![](https://secure.gravatar.com/avatar/1198e2d145718c841565712312e04227.jpg?s=120&d=mm&r=g)
Hi all
On Friday, Oct 19, 2018 at 10:28 AM, Matti Picus <matti.picus@gmail.com (mailto:matti.picus@gmail.com)> wrote:
Was there discussion around which of `asarray` or asanyarray` to prefer? PR 11162, https://github.com/numpy/numpy/pull/11162, proposes `asanyarray` in place of `asarray` at the entrance to `_quantile_ureduce_func` to preserve ndarray subclasses. Should we be looking into changing all the `asarray` calls into `asanyarray`?
I suspect that this will cause a large number of problems around np.matrix, so unless we deprecate that, this might cause a large amount of problems. The problem with np.matrix is that it’s a subclass, but it’s not substitutable for the base class, and so violates SOLID. There are efforts to remove np.matrix, with the largest consumer being scipy.sparse, so unless that’s revamped, deprecating np.matrix is kind of hard to do.
Best Regards, Hameer Abbasi
![](https://secure.gravatar.com/avatar/851ff10fbb1363b7d6111ac60194cc1c.jpg?s=120&d=mm&r=g)
There are exceptions for `matrix` in quite a few places, and there now is warning for `maxtrix` - it might not be bad to use `asanyarray` and add an exception for `maxtrix`. Indeed, I quite like the suggestion by Eric Wieser to just add the exception to `asanyarray` itself - that way when matrix is truly deprecated, it will be a very easy change. -- Marten
![](https://secure.gravatar.com/avatar/93a76a800ef6c5919baa8ba91120ee98.jpg?s=120&d=mm&r=g)
I don't think it makes much sense to change NumPy's existing usage of asarray() to asanyarray() unless we add subok=True arguments (which default to False). But this ends up cluttering NumPy's public API, which is also undesirable. The preferred way to override NumPy functions going forward should be __array_function__. On Fri, Oct 19, 2018 at 8:13 AM Marten van Kerkwijk < m.h.vankerkwijk@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/1198e2d145718c841565712312e04227.jpg?s=120&d=mm&r=g)
Hi!
On Friday, Oct 19, 2018 at 6:09 PM, Stephan Hoyer <shoyer@gmail.com (mailto:shoyer@gmail.com)> wrote: I don't think it makes much sense to change NumPy's existing usage of asarray() to asanyarray() unless we add subok=True arguments (which default to False). But this ends up cluttering NumPy's public API, which is also undesirable.
Agreed so far.
The preferred way to override NumPy functions going forward should be __array_function__.
I think we should “soft support” i.e. allow but consider unsupported, the case where one of NumPy’s functions is implemented in terms of others and “passing through” an array results in the correct behaviour for that array.
Best Regards, Hameer Abbasi
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 4:15 PM Hameer Abbasi <einstein.edison@gmail.com> wrote:
I'm not sure I agree. "subok" is very unpythonic; the average numpy library function should work fine for a well-behaved subclass (i.e. most things out there except np.matrix).
I don't think we have or want such a concept as "soft support". We intend to not break anything that now has asanyarray, i.e. it's supported and ideally we have regression tests for all such functions. For anything we transition over from asarray to asanyarray, PRs should come with new tests.
then may as well just rip np.matrix out straight away. What I suggested in the call about this issue is that it's not very effective to treat functions like percentile/quantile one by one without an overarching strategy. A way forward could be for someone to write an overview of which sets of functions now have asanyarray (and actually work with subclasses), which ones we can and want to change now, and which ones we can and want to change after np.matrix is gone. Also, some guidelines for new functions that we add to numpy would be handy. I suspect we've been adding new functions that use asarray rather than asanyarray, which is probably undesired. Cheers, Ralf
![](https://secure.gravatar.com/avatar/97c543aca1ac7bbcfb5279d0300c8330.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 3:28 PM, Ralf Gommers <ralf.gommers@gmail.com> wrote:
Masked arrays also tend to break code that's not expecting them (e.g. on a masked array, arr.sum()/arr.size will silently compute some meaningless nonsense instead of the mean, and there are lots of formulas out there that have some similarities with 'mean'). And people do all kinds of weird things in third-party array subclasses. Obviously we can't remove asanyarray or break existing code that assumes particular numpy functions use asanyarray, but fundamentally asanyarray is just not an API that makes sense or can be supported in a general way, and our overall goal is to get people to gradually transition away from using ndarray subclasses in general. That's why we're doing all this work to make duck arrays work. So extending asanyarray support doesn't seem like a good priority to spend our limited resources on, to me. -n -- Nathaniel J. Smith -- https://vorpus.org <http://vorpus.org>
![](https://secure.gravatar.com/avatar/93a76a800ef6c5919baa8ba91120ee98.jpg?s=120&d=mm&r=g)
The problem with asanyarray() is that there isn't any well defined subclass API for NumPy, beyond "mostly works like a NumPy array." If every NumPy subclass strictly obeyed the Liskov Substitution Principle asanyarray() would be fine, but in practice every subclass I've encountered deviates from the behavior of numpy.ndarray in some way. The means the NumPy codebase has ended up littered with hacks/workarounds to support various specific subclasses, and new subclasses still don't work reliably. This makes it challenging to change existing code. For an example of how bad this is gotten, look at all the work-arounds I had to add to support np.testing.assert_array_equal() on ndarray subclasses in this recent PR: https://github.com/numpy/numpy/pull/12119 My hope is that __array_function__ will finally let us put a stop to this by offering a better alternative to subclassing.
