Picard library modifies sklearn FastICA causing it to reject a default value for the fun parameter

[lmmx]

Initially mis-reported as a scikit-learn bug:

• [Solved] Picard library modifies sklearn FastICA causing it to reject a default value for the fun parameter scikit-learn/scikit-learn#32929

The following code errored while benchmarking and I couldn’t tell why: it turned out to be because Picard is overriding the underlying FastICA options(?)

picard/picard/dropin_sklearn.py

Lines 108 to 112 in bbeba2a

	# update parameters constraint dict
	self._parameter_constraints["fun"] = [
	StrOptions({"tanh", "exp", "cube"}),
	callable,
	]

import numpy as np
from picard import Picard
from sklearn.decomposition import FastICA

X = np.random.randn(1000, 384)

ica = Picard(n_components=10, random_state=42)
ica.fit_transform(X)

ica_sklearn = FastICA(n_components=10, random_state=42)
ica_sklearn.fit_transform(X)

  File "/home/louis/dev/arxiv_explorer/benchmark_ica_impls.py", line 84, in <module>                                                                                                                                                                          
    S_sklearn = ica_sklearn.fit_transform(X_whitened2)                                                                                                                                                                                                        
  File "/home/louis/dev/arxiv_explorer/.venv/lib/python3.13/site-packages/sklearn/utils/_set_output.py", line 316, in wrapped                                                                                                                                 
    data_to_wrap = f(self, X, *args, **kwargs)                                                                                                                                                                                                                
  File "/home/louis/dev/arxiv_explorer/.venv/lib/python3.13/site-packages/sklearn/base.py", line 1329, in wrapper                                                                                                                                             
    estimator._validate_params()
    ~~~~~~~~~~~~~~~~~~~~~~~~~~^^
  File "/home/louis/dev/arxiv_explorer/.venv/lib/python3.13/site-packages/sklearn/base.py", line 492, in _validate_params
    validate_parameter_constraints(      
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^                         
        self._parameter_constraints,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^                                                                                           
        self.get_params(deep=False),                        
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        caller_name=self.__class__.__name__,                
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )                      
    ^                      
  File "/home/louis/dev/arxiv_explorer/.venv/lib/python3.13/site-packages/sklearn/utils/_param_validation.py", line 98, in validate_parameter_constraints
    raise InvalidParameterError(
    ...<2 lines>...                                         
    )                          
sklearn.utils._param_validation.InvalidParameterError: The 'fun' parameter of FastICA must be a str among {'exp', 'cube', 'tanh'} or a callable. Got 'logcosh' instead.

Note that in this code the fun parameter is not passed, the FastICA is rendered unusable because of the modification Picard executes.

[agramfort]

fixed by #70