Don’t import numpy

[hoxbro]

The changes made in this PR are not to directly import non-essential imports (numpy in param) but to only check for the case when the user has already imported them.

This comes down to two things:

1. For param, this is to "hide" numpy imports in functions or generators. This is strictly not necessary, but people check import time with python -X importtime -c 'import param', and because we do an actual import of numpy, we are heavily penalized by it. The random modules (numpy and stdlib) also seem unnecessary to import, only to use the import for not outputting anything with pprint.

2. The second part involves using a new decorator, which overloads isinstance and issubclass and still makes it function like a generator with and without calling it. This approach can be used in our other libraries, e.g., HoloViews. Here, we currently import pandas to set up different class selectors, which heavily penalizes the import time. We currently require pandas, but that may not always be the case.

[codecov]

Codecov Report

Attention: Patch coverage is 92.85714% with 4 lines in your changes missing coverage. Please review.

Project coverage is 86.55%. Comparing base (0d191dc) to head (0e273cf).

Files with missing lines	Patch %	Lines
param/_utils.py	87.50%	2 Missing ⚠️
param/parameterized.py	96.29%	1 Missing ⚠️
param/parameters.py	92.30%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #966      +/-   ##
==========================================
+ Coverage   86.45%   86.55%   +0.09%     
==========================================
  Files          10       10              
  Lines        5177     5199      +22     
==========================================
+ Hits         4476     4500      +24     
+ Misses        701      699       -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[maximlt]

Could an approach like this be used to avoid creating these functions? https://peps.python.org/pep-3119/#overloading-isinstance-and-issubclass

[hoxbro]

I don't think so, though I haven't read the PEP in detail.

However, it seems pretty advanced to use for a simple conversion.

[maximlt]

It seems it'd be something like:

import sys

class GeneratorIsMeta(type):
    def __instancecheck__(cls, inst):
        return isinstance(inst, tuple(cls.gen_types()))

    def __subclasscheck__(cls, sub):
        return issubclass(sub, tuple(cls.gen_types()))


class DtTypes(metaclass=GeneratorIsMeta):
    @classmethod
    def gen_types(cls):
        yield dt.datetime
        yield dt.date
        if "numpy" in sys.modules:
            import numpy as np
            yield np.datetime64

class IntTypes(metaclass=GeneratorIsMeta):
    @classmethod
    def gen_types(cls):
        yield int
        if "numpy" in sys.modules:
            import numpy as np
            yield np.integer

[hoxbro]

Thank you.

I still thinks it is pretty advanced (even though you don't need it to be iterable in your example).

You could likely move the advanced logic into a decorator, but then you would need to import that into our other libraries. My main point with accepting iterator is not so much for param itself but for HoloViews.

[maximlt]

My main point with accepting iterator is not so much for param itself but for HoloViews.

Oh yeah I agree users shouldn't have to pass this custom class, a generator seems appropriate. Instead I was wondering if there couldn't be a way to use the metaclass approach internally to avoid the special anyisinstance and anyissubclass functions, they seem to be easy to forget in the future.

[hoxbro]

I have switched to a metaclass approach with a decorator for easy access.