NEP: array API standard adoption (NEP 47)
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
Hi all, Here is a NEP, written together with Stephan Hoyer and Aaron Meurer, for discussion on adoption of the array API standard ( https://data-apis.github.io/array-api/latest/). This will add a new numpy.array_api submodule containing that standardized API. The main purpose of this API is to be able to write code that is portable to other array/tensor libraries like CuPy, PyTorch, JAX, TensorFlow, Dask, and MXNet. We expect this NEP to remain in draft state for quite a while, while we're gaining experience with using it in downstream libraries, discuss adding it to other array libraries, and finishing some of the loose ends (e.g., specifications for linear algebra functions that aren't merged yet, see https://github.com/data-apis/array-api/pulls) in the API standard itself. See https://mail.python.org/pipermail/numpy-discussion/2020-November/081181.html for an initial discussion about this topic. Please keep high-level discussion here and detailed comments on https://github.com/numpy/numpy/pull/18456. Also, you can access a rendered version of the NEP from that PR (see PR description for how), which may be helpful. Cheers, Ralf Abstract -------- We propose to adopt the `Python array API standard`_, developed by the `Consortium for Python Data API Standards`_. Implementing this as a separate new namespace in NumPy will allow authors of libraries which depend on NumPy as well as end users to write code that is portable between NumPy and all other array/tensor libraries that adopt this standard. .. note:: We expect that this NEP will remain in a draft state for quite a while. Given the large scope we don't expect to propose it for acceptance any time soon; instead, we want to solicit feedback on both the high-level design and implementation, and learn what needs describing better in this NEP or changing in either the implementation or the array API standard itself. Motivation and Scope -------------------- Python users have a wealth of choice for libraries and frameworks for numerical computing, data science, machine learning, and deep learning. New frameworks pushing forward the state of the art in these fields are appearing every year. One unintended consequence of all this activity and creativity has been fragmentation in multidimensional array (a.k.a. tensor) libraries - which are the fundamental data structure for these fields. Choices include NumPy, Tensorflow, PyTorch, Dask, JAX, CuPy, MXNet, and others. The APIs of each of these libraries are largely similar, but with enough differences that it’s quite difficult to write code that works with multiple (or all) of these libraries. The array API standard aims to address that issue, by specifying an API for the most common ways arrays are constructed and used. The proposed API is quite similar to NumPy's API, and deviates mainly in places where (a) NumPy made design choices that are inherently not portable to other implementations, and (b) where other libraries consistently deviated from NumPy on purpose because NumPy's design turned out to have issues or unnecessary complexity. For a longer discussion on the purpose of the array API standard we refer to the `Purpose and Scope section of the array API standard < https://data-apis.github.io/array-api/latest/purpose_and_scope.html>`__ and the two blog posts announcing the formation of the Consortium [1]_ and the release of the first draft version of the standard for community review [2]_. The scope of this NEP includes: - Adopting the 2021 version of the array API standard - Adding a separate namespace, tentatively named ``numpy.array_api`` - Changes needed/desired outside of the new namespace, for example new dunder methods on the ``ndarray`` object - Implementation choices, and differences between functions in the new namespace with those in the main ``numpy`` namespace - A new array object conforming to the array API standard - Maintenance effort and testing strategy - Impact on NumPy's total exposed API surface and on other future and under-discussion design choices - Relation to existing and proposed NumPy array protocols (``__array_ufunc__``, ``__array_function__``, ``__array_module__``). - Required improvements to existing NumPy functionality Out of scope for this NEP are: - Changes in the array API standard itself. Those are likely to come up during review of this NEP, but should be upstreamed as needed and this NEP subsequently updated. Usage and Impact ---------------- *This section will be fleshed out later, for now we refer to the use cases given in* `the array API standard Use Cases section < https://data-apis.github.io/array-api/latest/use_cases.html>`__ In addition to those use cases, the new namespace contains functionality that is widely used and supported by many array libraries. As such, it is a good set of functions to teach to newcomers to NumPy and recommend as "best practice". That contrasts with NumPy's main namespace, which contains many functions and objects that have been superceded or we consider mistakes - but that we can't remove because of backwards compatibility reasons. The usage of the ``numpy.array_api`` namespace by downstream libraries is intended to enable them to consume multiple kinds of arrays, *without having to have a hard dependency on all of those array libraries*: .. image:: _static/nep-0047-library-dependencies.png Adoption in downstream libraries ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The prototype implementation of the ``array_api`` namespace will be used with SciPy, scikit-learn and other libraries of interest that depend on NumPy, in order to get more experience with the design and find out if any important parts are missing. The pattern to support multiple array libraries is intended to be something like:: def somefunc(x, y): # Retrieves standard namespace. Raises if x and y have different # namespaces. See Appendix for possible get_namespace implementation xp = get_namespace(x, y) out = xp.mean(x, axis=0) + 2*xp.std(y, axis=0) return out The ``get_namespace`` call is effectively the library author opting in to using the standard API namespace, and thereby explicitly supporting all conforming array libraries. The ``asarray`` / ``asanyarray`` pattern ```````````````````````````````````````` Many existing libraries use the same ``asarray`` (or ``asanyarray``) pattern as NumPy itself does; accepting any object that can be coerced into a ``np.ndarray``. We consider this design pattern problematic - keeping in mind the Zen of Python, *"explicit is better than implicit"*, as well as the pattern being historically problematic in the SciPy ecosystem for ``ndarray`` subclasses and with over-eager object creation. All other array/tensor libraries are more strict, and that works out fine in practice. We would advise authors of new libraries to avoid the ``asarray`` pattern. Instead they should either accept just NumPy arrays or, if they want to support multiple kinds of arrays, check if the incoming array object supports the array API standard by checking for ``__array_namespace__`` as shown in the example above. Existing libraries can do such a check as well, and only call ``asarray`` if the check fails. This is very similar to the ``__duckarray__`` idea in :ref:`NEP30`. .. _adoption-application-code: Adoption in application code ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The new namespace can be seen by end users as a cleaned up and slimmed down version of NumPy's main namespace. Encouraging end users to use this namespace like:: import numpy.array_api as xp x = xp.linspace(0, 2*xp.pi, num=100) y = xp.cos(x) seems perfectly reasonable, and potentially beneficial - users get offered only one function for each purpose (the one we consider best-practice), and they then write code that is more easily portable to other libraries. Backward compatibility ---------------------- No deprecations or removals of existing NumPy APIs or other backwards incompatible changes are proposed. High-level design ----------------- The array API standard consists of approximately 120 objects, all of which have a direct NumPy equivalent. This figure shows what is included at a high level: .. image:: _static/nep-0047-scope-of-array-API.png The most important changes compared to what NumPy currently offers are: - A new array object which: - conforms to the casting rules and indexing behaviour specified by the standard, - does not have methods other than dunder methods, - does not support the full range of NumPy indexing behaviour. Advanced indexing with integers is not supported. Only boolean indexing with a single (possibly multi-dimensional) boolean array is supported. An indexing expression that selects a single element returns a 0-D array rather than