in the same environment.

This document suggests just skimage2 as the single new name for scikit-image's API version 2.0, both for the import name and the name on PyPI, conda-forge and elsewhere. The following arguments were given in favor of this:

  • Only one new name is introduced with the project thereby keeping the number of associated names as low as possible.
  • With this change, the import and package name match.
  • Users might be confused whether they should install scikit-image2 or scikit-image-2. It was felt that skimage2 avoids this confusion.
  • Users who know what skimage is and see skimage2 in an install instruction somewhere, will likely be able to infer that it is a newer version of the package.
  • It is unlikely that users will be aware of the new API 2.0 but not of the new package name. A proposed release of scikit-image 1.1 might point users to skimage2 during the installation and update process and thereby clearly communicate the successors name.

The following arguments were made against naming the package skimage2:

  • According to the "Principle of least astonishment", scikit-image2 might be considered the least surprising evolution of the package name.
  • It breaks with the convention that is followed by other scikits including scikit-image. (It was pointed out that this convention has not been true for some time and introducing a version number in the name is a precedent anyway.)

The earlier section "Related Work" describes how other projects dealt with similar problems.

Discussion

This SKIP is the result of many evolving discussions among the core team, with fellow projects, and with our user base:

  • :ref:SKIP-3 <skip_3_transition_v1> was an earlier iteration of this SKIP. See the "Resolution" section of that document for further background on the motivation for this SKIP.
  • A pragmatic pathway towards skimage2 <https://discuss.scientific-python.org/t/a-pragmatic-pathway-towards-skimage2/530>_
  • Many discussions happened in issues and pull requests tagged as "Path to skimage2" <https://github.com/scikit-image/scikit-image/pulls?q=label%3A%22%3Ahiking_boot%3A+Path+to+skimage2%22+>_.
  • The meeting notes from the Sprint in Vienna <https://github.com/scikit-image/skimage-archive/blob/05e98d46b5d12466dcde84487fb75d710adb08b7/sprints/2025-08_sprint_notes_Vienna.md>_ captures part of the extensive discussion that informed this SKIP.

Resolution

Pending.

References and Footnotes

All SKIPs should be declared as dedicated to the public domain with the CC0 license [1], as in Copyright, below, with attribution encouraged with CC0+BY [2].

.. [1] CC0 1.0 Universal (CC0 1.0) Public Domain Dedication, https://creativecommons.org/publicdomain/zero/1.0/ .. [2] https://dancohen.org/2013/11/26/cc0-by/ .. [3] https://pandas.pydata.org/pandas-docs/stable/whatsnew/v1.0.0.html#backwards-incompatible-api-changes .. [4] https://docs.scipy.org/doc/scipy/reference/release.1.0.0.html .. [5] https://github.com/scipy/scipy/pull/12862 .. [6] https://semver.org/ .. [7] https://github.com/scikit-image/scikit-image/issues/5439

This document is dedicated to the public domain with the Creative Commons CC0 license [1]. Attribution to this source is encouraged where appropriate, as per CC0+BY [2].