scikit-learn · lesteve · Jan 19, 2023 · Dec 12, 2022 · Dec 15, 2022 · Dec 16, 2022
diff --git a/doc/whats_new/v1.2.rst b/doc/whats_new/v1.2.rst
@@ -30,6 +30,14 @@ Changelog
   The ARFF specs requires to ignore the leading space.
   :pr:`25312` by :user:`Guillaume Lemaitre <glemaitre>`.
 
+:mod:`sklearn.decomposition`
+............................
+
+- |Fix| :class:`decomposition.DictionaryLearning` better supports readonly NumPy
+  arrays. In particular, it better supports large datasets which are memory-mapped
+  when it is used with coordinate descent algorithms (i.e. when `fit_algorithm='cd'`).
+  :pr:`25172` by :user:`Julien Jerphanion <jjerphan>`.
+
 :mod:`sklearn.ensemble`
 .......................
 

diff --git a/sklearn/decomposition/_dict_learning.py b/sklearn/decomposition/_dict_learning.py
@@ -174,6 +174,13 @@ def _sparse_encode(
         )
 
         if init is not None:
+            # In some workflows using coordinate descent algorithms:
+            #  - users might provide NumPy arrays with read-only buffers
+            #  - `joblib` might memmap arrays making their buffer read-only
+            # TODO: move this handling (which is currently too broad)
+            # closer to the actual private function which need buffers to be writable.
+            if not init.flags["WRITEABLE"]:
+                init = np.array(init)
             clf.coef_ = init
 
         clf.fit(dictionary.T, X.T, check_input=check_input)

diff --git a/sklearn/decomposition/tests/test_dict_learning.py b/sklearn/decomposition/tests/test_dict_learning.py
@@ -5,6 +5,10 @@
 from functools import partial
 import itertools
 
+from joblib import Parallel
+
+import sklearn
+
 from sklearn.base import clone
 
 from sklearn.exceptions import ConvergenceWarning
@@ -990,6 +994,29 @@ def test_get_feature_names_out(estimator):
     )
 
 
+def test_cd_work_on_joblib_memmapped_data(monkeypatch):
+    monkeypatch.setattr(
+        sklearn.decomposition._dict_learning,
+        "Parallel",
+        partial(Parallel, max_nbytes=100),
+    )
+
+    rng = np.random.RandomState(0)
+    X_train = rng.randn(10, 10)
+
+    dict_learner = DictionaryLearning(
+        n_components=5,
+        random_state=0,
+        n_jobs=2,
+        fit_algorithm="cd",
+        max_iter=50,
+        verbose=True,
+    )
+
+    # This must run and complete without error.
+    dict_learner.fit(X_train)
+
+
 # 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

diff --git a/sklearn/linear_model/_cd_fast.pyx b/sklearn/linear_model/_cd_fast.pyx
@@ -90,13 +90,13 @@ cdef floating diff_abs_max(int n, floating* a, floating* b) nogil:
             m = d
     return m
 
-
+# TODO: use const fused typed memoryview where possible when Cython 0.29.33 is used.
 def enet_coordinate_descent(
-    floating[::1] w,
+    cnp.ndarray[floating, ndim=1, mode='c'] w,
     floating alpha,
     floating beta,
-    floating[::1, :] X,
-    floating[::1] y,
+    cnp.ndarray[floating, ndim=2, mode='fortran'] X,
+    cnp.ndarray[floating, ndim=1, mode='c'] y,
     unsigned int max_iter,
     floating tol,
     object rng,
@@ -273,16 +273,17 @@ def enet_coordinate_descent(
     return np.asarray(w), gap, tol, n_iter + 1
 
 
+# TODO: use const fused typed memoryview where possible when Cython 0.29.33 is used.
 def sparse_enet_coordinate_descent(
-    floating [::1] w,
+    cnp.ndarray[floating, ndim=1, mode='c'] w,
     floating alpha,
     floating beta,
-    floating[::1] X_data, # TODO: Make const after release of Cython 3 (#23147)
+    cnp.ndarray[floating, ndim=1, mode='c'] X_data,
     const int[::1] X_indices,
     const int[::1] X_indptr,
-    floating[::1] y,
-    floating[::1] sample_weight,
-    floating[::1] X_mean,
+    cnp.ndarray[floating, ndim=1, mode='c'] y,
+    cnp.ndarray[floating, ndim=1, mode='c'] sample_weight,
+    cnp.ndarray[floating, ndim=1, mode='c'] X_mean,
     unsigned int max_iter,
     floating tol,
     object rng,
@@ -564,8 +565,9 @@ def sparse_enet_coordinate_descent(
     return np.asarray(w), gap, tol, n_iter + 1
 
 
+# TODO: use const fused typed memoryview where possible when Cython 0.29.33 is used.
 def enet_coordinate_descent_gram(
-    floating[::1] w,
+    cnp.ndarray[floating, ndim=1, mode='c'] w,
     floating alpha,
     floating beta,
     cnp.ndarray[floating, ndim=2, mode='c'] Q,
@@ -630,9 +632,9 @@ def enet_coordinate_descent_gram(
     cdef UINT32_t* rand_r_state = &rand_r_state_seed
 
     cdef floating y_norm2 = np.dot(y, y)
-    cdef floating* w_ptr = <floating*>&w[0]
+    cdef floating* w_ptr = &w[0]
     cdef floating* Q_ptr = &Q[0, 0]
-    cdef floating* q_ptr = <floating*>q.data
+    cdef floating* q_ptr = &q[0]
     cdef floating* H_ptr = &H[0]
     cdef floating* XtA_ptr = &XtA[0]
     tol = tol * y_norm2
@@ -734,9 +736,9 @@ def enet_coordinate_descent_gram(
 
     return np.asarray(w), gap, tol, n_iter + 1
 
-
+# TODO: use const fused typed memoryview where possible when Cython 0.29.33 is used.
 def enet_coordinate_descent_multi_task(
-    floating[::1, :] W,
+    cnp.ndarray[floating, ndim=2, mode='fortran'] W,
     floating l1_reg,
     floating l2_reg,
     # TODO: use const qualified fused-typed memoryview when Cython 3.0 is used.

diff --git a/sklearn/linear_model/_coordinate_descent.py b/sklearn/linear_model/_coordinate_descent.py
@@ -30,7 +30,6 @@
     check_is_fitted,
     column_or_1d,
 )
-from ..utils._readonly_array_wrapper import ReadonlyArrayWrapper
 from ..utils.fixes import delayed
 
 # mypy error: Module 'sklearn.linear_model' has no attribute '_cd_fast'
@@ -594,9 +593,7 @@ def enet_path(
                 w=coef_,
                 alpha=l1_reg,
                 beta=l2_reg,
-                X_data=ReadonlyArrayWrapper(
-                    X.data
-                ),  # TODO: Remove after release of Cython 3 (#23147)
+                X_data=X.data,
                 X_indices=X.indices,
                 X_indptr=X.indptr,
                 y=y,