a scalar. - Functions in the ``array_api`` namespace: - do not accept ``array_like`` inputs, only NumPy arrays and Python scalars - do not support ``__array_ufunc__`` and ``__array_function__``, - use positional-only and keyword-only parameters in their signatures, - have inline type annotations, - may have minor changes to signatures and semantics of individual functions compared to their equivalents already present in NumPy, - only support dtype literals, not format strings or other ways of specifying dtypes - DLPack_ support will be added to NumPy, - New syntax for "device support" will be added, through a ``.device`` attribute on the new array object, and ``device=`` keywords in array creation functions in the ``array_api`` namespace, - Casting rules that differ from those NumPy currently has. Output dtypes can be derived from input dtypes (i.e. no value-based casting), and 0-D arrays are treated like >=1-D arrays. - Not all dtypes NumPy has are part of the standard. Only boolean, signed and unsigned integers, and floating-point dtypes up to ``float64`` are supported. Complex dtypes are expected to be added in the next version of the standard. Extended precision, string, void, object and datetime dtypes, as well as structured dtypes, are not included. Improvements to existing NumPy functionality that are needed include: - Add support for stacks of matrices to some functions in ``numpy.linalg`` that are currently missing such support. - Add the ``keepdims`` keyword to ``np.argmin`` and ``np.argmax``. - Add a "never copy" mode to ``np.asarray``. Functions in the ``array_api`` namespace ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Let's start with an example of a function implementation that shows the most important differences with the equivalent function in the main namespace:: def max(x: array, /, *, axis: Optional[Union[int, Tuple[int, ...]]] = None, keepdims: bool = False ) -> array: """ Array API compatible wrapper for :py:func:`np.max <numpy.max>`. """ return np.max._implementation(x, axis=axis, keepdims=keepdims) This function does not accept ``array_like`` inputs, only ``ndarray``. There are multiple reasons for this. Other array libraries all work like this. Letting the user do coercion of lists, generators, or other foreign objects separately results in a cleaner design with less unexpected behaviour. It's higher-performance - less overhead from ``asarray`` calls. Static typing is easier. Subclasses will work as expected. And the slight increase in verbosity because users have to explicitly coerce to ``ndarray`` on rare occasions seems like a small price to pay. This function does not support ``__array_ufunc__`` nor ``__array_function__``. These protocols serve a similar purpose as the array API standard module itself, but through a different mechanisms. Because only ``ndarray`` instances are accepted, dispatching via one of these protocols isn't useful anymore. This function uses positional-only parameters in its signature. This makes code more portable - writing ``max(x=x, ...)`` is no longer valid, hence if other libraries call the first parameter ``input`` rather than ``x``, that is fine. The rationale for keyword-only parameters (not shown in the above example) is two-fold: clarity of end user code, and it being easier to extend the signature in the future with keywords in the desired order. This function has inline type annotations. Inline annotations are far easier to maintain than separate stub files. And because the types are simple, this will not result in a large amount of clutter with type aliases or unions like in the current stub files NumPy has. DLPack support for zero-copy data interchange ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The ability to convert one kind of array into another kind is valuable, and indeed necessary when downstream libraries want to support multiple kinds of arrays. This requires a well-specified data exchange protocol. NumPy already supports two of these, namely the buffer protocol (i.e., PEP 3118), and the ``__array_interface__`` (Python side) / ``__array_struct__`` (C side) protocol. Both work similarly, letting the "producer" describe how the data is laid out in memory so the "consumer" can construct its own kind of array with a view on that data. DLPack works in a very similar way. The main reasons to prefer DLPack over the options already present in NumPy are: 1. DLPack is the only protocol with device support (e.g., GPUs using CUDA or ROCm drivers, or OpenCL devices). NumPy is CPU-only, but other array libraries are not. Having one protocol per device isn't tenable, hence device support is a must. 2. Widespread support. DLPack has the widest adoption of all protocols, only NumPy is missing support. And the experiences of other libraries with it are positive. This contrasts with the protocols NumPy does support, which are used very little - when other libraries want to interoperate with NumPy, they typically use the (more limited, and NumPy-specific) ``__array__`` protocol. Adding support for DLPack to NumPy entails: - Adding a ``ndarray.__dlpack__`` method - Adding a ``from_dlpack`` function, which takes as input an object supporting ``__dlpack__``, and returns an ``ndarray``. DLPack is currently a ~200 LoC header, and is meant to be included directly, so no external dependency is needed. Implementation should be straightforward. Syntax for device support ~~~~~~~~~~~~~~~~~~~~~~~~~ NumPy itself is CPU-only, so it clearly doesn't have a need for device support. However, other libraries (e.g. TensorFlow, PyTorch, JAX, MXNet) support multiple types of devices: CPU, GPU, TPU, and more exotic hardware. To write portable code on systems with multiple devices, it's often necessary to create new arrays on the same device as some other array, or check that two arrays live on the same device. Hence syntax for that is needed. The array object will have a ``.device`` attribute which enables comparing devices of different arrays (they only should compare equal if both arrays are from the same library and it's the same hardware device). Furthermore, ``device=`` keywords in array creation functions are needed. For example:: def empty(shape: Union[int, Tuple[int, ...]], /, *, dtype: Optional[dtype] = None, device: Optional[device] = None) -> array: """ Array API compatible wrapper for :py:func:`np.empty <numpy.empty>`. """ return np.empty(shape, dtype=dtype, device=device) The implementation for NumPy may be as simple as setting the device attribute to the string ``'cpu'`` and raising an exception if array creation functions encounter any other value. Dtypes and casting rules ~~~~~~~~~~~~~~~~~~~~~~~~ The supported dtypes in this namespace are boolean, 8/16/32/64-bit signed and unsigned integer, and 32/64-bit floating-point dtypes. These will be added to the namespace as dtype literals with the expected names (e.g., ``bool``, ``uint16``, ``float64``). The most obvious omissions are the complex dtypes. The rationale for the lack of complex support in the first version of the array API standard is that several libraries (PyTorch, MXNet) are still in the process of adding support for complex dtypes. The next version of the standard is expected to include ``complex64`` and ``complex128`` (see `this issue < https://github.com/data-apis/array-api/issues/102>`__ for more details). Specifying dtypes to functions, e.g. via the ``dtype=`` keyword, is expected to only use the dtype literals. Format strings, Python builtin dtypes, or string representations of the dtype literals are not accepted - this will improve readability and portability of code at little cost. Casting rules are only defined between different dtypes of the same kind. The rationale for this is that mixed-kind (e.g., integer to floating-point) casting behavior differs between libraries. NumPy's mixed-kind casting behavior doesn't need to be changed or restricted, it only needs to be documented that if users use mixed-kind casting, their code may not be portable. .. image:: _static/nep-0047-casting-rules-lattice.png *Type promotion diagram. Promotion between any two types is given by their join on this lattice. Only the types of participating arrays matter, not their values. Dashed lines indicate that behaviour for Python scalars is undefined on overflow. Boolean, integer and floating-point dtypes are not connected, indicating mixed-kind promotion is undefined.* The most important difference between the casting rules in NumPy and in the array API standard is how scalars and 0-dimensional arrays are handled. In the standard, array scalars do not exist and 0-dimensional arrays follow the same casting rules as higher-dimensional arrays. See the `Type Promotion Rules section of the array API standard < https://data-apis.github.io/array-api/latest/API_specification/type_promotio...
