__init__ arguments are being replaced by *args, **kwargs

[polakowo]

The following code

import plotly.graph_objects as go

class FigureWidget(go.FigureWidget):
    def __init__(self):
        """Documentation."""
        pass

Produces the following documentation

Although I clearly defined that the __init__ method should have no arguments, it shows *args and **kwargs in the signature. Are there any caveats in subclassing third-party classes? Also any ideas how to unveil the "real" signature?

[kernc]

pdoc delegates to inspect.signature():

pdoc/pdoc/__init__.py

Line 1214 in 0042043

signature = inspect.signature(doc_obj.obj)

Which, I think, apparently, prefers __new__().

>>> import plotly.graph_objects as go
                                                     
>>> class FigureWidget(go.FigureWidget): 
...     def __init__(self): 
...         """Documentation.""" 
...         pass                                                                              

>>> inspect.signature(FigureWidget)                                                           
<Signature (*args, **kwargs)>

>>> print(inspect.getsource(go.FigureWidget.__new__))                                         
    def __new__(cls, *args, **kwargs):
        # This is needed because object.__new__ only accepts
        # the cls argument.
        ...

You're experiencing a somewhat border case, and a likely violation of Liskov substitution principle, so I'm leaning to wontfix.

Try if overriding __new__() works for you.

[kernc]

Work around by overriding __new__; should work.

[RaduG]

I want to re-raise this as it is quite an annoying problem especially when using Generic. Take this code:

class A(Generic[T]):
    def __init__(self, a, b):
        ...

Because typing.Generic overrides __new, inspect.signature will always return its signature which clearly is not preferable. What we did to overcome this is make a change in Function._params to use inspect.signature(doc_obj.obj.__init__) if defined which fixed the problem. I don't know if there is a genuine use case for using __new__ in docs if __init__ is defined, but would you consider having a configuration parameter to drive this? I'm happy to do the PR if so.

[kernc]

@RaduG Sure, let's try it. I've no problem skipping the parameter and making it the default, if it works. 👍