DEPR deprecate n_iter in MiniBatchSparsePCA (scikit-learn#23726)

glemaitre · jeremiedbb · ogrisel · ogrisel · commit 67fb7fa80632 · 2022-07-11T15:15:25.000+02:00
Co-authored-by: Jérémie du Boisberranger &lt;34657725+jeremiedbb@users.noreply.github.com&gt;
Co-authored-by: jeremie du boisberranger &lt;jeremiedbb@yahoo.fr&gt;
Co-authored-by: Olivier Grisel &lt;olivier.grisel@ensta.org&gt;
diff --git a/doc/whats_new/v1.2.rst b/doc/whats_new/v1.2.rst
@@ -110,6 +110,16 @@ Changelog
   its memory footprint and runtime.
   :pr:`22268` by :user:`MohamedBsh <Bsh>`.
 
+:mod:`sklearn.decomposition`
+............................
+
+- |API| The `n_iter` parameter of :class:`decomposition.MiniBatchSparsePCA` is
+  deprecated and replaced by the parameters `max_iter`, `tol`, and
+  `max_no_improvement` to be consistent with
+  :class:`decomposition.MiniBatchDictionaryLearning`. `n_iter` will be removed
+  in version 1.3. :pr:`23726` by :user:`Guillaume Lemaitre <glemaitre>`.
+
+
 :mod:`sklearn.ensemble`
 .......................
 
diff --git a/sklearn/decomposition/_dict_learning.py b/sklearn/decomposition/_dict_learning.py
@@ -794,7 +794,7 @@ def dict_learning_online(
         Number of mini-batch iterations to perform.
 
         .. deprecated:: 1.1
-           `n_iter` is deprecated in 1.1 and will be removed in 1.3. Use
+           `n_iter` is deprecated in 1.1 and will be removed in 1.4. Use
            `max_iter` instead.
 
     max_iter : int, default=None
@@ -1758,7 +1758,7 @@ class MiniBatchDictionaryLearning(_BaseSparseCoding, BaseEstimator):
         Total number of iterations over data batches to perform.
 
         .. deprecated:: 1.1
-           ``n_iter`` is deprecated in 1.1 and will be removed in 1.3. Use
+           ``n_iter`` is deprecated in 1.1 and will be removed in 1.4. Use
            ``max_iter`` instead.
 
     max_iter : int, default=None
@@ -2251,6 +2251,17 @@ def fit(self, X, y=None):
         )
 
         self._check_params(X)
+
+        if self.n_iter != "deprecated":
+            warnings.warn(
+                "'n_iter' is deprecated in version 1.1 and will be removed "
+                "in version 1.4. Use 'max_iter' and let 'n_iter' to its default "
+                "value instead. 'n_iter' is also ignored if 'max_iter' is "
+                "specified.",
+                FutureWarning,
+            )
+            n_iter = self.n_iter
+
         self._random_state = check_random_state(self.random_state)
 
         dictionary = self._initialize_dict(X, self._random_state)
@@ -2310,15 +2321,7 @@ def fit(self, X, y=None):
             self.n_iter_ = np.ceil(self.n_steps_ / n_steps_per_iter)
         else:
             # TODO remove this branch in 1.3
-            if self.n_iter != "deprecated":
-                warnings.warn(
-                    "'n_iter' is deprecated in version 1.1 and will be removed"
-                    " in version 1.3. Use 'max_iter' instead.",
-                    FutureWarning,
-                )
-                n_iter = self.n_iter
-            else:
-                n_iter = 1000
+            n_iter = 1000 if self.n_iter == "deprecated" else self.n_iter
 
             batches = gen_batches(n_samples, self._batch_size)
             batches = itertools.cycle(batches)
diff --git a/sklearn/decomposition/_sparse_pca.py b/sklearn/decomposition/_sparse_pca.py
@@ -2,15 +2,13 @@
 # Author: Vlad Niculae, Gael Varoquaux, Alexandre Gramfort
 # License: BSD 3 clause
 
-import warnings
-
 import numpy as np
 
 from ..utils import check_random_state
 from ..utils.validation import check_is_fitted
 from ..linear_model import ridge_regression
 from ..base import BaseEstimator, TransformerMixin, _ClassNamePrefixFeaturesOutMixin
-from ._dict_learning import dict_learning, dict_learning_online
+from ._dict_learning import dict_learning, MiniBatchDictionaryLearning
 
 
 class SparsePCA(_ClassNamePrefixFeaturesOutMixin, TransformerMixin, BaseEstimator):
@@ -275,6 +273,17 @@ class MiniBatchSparsePCA(SparsePCA):
     n_iter : int, default=100
         Number of iterations to perform for each mini batch.
 
