typed-lazyimport

Provides type hinted lazy import for slow to load libraries

Usage

Simple example and principal behavior

from lazyimport import Libs
torch = Libs.torch         # does not trigger torch to be imported, but torch is still a typed variable

print(torch.__version__)   # causes torch to be imported and its version is then returned
print(torch.__version__)   # returns the same version from the already loaded library

Intended use in a real codebase

from lazyimport import Libs as L
torch = L.torch

def do_something_returning_a_tensor() -> torch.tensor:
    return torch.tensor(1)

# do various stuff that don't require torch and doesn't warrant loading it.
if condition:
    print(do_something_returning_a_tensor())  # only now do we pay the price for importing torch, since it adds value.
...

Included libraries

The following set of libraries are currently included in Libs out of the box, ready for lazy importing (provided you have any of them installed of course).

cv2 (Python OpenCV binding)
matplotlib
numpy
pandas
pytorch_lightning
sklearn
torch
torchvision

Check the source for the latest set of libraries bundled, since the above list may not always be updated in the readme when new packages are added to the Libs convenience library set.

Extending or lazily loading your own libraries

The provided Libs set is offered simply as a convenience, to offer close to zero-friction in order to gain lazy importing of the most popular libraries. However, if your specific set of libraries isn't included, you can easily create your own lazy-loaded library set.

Example:

# mylibs.py (for example)
from __future__ import annotations         # only needed if you want type hints
from typing import TYPE_CHECKING           # -"-
from lazyimport import Lib

if TYPE_CHECKING:                          # -"-. Skip this block if you don't need type hints.
    from PIL import Image
    from myown import libs

class Libraries:
    Image:Image = Lib('PIL.Image')         # the string the same format as "import" requires.
    mylibs:libs = Lib('myown.libs')

Then you can happily use it with the same ease as the bundled library set:

from mylibs import Libraries as L
flork = L.mylibs.flork
Image = L.Image

flork(Image.open('cat.png'))

All type hinted of course.

One point of note is that the type hinting comes from the explicit type annotations used when defining each class variable. Unfortunately those types are only present when the type-checker runs, and not at program runtime. This means that the nicer syntax option Image = lib(Image) isn't possible, because when the type-checker isn't running, the type Image is never imported from PIL, and the program would crash as a result. That's why we have to specify the import names using a little bit of (horrible) type-unsafe "string programming". No way around it, but it's minimized.

To ensure there's no spelling mistake, my recommendation is to first enter import <statement> in the code editor or interactive python, to verify the string is correct. Then copy the statement verbatim and quote it. That's the least risky option I found (absent a better python type system).

If you don't want/use any type hints, then creating your own lazy-loaded library set is about as DRY and concise as can be.

from lazyimport import Lib

class Libraries:
    Image = Lib('PIL.Image')
    mylibs = Lib('myown.libs')

Lazy import without library sets

If you want to do some one-off lazy imports, you can use the same proxy primitive the examples up to now have been using under the hood.

from lazyimport import Import

torch = Import('torch')      # no import is performed yet
...
print(torch.tensor(2) ** 2)  # now the import happens

The downside of this option is that if you want typing, you have to do the typing dance that PEP 484 forces you into, meaning importing the special annotation related primitives, making use of the type guard with the explicit library imports within it, then the type declarations etc..

That's the reason this example was provided last and not first, since this package is primarily aimed at offering typed lazy loading. That said, if all you need is to speed up a duck-typed CLI program or similar, there's nothing wrong with using the lazy loader (proxy) directly, as shown in this last example.

FAQ

Q1: Why?

Because many popular libraries are incredibly expensive (slow) to load. Especially ML related libraries.

Q2: Why another library?

Because none of the ones available was trivial to use, absent of magic string-programming or error-prone dict mapping.

This library provides:

Transparent type hinting for a set of popular libraries that are known to be very slow to import.
Minimal additional complexity. A single additional import, but otherwise the syntax is pretty much identical to using normal imports. In particular it allows the programmer to have module wide visibility of the imported types, just like with a normal import. It chucks away all the PEP 484 boilerplate that otherwise has to be added all across the code base, just to get some type hints. None of that needed here.

Q3: Why isn't library X included?

Most likely since I don't use it myself.

The library is first and foremost a convenience library aimed at addressing the slow import of the most popular libraries. That said, I'm open to adding additional ones. Open an issue, let's discuss and then a PR (since I probably don't have that library myself).

The cost of adding additional libraries is close to zero, so there's no practical reason for not expanding the set of libraries provided out of the box in-finitum, other than it becoming harder for users to get an overview of which libraries are made available.

If you have a special library or libraries you want to lazily load but that's not available in the Libs class, you can easily create your own library set. See Extending or lazily loading your own libraries as well as the unit-test in this project for an example that does just that. Take a look at the lazy_import.py implementation itself for how to get typing support. It's really trivial as can be seen by the minimal implementation of this package itself.

Q4: How can I see when the import happens?

Enable debug logging. lazy import: <module name> will be logged whenever a module is imported.

Development

To run the unit tests

python -m unittest

To simplify development, common actions are provided via Makefile targets:

test - default targets, runs pytest on the project
build - create a wheel package distribution, ready to be uploaded to pypi or given to someone else.
clean - removes temporary files generated as part of the package creation.

jojje/typed-lazyimport