FIX support scalar values in fit_params in SearchCV

doc/whats_new/v0.22.rst

@@ -35,6 +35,10 @@ Changelog
   Follow-up of :pr:`15898` by :user:`Shivam Gargsya <shivamgargsya>`.
   :pr:`15933` by :user:`Guillaume Lemaitre <glemaitre>` and `Olivier Grisel`_.
 
+- |Fix| :func:`inspection.plot_partial_dependence` and
+  :meth:`inspection.PartialDependenceDisplay.plot` now consistently checks
+  the number of axes passed in. :pr:`15760` by `Thomas Fan`_.
+
 :mod:`sklearn.metrics`
 ......................
 
@@ -50,19 +54,21 @@ Changelog
   value of the ``zero_division`` keyword argument. :pr:`15879`
   by :user:`Bibhash Chandra Mitra <Bibyutatsu>`.
 
+:mod:`sklearn.model_selection`
+..............................
+
+- |Fix| :class:`model_selection.GridSearchCV` and
+  :class:`model_selection.RandomizedSearchCV` will accept scalar provided in
+  `fit_params`. Change in 0.22 was breaking backward compatibility.
+  :pr:`15863` by :user:`Adrin Jalali <adrinjalali>` and
+  :user:`Guillaume Lemaitre <glemaitre>`.
+
 :mod:`sklearn.utils`
 ....................
 
 - |Fix| :func:`utils.check_array` now correctly converts pandas DataFrame with
   boolean columns to floats. :pr:`15797` by `Thomas Fan`_.
 
-:mod:`sklearn.inspection`
-.........................
-
-- |Fix| :func:`inspection.plot_partial_dependence` and
-  :meth:`inspection.PartialDependenceDisplay.plot` now consistently checks
-  the number of axes passed in. :pr:`15760` by `Thomas Fan`_.
-
 .. _changes_0_22:
 
 Version 0.22.0

sklearn/model_selection/_search.py

@@ -33,7 +33,7 @@
 from ..utils import check_random_state
 from ..utils.fixes import MaskedArray
 from ..utils.random import sample_without_replacement
-from ..utils.validation import indexable, check_is_fitted
+from ..utils.validation import indexable, check_is_fitted, _check_fit_params
 from ..utils.metaestimators import if_delegate_has_method
 from ..metrics._scorer import _check_multimetric_scoring
 from ..metrics import check_scoring
@@ -648,9 +648,7 @@ def fit(self, X, y=None, groups=None, **fit_params):
             refit_metric = 'score'
 
         X, y, groups = indexable(X, y, groups)
-        # make sure fit_params are sliceable
-        fit_params_values = indexable(*fit_params.values())
-        fit_params = dict(zip(fit_params.keys(), fit_params_values))
+        fit_params = _check_fit_params(fit_params)
 
         n_splits = cv.get_n_splits(X, y, groups)
 

sklearn/model_selection/tests/test_search.py

@@ -27,7 +27,7 @@
 
 from scipy.stats import bernoulli, expon, uniform
 
-from sklearn.base import BaseEstimator
+from sklearn.base import BaseEstimator, ClassifierMixin
 from sklearn.base import clone
 from sklearn.exceptions import NotFittedError
 from sklearn.datasets import make_classification
@@ -36,6 +36,7 @@
 
 from sklearn.model_selection import fit_grid_point
 from sklearn.model_selection import cross_val_score
+from sklearn.model_selection import train_test_split
 from sklearn.model_selection import KFold
 from sklearn.model_selection import StratifiedKFold
 from sklearn.model_selection import StratifiedShuffleSplit
@@ -1846,3 +1847,61 @@ def test_search_cv__pairwise_property_equivalence_of_precomputed():
 
