✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: b36c745. Link to the linter CI: here}

Yep the failure is tracked there: #29771

Not sure this is worth doing to be honest (LogisticRegression accept C=float('infinity') and has a test for it so I guess this is some kind of precedent?) ...

If we do it we may as well do it for LinearSVR and LinearSVC:

❯ git grep -p -P 'C":.+Inter'
sklearn/linear_model/_logistic.py=class LogisticRegression(LinearClassifierMixin, SparseCoefMixin, BaseEstimator):
sklearn/linear_model/_logistic.py:        "C": [Interval(Real, 0, None, closed="right")],
sklearn/linear_model/_passive_aggressive.py=class PassiveAggressiveClassifier(BaseSGDClassifier):
sklearn/linear_model/_passive_aggressive.py:        "C": [Interval(Real, 0, None, closed="right")],
sklearn/linear_model/_passive_aggressive.py=class PassiveAggressiveRegressor(BaseSGDRegressor):
sklearn/linear_model/_passive_aggressive.py:        "C": [Interval(Real, 0, None, closed="right")],
sklearn/svm/_base.py=class BaseLibSVM(BaseEstimator, metaclass=ABCMeta):
sklearn/svm/_base.py:        "C": [Interval(Real, 0.0, None, closed="neither")],
sklearn/svm/_classes.py=class LinearSVC(LinearClassifierMixin, SparseCoefMixin, BaseEstimator):
sklearn/svm/_classes.py:        "C": [Interval(Real, 0.0, None, closed="neither")],
sklearn/svm/_classes.py=class LinearSVR(RegressorMixin, LinearModel):
sklearn/svm/_classes.py:        "C": [Interval(Real, 0.0, None, closed="neither")],

I'll do for the other class.

Not sure this is worth doing to be honest

If it would not be a regression and that we never tolerated it, I would not make the change. But the fact that we introduced this small regression is a bit annoying :)

And it's good to be able to explore the unregularized limit if the underlying solver can support it. Although in practice, it should be very similar to C=1e30 or so in most cases.

@lesteve I quickly check what happen when setting C=float("inf") in LinearSVC and LinearSVR. It does seem to work: the computation never end and I had to kill the kernel, so we should be in the C code. I also make sure to reduce the number of iteration but still the same.

So this is probably not a good idea to tolerate this for both LinearSVM.