`__ for more details.
.. note:: It is not clear what the best way is to support the different casting rules for 0-dimensional arrays and no value-based casting. One option may be to implement this second set of casting rules, keep them private, mark the array API functions with a private attribute that says they adhere to these different rules, and let the casting machinery check whether for that attribute. This needs discussion. Indexing ~~~~~~~~ An indexing expression that would return a scalar with ``ndarray``, e.g. ``arr_2d[0, 0]``, will return a 0-D array with the new array object. There are several reasons for that: array scalars are largely considered a design mistake which no other array library copied; it works better for non-CPU libraries (typically arrays can live on the device, scalars live on the host); and it's simply a consistent design. To get a Python scalar out of a 0-D array, one can simply use the builtin for the type, e.g. ``float(arr_0d)``. The other `indexing modes in the standard < https://data-apis.github.io/array-api/latest/API_specification/indexing.html that kind of check can be expensive on accelerators. The lack of advanced indexing, and boolean indexing being limited to a single n-D boolean array, is due to those indexing modes not being suitable for all types of arrays or JIT compilation. Their absence does not seem to be problematic; if a user or library author wants to use them, they can do so through zero-copy conversion to ``numpy.ndarray``. This will signal correctly to whomever reads the code that it is then NumPy-specific rather than portable to all conforming array types. The array object ~~~~~~~~~~~~~~~~ The array object in the standard does not have methods other than dunder methods. The rationale for that is that not all array libraries have methods on their array object (e.g., TensorFlow does not). It also provides only a single way of doing something, rather than have functions and methods that are effectively duplicate. Mixing operations that may produce views (e.g., indexing, ``nonzero``) in combination with mutation (e.g., item or slice assignment) is `explicitly documented in the standard to not be supported < https://data-apis.github.io/array-api/latest/design_topics/copies_views_and_...
The standard current does not prescribe a name for the array object itself. We propose to simply name it ``ndarray``. This is the most obvious name, and because of the separate namespace should not clash with ``numpy.ndarray``. Implementation -------------- .. note:: This section needs a lot more detail, which will gradually be added when the implementation progresses. A prototype of the ``array_api`` namespace can be found in https://github.com/data-apis/numpy/tree/array-api/numpy/_array_api. The docstring in its ``__init__.py`` has notes on completeness of the implementation. The code for the wrapper functions also contains ``# Note:`` comments everywhere there is a difference with the NumPy API. Two important parts that are not implemented yet are the new array object and DLPack support. Functions may need changes to ensure the changed casting rules are respected. The array object ~~~~~~~~~~~~~~~~ Regarding the array object implementation, we plan to start with a regular Python class that wraps a ``numpy.ndarray`` instance. Attributes and methods can forward to that wrapped instance, applying input validation and implementing changed behaviour as needed. The casting rules are probably the most challenging part. The in-progress dtype system refactor (NEPs 40-43) should make implementing the correct casting behaviour easier - it is already moving away from value-based casting for example. The dtype objects ~~~~~~~~~~~~~~~~~ We must be able to compare dtypes for equality, and expressions like these must be possible:: np.array_api.some_func(..., dtype=x.dtype) The above implies it would be nice to have ``np.array_api.float32 == np.array_api.ndarray(...).dtype``. Dtypes should not be assumed to have a class hierarchy by users, however we are free to implement it with a class hierarchy if that's convenient. We considered the following options to implement dtype objects: 1. Alias dtypes to those in the main namespace. E.g., ``np.array_api.float32 = np.float32``. 2. Make the dtypes instances of ``np.dtype``. E.g., ``np.array_api.float32 = np.dtype(np.float32)``. 3. Create new singleton classes with only the required methods/attributes (currently just ``__eq__``). It seems like (2) would be easiest from the perspective of interacting with functions outside the main namespace. And (3) would adhere best to the standard. TBD: the standard does not yet have a good way to inspect properties of a dtype, to ask questions like "is this an integer dtype?". Perhaps this is easy enough to do for users, like so:: def _get_dtype(dt_or_arr): return dt_or_arr.dtype if hasattr(dt_or_arr, 'dtype') else dt_or_arr def is_floating(dtype_or_array): dtype = _get_dtype(dtype_or_array) return dtype in (float32, float64) def is_integer(dtype_or_array): dtype = _get_dtype(dtype_or_array) return dtype in (uint8, uint16, uint32, uint64, int8, int16, int32, int64) However it could make sense to add to the standard. Note that NumPy itself currently does not have a great for asking such questions, see `gh-17325 <https://github.com/numpy/numpy/issues/17325>`__.
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Sun, 2021-02-21 at 17:30 +0100, Ralf Gommers wrote:
There is too much to unpack in a day, I hope I did not miss something particularly important while reading. Do you have plans to try some of this outside of NumPy, or maybe make a repo in the numpy org for it? Some thoughts: The DLPack integration: I honestly think we can split that out, maybe even without or at least a very short NEP. It seems like a good addition to me. And a simple one, especially if we don't need to integrate it into `np.array(...)`. --- It seems the current idea is to create a new NumPy array subclass. That sounds good, but I am a bit worried how that is going to interact with actual NumPy arrays. Most SciPy users, will still use NumPy proper. So SciPy must not return this "random" subclass, when the user did pass in a NumPy array (or even by default). At that point, what will the SciPy dev do to juggle the fact that you would like to use the new API internally, but the interface must still default to NumPy (and may depend on the input)? This is probably not very tricky, but I am slightly worried about what happens when things get mixed up. Also if a user passes this "minimal array" into a current NumPy API function, it will often not work if it is a subclass. --- Related to that: how important is it to keep that namespace a "minimal" implementation, rather than a "conforming" one? For example, would you want to reject `numpy_api.array([1, 2, 3, 4], dtype="i,i")`? Or just `dtype="complex128"`? Maybe I got the wrong impression though. Is the aim for a minimal implementation, but you are OK as long as it is a conforming one? --- The implementation mentions bypassing `__array_function__` in the current implementation. That requires a semi-formalization of how `__array_function__` should be "bypassed". I think that is useful, but I have also been wondering about going the pytorch route of in-lining the check to avoid the current overheads. That avoids a bit jumping between C and python and multiple function calls, but might mean that `func._implementation` is a bit in the way. Besides `_implementation` usually does support array subclasses and may still dispatch again internally at this time! --- I am somewhat worried that getting the promotion (and other quirks) to where you want could be very tricky, unless we are patient enough to wait for NumPy proper to evolve. Hopefully I am just too pessimistic and e.g. a mild form of code duplication can solve all of that. Probably time and trial-and-error will be the judge on that... Cheers, Sebastian
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
Thanks for the feedback Sebastian! On Mon, Feb 22, 2021 at 7:49 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
I agree. I included it in the NEP because it's the standard exchange mechanism in the API standard, but it makes perfect sense to implement as a standalone feature.