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 10:28 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
Thanks Nathaniel and Stephan. Your comments on my other two points are both clear and correct (and have been made a number of times before). I think the "write an overview so we can stop making ad-hoc decisions and having these discussions" is the most important point I was trying to make though. If we had such a doc and it concluded "hence we don't change anything, __array_function__ is the only way to go" then we can just close PRs like https://github.com/numpy/numpy/pull/11162 straight away. Cheers, Ralf
![](https://secure.gravatar.com/avatar/851ff10fbb1363b7d6111ac60194cc1c.jpg?s=120&d=mm&r=g)
Hi All, It seems there are two extreme possibilities for general functions: 1. Put `asarray` everywhere. The main benefit that I can see is that even if people put in list instead of arrays, one is guaranteed to have shape, dtype, etc. But it seems a bit like calling `int` on everything that might get used as an index, instead of letting the actual indexing do the proper thing and call `__index__`. 2. Do not coerce at all, but rather write code assuming something is an array already. This will often, but not always, just work for array mimics, with coercion done only where necessary (e.g., in lower-lying C code such as that of the ufuncs which has a smaller API surface and can be overridden more easily). The current __array_function__ work may well provide us with a way to combine both, if we (over time) move the coercion inside `ndarray.__array_function__` so that the actual implementation *can* assume it deals with pure ndarray - then, when relevant, calling that implementation will be what subclasses/duck arrays can happily do (and it is up to them to ensure this works). Of course, the above does not really answer what to do in the meantime. But perhaps it helps in thinking of what we are actually aiming for. One last thing: could we please stop bashing subclasses? One can subclass essentially everything in python, often to great advantage. Subclasses such as MaskedArray and, yes, Quantity, are widely used, and if they cause problems perhaps that should be seen as a sign that ndarray subclassing should be made easier and clearer. All the best, Marten On Fri, Oct 19, 2018 at 7:02 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/209654202cde8ec709dee0a4d23c717d.jpg?s=120&d=mm&r=g)
Subclasses such as MaskedArray and, yes, Quantity, are widely used, and if they cause problems perhaps that should be seen as a sign that ndarray subclassing should be made easier and clearer. Both maskedarray and quantity seem like something that would make more sense at the dtype level if our dtype system was easier to extend. It might be good to compile a list of subclassing applications, and split them into “this ought to be a dtype” and “this ought to be a different type of container”. On Fri, 19 Oct 2018 at 18:24 Marten van Kerkwijk <m.h.vankerkwijk@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/96dd777e397ab128fedab46af97a3a4a.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 7:50 PM Eric Wieser <wieser.eric+numpy@gmail.com> wrote:
Wes Mckinney has been benchmarking masks vs sentinel values for arrow: http://wesmckinney.com/blog/bitmaps-vs-sentinel-values/. The (bit) masks are faster. I'm not convinced dtypes are the way to go. Chuck
![](https://secure.gravatar.com/avatar/97c543aca1ac7bbcfb5279d0300c8330.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 7:00 PM, Charles R Harris <charlesr.harris@gmail.com> wrote:
We need to add better support for both user-defined dtypes and for user-defined containers in any case. So we're going to support both missing value strategies regardless, and people will be able to choose based on engineering trade-offs. A missing value dtype is going to integrate much more easily into the rest of numpy than a new container where you have to reimplement indexing etc., but maybe custom containers can be faster. Okay, cool, they're both on PyPI, pick your favorite! Trying to wedge masks into *ndarray* seems like a non-starter, though, because it would require auditing and updating basically all code using the numpy C API. -n -- Nathaniel J. Smith -- https://vorpus.org
![](https://secure.gravatar.com/avatar/97c543aca1ac7bbcfb5279d0300c8330.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 6:23 PM, Marten van Kerkwijk <m.h.vankerkwijk@gmail.com> wrote:
Between these two options, Numpy's APIs are very firmly on the side of "option 1", and this is common in most public APIs I'm familiar with (e.g. scipy). I guess you could try to reopen the discussion, but you'd be pushing against 15+ years of precedent there...
We need some kind of asduckarray(), that coerces lists and similar but allows duck-arrays to pass through.
Who's bashing? I've spent years thinking about this and come to the conclusion that there are no viable solutions to the problems with subclassing ndarray, but that's not the same as bashing :-). If you've thought of something we've missed, you should share it... (I also know lots of senior Python devs who believe that using Python's subclassing support is pretty much always a mistake – this talk is popularly cited: https://www.youtube.com/watch?v=3MNVP9-hglc – but the issues with ndarray are much more severe than for the average Python class.) -n -- Nathaniel J. Smith -- https://vorpus.org
![](https://secure.gravatar.com/avatar/08bcb178fb791b7c8a5e9ad8d7121bed.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 8:24 PM Marten van Kerkwijk < m.h.vankerkwijk@gmail.com> wrote:
Yes, actually getting a proper "array protocol" into Python would be a fantastic approach. We have been working with Lenore Mullin who is a researcher on the mathematics of arrays on what it means to be an array and believe we can come up with an actual array protocol that perhaps could be put into Python itself (though that isn't our immediate goal right now).
