[REVIEW] Support and enable convert_dtype in estimator predict

CHANGELOG.md

@@ -4,10 +4,10 @@
 
 ## Improvements
 - PR #2735: Update seed to random_state in random forest and associated tests
-
 - PR #2739: Use cusparse_wrappers.h from RAFT
 - PR #2729: Replace `cupy.sparse` with `cupyx.scipy.sparse`
 - PR #2749: Correct docs for python version used in cuml_dev conda environment
+- PR #2723: Support and enable convert_dtype in estimator predict
 
 ## Bug Fixes
 

python/cuml/common/base.pyx

@@ -355,7 +355,7 @@ class RegressorMixin:
         else:
             handle = None
 
-        preds = self.predict(X)
+        preds = self.predict(X, **kwargs)
         return r2_score(y, preds, handle=handle)
 
 

python/cuml/linear_model/elastic_net.pyx

@@ -219,13 +219,13 @@ class ElasticNet(Base, RegressorMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X, convert_dtype=False):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts `y` values for `X`.
 
         """
 
-        return self.solver_model.predict(X)
+        return self.solver_model.predict(X, convert_dtype=convert_dtype)
 
     def get_params(self, deep=True):
         """

python/cuml/linear_model/lasso.pyx

@@ -180,7 +180,7 @@ class Lasso(Base, RegressorMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X, convert_dtype=False):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts the y for X.
 

python/cuml/linear_model/linear_regression.pyx

@@ -293,7 +293,7 @@ class LinearRegression(Base, RegressorMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X, convert_dtype=False):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts `y` values for `X`.
 

python/cuml/linear_model/logistic_regression.pyx

@@ -316,7 +316,7 @@ class LogisticRegression(Base, ClassifierMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X, convert_dtype=False):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts the y for X.
 
@@ -329,7 +329,7 @@ class LogisticRegression(Base, ClassifierMixin):
                                                        probabilities',
                                        'shape': '(n_samples, n_classes)'})
     @with_cupy_rmm
-    def predict_proba(self, X, convert_dtype=False):
+    def predict_proba(self, X, convert_dtype=True):
         """
         Predicts the class probabilities for each class in X
 
@@ -345,7 +345,7 @@ class LogisticRegression(Base, ClassifierMixin):
                                        'description': 'Logaright of predicted \
                                                        class probabilities',
                                        'shape': '(n_samples, n_classes)'})
-    def predict_log_proba(self, X, convert_dtype=False):
+    def predict_log_proba(self, X, convert_dtype=True):
         """
         Predicts the log class probabilities for each class in X
 

python/cuml/linear_model/ridge.pyx

@@ -334,7 +334,7 @@ class Ridge(Base, RegressorMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X, convert_dtype=False):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts the y for X.
 

python/cuml/metrics/regression.pyx

@@ -31,7 +31,7 @@ from cuml.common.input_utils import input_to_cuml_array
 from cuml.common.memory_utils import with_cupy_rmm
 
 
-def r2_score(y, y_hat, convert_dtype=False, handle=None):
+def r2_score(y, y_hat, convert_dtype=True, handle=None):
     """
     Calculates r2 score between y and y_hat
 

python/cuml/svm/svc.pyx

@@ -402,7 +402,7 @@ class SVC(SVMBase, ClassifierMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts the class labels for X. The returned y values are the class
         labels associated to sign(decision_function(X)).
@@ -416,7 +416,7 @@ class SVC(SVMBase, ClassifierMixin):
             # prob_svc has numpy output type, change it if it is necessary:
             return _to_output(preds, out_type)
         else:
-            return super(SVC, self).predict(X, True)
+            return super(SVC, self).predict(X, True, convert_dtype)
 
     @generate_docstring(skip_parameters_heading=True,
                         return_values={'name': 'preds',

python/cuml/svm/svm_base.pyx

@@ -453,7 +453,7 @@ class SVMBase(Base):
             else:
                 self._unique_labels = None
 
-    def predict(self, X, predict_class):
+    def predict(self, X, predict_class, convert_dtype=True):
         """
         Predicts the y for X, where y is either the decision function value
         (if predict_class == False), or the label associated with X.
@@ -483,7 +483,11 @@ class SVMBase(Base):
         self._check_is_fitted('_model')
 
         X_m, n_rows, n_cols, pred_dtype = \
-            input_to_cuml_array(X, check_dtype=self.dtype)
+            input_to_cuml_array(
+                X,
+                check_dtype=self.dtype,
+                convert_to_dtype=(self.dtype if convert_dtype else None))
+
         cdef uintptr_t X_ptr = X_m.ptr
 
         preds = CumlArray.zeros(n_rows, dtype=self.dtype)

python/cuml/svm/svr.pyx

@@ -292,10 +292,10 @@ class SVR(SVMBase, RegressorMixin):
                                        'type': 'dense',
                                        'description': 'Predicted values',
                                        'shape': '(n_samples, 1)'})
-    def predict(self, X):
+    def predict(self, X, convert_dtype=True):
         """
         Predicts the values for X.
 
