Add sparse input support to ElasticNet/Lasso

docs/source/cuml-accel/limitations.rst

@@ -269,7 +269,6 @@ ElasticNet
 - If ``positive=True``.
 - If ``warm_start=True``.
 - If ``precompute`` is not ``False``.
-- If ``X`` is sparse.
 
 Additionally, the following fitted attributes are currently not computed:
 
@@ -290,7 +289,6 @@ Lasso
 - If ``positive=True``.
 - If ``warm_start=True``.
 - If ``precompute`` is not ``False``.
-- If ``X`` is sparse.
 
 Additionally, the following fitted attributes are currently not computed:
 

python/cuml/cuml/accel/_overrides/sklearn/linear_model.py

@@ -10,6 +10,7 @@
 from cuml.accel.estimator_proxy import ProxyBase
 from cuml.common.sparse_utils import is_sparse
 from cuml.internals.array import CumlArray
+from cuml.internals.input_utils import input_to_cuml_array
 from cuml.internals.interop import UnsupportedOnGPU
 from cuml.internals.outputs import using_output_type
 
@@ -34,15 +35,6 @@ class LogisticRegression(ProxyBase):
     _gpu_class = cuml.linear_model.LogisticRegression
 
 
-class ElasticNet(ProxyBase):
-    _gpu_class = cuml.linear_model.ElasticNet
-    _not_implemented_attributes = frozenset(("dual_gap_",))
-
-    def _gpu_fit(self, X, y, sample_weight=None, check_input=True):
-        # Fixes signature mismatch with cuml.ElasticNet. check_input can be ignored.
-        return self._gpu.fit(X, y, sample_weight=sample_weight)
-
-
 class Ridge(ProxyBase):
     _gpu_class = cuml.linear_model.Ridge
 
@@ -74,10 +66,27 @@ def _gpu_fit(self, X, y, sample_weight=None):
         return self
 
 
+class ElasticNet(ProxyBase):
+    _gpu_class = cuml.linear_model.ElasticNet
+    _not_implemented_attributes = frozenset(("dual_gap_",))
+
+    def _gpu_fit(self, X, y, sample_weight=None, check_input=True):
+        # check_input is ignored, only here to fix signature mismatch with sklearn
+
+        y = input_to_cuml_array(y, convert_to_mem_type=False)[0]
+        if len(y.shape) > 1 and y.shape[1] > 1:
+            raise UnsupportedOnGPU("Multi-output targets are not supported")
+        return self._gpu.fit(X, y, sample_weight=sample_weight)
+
+
 class Lasso(ProxyBase):
     _gpu_class = cuml.linear_model.Lasso
     _not_implemented_attributes = frozenset(("dual_gap_",))
 
     def _gpu_fit(self, X, y, sample_weight=None, check_input=True):
-        # Fixes signature mismatch with cuml.Lasso. check_input can be ignored.
+        # check_input is ignored, only here to fix signature mismatch with sklearn
+
+        y = input_to_cuml_array(y, convert_to_mem_type=False)[0]
+        if len(y.shape) > 1 and y.shape[1] > 1:
+            raise UnsupportedOnGPU("Multi-output targets are not supported")
         return self._gpu.fit(X, y, sample_weight=sample_weight)

python/cuml/cuml/linear_model/elastic_net.py

@@ -2,8 +2,11 @@
 # SPDX-FileCopyrightText: Copyright (c) 2019-2026, NVIDIA CORPORATION.
 # SPDX-License-Identifier: Apache-2.0
 #
+import cupyx.scipy.sparse
+
 from cuml.common.array_descriptor import CumlArrayDescriptor
 from cuml.common.doc_utils import generate_docstring
+from cuml.common.sparse_utils import is_sparse
 from cuml.internals.array import CumlArray
 from cuml.internals.base import Base
 from cuml.internals.interop import (
@@ -12,15 +15,24 @@
     to_cpu,
     to_gpu,
 )
-from cuml.internals.mixins import FMajorInputTagMixin, RegressorMixin
+from cuml.internals.mixins import (
+    FMajorInputTagMixin,
+    RegressorMixin,
+    SparseInputTagMixin,
+)
 from cuml.internals.outputs import reflect
 from cuml.linear_model.base import LinearPredictMixin
 from cuml.solvers.cd import fit_coordinate_descent
 from cuml.solvers.qn import fit_qn
 