Also, we could get rid of asarray entirely by changing expectations. This automatic conversion code throughout NumPy and SciPy is an example of the confusion in both of these libraries between "user-oriented interfaces" and "developer-oriented interfaces". A developer just wants the library to use duck-typing and then raise errors if you don't provide the right type (i.e. a list instead of an array). The user-interface could happen in Jupyter, or be isolated to a high-level library or meta-code approach (of which there are several possibilities for Python).
Of course, the above does not really answer what to do in the meantime. But perhaps it helps in thinking of what we are actually aiming for.
I agree that we can stop bashing subclasses in general. The problem with numpy subclasses is that they were made without adherence to SOLID: https://en.wikipedia.org/wiki/SOLID. In particular the Liskov substitution principle: https://en.wikipedia.org/wiki/Liskov_substitution_principle . Much of this is my fault. Being a scientist/engineer more than a computer scientist, I had no idea what these principles were and did not properly apply them in creating np.matrix which clearly violates the substitution principle. We can clean all this and more up. But, we really need to start talking about NumPy 2.0 to do it. Now that Python 3.x is really here, we can raise the money for it and get it done. We don't have to just rely on volunteer time. The world will thank us for actually pushing NumPy 2.0. I know not everyone agrees, but for whatever its worth, I feel very, very strongly about this, and despite not being very active on this list for the past years, I do have a lot of understanding about how the current code actually works (and where and why its warts are). -Travis
![](https://secure.gravatar.com/avatar/5dde29b54a3f1b76b2541d0a4a9b232c.jpg?s=120&d=mm&r=g)
On Fri, Oct 26, 2018 at 7:12 PM, Travis Oliphant <teoliphant@gmail.com> wrote:
...
did not properly apply them in creating np.matrix which clearly violates the substitution principle.
So -- could a matrix subclass be made "properly"? or is that an example of something that should not have been a subclass? -CHB -- Christopher Barker, Ph.D. Oceanographer Emergency Response Division NOAA/NOS/OR&R (206) 526-6959 voice 7600 Sand Point Way NE (206) 526-6329 fax Seattle, WA 98115 (206) 526-6317 main reception Chris.Barker@noaa.gov
![](https://secure.gravatar.com/avatar/209654202cde8ec709dee0a4d23c717d.jpg?s=120&d=mm&r=g)
The latter - changing the behavior of multiplication breaks the principle. But this is not the main reason for deprecating matrix - almost all of the problems I’ve seen have been caused by the way that matrices behave when sliced. The way that m[i][j] and m[i,j] are different is just one example of this, the fact that they must be 2d is another. Matrices behaving differently on multiplication isn’t super different in my mind to how string arrays fail to multiply at all. Eric On Mon, 29 Oct 2018 at 20:54 Ralf Gommers <ralf.gommers@gmail.com> wrote: On Mon, Oct 29, 2018 at 4:31 PM Chris Barker <chris.barker@noaa.gov> wrote:
![](https://secure.gravatar.com/avatar/79653c343e58afbfbedcfd11e7b668d7.jpg?s=120&d=mm&r=g)
On Tue, Oct 30, 2018 at 12:49 AM Eric Wieser <wieser.eric+numpy@gmail.com> wrote:
The difference is that string arrays are not numeric. This is an issue since people want to pass a matrix Into places that want to multiple element wise but that then breaks that code unless special provisions are taken. Numerical codes don’t work on string arrays anyway. Eric Eric
![](https://secure.gravatar.com/avatar/93a76a800ef6c5919baa8ba91120ee98.jpg?s=120&d=mm&r=g)
On Mon, Oct 29, 2018 at 9:49 PM Eric Wieser <wieser.eric+numpy@gmail.com> wrote:
It's certainly fine for arithmetic to work differently on an element-wise basis or even to error. But np.matrix changes the shape of results from various ndarray operations (e.g., both multiplication and indexing), which is more than any dtype can do. The Liskov substitution principle (LSP) suggests that the set of reasonable ndarray subclasses are exactly those that could also in principle correspond to a new dtype. Of np.ndarray subclasses in wide-spread use, I think only the various "array with units" types come close satisfying this criteria. They only fall short insofar as they present a misleading dtype (without unit information). The main problem with subclassing for numpy.ndarray is that it guarantees too much: a large set of operations/methods along with a specific memory layout exposed as part of its public API. Worse, ndarray itself is a little quirky (e.g., with indexing, and its handling of scalars vs. 0d arrays). In practice, it's basically impossible to layer on complex behavior with these exact semantics, so only extremely minimal ndarray subclasses don't violate LSP. Once we have more easily extended dtypes, I suspect most of the good use cases for subclassing will have gone away.