Not a subclass! If you got that impression, I should clarify the text. The idea is a standalone class that doesn't inherit from anything, and has only the methods, attributes and semantics described in the API standard. It just uses np.asarray under the hood. This will be clearer when we have a prototype for it, probably within two weeks. Most SciPy users, will still use NumPy proper. So SciPy must not
return this "random" subclass, when the user did pass in a NumPy array (or even by default).
Agreed. ndarray in = ndarray out; new array object in = new array object out.
Yes, I'd definitely want to reject that. Format strings are terrible. Maybe I got the wrong impression though. Is the aim for a minimal
implementation, but you are OK as long as it is a conforming one?
In principle we're okay with a conforming one that's a superset of what is in the standard. But I think we'd only want to do that if creating a minimal one turns out to be difficult. Having the minimal required set is much nicer when one wants to write portable code. Because then you can do so without checking the docs whether any object/method is in "minimal" or in "extended".
I think simply: def somefunc(x): # do whatever checks needed here for, e.g., input validation # then call the native numpy implementation: return np.somefunc._implementation(x) I think that is useful, but
What, it dispatches again? That seems very suboptimal. If there's no clean way to avoid a dispatch, it may make sense to just check array inputs for the presence of __array_function__ and raise an exception if it's present. It's not just about overhead (that's a minor thing), it's that the feature does not make much sense in combination with the array_api namespace. The "get a hold of a new namespace" approach is like __array_module__, which was an alternative to __array_function__ not an addition to it.
I do agree that the different casting rules are the single most tricky issue implementation wise. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Mon, 2021-02-22 at 20:16 +0100, Ralf Gommers wrote: <snip>
Sorry, you do write "wraps a numpy.ndarray", I am not sure why I got the subclass idea when reading it yesterday. But, in that case maybe you should just implement it as: def somefunc(x): # do whatever checks needed here for, e.g., input validation # then call the native numpy implementation: x = as_numpy_array(x) result = np.somefunc(x) return as_minimal(x) and not even use the `._implementation`? I guess small issue is that we don't have a way to call `as_ndarray` on all relevant inputs conveniently. In most cases it will be straight forward though. I assume you need your own `as_numpy_array` call, to reject ndarray- subclasses, that `_implementation` will allow to pass through.
Then I should wait for the prototype, for more discussion :).
Agreed, I guess I am wondering whether we can find a good solution that does not involve writing stubs around 140 functions with more strict input validation. But maybe it is also not particularly difficult or churn to do... Or even automate, e.g. from the typing stubs.
Right, you would like to have a minimal implementation somewhere. Having it in NumPy could be convenient, although not strictly necessary. <snip>
I do not think our functions where ever rewritten to only use e.g. `._implementation()` internally. I am not even quite certain that would be correct for subclasses. It is annoying that you may have to struggle with it here to do something that is different from the implicit dispatchers. But on the up-side a clear solution would be helpful in any case. Cheers, Sebastian
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Mon, Feb 22, 2021 at 8:52 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
That makes sense. It's like doing a double asarray, but you get the fast path in __array_function__ I guess small issue is that
we don't have a way to call `as_ndarray` on all relevant inputs conveniently.
Why not? The input should be very well-defined, basically just instances of the new array object. Note that you cannot pass lists, generators, or other such types. In most cases it will be straight forward though.
I think it's easy to do, and better than something "smart". Also note that there are no typing stubs, the type annotations are clean enough that they can be added inline, which is much nicer than stubs.
That was my original thinking - just reuse `np.ndarray`, and have the "minimal" thing as a standalone implementation in a new package. But that's more work, and less nice. After getting used to the idea of a second array object, I'm actually much happier with having it in numpy.
Agreed. Cheers, Ralf
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Mon, Feb 22, 2021 at 7:49 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
Sorry, I forgot to answer this question. That is what we're doing now, the current prototype is at https://github.com/data-apis/numpy/tree/array-api/numpy/_array_api. I do expect that as soon we need any changes in C code, that becomes impractical. I think merging as a private submodule (numpy._array_api) makes sense. That will help with WIP PRs to other libraries - then we can use the "test against master" CI for that, rather than having to make a mess injecting things inside CI. Also, there are a few parts of the NEP that are improvements outside of the new submodule. Not only DLPack, but also consistency in "stacks of matrices" in linalg functions, adding a missing keepdims keyword, the never-copy mode for asarray, and improving the API for inspecting dtype families (https://github.com/numpy/numpy/issues/17325). Those things can all be pushed forward. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
Top Posting, to discuss post specific questions about NEP 47 and partially the start on implementing it in: https://github.com/numpy/numpy/pull/18585 There are probably many more that will crop up. But for me, each of these is a pretty major difficulty without a clear answer as of now. 1. I still need clarity how a library is supposed to use this namespace when the user passes in a NumPy array (mentioned before). The user must get back a NumPy array after all. Maybe that is just a decorator, but it seems important. 2. `np.result_type` special cases array-scalars (the current PR), NEP 47 promises it will not. The PR could attempt to work around that using `arr.dtype` int `result_type`, I expect there are more details to fight with there, but I am not sure. 3. For all other function, the same problem applies. You don't actually have anything to fix NumPy promotion rules. You could bake your own cake here for numeric types, but I am not sure, you might also need NEP 43 in all its promotion power to pull it off. 4. Now that I looked at the above, I do not feel its reasonable to limit this functionality to numeric dtypes. If someone uses a NumPy rational-dtype, why should a SciPy function currently implemented in pure NumPy reject that? In other words, I think this is the point where trying to be "minimal" is counterproductive. 4. The PR makes no attempt at handling binary operators in any way aside from greedily coercing the other operand. 5. What happens with a mix of array-likes or even array subclasses like `astropy.quantity`? 6. Is there any provision on how to deal with mixed array-like inputs? CuPy+numpy, etc.? I don't think we have to figure out everything up-front, but I do think there are a few very fundamental questions still open, at least for me personally. Cheers, Sebastian On Sun, 2021-02-21 at 17:30 +0100, Ralf Gommers wrote:
![](https://secure.gravatar.com/avatar/0383e4cae325f65a1bbd906be4be2276.jpg?s=120&d=mm&r=g)
On Wed, Mar 10, 2021 at 10:42 AM Sebastian Berg <sebastian@sipsolutions.net> wrote:
The idea is to work around it everywhere, so that it follows the rules in the spec (no array scalars, no value-based casting). I haven't started it yet, though, so I don't know yet how hard it will be. If it ends up being too hard we could put it in the same camp as device support and dlpack support where it needs some basic implementation in numpy itself first before we can properly do it in the array API namespace.
