asarray array disagrees with isscalar about what is a scalar (at least for sets)

[allComputableThings]

np.asarray believes sets are scalar:

>>> np.asarray(1).shape
()   # scalar
>>> np.asarray({1,2,3}).shape
()  # scalar

np.isscalar does not:

>>> np.isscalar(1)
True
>>> np.isscalar({1,2,3})
False

I'd argues that asarray should always returns an scalar if a scalar is passed, and a non-scalar if a non-scalar is passed. (to be consistent with its treatment when a plain number is passed in).

Related to #8538 ?

[seberg]

isscalar is a pretty buggy/limited function I guess. np.ndim() == 0 would give you a better result probably (but will coerce to array as an intermediate if necessary).

Yes, that issue (and I am pretty sure one more), is basically a duplicate. isscalar uses its full own set of rules to decide what a scalar is, and that just cannot work, I guess.

[eric-wieser]

Perhaps we should try to deprecate isscalar - we have a very small number of internal uses left.

[rgommers]

That's going to be very noisy, there's a large amount of usage in SciPy, and I checked Dask, scikit-image and scikit-image, that turned up 5-20 usages per package.

So if we really want to deprecate isscalar, then I'd rather do it longer-term (say in 1-2 years) and first push out fixes to all those libraries. And given the widespread usage, removal before a NumPy 2.0 seems to be a bad idea.

[allComputableThings]

Does isscalar need to be deprecated?
It seems like a useful function -- its just that numpy doesn't have consistent rules for scalarness. My use was to decide whether to call f(x): float or g(x): floats depends on whether x was a collection. f for numbers, g for things that can be asarray.

I'd say if anything should change it might be that np.asarray(set) should return a non-scalar array.

What do others think should happen if I ask for np.asarray(set) ?

[seberg]

The careful consensus right now is probably that np.asarray(set) should just raise an error, because returning np.asarray(list(set)) may be a bit strange, but returning a 0D array is just weird/unexpected. To be honest, my gut feeling is that it might be the best way to deprecate it. Sure, make it a super long deprecation. However, we may need a good replacement before we do so, and I am not sure np.ndim(x) == 0 is one.

EDIT: To be clear in the error scenario np.asarray(set, dtype=object) would still return a 0D array probably.

[allComputableThings]

+1 I would have preferred an exception also (when passing a set).

…

On Fri, Aug 30, 2019 at 9:03 AM Sebastian Berg ***@***.***> wrote: The careful consensus right now is probably that np.asarray(set) should just raise an error, because returning np.asarray(list(set)) may be a bit strange, but returning a 0D array is just weird/unexpected. To be honest, my gut feeling is that it might be the best way to deprecate it. Sure, make it a super long deprecation. However, we may need a good replacement before we do so, and I am not sure np.ndim(x) == 0 is one. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#14396?email_source=notifications&email_token=AB3QJLALQJTFWXGCCBGLBP3QHFAEZA5CNFSM4ISDCCO2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5SCLAA#issuecomment-526656896>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB3QJLCI2I4G3ZNEACOVWDLQHFAEZANCNFSM4ISDCCOQ> .