![](https://secure.gravatar.com/avatar/5dde29b54a3f1b76b2541d0a4a9b232c.jpg?s=120&d=mm&r=g)
On Tue, Oct 30, 2018 at 2:22 PM, Stephan Hoyer <shoyer@gmail.com> wrote:
How about subclasses that only add functionality? My only use case of subclassing is exactly that: I have a "bounding box" object (probably could have been called a rectangle) that is a subclass of ndarray, is always shape (2,2), and has various methods for merging two such boxes, etc, adding a point, etc. I did it that way, 'cause I had a lot of code already that simply used a (2,2) array to represent a bounding box, and I wanted all that code to still work. I have had zero problems with it. Maybe that's too trivial to be worth talking about, but this kind of use case can be handy. It is a bit awkward to write the code, though -- it would be nice to have a cleaner API for this sort of subclassing (not that I have any idea how to do that) The main problem with subclassing for numpy.ndarray is that it guarantees
too much: a large set of operations/methods along with a specific memory layout exposed as part of its public API.
This is a big deal -- we really have two concepts here: - a Python class (type) with certain behaviors in Python code - a wrapper around a strided memory block. maybe it's possible to be clear about that distinction: "Duck Arrays" are the Python API Maybe a C-API object would be useful, that shares the memory layout, but could have completely different functionality at the Python level. - CHB -- Christopher Barker, Ph.D. Oceanographer Emergency Response Division NOAA/NOS/OR&R (206) 526-6959 voice 7600 Sand Point Way NE (206) 526-6329 fax Seattle, WA 98115 (206) 526-6317 main reception Chris.Barker@noaa.gov
![](https://secure.gravatar.com/avatar/5e44e61af952612c3c45385b62806c67.jpg?s=120&d=mm&r=g)
Would the extended dtypes also violate the Liskov substitution principle? In place operations which would mutate the dtype are one potential issue. Would a single dtype for an array be sufficient, i.e. np.polynomial coefficients? Compared to ndarray subclasses, the memory layout issue goes away, but there is still a large set of operations exposed as part of a public API with various quirks. I can imagine a new function "asunitless" scattered around downstream projects. On Tue, Oct 30, 2018 at 5:23 PM Stephan Hoyer <shoyer@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Tue, Oct 30, 2018 at 2:22 PM Stephan Hoyer <shoyer@gmail.com> wrote:
I don't think so. Dtypes have nothing to do with a whole set of use cases that add extra methods or attributes. Random made-up example: user has a system with 1000 sensor signals, some of which should be treated with robust statistics for <reasons like unreliable hardware>. So user writes a subclass robust_ndarray, adds a bunch of methods like median/iqr/mad, and uses isinstance checks in functions that accept both ndarray and robust_ndarray to figure out how to preprocess sensor signals. Of course you can do everything you can do with subclasses also in other ways, but such "let's add some methods or attributes" are much more common (I think, hard to prove) than "let's change how indexing or multiplication works" in end user code. Cheers, Ralf
![](https://secure.gravatar.com/avatar/5e44e61af952612c3c45385b62806c67.jpg?s=120&d=mm&r=g)
The build on Ralf's thought, a common subclass use case would be to add logging to various methods and attributes. That might actually be useful for ndarray for understanding what is under the hood of some function in a downstream project. It would satisfy SOLID and not be related at all to dtype subclasses. On Wed, Oct 31, 2018 at 8:28 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/851ff10fbb1363b7d6111ac60194cc1c.jpg?s=120&d=mm&r=g)
The substitution principle is interesting (and, being trained as an astronomer, not a computer scientist, I had not heard of it before). I think matrix is indeed obviously wrong here (with indexing being more annoying, but multiplication being a good example as well). Perhaps more interesting as an example to consider is MaskedArray, which is much closer to a sensible subclass, though different from Quantity in that what is masked can itself be an ndarray subclass. In a sense, it is more of a container class, in which the operations are done on what is inside it, with some care taken about which elements are fixed. This becomes quite clear when one thinks of implementing __array_ufunc__ or __array_function__: for Quantity, calling super after dealing with the units is very logical, for MaskedArray, it makes more sense to call the (universal) function again on the contents [1]. For this particular class, if reimplemented, it might make most sense as a "mixin" since its attributes depend both on the masked class (.mask, etc.) and on what is being masked (say, .unit for a quantity). Thus, the final class might be an auto-generated new class (e.g., MaskedQuantity(MaskedArray, Quantity)). We have just added a new Distribution class to astropy which is based on this idea [2] (since this uses casting from structured dtypes which hold the samples to real arrays on which functions are evaluated, this probably could be done just as well or better with more flexible dtypes, but we have to deal with what's available in the real world, not the ideal one...). -- Marten [1] http://www.numpy.org/neps/nep-0013-ufunc-overrides.html#subclass-hierarchies [2] https://github.com/astropy/astropy/pull/6945
![](https://secure.gravatar.com/avatar/1198e2d145718c841565712312e04227.jpg?s=120&d=mm&r=g)
Hi all
On Friday, Oct 19, 2018 at 10:28 AM, Matti Picus <matti.picus@gmail.com (mailto:matti.picus@gmail.com)> wrote:
Was there discussion around which of `asarray` or asanyarray` to prefer? PR 11162, https://github.com/numpy/numpy/pull/11162, proposes `asanyarray` in place of `asarray` at the entrance to `_quantile_ureduce_func` to preserve ndarray subclasses. Should we be looking into changing all the `asarray` calls into `asanyarray`?