The idea of minimality is to make it so users can be sure they will be able to use other libraries, once they also have array API compliant namespaces. A rational-dtype wouldn't ever be implemented in those other libraries, because it isn't part of the standard, so if a user is using those, that is a sign they are using things that aren't in the array API, so they can't expect to be able to swap out their dtypes. If a user wants to use something that's only in NumPy, then they should just use NumPy.
Neither of these are defined in the spec. The spec only deals with staying inside of the compliant namespace. It doesn't require any behavior mixing things from other namespaces. That's generally considered a much harder problem, and there is the data interchange protocol to deal with it (https://data-apis.github.io/array-api/latest/design_topics/data_interchange....). Aaron Meurer
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Wed, 2021-03-10 at 13:44 -0700, Aaron Meurer wrote:
Quite frankly. If you really want to implement a minimal API, it may be best to just write it yourself and ditch NumPy. (Of course I currently doubt that the NEP 47 implementation should be minimal.) About doing promotion yourself ("promotion" as in what ufuncs do; I call `np.result_type` "common DType", because it is used e.g. in `concatenate`): Ufuncs have at least one more rule for true-division, plus there may be mixed float-int loops, etc. Since the standard is very limited and you only have numeric dtypes that might be all though. In any case, my point is: If NumPy does strange things (and it does with 0-D arrays currently). You could cook your own soup there also, and implement it in NumPy by using `signature=...` in the ufunc call.
This is not about the "user", in your scenario the end-user does use NumPy. The way I understand this is not a prerequisite. If it is, a lot of things will be simpler though, and most of my doubts will go away (but be replaced with uncertainty about the usefulness). The problem is that SciPy as the "library author" wants to to use NEP 47 without limiting the end-user (or the end-user even noticing!). The distinction between end-user and library author (someone who writes a function that should work with numpy, pytorch, etc.) is very important here and too all of these "protocol" discussions. I assume that SciPy should be able to have the cake and eat it to: * Uses the limited array-api and make sure to only rely on the minimal subset. * Not artificially limit end-users who pass in NumPy arrays. The second point can also be read as: SciPy would be able to support practically all current NumPy array use cases without jumping through any additional hoops (or well, maybe a bit of churn, but churn that is made easy by as of now undefined API).
OK, maybe you can get away with it, since the current proposal seems to be that `get_namespace()` raises on mixed input. Still seems like something that should probably raise an error rather than coerce to NumPy when calling: `nep47_array_object + dask_array`. Cheers, Sebastian
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Wed, Mar 10, 2021 at 11:41 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
I'm not really sure what to say other than that I don't think anyone will be served by "ditching NumPy". The goal for this "minimal" part is to provide an API that you can write code against that will work portably across other array libraries. That seems like a valuable goal, right? And if you want NumPy-specific things that other libraries don't commonly (or at all) implement and are not supported by array_api, then you don't use this API but the existing main numpy namespace.
Interesting idea.
SciPy would still be free to implement *both* a portable code path and a numpy-specific path (if that makes sense, which I doubt in many cases). There's just no way those two code paths can be 100% common, because no other library implements a rational dtype.
The example feels a little forced. >99% of end user code written against libraries like SciPy uses standard numerical dtypes. Things like a rational dtype are very niche. A rationale dtype works with most NumPy functions, but is not at all guaranteed to work with SciPy functions - and if it does it's accidental, untested and may break if SciPy would change its implementation (e.g. move from pure Python + NumPy to Cython or C++).
I suspect you have things in mind that are not actually supported by SciPy today. The rational dtype is one example, but so are ndarray subclasses. Take masked arrays as an example - these are not supported today, except for scipy.stats.mstats functionality - where support is intentional, special-cases and tested. For masked arrays as well as other arbitrary fancy subclasses, there's some not-well-defined subset of functionality that may work today, but that is fragile, untested and can break without warning in any release. Only Liskov-substitutable ndarray subclasses are not fragile - those are simply coerced to ndarray via the ubiquitous `np.asarray` pattern, and ndarrays are returned. That must and will remain working. This is a complex topic, and it's possible that I'm missing other use cases you have in mind, so I thought I'd make a diagram to explain the difference between the custom dtypes & subclasses that are supported by NumPy itself but not by downstream libraries: https://github.com/rgommers/numpy/blob/numpy-scipy-custom-inputs/doc/neps/_s...
Agreed, this must raise too. Cheers, Ralf
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Wed, Mar 10, 2021 at 6:41 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
All great questions, that Sebastian. Let me reply to the questions that Aaron didn't reply to inline below.
I agree that it will be a common pattern that libraries will accept all standard-compliant array types plus numpy.ndarray. And the output array type should match the input type. In Aaron's implementation the new array object has a numpy.ndarray as private attribute, so that's the instance that should be returned. A decorator seems like a sensible way to handle that. Or a simple utility function, something like `return correct_arraytype(out)`. Either way, that pattern should be added to NEP 47. I don't see a fundamental problem here, we just need to find the nicest UX for it. 3. For all other function, the same problem applies. You don't actually
This is probably the single most difficult question implementation-wise. Note that there are only numerical dtypes (plus boolean), so dealing with string, datetime, object or third-party dtypes is a non-issue. 4. The PR makes no attempt at handling binary operators in any way
aside from greedily coercing the other operand.
Agreed. This is the same point as (3) I think - how to handle dtype promotion is the main open question.