 
 class ElasticNet(
-    Base, InteropMixin, LinearPredictMixin, RegressorMixin, FMajorInputTagMixin
+    Base,
+    InteropMixin,
+    LinearPredictMixin,
+    RegressorMixin,
+    SparseInputTagMixin,
+    FMajorInputTagMixin,
 ):
     """
     Linear regression with combined L1 and L2 priors as regularizer.
@@ -48,11 +60,12 @@ class ElasticNet(
         The tolerance for the optimization: if the updates are smaller than
         tol, the optimization code checks the dual gap for optimality and
         continues until it is smaller than tol.
-    solver : {'cd', 'qn'}, default='cd'
-        Choose an algorithm:
+    solver : {'auto', 'cd', 'qn'}, default='auto'
+        The solver to use.
 
-          * 'cd' - coordinate descent
-          * 'qn' - quasi-newton
+        - 'auto': uses 'cd' for dense inputs, and 'qn' for sparse inputs
+        - 'cd': uses coordinate descent. Only supports dense inputs.
+        - 'qn': uses quasi-newton methods. Supports sparse and dense inputs.
 
         You may find the alternative 'qn' algorithm is faster when the number
         of features is sufficiently large but the sample size is small.
@@ -76,6 +89,8 @@ class ElasticNet(
     ----------
     coef_ : array, shape (n_features)
         The estimated coefficients for the linear regression model.
+    sparse_coef_ : sparse matrix, shape (n_targets, n_features)
+        Sparse matrix representation of `coef_`.
     intercept_ : float
         The independent term, will be 0 if `fit_intercept` is False.
     n_iter_ : int
@@ -192,7 +207,7 @@ def __init__(
         fit_intercept=True,
         max_iter=1000,
         tol=1e-3,
-        solver="cd",
+        solver="auto",
         selection="cyclic",
         output_type=None,
         verbose=False,
@@ -207,6 +222,12 @@ def __init__(
         self.solver = solver
         self.selection = selection
 
+    @property
+    @reflect
+    def sparse_coef_(self):
+        """Sparse representation of the fitted `coef_`."""
+        return cupyx.scipy.sparse.csr_matrix(self.coef_.to_output("cupy"))
+
     @generate_docstring()
     @reflect(reset=True)
     def fit(
@@ -225,7 +246,11 @@ def fit(
                 f"Expected 0.0 <= l1_ratio <= 1.0, got {self.l1_ratio}"
             )
 
-        if self.solver == "qn":
+        solver = self.solver
+        if solver == "auto":
+            solver = "qn" if is_sparse(X) else "cd"
+
+        if solver == "qn":
             coef, intercept, n_iter, _ = fit_qn(
                 X,
                 y,
@@ -242,7 +267,12 @@ def fit(
             )
             coef = CumlArray(data=coef.to_output("cupy").flatten())
             intercept = intercept.item()
-        elif self.solver == "cd":
+        elif solver == "cd":
+            if is_sparse(X):
+                raise ValueError(
+                    "solver='cd' doesn't support sparse inputs, please use "
+                    "solver='auto' or solver='qn' instead"
+                )
             coef, intercept, n_iter = fit_coordinate_descent(
                 X,
                 y,
@@ -256,7 +286,7 @@ def fit(
                 tol=self.tol,
             )
         else:
-            raise ValueError(f"solver {self.solver} is not supported")
+            raise ValueError(f"solver={solver!r} is not supported")
 
         self.coef_ = coef
         self.intercept_ = intercept

python/cuml/cuml/linear_model/lasso.py

@@ -30,14 +30,15 @@ class Lasso(ElasticNet):
         The tolerance for the optimization: if the updates are smaller than
         tol, the optimization code checks the dual gap for optimality and
         continues until it is smaller than tol.
-    solver : {'cd', 'qn'} (default='cd')
-        Choose an algorithm:
+    solver : {'auto', 'cd', 'qn'}, default='auto'
+        The solver to use.
 
-          * 'cd' - coordinate descent
-          * 'qn' - quasi-newton
+        - 'auto': uses 'cd' for dense inputs, and 'qn' for sparse inputs
+        - 'cd': uses coordinate descent. Only supports dense inputs.
+        - 'qn': uses quasi-newton methods. Supports sparse and dense inputs.
 
