Explain why we subclass ImageQt with Any

[Avasam]

These are way too complex, with 4 different possible sources (2 deprecated)
And we don't want to force the user to install PyQt or Pyside when they may not even use it.

Fixes 6 subclassing errors (the same error in Pillow, D3DShot, python-xlib, fpdf2, PyScreeze, and PyAutoGUI stubs)

Ref: #9491

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

[hauntsaninja]

Maybe this should be a protocol?

[Avasam]

Maybe this should be a protocol?

I thought about it, but ImageQt is used as a return type, so it'd have to completely match QtImage from all 4 implementations. I did a quick try and it got messy real quick with thousands of lines copied from PyQt6.

Protocol could probably be simplified by using Any when it starts referencing other types (so at least we have all methods and members). But would a protocol still work as a return type?

[AlexWaygood]

Protocol could probably be simplified by using Any when it starts referencing other types (so at least we have all methods and members). But would a protocol still work as a return type?

Protocols can be used as a return type! I agree with @hauntsaninja that we can probably at least try to roughly describe the shape of the superclass here.

[Avasam]

I meant that I'm concerned about the variance:

from typing import Protocol

class Abc:
    foo: str
    
class AbcProtocol(Abc, Protocol): ...  # type: ignore[misc]

reveal_type(AbcProtocol().foo). # str

class Abcd:
    foo: str
   
abcd: Abcd
abcd = AbcProtocol()  # Incompatible types in assignment (
abcproto: AbcProtocol = Abcd()

[hauntsaninja]

I'm not sure I follow, wouldn't users just use the QImage type hint?
(Or is it a case where users do actually know what type they're getting back and e.g. do nominal checking or the cases where the types' APIs are non-overlapping are actually important to users?)

[Avasam]

Here's a minimal example in PyQt6. This is what we need to support (without actually having to install PyQt6, PyQt5, PySide6 or PySide2 when installing types-pillow). Would using a protocol for ImageQt's base work thanks to duck typing? And how complex would that protocol have to be? (QImages have lots of methods, members and bases).

from PyQt6 import QtGui
from PIL import Image, ImageQt
data: memoryview  # Some graphics data from elsewhere in the code
pil_image: Image.Image  # Some pillow image from elsewhere in the code

def show_qimage(image: QtGui.QImage):
    # Show the image someplace in the GUI
    pass

# With PyQt directly
qimage = QtGui.QImage(data, 640, 480, 640 * 4, QtGui.QImage.Format.Format_RGBA8888)
show_qimage(qimage)

# using the pillow wrapper instead
qimage = ImageQt.ImageQt(pil_image)
show_qimage(qimage) # Argument of type "ImageQt" cannot be assigned to parameter "image" of type "QImage" in function "show_qimage" "ImageQt" is incompatible with "QImage"

---

Would using a protocol for ImageQt's base work thanks to duck typing?

Edit: Don't think so

class A:
    def test(self, foo: str) -> int: ...


class ProtocolA(Protocol):
    def test(self, foo: str) -> int: ...


protoA = cast(ProtocolA, A())

a: A = protoA  # Expression of type "ProtocolA" cannot be assigned to declared type "A"  "ProtocolA" is incompatible with "A"

[Avasam]

@AlexWaygood , I don't see an "easy" solution for this (or any good solution at all). Since type-checkers can't have a conditional type based on if a module exists or not. (without having to check for a full ImportError, which can't be done statically, I feel like checking for a module's existence could be a beneficial proposition)

Can I move this to #9491 ?

[AlexWaygood]

@AlexWaygood , I don't see an "easy" solution for this (or any good solution at all). Since type-checkers can't have a conditional type based on if a module exists or not. (without having to check for a full ImportError, which can't be done statically, I feel like checking for a module's existence could be a beneficial proposition)

Can I move this to #9491 ?

I'm fine with that. It's not like we're introducing a regression here, after all -- we're already sbuclassing Any in the stub. If anybody wants to invest some time into working out a more principled solution, they're always welcome to do so at a later date!

[Avasam]

Closing in favor of #9491