I suspect that this will cause a large number of problems around np.matrix, so unless we deprecate that, this might cause a large amount of problems. The problem with np.matrix is that it’s a subclass, but it’s not substitutable for the base class, and so violates SOLID. There are efforts to remove np.matrix, with the largest consumer being scipy.sparse, so unless that’s revamped, deprecating np.matrix is kind of hard to do.
Best Regards, Hameer Abbasi
![](https://secure.gravatar.com/avatar/851ff10fbb1363b7d6111ac60194cc1c.jpg?s=120&d=mm&r=g)
There are exceptions for `matrix` in quite a few places, and there now is warning for `maxtrix` - it might not be bad to use `asanyarray` and add an exception for `maxtrix`. Indeed, I quite like the suggestion by Eric Wieser to just add the exception to `asanyarray` itself - that way when matrix is truly deprecated, it will be a very easy change. -- Marten
![](https://secure.gravatar.com/avatar/93a76a800ef6c5919baa8ba91120ee98.jpg?s=120&d=mm&r=g)
I don't think it makes much sense to change NumPy's existing usage of asarray() to asanyarray() unless we add subok=True arguments (which default to False). But this ends up cluttering NumPy's public API, which is also undesirable. The preferred way to override NumPy functions going forward should be __array_function__. On Fri, Oct 19, 2018 at 8:13 AM Marten van Kerkwijk < m.h.vankerkwijk@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/1198e2d145718c841565712312e04227.jpg?s=120&d=mm&r=g)
Hi!
On Friday, Oct 19, 2018 at 6:09 PM, Stephan Hoyer <shoyer@gmail.com (mailto:shoyer@gmail.com)> wrote: I don't think it makes much sense to change NumPy's existing usage of asarray() to asanyarray() unless we add subok=True arguments (which default to False). But this ends up cluttering NumPy's public API, which is also undesirable.
Agreed so far.
The preferred way to override NumPy functions going forward should be __array_function__.
I think we should “soft support” i.e. allow but consider unsupported, the case where one of NumPy’s functions is implemented in terms of others and “passing through” an array results in the correct behaviour for that array.
Best Regards, Hameer Abbasi
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 4:15 PM Hameer Abbasi <einstein.edison@gmail.com> wrote:
I'm not sure I agree. "subok" is very unpythonic; the average numpy library function should work fine for a well-behaved subclass (i.e. most things out there except np.matrix).
I don't think we have or want such a concept as "soft support". We intend to not break anything that now has asanyarray, i.e. it's supported and ideally we have regression tests for all such functions. For anything we transition over from asarray to asanyarray, PRs should come with new tests.
then may as well just rip np.matrix out straight away. What I suggested in the call about this issue is that it's not very effective to treat functions like percentile/quantile one by one without an overarching strategy. A way forward could be for someone to write an overview of which sets of functions now have asanyarray (and actually work with subclasses), which ones we can and want to change now, and which ones we can and want to change after np.matrix is gone. Also, some guidelines for new functions that we add to numpy would be handy. I suspect we've been adding new functions that use asarray rather than asanyarray, which is probably undesired. Cheers, Ralf
![](https://secure.gravatar.com/avatar/97c543aca1ac7bbcfb5279d0300c8330.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 3:28 PM, Ralf Gommers <ralf.gommers@gmail.com> wrote:
Masked arrays also tend to break code that's not expecting them (e.g. on a masked array, arr.sum()/arr.size will silently compute some meaningless nonsense instead of the mean, and there are lots of formulas out there that have some similarities with 'mean'). And people do all kinds of weird things in third-party array subclasses. Obviously we can't remove asanyarray or break existing code that assumes particular numpy functions use asanyarray, but fundamentally asanyarray is just not an API that makes sense or can be supported in a general way, and our overall goal is to get people to gradually transition away from using ndarray subclasses in general. That's why we're doing all this work to make duck arrays work. So extending asanyarray support doesn't seem like a good priority to spend our limited resources on, to me. -n -- Nathaniel J. Smith -- https://vorpus.org <http://vorpus.org>
![](https://secure.gravatar.com/avatar/93a76a800ef6c5919baa8ba91120ee98.jpg?s=120&d=mm&r=g)
The problem with asanyarray() is that there isn't any well defined subclass API for NumPy, beyond "mostly works like a NumPy array." If every NumPy subclass strictly obeyed the Liskov Substitution Principle asanyarray() would be fine, but in practice every subclass I've encountered deviates from the behavior of numpy.ndarray in some way. The means the NumPy codebase has ended up littered with hacks/workarounds to support various specific subclasses, and new subclasses still don't work reliably. This makes it challenging to change existing code. For an example of how bad this is gotten, look at all the work-arounds I had to add to support np.testing.assert_array_equal() on ndarray subclasses in this recent PR: https://github.com/numpy/numpy/pull/12119 My hope is that __array_function__ will finally let us put a stop to this by offering a better alternative to subclassing.