         You may find the alternative 'qn' algorithm is faster when the number
-        of features is sufficiently large, but the sample size is small.
+        of features is sufficiently large but the sample size is small.
     selection : {'cyclic', 'random'} (default='cyclic')
         If set to 'random', a random coefficient is updated every iteration
         rather than looping over features sequentially by default.
@@ -57,6 +58,8 @@ class Lasso(ElasticNet):
     ----------
     coef_ : array, shape (n_features)
         The estimated coefficients for the linear regression model.
+    sparse_coef_ : sparse matrix, shape (n_targets, n_features)
+        Sparse matrix representation of `coef_`.
     intercept_ : array
         The independent term. If `fit_intercept` is False, will be 0.
     n_iter_ : int
@@ -119,7 +122,7 @@ def __init__(
         fit_intercept=True,
         max_iter=1000,
         tol=1e-3,
-        solver="cd",
+        solver="auto",
         selection="cyclic",
         output_type=None,
         verbose=False,

python/cuml/cuml_accel_tests/upstream/scikit-learn/xfail-list.yaml

@@ -33,10 +33,16 @@
   tests:
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_coordinate_descent[Lasso-1-kwargs1]"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_warm_start_convergence"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-24-6-False-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-24-6-True-Lasso]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-6-24-False-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-6-24-True-ElasticNet]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-24-6-False-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-24-6-True-Lasso]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-6-24-False-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-6-24-True-ElasticNet]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_enet_coordinate_descent[csc_array]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_enet_coordinate_descent[csc_matrix]"
   - "sklearn.manifold.tests.test_t_sne::test_pca_initialization_not_compatible_with_sparse_input[csr_array]"
   - "sklearn.manifold.tests.test_t_sne::test_pca_initialization_not_compatible_with_sparse_input[csr_matrix]"
 - reason: Test should fail with cuml.accel
@@ -235,14 +241,17 @@
   - "sklearn.linear_model.tests.test_base::test_linear_regression"
   - "sklearn.linear_model.tests.test_base::test_linear_regression_pd_sparse_dataframe_warning"
   - "sklearn.linear_model.tests.test_common::test_balance_property[42-True-LogisticRegression]"
-  - "sklearn.linear_model.tests.test_coordinate_descent::test_check_input_false"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_elasticnet_precompute_gram_weighted_samples"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_copy_X_False_check_input_False"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_float_precision"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_multitarget"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_nonfinite_params"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_sample_weight_consistency[42-None-False-0.01-False]"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_sample_weight_consistency[42-None-False-0.01-True]"
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_sample_weight_consistency[42-csr_array-False-0.01-False]"
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_sample_weight_consistency[42-csr_array-False-0.01-True]"
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_sample_weight_consistency[42-csr_matrix-False-0.01-False]"
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_sample_weight_consistency[42-csr_matrix-False-0.01-True]"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_toy"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_lassoCV_does_not_set_precompute[False-False]"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_lassoCV_does_not_set_precompute[auto-False]"
@@ -251,6 +260,8 @@
   - "sklearn.linear_model.tests.test_coordinate_descent::test_lasso_readonly_data"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_lasso_toy"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_lasso_zero"
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_sparse_input_convergence_warning[csr_array]"
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_sparse_input_convergence_warning[csr_matrix]"
   - "sklearn.linear_model.tests.test_coordinate_descent::test_warm_start_convergence_with_regularizer_decrement"
   - "sklearn.linear_model.tests.test_ransac::test_perfect_horizontal_line"
   - "sklearn.linear_model.tests.test_ransac::test_ransac_exceed_max_skips"
@@ -273,16 +284,36 @@
   - "sklearn.linear_model.tests.test_ridge::test_ridge_sample_weight_consistency[42-saga-wide-csr_matrix-False]"
   - "sklearn.linear_model.tests.test_ridge::test_ridgecv_sample_weight"
   - "sklearn.linear_model.tests.test_sag::test_step_size_alpha_error"
-  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_same_multiple_output_sparse_dense[coo_array]"