5. What happens with a mix of array-likes or even array subclasses like `astropy.quantity`?
Array-likes (e.g. list) should raise an exception, the NEP clearly says "do not accept array_like dtypes". This is what every other array/tensor library already does. Array subclasses should work as expected, assuming they're valid subclasses and not things like np.matrix. Using Mypy will help avoid writing more subclasses that break the Liskov substitution principle. More comments in https://numpy.org/neps/nep-0047-array-api-standard.html#the-asarray-asanyarr... Mixing two different types of arrays into a single function call should raise an exception. A design goal is: enable writing functions `somefunc(x1, x2)` that work for any type of array where `x1, x2` come from the same library = so they're either the same type, or two types for which the library itself knows how to mix them. If x1 and x2 are from different libraries, this will raise an exception. To be clear, it is not intended that `np.array_api.somefunc(x_cupy)` works - this will raise an exception. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Thu, 2021-03-11 at 12:37 +0100, Ralf Gommers wrote:
To be clear, I do not expect complete answers to these questions right now. (Although being unsure about some of them does make me slightly reluctant to merge the work-in-progress into NumPy proper as opposed to a separate repo.) Also, yes, most/all questions are hopefully are just trivialities to check of (or no more than seeds for thought). Or even just a starting point for making NEP 47's "Usage and Impact" section more complete including them as either "example usage patterns" or "limitations". My second takeaway from the questions is that I have doubts the "minimal" version will pan out, it feels like many of the questions might disappear if you drop that part. So, from my current thinking, the minimal implementation may not be a good "NEP 47" implementation. That does _not_ mean that I think you should pause and reconsider or even worry about pleasing me with good answers! Just continue under whatever assumption you prefer and if it turns out that "minimal" won't work for NEP 47: no harm done! We need a "minimal implementation" in any case. Cheers, Sebastian [1] If SciPy needs an additional NumPy code path to keep support `object` arrays or other dtypes – right now even complex –, then the reader needs to be aware of that to make a decision if NEP 47 will actually help for their library. Will AstroPy have to reimplement `astropy.units.Quantity` to be "standard conform" (is that even possible!?) before it can easily adopt it any of its API that currently works with `astropy.units.Quantity`?
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Thu, Mar 11, 2021 at 6:08 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
Yes, those are always good to have more of.
My impression is that a strictly compliant (or "minimal") version is *more* useful than something that's a mix between portable and non-portable functionality. The reason to add more than the minimum required functionality would be that it's too hard to hide the numpy-specific extras. E.g., if we'd do `np.array_api.int32 = np.int32` then that dtype would have methods and behavior that's NumPy-specific. But it'd be hard to hide, so we'd accept it. It's maybe easier to discuss in a call, I've put it on the community meeting agenda.
Yes, I agree.
Clearly. This is why we'd like to have some WIP PRs for other libraries, actual code to review will be more helpful than only a proposal.
I'm not sure if the question is well-defined, so let me answer both cases: 1. If the APIs in question require units, then there's no other array/tensor types that have unit support, so those APIs accept *only* Quantity. Adopting the standard isn't possible. 2. If the units are unnecessary/optional, then Quantity is not special and can be treated exactly the same as a `numpy.ndarray`. We don't intend to make any changes to how ndarray subclasses work, so if ndarray works with that API after adoption of the standard then Quantity works too. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Sun, 2021-02-21 at 17:30 +0100, Ralf Gommers wrote:
There is too much to unpack in a day, I hope I did not miss something particularly important while reading. Do you have plans to try some of this outside of NumPy, or maybe make a repo in the numpy org for it? Some thoughts: The DLPack integration: I honestly think we can split that out, maybe even without or at least a very short NEP. It seems like a good addition to me. And a simple one, especially if we don't need to integrate it into `np.array(...)`. --- It seems the current idea is to create a new NumPy array subclass. That sounds good, but I am a bit worried how that is going to interact with actual NumPy arrays. Most SciPy users, will still use NumPy proper. So SciPy must not return this "random" subclass, when the user did pass in a NumPy array (or even by default). At that point, what will the SciPy dev do to juggle the fact that you would like to use the new API internally, but the interface must still default to NumPy (and may depend on the input)? This is probably not very tricky, but I am slightly worried about what happens when things get mixed up. Also if a user passes this "minimal array" into a current NumPy API function, it will often not work if it is a subclass. --- Related to that: how important is it to keep that namespace a "minimal" implementation, rather than a "conforming" one? For example, would you want to reject `numpy_api.array([1, 2, 3, 4], dtype="i,i")`? Or just `dtype="complex128"`? Maybe I got the wrong impression though. Is the aim for a minimal implementation, but you are OK as long as it is a conforming one? --- The implementation mentions bypassing `__array_function__` in the current implementation. That requires a semi-formalization of how `__array_function__` should be "bypassed". I think that is useful, but I have also been wondering about going the pytorch route of in-lining the check to avoid the current overheads. That avoids a bit jumping between C and python and multiple function calls, but might mean that `func._implementation` is a bit in the way. Besides `_implementation` usually does support array subclasses and may still dispatch again internally at this time! --- I am somewhat worried that getting the promotion (and other quirks) to where you want could be very tricky, unless we are patient enough to wait for NumPy proper to evolve. Hopefully I am just too pessimistic and e.g. a mild form of code duplication can solve all of that. Probably time and trial-and-error will be the judge on that... Cheers, Sebastian
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
Thanks for the feedback Sebastian! On Mon, Feb 22, 2021 at 7:49 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
I agree. I included it in the NEP because it's the standard exchange mechanism in the API standard, but it makes perfect sense to implement as a standalone feature.
Not a subclass! If you got that impression, I should clarify the text. The idea is a standalone class that doesn't inherit from anything, and has only the methods, attributes and semantics described in the API standard. It just uses np.asarray under the hood. This will be clearer when we have a prototype for it, probably within two weeks. Most SciPy users, will still use NumPy proper. So SciPy must not
return this "random" subclass, when the user did pass in a NumPy array (or even by default).
Agreed. ndarray in = ndarray out; new array object in = new array object out.
Yes, I'd definitely want to reject that. Format strings are terrible. Maybe I got the wrong impression though. Is the aim for a minimal
implementation, but you are OK as long as it is a conforming one?
In principle we're okay with a conforming one that's a superset of what is in the standard. But I think we'd only want to do that if creating a minimal one turns out to be difficult. Having the minimal required set is much nicer when one wants to write portable code. Because then you can do so without checking the docs whether any object/method is in "minimal" or in "extended".
I think simply: def somefunc(x): # do whatever checks needed here for, e.g., input validation # then call the native numpy implementation: return np.somefunc._implementation(x) I think that is useful, but
What, it dispatches again? That seems very suboptimal. If there's no clean way to avoid a dispatch, it may make sense to just check array inputs for the presence of __array_function__ and raise an exception if it's present. It's not just about overhead (that's a minor thing), it's that the feature does not make much sense in combination with the array_api namespace. The "get a hold of a new namespace" approach is like __array_module__, which was an alternative to __array_function__ not an addition to it.
I do agree that the different casting rules are the single most tricky issue implementation wise. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Mon, 2021-02-22 at 20:16 +0100, Ralf Gommers wrote: <snip>
Sorry, you do write "wraps a numpy.ndarray", I am not sure why I got the subclass idea when reading it yesterday. But, in that case maybe you should just implement it as: def somefunc(x): # do whatever checks needed here for, e.g., input validation # then call the native numpy implementation: x = as_numpy_array(x) result = np.somefunc(x) return as_minimal(x) and not even use the `._implementation`? I guess small issue is that we don't have a way to call `as_ndarray` on all relevant inputs conveniently. In most cases it will be straight forward though. I assume you need your own `as_numpy_array` call, to reject ndarray- subclasses, that `_implementation` will allow to pass through.