+        .. deprecated:: 1.2
+           `n_iter` is deprecated in 1.2 and will be removed in 1.4. Use
+           `max_iter` instead.
+
+    max_iter : int, default=None
+        Maximum number of iterations over the complete dataset before
+        stopping independently of any early stopping criterion heuristics.
+        If `max_iter` is not `None`, `n_iter` is ignored.
+
+        .. versionadded:: 1.2
+
     callback : callable, default=None
         Callable that gets invoked every five iterations.
 
@@ -307,6 +316,25 @@ class MiniBatchSparsePCA(SparsePCA):
         across multiple function calls.
         See :term:`Glossary <random_state>`.
 
+    tol : float, default=1e-3
+        Control early stopping based on the norm of the differences in the
+        dictionary between 2 steps. Used only if `max_iter` is not None.
+
+        To disable early stopping based on changes in the dictionary, set
+        `tol` to 0.0.
+
+        .. versionadded:: 1.1
+
+    max_no_improvement : int, default=10
+        Control early stopping based on the consecutive number of mini batches
+        that does not yield an improvement on the smoothed cost function. Used only if
+        `max_iter` is not None.
+
+        To disable convergence detection based on cost function, set
+        `max_no_improvement` to `None`.
+
+        .. versionadded:: 1.1
+
     Attributes
     ----------
     components_ : ndarray of shape (n_components, n_features)
@@ -350,15 +378,15 @@ class MiniBatchSparsePCA(SparsePCA):
     >>> from sklearn.decomposition import MiniBatchSparsePCA
     >>> X, _ = make_friedman1(n_samples=200, n_features=30, random_state=0)
     >>> transformer = MiniBatchSparsePCA(n_components=5, batch_size=50,
-    ...                                  random_state=0)
+    ...                                  max_iter=10, random_state=0)
     >>> transformer.fit(X)
     MiniBatchSparsePCA(...)
     >>> X_transformed = transformer.transform(X)
     >>> X_transformed.shape
     (200, 5)
     >>> # most values in the components_ are zero (sparsity)
     >>> np.mean(transformer.components_ == 0)
-    0.94
+    0.9...
     """
 
     def __init__(
@@ -367,14 +395,17 @@ def __init__(
         *,
         alpha=1,
         ridge_alpha=0.01,
-        n_iter=100,
+        n_iter="deprecated",
+        max_iter=None,
         callback=None,
         batch_size=3,
         verbose=False,
         shuffle=True,
         n_jobs=None,
         method="lars",
         random_state=None,
+        tol=1e-3,
+        max_no_improvement=10,
     ):
         super().__init__(
             n_components=n_components,
@@ -386,9 +417,12 @@ def __init__(
             random_state=random_state,
         )
         self.n_iter = n_iter
+        self.max_iter = max_iter
         self.callback = callback
         self.batch_size = batch_size
         self.shuffle = shuffle
+        self.tol = tol
+        self.max_no_improvement = max_no_improvement
 
     def fit(self, X, y=None):
         """Fit the model from data in X.
@@ -418,44 +452,27 @@ def fit(self, X, y=None):
         else:
             n_components = self.n_components
 
-        with warnings.catch_warnings():
-            # return_n_iter and n_iter are deprecated. TODO Remove in 1.3
-            warnings.filterwarnings(
-                "ignore",
-                message=(
-                    "'return_n_iter' is deprecated in version 1.1 and will be "
-                    "removed in version 1.3. From 1.3 'n_iter' will never be "
-                    "returned. Refer to the 'n_iter_' and 'n_steps_' attributes "
-                    "of the MiniBatchDictionaryLearning object instead."
-                ),
-                category=FutureWarning,
-            )
-            warnings.filterwarnings(
-                "ignore",
-                message=(
-                    "'n_iter' is deprecated in version 1.1 and will be removed in "
-                    "version 1.3. Use 'max_iter' instead."
-                ),
-                category=FutureWarning,
-            )
-            Vt, _, self.n_iter_ = dict_learning_online(
-                X.T,
-                n_components,
-                alpha=self.alpha,
-                n_iter=self.n_iter,
-                return_code=True,
-                dict_init=None,
-                verbose=self.verbose,
-                callback=self.callback,
-                batch_size=self.batch_size,
-                shuffle=self.shuffle,
-                n_jobs=self.n_jobs,
-                method=self.method,
-                random_state=random_state,
-                return_n_iter=True,
-            )
+        transform_algorithm = "lasso_" + self.method
+        est = MiniBatchDictionaryLearning(
+            n_components=n_components,
+            alpha=self.alpha,
+            n_iter=self.n_iter,
+            max_iter=self.max_iter,
+            dict_init=None,
+            batch_size=self.batch_size,
+            shuffle=self.shuffle,
+            n_jobs=self.n_jobs,
+            fit_algorithm=self.method,
+            random_state=random_state,
+            transform_algorithm=transform_algorithm,
+            transform_alpha=self.alpha,
+            verbose=self.verbose,
+            callback=self.callback,
+            tol=self.tol,
+            max_no_improvement=self.max_no_improvement,
+        ).fit(X.T)
 