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 10:28 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
Thanks Nathaniel and Stephan. Your comments on my other two points are both clear and correct (and have been made a number of times before). I think the "write an overview so we can stop making ad-hoc decisions and having these discussions" is the most important point I was trying to make though. If we had such a doc and it concluded "hence we don't change anything, __array_function__ is the only way to go" then we can just close PRs like https://github.com/numpy/numpy/pull/11162 straight away. Cheers, Ralf
![](https://secure.gravatar.com/avatar/851ff10fbb1363b7d6111ac60194cc1c.jpg?s=120&d=mm&r=g)
Hi All, It seems there are two extreme possibilities for general functions: 1. Put `asarray` everywhere. The main benefit that I can see is that even if people put in list instead of arrays, one is guaranteed to have shape, dtype, etc. But it seems a bit like calling `int` on everything that might get used as an index, instead of letting the actual indexing do the proper thing and call `__index__`. 2. Do not coerce at all, but rather write code assuming something is an array already. This will often, but not always, just work for array mimics, with coercion done only where necessary (e.g., in lower-lying C code such as that of the ufuncs which has a smaller API surface and can be overridden more easily). The current __array_function__ work may well provide us with a way to combine both, if we (over time) move the coercion inside `ndarray.__array_function__` so that the actual implementation *can* assume it deals with pure ndarray - then, when relevant, calling that implementation will be what subclasses/duck arrays can happily do (and it is up to them to ensure this works). Of course, the above does not really answer what to do in the meantime. But perhaps it helps in thinking of what we are actually aiming for. One last thing: could we please stop bashing subclasses? One can subclass essentially everything in python, often to great advantage. Subclasses such as MaskedArray and, yes, Quantity, are widely used, and if they cause problems perhaps that should be seen as a sign that ndarray subclassing should be made easier and clearer. All the best, Marten On Fri, Oct 19, 2018 at 7:02 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/209654202cde8ec709dee0a4d23c717d.jpg?s=120&d=mm&r=g)
Subclasses such as MaskedArray and, yes, Quantity, are widely used, and if they cause problems perhaps that should be seen as a sign that ndarray subclassing should be made easier and clearer. Both maskedarray and quantity seem like something that would make more sense at the dtype level if our dtype system was easier to extend. It might be good to compile a list of subclassing applications, and split them into “this ought to be a dtype” and “this ought to be a different type of container”. On Fri, 19 Oct 2018 at 18:24 Marten van Kerkwijk <m.h.vankerkwijk@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/96dd777e397ab128fedab46af97a3a4a.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 7:50 PM Eric Wieser <wieser.eric+numpy@gmail.com> wrote:
Wes Mckinney has been benchmarking masks vs sentinel values for arrow: http://wesmckinney.com/blog/bitmaps-vs-sentinel-values/. The (bit) masks are faster. I'm not convinced dtypes are the way to go. Chuck
![](https://secure.gravatar.com/avatar/97c543aca1ac7bbcfb5279d0300c8330.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 7:00 PM, Charles R Harris <charlesr.harris@gmail.com> wrote:
We need to add better support for both user-defined dtypes and for user-defined containers in any case. So we're going to support both missing value strategies regardless, and people will be able to choose based on engineering trade-offs. A missing value dtype is going to integrate much more easily into the rest of numpy than a new container where you have to reimplement indexing etc., but maybe custom containers can be faster. Okay, cool, they're both on PyPI, pick your favorite! Trying to wedge masks into *ndarray* seems like a non-starter, though, because it would require auditing and updating basically all code using the numpy C API. -n -- Nathaniel J. Smith -- https://vorpus.org
![](https://secure.gravatar.com/avatar/97c543aca1ac7bbcfb5279d0300c8330.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 6:23 PM, Marten van Kerkwijk <m.h.vankerkwijk@gmail.com> wrote:
Between these two options, Numpy's APIs are very firmly on the side of "option 1", and this is common in most public APIs I'm familiar with (e.g. scipy). I guess you could try to reopen the discussion, but you'd be pushing against 15+ years of precedent there...
We need some kind of asduckarray(), that coerces lists and similar but allows duck-arrays to pass through.
Who's bashing? I've spent years thinking about this and come to the conclusion that there are no viable solutions to the problems with subclassing ndarray, but that's not the same as bashing :-). If you've thought of something we've missed, you should share it... (I also know lots of senior Python devs who believe that using Python's subclassing support is pretty much always a mistake – this talk is popularly cited: https://www.youtube.com/watch?v=3MNVP9-hglc – but the issues with ndarray are much more severe than for the average Python class.) -n -- Nathaniel J. Smith -- https://vorpus.org
![](https://secure.gravatar.com/avatar/08bcb178fb791b7c8a5e9ad8d7121bed.jpg?s=120&d=mm&r=g)
On Fri, Oct 19, 2018 at 8:24 PM Marten van Kerkwijk < m.h.vankerkwijk@gmail.com> wrote:
Yes, actually getting a proper "array protocol" into Python would be a fantastic approach. We have been working with Lenore Mullin who is a researcher on the mathematics of arrays on what it means to be an array and believe we can come up with an actual array protocol that perhaps could be put into Python itself (though that isn't our immediate goal right now).
