SplineTransformer raises when passing knots explicity for n_knots=None

On master,

import numpy as np

from sklearn.preprocessing import SplineTransformer

X = np.linspace(0, 1, 10)[:, None]
splt = SplineTransformer(
   knots=[[0], [1], [2], [3], [4]],
   n_knots=None,
   degree=3
)

splt.fit(X)

currently raises

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-5-86bffa93d84e> in <module>
     10 )
     11 
---> 12 splt.fit(X)

~/code/scikit-learn/sklearn/preprocessing/_polynomial.py in fit(self, X, y)
    226             isinstance(self.n_knots, numbers.Integral) and self.n_knots >= 2
    227         ):
--> 228             raise ValueError("n_knots must be a positive integer >= 2.")
    229 
    230         if isinstance(self.knots, str) and self.knots in [

ValueError: n_knots must be a positive integer >= 2.

IMO the n_knots argument should be allowed to be None (or any other value) if the knots are passed explicity.

cc: @lorentzenchr @ogrisel xref #18368 and #19483

[lorentzenchr]

If knots in ["uniform", "quantile"], then n_knots specifies the number if knots. If knots=2d array specifying the knot positions manually, then n_knots is ignored. So they can’t be collapsed, but that would have been nice!

[mlondschien]

Currently

import numpy as np

from sklearn.preprocessing import SplineTransformer

X = np.linspace(0, 1, 10)[:, None]
splt = SplineTransformer(
   knots=[[0], [1], [2], [3], [4]],
   n_knots=2,
   degree=3
)

splt.fit(X)

works. IMO this is “more incorrect” than n_knots=None.

Raising an error that n_knots needs to be in integer when knots is passed explicitly is confusing. This suggests that the value in n_knots get’s used somewhere. I would generally pass n_knots=None explicitly to make sure that the argument does not get used.