Fix CastTransformer output type (#1200)

* update changelogs

Signed-off-by: xadupre <[email protected]>

* Fix unknown_value=np.nan in OrdinalEncoder

Signed-off-by: xadupre <[email protected]>

* changelogs

Signed-off-by: xadupre <[email protected]>

* Fix CastTransformer output type

Signed-off-by: xadupre <[email protected]>

* changes

Signed-off-by: xadupre <[email protected]>

* disable the test for old version of scikit-learn

---------

Signed-off-by: xadupre <[email protected]>

Author: xadupre

CHANGELOGS.md

@@ -2,6 +2,8 @@
 
 ## 1.20.0
 
+* Fixes CastTransformer output type
+  [#1200](https://github.com/onnx/sklearn-onnx/issues/1200)
 * Fixes unknown_value=np.nan in OrdinalEncoder
   [#1198](https://github.com/onnx/sklearn-onnx/issues/1198)
 * Enhance OrdinalEncoder conversion to handle infrequent categories

skl2onnx/_parse.py

@@ -52,6 +52,7 @@ class OutlierMixin:
 from .common.data_types import (
     DictionaryType,
     Int64TensorType,
+    DoubleTensorType,
     SequenceType,
     StringTensorType,
     TensorType,
@@ -62,6 +63,7 @@ class OutlierMixin:
 from .common.utils_checking import check_signature
 from .common.utils_classifier import get_label_classes
 from .common.utils_sklearn import _process_options
+from .sklapi import CastTransformer
 
 
 do_not_merge_columns = tuple(
@@ -250,6 +252,18 @@ def _parse_sklearn_simple_model(scope, model, inputs, custom_parsers=None, alias
             otype = guess_tensor_type(inputs[0].type)
         variable = scope.declare_local_variable("variable", otype)
         this_operator.outputs.append(variable)
+    elif type(model) in {CastTransformer}:
+        dtype = model.dtype
+        if dtype == np.float32:
+            cls = FloatTensorType
+        elif dtype == np.float64:
+            cls = DoubleTensorType
+        elif dtype == np.int64:
+            cls = Int64TensorType
+        else:
+            raise NotImplementedError(f"Unexpected dtype={dtype} for model={model}")
+        variable = scope.declare_local_variable("cast", cls())
+        this_operator.outputs.append(variable)
     else:
         if hasattr(model, "get_feature_names_out"):
             try:

tests/test_sklearn_imputer_converter.py

@@ -10,6 +10,8 @@
 import pandas as pd
 from numpy.testing import assert_almost_equal
 import sklearn
+from sklearn.tree import DecisionTreeRegressor
+from sklearn.pipeline import Pipeline
 
 try:
     from sklearn.preprocessing import Imputer
@@ -24,8 +26,10 @@
 
 from onnxruntime import __version__ as ort_version
 
+from skl2onnx.sklapi import CastTransformer
 from skl2onnx import convert_sklearn
 from skl2onnx.common.data_types import (
+    DoubleTensorType,
     FloatTensorType,
     Int64TensorType,
     StringTensorType,
@@ -125,6 +129,37 @@ def test_simple_imputer_float_inputs(self):
             basename="SklearnSimpleImputerMeanFloat32",
         )
 
+    @unittest.skipIf(SimpleImputer is None, reason="SimpleImputer changed in 0.20")
+    @unittest.skipIf(
+        pv.Version(skl_ver) <= pv.Version("1.4.0"),
+        reason="unexpected pipeline + transform",
+    )
+    def test_simple_imputer_double_inputs(self):
+        model = Pipeline(
+            [
+                ("cast32", CastTransformer(dtype=np.float32)),
+                ("imputer", SimpleImputer()),
+                ("dt", DecisionTreeRegressor(max_depth=2)),
+            ]
+        )
+        data = np.array([[1, 2], [np.nan, 3], [7, 6]], dtype=np.float64)
+        y = [0, 1, 0]
+        model.fit(data, y)
+
+        model_onnx = convert_sklearn(
+            model,
+            "double",
+            [("input", DoubleTensorType([None, 2]))],
+            final_types=[("y", FloatTensorType([None, 1]))],
+            target_opset=TARGET_OPSET,
+        )
+        dump_data_and_model(
+            np.array(data, dtype=np.float64),
+            model,
+            model_onnx,
+            basename="SklearnSimpleImputerDoubleInputs",
+        )
+
     @unittest.skipIf(SimpleImputer is None, reason="SimpleImputer changed in 0.20")
     @unittest.skipIf(
         pv.Version(ort_version) <= pv.Version("1.11.0"),