Possible regression in 1.3.0?

[kmnhan]

I was exploiting the behavior in

lmfit-py/lmfit/model.py

Lines 516 to 520 in cf2626b

	# 1. limited support for asteval functions as the model functions:
	if hasattr(self.func, 'argnames') and hasattr(self.func, 'kwargs'):
	pos_args = self.func.argnames[:]
	for name, defval in self.func.kwargs:
	kw_args[name] = defval

to create a function with variable number of input parameters, by passing in a callable class with argnames and kwargs properties.
For a minimal example, I could make a polynomial function with an arbitrary degree with the following code:

import lmfit
import numpy as np


class PolynomialFunction:
    def __init__(self, degree: int = 1):
        self.degree = degree

    @property
    def __name__(self) -> str:
        return self.__class__.__name__

    @property
    def argnames(self) -> list[str]:
        return ["x"] + [f"c{i}" for i in range(self.degree + 1)]

    @property
    def kwargs(self):
        return {}

    def __call__(self, x, *coeffs, **params):
        if len(coeffs) != self.degree + 1:
            coeffs = [params[f"c{d}"] for d in range(self.degree + 1)]
        return np.polynomial.polynomial.polyval(x, coeffs)


model = lmfit.Model(PolynomialFunction(degree=2))
model.make_params()

This broke with 1.3.0 due to changes made in #941 because default_vals is not defined for the first if block. It looks like a simple fix, but I am unsure whether this is supported public API. Will this be removed entirely in the future? If so, how should I write a function/model that accomplishes the same thing?

[newville]

@kmnhan Ah, sorry for the trouble. Well, I would say both that
a) we never really promised this was part of a supported API that would never change ;)
b) but also: variadic functions are an important-if-maybe-challenging use case, and we want them to work!

As you noticed, the changes for 1.3.0 from #941 were a fairly large change to the internals of parsing and converting function arguments to parameters and independent variables. I do not expect this to be something we change very often, but we may have to make some tweaks from those changes to fix bugs. See #950, for example.

I think we could (and should) probably fix the code so your use case works too. I guess we weren't testing that portion of the code (or at least not very thoroughly!). We should be able to fold fixes for this into #950, and then push that out as 1.3.1 fairly soon.

But also: we do have a PolynomialModel builtin(lmfit.models.PolynomialModel), and also lmfit.models.SplineModel that are both variadic. Would it be reasonable to use that builtin? if not, are there changes we should make?

[kmnhan]

Thank you for the response! I will keep an eye out for the 1.3.1 release.
I've been using this approach for more complex models, but looking at the implementation of the builtin models, I think I can start refactoring my models so that they are less.. hacky. Thanks for the suggestion!