Also, we could get rid of asarray entirely by changing expectations. This automatic conversion code throughout NumPy and SciPy is an example of the confusion in both of these libraries between "user-oriented interfaces" and "developer-oriented interfaces". A developer just wants the library to use duck-typing and then raise errors if you don't provide the right type (i.e. a list instead of an array). The user-interface could happen in Jupyter, or be isolated to a high-level library or meta-code approach (of which there are several possibilities for Python).
Of course, the above does not really answer what to do in the meantime. But perhaps it helps in thinking of what we are actually aiming for.
I agree that we can stop bashing subclasses in general. The problem with numpy subclasses is that they were made without adherence to SOLID: https://en.wikipedia.org/wiki/SOLID. In particular the Liskov substitution principle: https://en.wikipedia.org/wiki/Liskov_substitution_principle . Much of this is my fault. Being a scientist/engineer more than a computer scientist, I had no idea what these principles were and did not properly apply them in creating np.matrix which clearly violates the substitution principle. We can clean all this and more up. But, we really need to start talking about NumPy 2.0 to do it. Now that Python 3.x is really here, we can raise the money for it and get it done. We don't have to just rely on volunteer time. The world will thank us for actually pushing NumPy 2.0. I know not everyone agrees, but for whatever its worth, I feel very, very strongly about this, and despite not being very active on this list for the past years, I do have a lot of understanding about how the current code actually works (and where and why its warts are). -Travis
![](https://secure.gravatar.com/avatar/5dde29b54a3f1b76b2541d0a4a9b232c.jpg?s=120&d=mm&r=g)
On Fri, Oct 26, 2018 at 7:12 PM, Travis Oliphant <teoliphant@gmail.com> wrote:
...
did not properly apply them in creating np.matrix which clearly violates the substitution principle.
So -- could a matrix subclass be made "properly"? or is that an example of something that should not have been a subclass? -CHB -- Christopher Barker, Ph.D. Oceanographer Emergency Response Division NOAA/NOS/OR&R (206) 526-6959 voice 7600 Sand Point Way NE (206) 526-6329 fax Seattle, WA 98115 (206) 526-6317 main reception Chris.Barker@noaa.gov
![](https://secure.gravatar.com/avatar/209654202cde8ec709dee0a4d23c717d.jpg?s=120&d=mm&r=g)
The latter - changing the behavior of multiplication breaks the principle. But this is not the main reason for deprecating matrix - almost all of the problems I’ve seen have been caused by the way that matrices behave when sliced. The way that m[i][j] and m[i,j] are different is just one example of this, the fact that they must be 2d is another. Matrices behaving differently on multiplication isn’t super different in my mind to how string arrays fail to multiply at all. Eric On Mon, 29 Oct 2018 at 20:54 Ralf Gommers <ralf.gommers@gmail.com> wrote: On Mon, Oct 29, 2018 at 4:31 PM Chris Barker <chris.barker@noaa.gov> wrote:
![](https://secure.gravatar.com/avatar/79653c343e58afbfbedcfd11e7b668d7.jpg?s=120&d=mm&r=g)
On Tue, Oct 30, 2018 at 12:49 AM Eric Wieser <wieser.eric+numpy@gmail.com> wrote:
The difference is that string arrays are not numeric. This is an issue since people want to pass a matrix Into places that want to multiple element wise but that then breaks that code unless special provisions are taken. Numerical codes don’t work on string arrays anyway. Eric Eric
![](https://secure.gravatar.com/avatar/93a76a800ef6c5919baa8ba91120ee98.jpg?s=120&d=mm&r=g)
On Mon, Oct 29, 2018 at 9:49 PM Eric Wieser <wieser.eric+numpy@gmail.com> wrote:
It's certainly fine for arithmetic to work differently on an element-wise basis or even to error. But np.matrix changes the shape of results from various ndarray operations (e.g., both multiplication and indexing), which is more than any dtype can do. The Liskov substitution principle (LSP) suggests that the set of reasonable ndarray subclasses are exactly those that could also in principle correspond to a new dtype. Of np.ndarray subclasses in wide-spread use, I think only the various "array with units" types come close satisfying this criteria. They only fall short insofar as they present a misleading dtype (without unit information). The main problem with subclassing for numpy.ndarray is that it guarantees too much: a large set of operations/methods along with a specific memory layout exposed as part of its public API. Worse, ndarray itself is a little quirky (e.g., with indexing, and its handling of scalars vs. 0d arrays). In practice, it's basically impossible to layer on complex behavior with these exact semantics, so only extremely minimal ndarray subclasses don't violate LSP. Once we have more easily extended dtypes, I suspect most of the good use cases for subclassing will have gone away.
