Nicer structured-dtype constructor

[anntzer]

It would be nice if a structured-stype constructor taking **kwargs was available, ie.

np.dtype(foo=int, bar=(float, 2))

being equivalent to

np.dtype([("foo", int), ("bar", float, 2)])

the latter being quite a mouthful to type (I realize that this can be made into a one-line helper function too..).

It is true that further overloading the (already very overloaded) np.dtype constructor results in an ambiguity, because the following is currently legal:

In [1]: np.dtype(dtype=np.uint16)
Out[1]: dtype('uint16')

(even though the docs say that the first argument is named "obj" -- but it's actually named "dtype"...)

However, I think it would make sense to make the first argument (when used in that semantic) positional only, so that (after some deprecation period) np.dtype(dtype=np.uint16) means np.dtype([("dtype", np.uint16)]).
Obviously, if there are no fields named "dtype", there is no ambiguity even as of now.

Numpy/Python version information:

1.16.3 3.7.3 (default, Mar 27 2019, 22:11:17)
[GCC 7.3.0]

[seberg]

Sounds like something to think about when/if we split up the dtypes into a true type hierarchy, so that it might be np.structured(...) or similar.

[eric-wieser]

I would be very against adding more overloads to the dtype constructor, it has too many overloads already. A different helper function or static factory function might be fine though. Elsewhere I show how to declare structured types using class syntax

Note also that your proposed syntax results in a non-deterministic field order until python 3.6

[anntzer]

That's definitely intended for Py3.6+ (I'm fine with waiting until numpy only supports Py3.6+, it's not really urgent...).
Making this a separate function is fine too.