Demo:

>>> from sklearn.svm import LinearSVC
>>> clf = LinearSVC()
>>> from sklearn.datasets import load_digits
>>> digits = load_digits()
>>> from sklearn.grid_search import GridSearchCV
>>> GridSearchCV(clf, {}, cv=2).fit(digits.data, digits.target)
GridSearchCV(cv=2,
       estimator=LinearSVC(C=1.0, class_weight=None, dual=True, fit_intercept=True,
     intercept_scaling=1, loss='l2', multi_class='ovr', penalty='l2',
     random_state=None, tol=0.0001, verbose=0),
       fit_params={}, iid=True, loss_func=None, n_jobs=1, param_grid={},
       pre_dispatch='2*n_jobs', refit=True, score_func=None, scoring=None,
       verbose=0)

That's exactly the default settings for LinearSVC:

>>> LinearSVC()
LinearSVC(C=1.0, class_weight=None, dual=True, fit_intercept=True,
     intercept_scaling=1, loss='l2', multi_class='ovr', penalty='l2',
     random_state=None, tol=0.0001, verbose=0)

Hmm. While I've had a case that would benefit from this (only because the extension of ParameterGrid to handle lists of grids, together with #1769), I think this is a bit unintuitive. Is it better to use None for this than change the semantics of dict-qua-grid?

IMHO, it hardly changes the semantics. A grid like {"C": [1, 10]} searches C but with the default value for all other parameters. It seems natural to me that an empty grid would mean the default value for all parameters. I would find an explicit None rather unintuitive.

I see your point. The number of candidates is less intuitive, but if from {'C': [1, 10], 'gamma': [1, 2]} to {'C': [1, 10]} you divide by 2, divide by two again from there to {}!

In any case, please add documentation and fix/test for __len__. (I also note test_parameter_grid doesn't cover the list of dicts case.)

And please refer to it as a "special case" rather than a "corner case", even if you find it an intuitive special case.

There. I added a commit that finally explains the behavior of ParameterGrid on lists.

LGTM, then. But I'd like another opinion on whether the semantics are reasonable. @amueller? @ogrisel?

Ping @mblondel.

I think this comment could also be removed:
https://github.com/larsmans/scikit-learn/blob/cb4b7e9740d0dbf0d470fd7d5055125b07a167bb/sklearn/grid_search.py#L82

Other than that, LGTM

I figured it would be smart to explain what the example does, but I can remove it.