-  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_same_multiple_output_sparse_dense[coo_matrix]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_multitarget[csc_array]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_multitarget[csc_matrix]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_toy_explicit_sparse_input[lil_array]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_toy_explicit_sparse_input[lil_matrix]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_toy_list_input[csc_array-False]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_toy_list_input[csc_array-True]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_toy_list_input[csc_matrix-False]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_enet_toy_list_input[csc_matrix-True]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_lasso_zero[csc_array]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_lasso_zero[csc_matrix]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-6-24-False-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-6-24-False-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-6-24-True-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-False-6-24-True-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-24-6-False-ElasticNet]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-24-6-False-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-24-6-True-ElasticNet]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-24-6-True-Lasso]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-6-24-False-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-6-24-False-Lasso]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-6-24-True-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_array-True-6-24-True-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-6-24-False-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-6-24-False-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-6-24-True-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-False-6-24-True-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-24-6-False-ElasticNet]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-24-6-False-Lasso]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-24-6-True-ElasticNet]"
+  - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-24-6-True-Lasso]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-6-24-False-ElasticNet]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-6-24-False-Lasso]"
   - "sklearn.linear_model.tests.test_sparse_coordinate_descent::test_sparse_dense_equality[csc_matrix-True-6-24-True-ElasticNet]"
@@ -728,6 +759,8 @@
   - "sklearn.neighbors.tests.test_neighbors::test_neighbor_classifiers_loocv[ball_tree-nn_model0]"
   - "sklearn.neighbors.tests.test_neighbors::test_neighbor_classifiers_loocv[brute-nn_model0]"
   - "sklearn.neighbors.tests.test_neighbors::test_neighbor_classifiers_loocv[kd_tree-nn_model0]"
+  - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_sample_weight_equivalence_on_sparse_data]"
+  - "sklearn.tests.test_common::test_estimators[Lasso()-check_sample_weight_equivalence_on_sparse_data]"
   - "sklearn.tests.test_common::test_estimators[LogisticRegression()-check_sample_weight_equivalence_on_dense_data]"
   - "sklearn.tests.test_common::test_estimators[LogisticRegression()-check_sample_weight_equivalence_on_sparse_data]"
   - "sklearn.tests.test_common::test_estimators[Ridge()-check_non_transformer_estimators_n_iter]"
@@ -1067,7 +1100,6 @@
   - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_dtype_object]"
   - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_estimators_nan_inf]"
   - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_regressor_data_not_an_array]"
-  - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_regressor_multioutput]"
   - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_requires_y_none]"
   - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_sample_weights_not_an_array]"
   - "sklearn.tests.test_common::test_estimators[ElasticNet()-check_supervised_y_no_nan]"
@@ -1093,7 +1125,6 @@
   - "sklearn.tests.test_common::test_estimators[Lasso()-check_dtype_object]"
   - "sklearn.tests.test_common::test_estimators[Lasso()-check_estimators_nan_inf]"
   - "sklearn.tests.test_common::test_estimators[Lasso()-check_regressor_data_not_an_array]"
-  - "sklearn.tests.test_common::test_estimators[Lasso()-check_regressor_multioutput]"
   - "sklearn.tests.test_common::test_estimators[Lasso()-check_requires_y_none]"
   - "sklearn.tests.test_common::test_estimators[Lasso()-check_sample_weights_not_an_array]"
   - "sklearn.tests.test_common::test_estimators[Lasso()-check_supervised_y_no_nan]"
@@ -1266,6 +1297,14 @@
   tests:
   - "sklearn.tests.test_calibration::test_calibrated_classifier_cv_double_sample_weights_equivalence[False-temperature]"
   - "sklearn.tests.test_calibration::test_calibrated_classifier_cv_double_sample_weights_equivalence[True-temperature]"
+- reason: ElasticNet sample_weight handling is a bit off
+  condition: scikit-learn>=1.6,<1.8
+  tests:
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_alpha_max_sample_weight[sample_weight0-False-True]"
+- reason: ElasticNet sample_weight handling is a bit off
+  condition: scikit-learn>=1.8
+  tests:
+  - "sklearn.linear_model.tests.test_coordinate_descent::test_enet_alpha_max[sample_weight0-False-True]"
 - reason: Elasticnet scores attribute layout differs with cuml.accel in sklearn 1.8
   condition: scikit-learn>=1.8
   tests: