Magic factory

[tlambert03]

As discussed on zulip, this adds a @magic_factory decorator that, instead of returning a FunctionGui, returns a MagicFactory instance (basically just a partial subclass that we can type check) which will return the actual FunctionGui instance when called.

still need to add tests and take another pass through it, but this can be evaluated if you want @jni

update (1/21)

Here is where the API has settled and what it can do:

from magicgui import magicgui, magic_factory
from magicgui.widgets import FunctionGui, Slider, SpinBox
from magicgui._magicgui import MagicFactory

# the magicgui decorator creates a single widget instance
@magicgui
def func():
    pass

assert isinstance(func, FunctionGui)

# the magic_factory decorator creates a factory function
# that must be called to produce the widget

@magic_factory
def factory(x: int):
    pass

assert isinstance(factory, MagicFactory)
widget = factory()
assert isinstance(widget, FunctionGui)
assert isinstance(widget.x, SpinBox)

# While the `MagicFactory` will remember the kwargs provided
# to the `magic_factory` decorator, you can also override them
# individually when creating the widget

widget2 = factory(x={'widget_type': 'Slider'})
assert isinstance(widget2.x, Slider)

self-referencing also works with the @magic_factory decorator (meaning that inside the body of the function, the name of the function will be pointing to the ultimate widget instance)... with the main caveat being that this only works when decorating a module-level function:

# this works

@magic_factory
def factory(x: int):
    assert isinstance(factory, FunctionGui)
    print(factory)

widget = factory()
widget()  # prints "<FunctionGui factory(x: int = 0)>"

this, however, does not work:

def outer():
    @magic_factory
    def inner(x: int):
        assert isinstance(inner, FunctionGui)
        print(factory)
    return inner

factory = outer()
assert isinstance(factory, MagicFactory)
widget = factory()
widget()  # AssertionError

personally, I'm just fine with this caveat.

[codecov]

Codecov Report

Merging #117 (7c8f17b) into master (813e9a7) will increase coverage by 0.08%.
The diff coverage is 93.90%.

@@            Coverage Diff             @@
##           master     #117      +/-   ##
==========================================
+ Coverage   90.23%   90.31%   +0.08%     
==========================================
  Files          25       25              
  Lines        2416     2437      +21     
==========================================
+ Hits         2180     2201      +21     
  Misses        236      236              

Impacted Files	Coverage Δ
magicgui/widgets/_function_gui.py	96.18% <88.88%> (+0.91%)	⬆️
magicgui/_magicgui.py	96.61% <96.36%> (-3.39%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 813e9a7...d3a6f4d. Read the comment docs.

[jni]

Ha, the failing tests in 3.7:

  E     File "/home/runner/work/magicgui/magicgui/magicgui/_magicgui.py", line 27
  E       def __call__(self, /, *args, **keywords) -> FunctionGui:
  E                          ^
  E   SyntaxError: invalid syntax

You're just too damn clever, @tlambert03! 😂 At any rate that seems like a super unnecessary use of /?

[jni]

@tlambert03 as I started pondering about this in my head, I had the dreadful feeling that it wouldn't work, and indeed that feeling has been realised:

    mode = start_affinder._call_button.text  # can be "Start" or "Finish"
AttributeError: 'MagicFactory' object has no attribute '_call_button'

All of the "reflexive" tricks we do no longer work... Not sure how to deal with it though. =\

[tlambert03]

You're just too damn clever, @tlambert03! 😂 At any rate that seems like a super unnecessary use of /?

Hah, good point, I actually just copied that straight from functools.partial.__call__ but forgot about that!

[tlambert03]

All of the "reflexive" tricks we do no longer work... Not sure how to deal with it though. =\

Ohhhh nooo :( forgot about that stuff. My first thought is that maybe we try to use a modified locals() in magic_factory, that forwards just the name of the function? (But haven't tried to actually do that)

[tlambert03]

ok, @jni, I think I've got a solution I'm happy with, read the update in the main post for details... the main point is that self-reference will now work fine for factories decorated at the module level, but not for factories decorated in local scopes. (which is fine with me... I'm not going to bother at the moment detecting that probably-rare scenario to raise a warning).

If you want to see the guts of how I did that see the context manager used here and defined here

[tlambert03]

@brisvag, you may also want to look at the API in the updated first post. Let me know if you have suggestions or questions.

[brisvag]

That looks great! Overriding kwargs is perfect 👍

What I would like to be able to do, and tried to (poorly) describe on zulip was something a bit more complex: accepting arbitrary amounts and combinations of kwargs at runtime. After thinking about it a bit more, I think it's a very different thing from what's happening here, and might be entirely different from the goal of magicgui. Or it might be already possible, and I'm thinking about it the wrong way.
Here's some mockup of what I envisioned:

@magicgui
def factory(**kwargs: Union[int, float]):
    for k, v in kwargs.items():
        if isinstance(v, int):
            # do something
        if isinstance(v, float):
            # do something else

factory(x=1, y=2.0)

And this would generate a widget equivalent to:

@magicgui
def factory(x: int, y: float):
    # stuff

[tlambert03]

(chatted with @brisvag on zulip and have a better idea of how we might help there... probably not the factory)

[tlambert03]

I'm not going to bother at the moment detecting that probably-rare scenario to raise a warning

oh I bothered... got a nice warning now :)

[jni]

In what situations does function=None make sense? Is this that weird trick that could be avoided with toolz.curry? =)

[tlambert03]

yeah, we've had this discussion before! 😂 ... but I don't want to depend on toolz

[tlambert03]

... and function=None makes sense when you provide parameters to the decorator:

@magicgui(call_button=True)  # magicgui is being called with function=None
def function():
    ...

[jni]

Fine. 🤷 I think the toolz logic ends up so much cleaner, but I'll accept this for now. 😂

[tlambert03]

i agree. looks much better that way.

[jni]

This is just so insane. 😄

[jni]

Indeed, I would suggest adding a comment above the freevars = line to explain wth you are doing here. I would have exactly zero idea if I hadn't been the cause of this thing. 😂

[tlambert03]

??

[jni]

Q for my own education: does Literal work if it's not actually a literal? ie will this code pass type checking?

main_win = True
f = magic_factory(func, main_window=main_win)()
f.<something_main_function_gui_provides>()

?

[jni]

Or even worse, something where main_win is determined from external input...

[tlambert03]

yep, that works

[jni]

Wowza.

[jni]

@tlambert03 all of my comments are for clarification, I'll let you use your judgement regarding which/how to address. Amazing PR. 😀