-        self.components_ = Vt.T
+        self.components_, self.n_iter_ = est.transform(X.T).T, est.n_iter_
 
         components_norm = np.linalg.norm(self.components_, axis=1)[:, np.newaxis]
         components_norm[components_norm == 0] = 1
diff --git a/sklearn/decomposition/tests/test_dict_learning.py b/sklearn/decomposition/tests/test_dict_learning.py
@@ -721,7 +721,7 @@ def test_minibatch_dict_learning_n_iter_deprecated():
     # check the deprecation warning of n_iter
     # FIXME: remove in 1.3
     depr_msg = (
-        "'n_iter' is deprecated in version 1.1 and will be removed in version 1.3"
+        "'n_iter' is deprecated in version 1.1 and will be removed in version 1.4"
     )
     est = MiniBatchDictionaryLearning(
         n_components=2, batch_size=4, n_iter=5, random_state=0
@@ -1072,3 +1072,14 @@ def test_get_feature_names_out(estimator):
         feature_names_out,
         [f"{estimator_name}{i}" for i in range(n_components)],
     )
+
+
+# TODO(1.4) remove
+def test_minibatch_dictionary_learning_warns_and_ignore_n_iter():
+    """Check that we always raise a warning when `n_iter` is set even if it is
+    ignored if `max_iter` is set.
+    """
+    warn_msg = "'n_iter' is deprecated in version 1.1"
+    with pytest.warns(FutureWarning, match=warn_msg):
+        model = MiniBatchDictionaryLearning(batch_size=256, n_iter=2, max_iter=2).fit(X)
+    assert model.n_iter_ == 2
diff --git a/sklearn/decomposition/tests/test_sparse_pca.py b/sklearn/decomposition/tests/test_sparse_pca.py
@@ -265,3 +265,51 @@ def test_spca_feature_names_out(SPCA):
 
     estimator_name = SPCA.__name__.lower()
     assert_array_equal([f"{estimator_name}{i}" for i in range(4)], names)
+
+
+# TODO (1.4): remove this test
+def test_spca_n_iter_deprecation():
+    """Check that we raise a warning for the deprecation of `n_iter` and it is ignored
+    when `max_iter` is specified.
+    """
+    rng = np.random.RandomState(0)
+    n_samples, n_features = 12, 10
+    X = rng.randn(n_samples, n_features)
+
+    warn_msg = "'n_iter' is deprecated in version 1.1 and will be removed"
+    with pytest.warns(FutureWarning, match=warn_msg):
+        MiniBatchSparsePCA(n_iter=2).fit(X)
+
+    n_iter, max_iter = 1, 100
+    with pytest.warns(FutureWarning, match=warn_msg):
+        model = MiniBatchSparsePCA(
+            n_iter=n_iter, max_iter=max_iter, random_state=0
+        ).fit(X)
+    assert model.n_iter_ > 1
+    assert model.n_iter_ <= max_iter
+
+
+def test_spca_early_stopping(global_random_seed):
+    """Check that `tol` and `max_no_improvement` act as early stopping."""
+    rng = np.random.RandomState(global_random_seed)
+    n_samples, n_features = 50, 10
+    X = rng.randn(n_samples, n_features)
+
+    # vary the tolerance to force the early stopping of one of the model
+    model_early_stopped = MiniBatchSparsePCA(
+        max_iter=100, tol=0.5, random_state=global_random_seed
+    ).fit(X)
+    model_not_early_stopped = MiniBatchSparsePCA(
+        max_iter=100, tol=1e-3, random_state=global_random_seed
+    ).fit(X)
+    assert model_early_stopped.n_iter_ < model_not_early_stopped.n_iter_
+
+    # force the max number of no improvement to a large value to check that
+    # it does help to early stop
+    model_early_stopped = MiniBatchSparsePCA(
+        max_iter=100, tol=1e-6, max_no_improvement=2, random_state=global_random_seed
+    ).fit(X)
+    model_not_early_stopped = MiniBatchSparsePCA(
+        max_iter=100, tol=1e-6, max_no_improvement=100, random_state=global_random_seed
+    ).fit(X)
+    assert model_early_stopped.n_iter_ < model_not_early_stopped.n_iter_