     attr_message = "GridSearchCV not identical with precomputed metric"
     assert (preds_original == preds_precomputed).all(), attr_message
+
+
+@pytest.mark.parametrize(
+    "SearchCV, param_search",
+    [(GridSearchCV, {'a': [0.1, 0.01]}),
+     (RandomizedSearchCV, {'a': np.random.randint(1, 3, size=2)})]
+)
+def test_scalar_fit_param(SearchCV, param_search):
+    # check general support for scalar in fit_params
+    # non-regression test for:
+    # https://github.com/scikit-learn/scikit-learn/issues/15805
+    class TestEstimator(BaseEstimator, ClassifierMixin):
+        def __init__(self, a=None):
+            self.a = a
+
+        def fit(self, X, y, r):
+            assert r == 42
+
+        def predict(self, X):
+            return np.zeros(shape=(len(X)))
+
+    model = SearchCV(TestEstimator(), param_search)
+    X, y = make_classification(random_state=42)
+    model.fit(X, y, r=42)
+
+
+def _custom_lgbm_metric(y_test, y_pred):
+    # y_pred are probablities which need to be thresholded
+    y_pred = (y_pred > 0.5).astype(int)
+    acc = accuracy_score(y_test, y_pred)
+    # required output of format: (eval_name, eval_result, is_higher_better)
+    return ('accuracy', acc, True)
+
+
+@pytest.mark.parametrize("metric", ['auc', _custom_lgbm_metric])
+@pytest.mark.parametrize(
+    "SearchCV, param_search",
+    [(GridSearchCV, {'learning_rate': [0.1, 0.01]}),
+     (RandomizedSearchCV, {'learning_rate': uniform(0.01, 0.1)})]
+)
+def test_scalar_fit_param_lgbm(metric, SearchCV, param_search):
+    # check support for scalar in fit_params in LightGBM
+    # non-regression test for:
+    # https://github.com/scikit-learn/scikit-learn/issues/15805
+    lgbm = pytest.importorskip("lightgbm")
+    X_train, X_valid, y_train, y_valid = train_test_split(
+        *make_classification(random_state=42), random_state=42
+    )
+    model = SearchCV(
+        lgbm.LGBMClassifier(n_estimators=5), param_search
+    )
+    fit_params = {
+        'eval_set': [(X_valid, y_valid)],
+        'eval_metric': metric,
+        'early_stopping_rounds': 5,
+        'verbose': False
+    }
+    model.fit(X_train, y_train, **fit_params)

sklearn/utils/validation.py

@@ -16,7 +16,8 @@
 import numpy as np
 import scipy.sparse as sp
 from distutils.version import LooseVersion
-from inspect import signature, isclass, Parameter
+from inspect import signature, Parameter
+from inspect import isclass, isfunction, ismethod, ismodule
 
 from numpy.core.numeric import ComplexWarning
 import joblib
@@ -212,6 +213,26 @@ def check_consistent_length(*arrays):
                          " samples: %r" % [int(l) for l in lengths])
 
 
+def _convert_iterable(iterable):
+    """Helper convert iterable to arrays of sparse matrices.
+
+    Convert sparse matrices to csr and non-interable objects to arrays.
+    Let passes `None`.
+
+    Parameters
+    ----------
+    iterable : {list, dataframe, array, sparse} or None
+        Object to be converted to a sliceable iterable.
+    """
+    if sp.issparse(iterable):
+        return iterable.tocsr()
+    elif hasattr(iterable, "__getitem__") or hasattr(iterable, "iloc"):
+        return iterable
+    elif iterable is None:
+        return iterable
+    return np.array(iterable)
+
+
 def indexable(*iterables):
     """Make arrays indexable for cross-validation.
 
@@ -224,16 +245,7 @@ def indexable(*iterables):
     *iterables : lists, dataframes, arrays, sparse matrices
         List of objects to ensure sliceability.
     """
-    result = []
-    for X in iterables:
-        if sp.issparse(X):
-            result.append(X.tocsr())
-        elif hasattr(X, "__getitem__") or hasattr(X, "iloc"):
-            result.append(X)
-        elif X is None:
-            result.append(X)
-        else:
-            result.append(np.array(X))
+    result = [_convert_iterable(X) for X in iterables]
     check_consistent_length(*result)
     return result
 
@@ -1259,3 +1271,33 @@ def inner_f(*args, **kwargs):
         kwargs.update({k: arg for k, arg in zip(all_args, args)})
         return f(**kwargs)
     return inner_f
+
+
+def _check_fit_params(fit_params):
+    """Check and validate the parameters passed during `fit`.
+
+    Parameters
+    ----------
+    fit_params : dict
+        Dictionary containing the parameters passed at fit.
+
+    Returns
+    -------
+    fit_params_validated : dict
+        Validated parameters. We ensure that the values are iterable.
+    """
+    fit_params_validated = {}
+    for param_key, param_value in fit_params.items():
+        is_scalar = [
+            check(param_value)
+            for check in [np.isscalar, ismodule, isclass, ismethod, isfunction]
+        ]
+        if any(is_scalar):
+            # keep scalar as is for backward-compatibility
+            # https://github.com/scikit-learn/scikit-learn/issues/15805
+            fit_params_validated[param_key] = param_value
+        else:
+            # ensure iterable will be sliceable
+            fit_params_validated[param_key] = _convert_iterable(param_value)
+
+    return fit_params_validated