         """
 
-        return super(SVR, self).predict(X, False)
+        return super(SVR, self).predict(X, False, convert_dtype)

python/cuml/test/test_coordinate_descent.py

@@ -159,3 +159,33 @@ def test_elastic_net_default(datatype, nrows, column_info):
     sk_predict = elastic_sk.predict(X_test)
     sk_r2 = r2_score(y_test, sk_predict)
     assert cu_r2 >= sk_r2 - 0.07
+
+
+@pytest.mark.parametrize('train_dtype', [np.float32, np.float64])
+@pytest.mark.parametrize('test_dtype', [np.float64, np.float32])
+def test_elastic_net_predict_convert_dtype(train_dtype, test_dtype):
+    X, y = make_regression(n_samples=50, n_features=10,
+                           n_informative=5, random_state=0)
+    X = X.astype(train_dtype)
+    y = y.astype(train_dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
+                                                        random_state=0)
+
+    clf = cuElasticNet()
+    clf.fit(X_train, y_train)
+    clf.predict(X_test.astype(test_dtype))
+
+
+@pytest.mark.parametrize('train_dtype', [np.float32, np.float64])
+@pytest.mark.parametrize('test_dtype', [np.float64, np.float32])
+def test_lasso_predict_convert_dtype(train_dtype, test_dtype):
+    X, y = make_regression(n_samples=50, n_features=10,
+                           n_informative=5, random_state=0)
+    X = X.astype(train_dtype)
+    y = y.astype(train_dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
+                                                        random_state=0)
+
+    clf = cuLasso()
+    clf.fit(X_train, y_train)
+    clf.predict(X_test.astype(test_dtype))

python/cuml/test/test_linear_model.py

@@ -400,3 +400,46 @@ def test_logistic_regression_input_type_consistency(constructor, dtype):
 
     assert isinstance(clf.predict_proba(X), original_type)
     assert isinstance(clf.predict(X), original_type)
+
+
+@pytest.mark.parametrize('train_dtype', [np.float32, np.float64])
+@pytest.mark.parametrize('test_dtype', [np.float64, np.float32])
+def test_linreg_predict_convert_dtype(train_dtype, test_dtype):
+    X, y = make_regression(n_samples=50, n_features=10,
+                           n_informative=5, random_state=0)
+    X = X.astype(train_dtype)
+    y = y.astype(train_dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
+                                                        random_state=0)
+
+    clf = cuLinearRegression()
+    clf.fit(X_train, y_train)
+    clf.predict(X_test.astype(test_dtype))
+
+
+@pytest.mark.parametrize('train_dtype', [np.float32, np.float64])
+@pytest.mark.parametrize('test_dtype', [np.float64, np.float32])
+def test_ridge_predict_convert_dtype(train_dtype, test_dtype):
+    X, y = make_regression(n_samples=50, n_features=10,
+                           n_informative=5, random_state=0)
+    X = X.astype(train_dtype)
+    y = y.astype(train_dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
+                                                        random_state=0)
+
+    clf = cuRidge()
+    clf.fit(X_train, y_train)
+    clf.predict(X_test.astype(test_dtype))
+
+
+@pytest.mark.parametrize('train_dtype', [np.float32, np.float64])
+@pytest.mark.parametrize('test_dtype', [np.float64, np.float32])
+def test_logistic_predict_convert_dtype(train_dtype, test_dtype):
+    X, y = make_classification(n_samples=50, n_features=10, random_state=0)
+    X = X.astype(train_dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
+                                                        random_state=0)
+
+    clf = cuLog()
+    clf.fit(X_train, y_train)
+    clf.predict(X_test.astype(test_dtype))

python/cuml/test/test_svm.py

@@ -623,3 +623,22 @@ def test_svr_skl_cmp_weighted():
     sklSVR.fit(X, y, sample_weights)
 
     compare_svr(cuSVR, sklSVR, X, y)
+
+
+@pytest.mark.parametrize('classifier', [True, False])
+@pytest.mark.parametrize('train_dtype', [np.float32, np.float64])
+@pytest.mark.parametrize('test_dtype', [np.float64, np.float32])
+def test_svm_predict_convert_dtype(train_dtype, test_dtype, classifier):
+    X, y = make_classification(n_samples=50, random_state=0)
+
+    X = X.astype(train_dtype)
+    y = y.astype(train_dtype)
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
+                                                        random_state=0)
+
+    if classifier:
+        clf = cu_svm.SVC()
+    else:
+        clf = cu_svm.SVR()
+    clf.fit(X_train, y_train)
+    clf.predict(X_test.astype(test_dtype))