NEP 52 — Python API cleanup for NumPy 2.0#

Author:

Ralf Gommers <ralf.gommers@gmail.com>

Author:

Stéfan van der Walt <stefanv@berkeley.edu>

Author:

Nathan Goldbaum <ngoldbaum@quansight.com>

Author:

Mateusz Sokół <msokol@quansight.com>

Status:

Final

Type:

Standards Track

Created:

2023-03-28

Resolution:

https://mail.python.org/archives/list/numpy-discussion@python.org/thread/QLMPFTWA67DXE3JCUQT2RIRLQ44INS4F/

Abstract#

We propose to clean up NumPy’s Python API for the NumPy 2.0 release. This includes a more clearly defined split between what is public and what is private, and reducing the size of the main namespace by removing aliases and functions that have better alternatives. Furthermore, each function is meant to be accessible from only one place, so all duplicates also need to be dropped.

Motivation and scope#

NumPy has a large API surface that evolved organically over many years:

>>> objects_in_api = [s for s in dir(np) if not s.startswith('_')]
>>> len(objects_in_api)
562
>>> modules = [s for s in objects_in_api if inspect.ismodule(eval(f'np.{s}'))]
>>> modules
['char', 'compat', 'ctypeslib', 'emath', 'fft', 'lib', 'linalg', 'ma', 'math', 'polynomial', 'random', 'rec', 'testing', 'version']
>>> len(modules)
14

The above doesn’t even include items that are public but have been hidden from __dir__. A particularly problematic example of that is np.core, which is technically private but heavily used in practice. For a full overview of what’s considered public, private or a bit in between, see numpy/numpy.

The size of the API and the lacking definition of its boundaries incur significant costs:

  • Users find it hard to disambiguate between similarly named functions.

    Looking for functions with tab completion in IPython, a notebook, or an IDE is a challenge. E.g., type np.<TAB> and look at the first six items offered: two ufuncs (abs, add), one alias (absolute), and three functions that are not intended for end-users (add_docstring, add_newdoc, add_newdoc_ufunc). As a result, the learning curve for NumPy is steeper than it has to be.

  • Libraries that mimic the NumPy API face significant implementation barriers.

    For maintainers of NumPy API-compatible array libraries (Dask, CuPy, JAX, PyTorch, TensorFlow, cuNumeric, etc.) and compilers/transpilers (Numba, Pythran, Cython, etc.) there is an implementation cost to each object in the namespace. In practice, no other library has full support for the entire NumPy API, partly because it is so hard to know what to include when faced with a slew of aliases and legacy objects.

  • Teaching NumPy is more complicated than it needs to be.

    Similarly, a larger API is confusing to learners, who not only have to find functions but have to choose which functions to use.

  • Developers are hesitant to grow the API surface.

    This happens even when the changes are warranted, because they are aware of the above concerns.

The scope of this NEP includes:

  • Deprecating or removing functionality that is too niche for NumPy, not well-designed, superseded by better alternatives, an unnecessary alias, or otherwise a candidate for removal.

  • Clearly separating public from private NumPy API by use of underscores.

  • Restructuring the NumPy namespaces to be easier to understand and navigate.

Out of scope for this NEP are:

  • Introducing new functionality or performance enhancements.

Usage and impact#

A key principle of this API refactor is to ensure that, when code has been adapted to the changes and is 2.0-compatible, that code then also works with NumPy 1.2x.x. This keeps the burden on users and downstream library maintainers low by not having to carry duplicate code which switches on the NumPy major version number.

Backward compatibility#

As mentioned above, while the new (or cleaned up, NumPy 2.0) API should be backward compatible, there is no guarantee of forward compatibility from 1.25.X to 2.0. Code will have to be updated to account for deprecated, moved, or removed functions/classes, as well as for more strictly enforced private APIs.

In order to make it easier to adopt the changes in this NEP, we will:

  1. Provide a transition guide that lists each API change and its replacement.

  2. Explicitly flag all expired attributes with a meaningful AttributeError that points out to the new place or recommends an alternative.

  3. Provide a script to automate the migration wherever possible. This will be similar to tools/replace_old_macros.sed (which adapts code for a previous C API naming scheme change). This will be sed (or equivalent) based rather than attempting AST analysis, so it won’t cover everything.

Detailed description#

Cleaning up the main namespace#

We expect to reduce the main namespace by a large number of entries, on the order of 100. Here is a representative set of examples:

  • np.inf and np.nan have 8 aliases between them, of which most can be removed.

  • A collection of random and undocumented functions (e.g., byte_bounds, disp, safe_eval, who) listed in gh-12385 can be deprecated and removed.

  • All *sctype functions can be deprecated and removed, they (see gh-17325, gh-12334, and other issues for maximum_sctype and related functions).

  • The np.compat namespace, used during the Python 2 to 3 transition, will be removed.

  • Functions that are narrow in scope, with very few public use-cases, will be removed. These will have to be identified manually and by issue triage.

New namespaces are introduced for warnings/exceptions (np.exceptions) and for dtype-related functionality (np.dtypes). NumPy 2.0 is a good opportunity to populate these submodules from the main namespace.