![](https://secure.gravatar.com/avatar/5dde29b54a3f1b76b2541d0a4a9b232c.jpg?s=120&d=mm&r=g)
On Tue, Oct 30, 2018 at 2:22 PM, Stephan Hoyer <shoyer@gmail.com> wrote:
How about subclasses that only add functionality? My only use case of subclassing is exactly that: I have a "bounding box" object (probably could have been called a rectangle) that is a subclass of ndarray, is always shape (2,2), and has various methods for merging two such boxes, etc, adding a point, etc. I did it that way, 'cause I had a lot of code already that simply used a (2,2) array to represent a bounding box, and I wanted all that code to still work. I have had zero problems with it. Maybe that's too trivial to be worth talking about, but this kind of use case can be handy. It is a bit awkward to write the code, though -- it would be nice to have a cleaner API for this sort of subclassing (not that I have any idea how to do that) The main problem with subclassing for numpy.ndarray is that it guarantees
too much: a large set of operations/methods along with a specific memory layout exposed as part of its public API.
This is a big deal -- we really have two concepts here: - a Python class (type) with certain behaviors in Python code - a wrapper around a strided memory block. maybe it's possible to be clear about that distinction: "Duck Arrays" are the Python API Maybe a C-API object would be useful, that shares the memory layout, but could have completely different functionality at the Python level. - CHB -- Christopher Barker, Ph.D. Oceanographer Emergency Response Division NOAA/NOS/OR&R (206) 526-6959 voice 7600 Sand Point Way NE (206) 526-6329 fax Seattle, WA 98115 (206) 526-6317 main reception Chris.Barker@noaa.gov
![](https://secure.gravatar.com/avatar/5e44e61af952612c3c45385b62806c67.jpg?s=120&d=mm&r=g)
Would the extended dtypes also violate the Liskov substitution principle? In place operations which would mutate the dtype are one potential issue. Would a single dtype for an array be sufficient, i.e. np.polynomial coefficients? Compared to ndarray subclasses, the memory layout issue goes away, but there is still a large set of operations exposed as part of a public API with various quirks. I can imagine a new function "asunitless" scattered around downstream projects. On Tue, Oct 30, 2018 at 5:23 PM Stephan Hoyer <shoyer@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/5f88830d19f9c83e2ddfd913496c5025.jpg?s=120&d=mm&r=g)
On Tue, Oct 30, 2018 at 2:22 PM Stephan Hoyer <shoyer@gmail.com> wrote:
I don't think so. Dtypes have nothing to do with a whole set of use cases that add extra methods or attributes. Random made-up example: user has a system with 1000 sensor signals, some of which should be treated with robust statistics for <reasons like unreliable hardware>. So user writes a subclass robust_ndarray, adds a bunch of methods like median/iqr/mad, and uses isinstance checks in functions that accept both ndarray and robust_ndarray to figure out how to preprocess sensor signals. Of course you can do everything you can do with subclasses also in other ways, but such "let's add some methods or attributes" are much more common (I think, hard to prove) than "let's change how indexing or multiplication works" in end user code. Cheers, Ralf
![](https://secure.gravatar.com/avatar/5e44e61af952612c3c45385b62806c67.jpg?s=120&d=mm&r=g)
The build on Ralf's thought, a common subclass use case would be to add logging to various methods and attributes. That might actually be useful for ndarray for understanding what is under the hood of some function in a downstream project. It would satisfy SOLID and not be related at all to dtype subclasses. On Wed, Oct 31, 2018 at 8:28 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:
![](https://secure.gravatar.com/avatar/851ff10fbb1363b7d6111ac60194cc1c.jpg?s=120&d=mm&r=g)
The substitution principle is interesting (and, being trained as an astronomer, not a computer scientist, I had not heard of it before). I think matrix is indeed obviously wrong here (with indexing being more annoying, but multiplication being a good example as well). Perhaps more interesting as an example to consider is MaskedArray, which is much closer to a sensible subclass, though different from Quantity in that what is masked can itself be an ndarray subclass. In a sense, it is more of a container class, in which the operations are done on what is inside it, with some care taken about which elements are fixed. This becomes quite clear when one thinks of implementing __array_ufunc__ or __array_function__: for Quantity, calling super after dealing with the units is very logical, for MaskedArray, it makes more sense to call the (universal) function again on the contents [1]. For this particular class, if reimplemented, it might make most sense as a "mixin" since its attributes depend both on the masked class (.mask, etc.) and on what is being masked (say, .unit for a quantity). Thus, the final class might be an auto-generated new class (e.g., MaskedQuantity(MaskedArray, Quantity)). We have just added a new Distribution class to astropy which is based on this idea [2] (since this uses casting from structured dtypes which hold the samples to real arrays on which functions are evaluated, this probably could be done just as well or better with more flexible dtypes, but we have to deal with what's available in the real world, not the ideal one...). -- Marten [1] http://www.numpy.org/neps/nep-0013-ufunc-overrides.html#subclass-hierarchies [2] https://github.com/astropy/astropy/pull/6945
participants (12)
-
Charles R Harris
-
Chris Barker
-
Eric Moore
-
Eric Wieser
-
Hameer Abbasi
-
Marten van Kerkwijk
-
Matthew Harrigan
-
Matti Picus
-
Nathaniel Smith
-
Ralf Gommers
-
Stephan Hoyer
-
Travis Oliphant