ENH: Improve object fully qualified names.

[Carreau]

Feature

For reasons related to documentation and showing it in IPython/Jupyter I have some questions/suggestions wrt fully qualified names of object and improvement.

For example if a user does:

import <whatever> as x
x?

I want to be able to say "view online docs at https://numpy.org/doc/stable/reference/generated/{fully_qual}.html".format(full_qual(x))" with roughly:

def full_qual(x):
    return x.__module__+'.'+x.__qualname__

As well as the "inverse" operation: from a str get the object back:

parts = full_qual_name.split('.')
target = parts[0]
for k in parts[1:]:
    target = getattr(target, k)

And I have some issues with a few object, the mostly fell in 2 categories:

1. Object with no module

# full_qual(np.exp) # AttrError , no attribute __module__, (no __qualname__, but have __name__) same with sin, cos...

Could it be possible to add a module to those ?

full_qual(np.memmap) # 'numpy.memmap' is Ok
full_qual(np.memmap.flush)  # 'numpy.core.memmap.memmap.flush' is problematic:

target = numpy
for k in ['core', 'memmap', 'memmap', 'flush']:
    target = getattr(target, k) # fails with AttributeError: type object 'memmap' has no attribute 'memmap'

For this second one this is due to numpy.core.memmap being ambiguous as core/__init__.py contain from .memmap import memmap so the class shadow the module in which it's defined.

Yes I know I can access .flush via np.memmap.flush it's just one example.

Would it be ok to have a consistent way to go from objects to fully_qualified name and vice versa, and should I open issues when I find such functions/methods/class ?

[eric-wieser]

#4952 I think is most of your problem here

[seberg]

I am happy to add a module and/or qualname to ufuncs, we just need to decide on how...

Currently, the name is basically passed to the C creation function. We could do it like Python's static types and just assume everything before the first . is part of the module, that is not ideal (for one some ufuncs couldrequire a __qualname__ technically), but likely "just works" (have to check if pickle is fine with getting __module__ + "."+ __qualname__, but I expect it is).

The other option, is to make __module__ and __qualname__ writeable attribute (or writeable exactly once, with the option of disallowing it when used with future API).

---

np.memmap thing seems like the @set_module('numpy') decorator can't or doesn't properly set the module of the attributes. I guess the decorator could try to be smart about methods or is there a better solution?