Functionality that is widely used but has a preferred alternative may either be deprecated (with the deprecation message pointing out what to use instead) or be hidden by not including it in __dir__. In case of hiding, a .. legacy:: directory may be used to mark such functionality in the documentation.

A test will be added to ensure limited future growth of all namespaces; i.e., every new entry will need to be explicitly added to an allow-list.

Cleaning up the submodule structure#

We will clean up the NumPy submodule structure, so it is easier to navigate. When this was discussed before (see MAINT: Hide internals of np.lib to only show submodules) there was already rough consensus on that - however it was hard to pull off in a minor release.

A basic principle we will adhere to is “one function, one location”. Functions that are exposed in more than one namespace (e.g., many functions are present in numpy and numpy.lib) need to find a single home.

We will reorganize the API reference guide along main and submodule namespaces, and only within the main namespace use the current subdivision along functionality groupings. Also by “mainstream” and special-purpose namespaces:

# Regular/recommended user-facing namespaces for general use. Present these
# as the primary set of namespaces to the users.
numpy
numpy.exceptions
numpy.fft
numpy.linalg
numpy.polynomial
numpy.random
numpy.testing
numpy.typing

# Special-purpose namespaces. Keep these, but document them in a separate
# grouping in the reference guide and explain their purpose.
numpy.array_api
numpy.ctypeslib
numpy.emath
numpy.f2py  # only a couple of public functions, like `compile` and `get_include`
numpy.lib.stride_tricks
numpy.lib.npyio
numpy.rec
numpy.dtypes
numpy.array_utils

# Legacy (prefer not to use, there are better alternatives and/or this code
# is deprecated or isn't reliable). This will be a third grouping in the
# reference guide; it's still there, but de-emphasized and the problems
# with it or better alternatives are explained in the docs.
numpy.char
numpy.distutils
numpy.ma
numpy.matlib

# To remove
numpy.compat
numpy.core  # rename to _core
numpy.doc
numpy.math
numpy.version  # rename to _version
numpy.matrixlib

# To clean out or somehow deal with: everything in `numpy.lib`

Note

TBD: will we preserve np.lib or not? It only has a couple of unique functions/objects, like Arrayterator (a candidate for removal), NumPyVersion, and the stride_tricks, mixins and format subsubmodules. numpy.lib itself is not a coherent namespace, and does not even have a reference guide page.

We will make all submodules available lazily, so that users don’t have to type import numpy.xxx but can use import numpy as np; np.xxx.*, while at the same time not negatively impacting the overhead of import numpy. This has been very helpful for teaching scikit-image and SciPy, and it resolves a potential issue for Spyder users because Spyder already makes all submodules available - so code using the above import pattern then works in Spyder but not outside it.

Reducing the number of ways to select dtypes#

The many dtype classes, instances, aliases and ways to select them are one of the larger usability problems in the NumPy API. E.g.:

>>> # np.intp is different, but compares equal too
>>> np.int64 == np.int_ == np.dtype('i8') == np.sctypeDict['i8']
True
>>> np.float64 == np.double == np.float_ == np.dtype('f8') == np.sctypeDict['f8']
True
### Really?
>>> np.clongdouble == np.clongfloat == np.longcomplex == np.complex256
True

These aliases can go: https://numpy.org/devdocs/reference/arrays.scalars.html#other-aliases

All one-character type code strings and related routines like mintypecode will be marked as legacy.

To discuss:

  • move all dtype-related classes to np.dtypes?

  • canonical way to compare/select dtypes: np.isdtype (new, xref array API NEP), leaving np.issubdtype for the more niche use of numpy’s dtype class hierarchy, and hide most other stuff.

  • possibly remove float96/float128? they’re aliases that may not exist, and are too easy to shoot yourself in the foot with.

Cleaning up the niche methods on numpy.ndarray#

The ndarray object has a lot of attributes and methods, some of which are too niche to be that prominent, all that does is distract the average user. E.g.:

  • .itemset (already discouraged)

  • .newbyteorder (too niche)

  • .ptp (niche, use np.ptp function instead)

API changes considered and rejected#

For some functions and submodules it turned out that removing them would cause too much disruption or would require an amount of work disproportional to the actual gain. We arrived at this conclusion for such items:

  • Removing business day functions: np.busday_count, np.busday_offset, np.busdaycalendar.

  • Removing np.nan* functions and introducing new nan_mode argument to the related base functions.

  • Hiding histogram functions in the np.histograms submodule.

  • Hiding c_, r_ and s_ in the np.lib.index_tricks submodule.

  • Functions that looked niche but are present in the Array API (for example np.can_cast).

  • Removing .repeat and .ctypes from ndarray object.

Implementation#

The implementation has been split over many different PRs, each touching on a single API or a set of related APIs. Here’s a sample of the most impactful PRs:

The complete list of cleanup work done in the 2.0 release can be found by searching a dedicated label:

Some PRs has already been merged and shipped with the 1.25.0 release. For example, deprecating non-preferred aliases:

Hiding or removing objects that are accidentally made public or not even NumPy objects at all:

Creation of new namespaces to make it easier to navigate the module structure:

Alternatives#

Discussion#

References and footnotes#