[enhancement] add assume_centered capability to EmpiricalCovariance (#1796)

* Update covariance.cpp

* Update covariance.py

* Update covariance.py

* Update covariance.py

* Update test_covariance.py

* import daal_check_version

* Update covariance.py

* Update covariance.py

Author: icfaust

onedal/covariance/covariance.cpp

@@ -52,6 +52,11 @@ struct params2desc {
             desc.set_bias(params["bias"].cast<bool>());
         }
 #endif // defined(ONEDAL_VERSION) && ONEDAL_VERSION>=20240001
+#if defined(ONEDAL_VERSION) && ONEDAL_VERSION >= 20240400
+        if (params.contains("assumeCentered")) {
+            desc.set_assume_centered(params["assumeCentered"].cast<bool>());
+        }
+#endif // defined(ONEDAL_VERSION) && ONEDAL_VERSION>=20240400
         return desc;
     }
 };

onedal/covariance/covariance.py

@@ -26,9 +26,10 @@
 
 
 class BaseEmpiricalCovariance(BaseEstimator, metaclass=ABCMeta):
-    def __init__(self, method="dense", bias=False):
+    def __init__(self, method="dense", bias=False, assume_centered=False):
         self.method = method
         self.bias = bias
+        self.assume_centered = assume_centered
 
     def _get_onedal_params(self, dtype=np.float32):
         params = {
@@ -37,6 +38,8 @@ def _get_onedal_params(self, dtype=np.float32):
         }
         if daal_check_version((2024, "P", 1)):
             params["bias"] = self.bias
+        if daal_check_version((2024, "P", 400)):
+            params["assumeCentered"] = self.assume_centered
 
         return params
 
@@ -55,6 +58,12 @@ class EmpiricalCovariance(BaseEmpiricalCovariance):
         If True biased estimation of covariance is computed which equals to
         the unbiased one multiplied by (n_samples - 1) / n_samples.
 
+    assume_centered : bool, default=False
+        If True, data are not centered before computation.
+        Useful when working with data whose mean is almost, but not exactly
+        zero.
+        If False (default), data are centered before computation.
+
     Attributes
     ----------
     location_ : ndarray of shape (n_features,)

sklearnex/preview/covariance/covariance.py

@@ -22,7 +22,7 @@
 from sklearn.utils import check_array
 
 from daal4py.sklearn._n_jobs_support import control_n_jobs
-from daal4py.sklearn._utils import sklearn_check_version
+from daal4py.sklearn._utils import daal_check_version, sklearn_check_version
 from onedal.common.hyperparameters import get_hyperparameters
 from onedal.covariance import EmpiricalCovariance as onedal_EmpiricalCovariance
 from sklearnex import config_context
@@ -44,6 +44,10 @@ class EmpiricalCovariance(sklearn_EmpiricalCovariance):
 
     def _save_attributes(self):
         assert hasattr(self, "_onedal_estimator")
+        if not daal_check_version((2024, "P", 400)) and self.assume_centered:
+            location = self._onedal_estimator.location_[None, :]
+            self._onedal_estimator.covariance_ += np.dot(location.T, location)
+            self._onedal_estimator.location_ = np.zeros_like(np.squeeze(location))
         self._set_covariance(self._onedal_estimator.covariance_)
         self.location_ = self._onedal_estimator.location_
 
@@ -58,6 +62,7 @@ def _onedal_fit(self, X, queue=None):
         onedal_params = {
             "method": "dense",
             "bias": True,
+            "assume_centered": self.assume_centered,
         }
 
         self._onedal_estimator = self._onedal_covariance(**onedal_params)
@@ -73,10 +78,6 @@ def _onedal_supported(self, method_name, *data):
             (X,) = data
             patching_status.and_conditions(
                 [
-                    (
-                        self.assume_centered == False,
-                        "assume_centered parameter is not supported on oneDAL side",
-                    ),
                     (not sp.issparse(X), "X is sparse. Sparse input is not supported."),
                 ]
             )

sklearnex/preview/covariance/tests/test_covariance.py

@@ -27,27 +27,40 @@
 
 @pytest.mark.parametrize("dataframe,queue", get_dataframes_and_queues())
 @pytest.mark.parametrize("macro_block", [None, 1024])
-def test_sklearnex_import_covariance(dataframe, queue, macro_block):
+@pytest.mark.parametrize("assume_centered", [True, False])
+def test_sklearnex_import_covariance(dataframe, queue, macro_block, assume_centered):
     from sklearnex.preview.covariance import EmpiricalCovariance
 
     X = np.array([[0, 1], [0, 1]])
+
     X = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
-    empcov = EmpiricalCovariance()
+    empcov = EmpiricalCovariance(assume_centered=assume_centered)
     if daal_check_version((2024, "P", 0)) and macro_block is not None:
         hparams = empcov.get_hyperparameters("fit")
         hparams.cpu_macro_block = macro_block
     result = empcov.fit(X)
+
     expected_covariance = np.array([[0, 0], [0, 0]])
-    expected_means = np.array([0, 1])
+    expected_means = np.array([0, 0])
+
+    if assume_centered:
+        expected_covariance = np.array([[0, 0], [0, 1]])
+    else:
+        expected_means = np.array([0, 1])
 
     assert_allclose(expected_covariance, result.covariance_)
     assert_allclose(expected_means, result.location_)
 
     X = np.array([[1, 2], [3, 6]])
+
     X = _convert_to_dataframe(X, sycl_queue=queue, target_df=dataframe)
     result = empcov.fit(X)
-    expected_covariance = np.array([[1, 2], [2, 4]])
-    expected_means = np.array([2, 4])
+
+    if assume_centered:
+        expected_covariance = np.array([[5, 10], [10, 20]])
+    else:
+        expected_covariance = np.array([[1, 2], [2, 4]])
+        expected_means = np.array([2, 4])
 
     assert_allclose(expected_covariance, result.covariance_)
     assert_allclose(expected_means, result.location_)