Then I should wait for the prototype, for more discussion :).
Agreed, I guess I am wondering whether we can find a good solution that does not involve writing stubs around 140 functions with more strict input validation. But maybe it is also not particularly difficult or churn to do... Or even automate, e.g. from the typing stubs.
Right, you would like to have a minimal implementation somewhere. Having it in NumPy could be convenient, although not strictly necessary. <snip>
I do not think our functions where ever rewritten to only use e.g. `._implementation()` internally. I am not even quite certain that would be correct for subclasses. It is annoying that you may have to struggle with it here to do something that is different from the implicit dispatchers. But on the up-side a clear solution would be helpful in any case. Cheers, Sebastian
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Mon, Feb 22, 2021 at 8:52 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
That makes sense. It's like doing a double asarray, but you get the fast path in __array_function__ I guess small issue is that
we don't have a way to call `as_ndarray` on all relevant inputs conveniently.
Why not? The input should be very well-defined, basically just instances of the new array object. Note that you cannot pass lists, generators, or other such types. In most cases it will be straight forward though.
I think it's easy to do, and better than something "smart". Also note that there are no typing stubs, the type annotations are clean enough that they can be added inline, which is much nicer than stubs.
That was my original thinking - just reuse `np.ndarray`, and have the "minimal" thing as a standalone implementation in a new package. But that's more work, and less nice. After getting used to the idea of a second array object, I'm actually much happier with having it in numpy.
Agreed. Cheers, Ralf
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Mon, Feb 22, 2021 at 7:49 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
Sorry, I forgot to answer this question. That is what we're doing now, the current prototype is at https://github.com/data-apis/numpy/tree/array-api/numpy/_array_api. I do expect that as soon we need any changes in C code, that becomes impractical. I think merging as a private submodule (numpy._array_api) makes sense. That will help with WIP PRs to other libraries - then we can use the "test against master" CI for that, rather than having to make a mess injecting things inside CI. Also, there are a few parts of the NEP that are improvements outside of the new submodule. Not only DLPack, but also consistency in "stacks of matrices" in linalg functions, adding a missing keepdims keyword, the never-copy mode for asarray, and improving the API for inspecting dtype families (https://github.com/numpy/numpy/issues/17325). Those things can all be pushed forward. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
Top Posting, to discuss post specific questions about NEP 47 and partially the start on implementing it in: https://github.com/numpy/numpy/pull/18585 There are probably many more that will crop up. But for me, each of these is a pretty major difficulty without a clear answer as of now. 1. I still need clarity how a library is supposed to use this namespace when the user passes in a NumPy array (mentioned before). The user must get back a NumPy array after all. Maybe that is just a decorator, but it seems important. 2. `np.result_type` special cases array-scalars (the current PR), NEP 47 promises it will not. The PR could attempt to work around that using `arr.dtype` int `result_type`, I expect there are more details to fight with there, but I am not sure. 3. For all other function, the same problem applies. You don't actually have anything to fix NumPy promotion rules. You could bake your own cake here for numeric types, but I am not sure, you might also need NEP 43 in all its promotion power to pull it off. 4. Now that I looked at the above, I do not feel its reasonable to limit this functionality to numeric dtypes. If someone uses a NumPy rational-dtype, why should a SciPy function currently implemented in pure NumPy reject that? In other words, I think this is the point where trying to be "minimal" is counterproductive. 4. The PR makes no attempt at handling binary operators in any way aside from greedily coercing the other operand. 5. What happens with a mix of array-likes or even array subclasses like `astropy.quantity`? 6. Is there any provision on how to deal with mixed array-like inputs? CuPy+numpy, etc.? I don't think we have to figure out everything up-front, but I do think there are a few very fundamental questions still open, at least for me personally. Cheers, Sebastian On Sun, 2021-02-21 at 17:30 +0100, Ralf Gommers wrote:
![](https://secure.gravatar.com/avatar/0383e4cae325f65a1bbd906be4be2276.jpg?s=120&d=mm&r=g)
On Wed, Mar 10, 2021 at 10:42 AM Sebastian Berg <sebastian@sipsolutions.net> wrote:
The idea is to work around it everywhere, so that it follows the rules in the spec (no array scalars, no value-based casting). I haven't started it yet, though, so I don't know yet how hard it will be. If it ends up being too hard we could put it in the same camp as device support and dlpack support where it needs some basic implementation in numpy itself first before we can properly do it in the array API namespace.
The idea of minimality is to make it so users can be sure they will be able to use other libraries, once they also have array API compliant namespaces. A rational-dtype wouldn't ever be implemented in those other libraries, because it isn't part of the standard, so if a user is using those, that is a sign they are using things that aren't in the array API, so they can't expect to be able to swap out their dtypes. If a user wants to use something that's only in NumPy, then they should just use NumPy.
Neither of these are defined in the spec. The spec only deals with staying inside of the compliant namespace. It doesn't require any behavior mixing things from other namespaces. That's generally considered a much harder problem, and there is the data interchange protocol to deal with it (https://data-apis.github.io/array-api/latest/design_topics/data_interchange....). Aaron Meurer
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Wed, 2021-03-10 at 13:44 -0700, Aaron Meurer wrote:
Quite frankly. If you really want to implement a minimal API, it may be best to just write it yourself and ditch NumPy. (Of course I currently doubt that the NEP 47 implementation should be minimal.) About doing promotion yourself ("promotion" as in what ufuncs do; I call `np.result_type` "common DType", because it is used e.g. in `concatenate`): Ufuncs have at least one more rule for true-division, plus there may be mixed float-int loops, etc. Since the standard is very limited and you only have numeric dtypes that might be all though. In any case, my point is: If NumPy does strange things (and it does with 0-D arrays currently). You could cook your own soup there also, and implement it in NumPy by using `signature=...` in the ufunc call.
This is not about the "user", in your scenario the end-user does use NumPy. The way I understand this is not a prerequisite. If it is, a lot of things will be simpler though, and most of my doubts will go away (but be replaced with uncertainty about the usefulness). The problem is that SciPy as the "library author" wants to to use NEP 47 without limiting the end-user (or the end-user even noticing!). The distinction between end-user and library author (someone who writes a function that should work with numpy, pytorch, etc.) is very important here and too all of these "protocol" discussions. I assume that SciPy should be able to have the cake and eat it to: * Uses the limited array-api and make sure to only rely on the minimal subset. * Not artificially limit end-users who pass in NumPy arrays. The second point can also be read as: SciPy would be able to support practically all current NumPy array use cases without jumping through any additional hoops (or well, maybe a bit of churn, but churn that is made easy by as of now undefined API).
OK, maybe you can get away with it, since the current proposal seems to be that `get_namespace()` raises on mixed input. Still seems like something that should probably raise an error rather than coerce to NumPy when calling: `nep47_array_object + dask_array`. Cheers, Sebastian
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Wed, Mar 10, 2021 at 11:41 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
I'm not really sure what to say other than that I don't think anyone will be served by "ditching NumPy". The goal for this "minimal" part is to provide an API that you can write code against that will work portably across other array libraries. That seems like a valuable goal, right? And if you want NumPy-specific things that other libraries don't commonly (or at all) implement and are not supported by array_api, then you don't use this API but the existing main numpy namespace.
Interesting idea.
SciPy would still be free to implement *both* a portable code path and a numpy-specific path (if that makes sense, which I doubt in many cases). There's just no way those two code paths can be 100% common, because no other library implements a rational dtype.
The example feels a little forced. >99% of end user code written against libraries like SciPy uses standard numerical dtypes. Things like a rational dtype are very niche. A rationale dtype works with most NumPy functions, but is not at all guaranteed to work with SciPy functions - and if it does it's accidental, untested and may break if SciPy would change its implementation (e.g. move from pure Python + NumPy to Cython or C++).
I suspect you have things in mind that are not actually supported by SciPy today. The rational dtype is one example, but so are ndarray subclasses. Take masked arrays as an example - these are not supported today, except for scipy.stats.mstats functionality - where support is intentional, special-cases and tested. For masked arrays as well as other arbitrary fancy subclasses, there's some not-well-defined subset of functionality that may work today, but that is fragile, untested and can break without warning in any release. Only Liskov-substitutable ndarray subclasses are not fragile - those are simply coerced to ndarray via the ubiquitous `np.asarray` pattern, and ndarrays are returned. That must and will remain working. This is a complex topic, and it's possible that I'm missing other use cases you have in mind, so I thought I'd make a diagram to explain the difference between the custom dtypes & subclasses that are supported by NumPy itself but not by downstream libraries: https://github.com/rgommers/numpy/blob/numpy-scipy-custom-inputs/doc/neps/_s...
Agreed, this must raise too. Cheers, Ralf
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Wed, Mar 10, 2021 at 6:41 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
All great questions, that Sebastian. Let me reply to the questions that Aaron didn't reply to inline below.
I agree that it will be a common pattern that libraries will accept all standard-compliant array types plus numpy.ndarray. And the output array type should match the input type. In Aaron's implementation the new array object has a numpy.ndarray as private attribute, so that's the instance that should be returned. A decorator seems like a sensible way to handle that. Or a simple utility function, something like `return correct_arraytype(out)`. Either way, that pattern should be added to NEP 47. I don't see a fundamental problem here, we just need to find the nicest UX for it. 3. For all other function, the same problem applies. You don't actually
This is probably the single most difficult question implementation-wise. Note that there are only numerical dtypes (plus boolean), so dealing with string, datetime, object or third-party dtypes is a non-issue. 4. The PR makes no attempt at handling binary operators in any way
aside from greedily coercing the other operand.
Agreed. This is the same point as (3) I think - how to handle dtype promotion is the main open question.
5. What happens with a mix of array-likes or even array subclasses like `astropy.quantity`?
Array-likes (e.g. list) should raise an exception, the NEP clearly says "do not accept array_like dtypes". This is what every other array/tensor library already does. Array subclasses should work as expected, assuming they're valid subclasses and not things like np.matrix. Using Mypy will help avoid writing more subclasses that break the Liskov substitution principle. More comments in https://numpy.org/neps/nep-0047-array-api-standard.html#the-asarray-asanyarr... Mixing two different types of arrays into a single function call should raise an exception. A design goal is: enable writing functions `somefunc(x1, x2)` that work for any type of array where `x1, x2` come from the same library = so they're either the same type, or two types for which the library itself knows how to mix them. If x1 and x2 are from different libraries, this will raise an exception. To be clear, it is not intended that `np.array_api.somefunc(x_cupy)` works - this will raise an exception. Cheers, Ralf
![](https://secure.gravatar.com/avatar/b4f6d4f8b501cb05fd054944a166a121.jpg?s=120&d=mm&r=g)
On Thu, 2021-03-11 at 12:37 +0100, Ralf Gommers wrote:
To be clear, I do not expect complete answers to these questions right now. (Although being unsure about some of them does make me slightly reluctant to merge the work-in-progress into NumPy proper as opposed to a separate repo.) Also, yes, most/all questions are hopefully are just trivialities to check of (or no more than seeds for thought). Or even just a starting point for making NEP 47's "Usage and Impact" section more complete including them as either "example usage patterns" or "limitations". My second takeaway from the questions is that I have doubts the "minimal" version will pan out, it feels like many of the questions might disappear if you drop that part. So, from my current thinking, the minimal implementation may not be a good "NEP 47" implementation. That does _not_ mean that I think you should pause and reconsider or even worry about pleasing me with good answers! Just continue under whatever assumption you prefer and if it turns out that "minimal" won't work for NEP 47: no harm done! We need a "minimal implementation" in any case. Cheers, Sebastian [1] If SciPy needs an additional NumPy code path to keep support `object` arrays or other dtypes – right now even complex –, then the reader needs to be aware of that to make a decision if NEP 47 will actually help for their library. Will AstroPy have to reimplement `astropy.units.Quantity` to be "standard conform" (is that even possible!?) before it can easily adopt it any of its API that currently works with `astropy.units.Quantity`?
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Thu, Mar 11, 2021 at 6:08 PM Sebastian Berg <sebastian@sipsolutions.net> wrote:
Yes, those are always good to have more of.
My impression is that a strictly compliant (or "minimal") version is *more* useful than something that's a mix between portable and non-portable functionality. The reason to add more than the minimum required functionality would be that it's too hard to hide the numpy-specific extras. E.g., if we'd do `np.array_api.int32 = np.int32` then that dtype would have methods and behavior that's NumPy-specific. But it'd be hard to hide, so we'd accept it. It's maybe easier to discuss in a call, I've put it on the community meeting agenda.
Yes, I agree.
Clearly. This is why we'd like to have some WIP PRs for other libraries, actual code to review will be more helpful than only a proposal.
I'm not sure if the question is well-defined, so let me answer both cases: 1. If the APIs in question require units, then there's no other array/tensor types that have unit support, so those APIs accept *only* Quantity. Adopting the standard isn't possible. 2. If the units are unnecessary/optional, then Quantity is not special and can be treated exactly the same as a `numpy.ndarray`. We don't intend to make any changes to how ndarray subclasses work, so if ndarray works with that API after adoption of the standard then Quantity works too. Cheers, Ralf
participants (3)
-
Aaron Meurer
-
Ralf Gommers
-
Sebastian Berg