Improve evaluation api

mlflow/models/__init__.py

@@ -24,14 +24,16 @@
 from .model import Model
 from .flavor_backend import FlavorBackend
 from ..utils.environment import infer_pip_requirements
-from .evaluation import evaluate, EvaluationDataset
+from .evaluation import evaluate, EvaluationArtifact, EvaluationResult, list_evaluators
 
 __all__ = [
     "Model",
     "FlavorBackend",
     "infer_pip_requirements",
     "evaluate",
-    "EvaluationDataset",
+    "EvaluationArtifact",
+    "EvaluationResult",
+    "list_evaluators",
 ]
 
 

mlflow/models/evaluation/__init__.py

@@ -1,18 +1,14 @@
 from mlflow.models.evaluation.base import (
     ModelEvaluator,
-    EvaluationDataset,
     EvaluationResult,
-    EvaluationMetrics,
     EvaluationArtifact,
     evaluate,
     list_evaluators,
 )
 
 __all__ = [
     "ModelEvaluator",
-    "EvaluationDataset",
     "EvaluationResult",
-    "EvaluationMetrics",
     "EvaluationArtifact",
     "evaluate",
     "list_evaluators",

mlflow/models/evaluation/default_evaluator.py

@@ -1,7 +1,6 @@
 import mlflow
 from mlflow.models.evaluation.base import (
     ModelEvaluator,
-    EvaluationMetrics,
     EvaluationResult,
 )
 from mlflow.entities.metric import Metric
@@ -326,22 +325,31 @@ def _log_model_explainability(self):
             #  but spark model input dataframe contains Vector type feature column
             #  which shap explainer does not support.
             #  To support this, we need expand the Vector type feature column into
-            #  multiple scaler feature columns and pass it to shap explainer.
+            #  multiple scalar feature columns and pass it to shap explainer.
             _logger.warning(
                 "Logging model explainability insights is not currently supported for PySpark "
                 "models."
             )
             return
 
-        if self.model_type == "classifier" and not all(
-            [isinstance(label, (numbers.Number, np.bool_)) for label in self.label_list]
-        ):
+        if not (np.issubdtype(self.y.dtype, np.number) or self.y.dtype == np.bool_):
+            # Note: python bool type inherits number type but np.bool_ does not inherit np.number.
             _logger.warning(
                 "Skip logging model explainability insights because it requires all label "
-                "values to be Number type."
+                "values to be number type or bool type."
             )
             return
 
+        feature_dtypes = list(self.X.dtypes) if isinstance(self.X, pd.DataFrame) else [self.X.dtype]
+        for feature_dtype in feature_dtypes:
+            if not np.issubdtype(feature_dtype, np.number):
+                _logger.warning(
+                    "Skip logging model explainability insights because it requires all feature "
+                    "values to be number type, and each feature column must only contain scalar "
+                    "values."
+                )
+                return
+
         try:
             import shap
             import matplotlib.pyplot as pyplot
@@ -652,7 +660,7 @@ def evaluate(
 
             self.X = dataset.features_data
             self.y = dataset.labels_data
-            self.metrics = EvaluationMetrics()
+            self.metrics = dict()
             self.artifacts = {}
 
             infered_model_type = _infer_model_type_by_labels(self.y)

tests/models/test_default_evaluator.py

@@ -40,9 +40,11 @@ def assert_dict_equal(d1, d2, rtol):
 def test_regressor_evaluation(linear_regressor_model_uri, diabetes_dataset):
     with mlflow.start_run() as run:
         result = evaluate(
-            model=linear_regressor_model_uri,
+            linear_regressor_model_uri,
+            diabetes_dataset._constructor_args["data"],
             model_type="regressor",
-            dataset=diabetes_dataset,
+            targets=diabetes_dataset._constructor_args["targets"],
+            dataset_name=diabetes_dataset.name,
             evaluators="default",
         )
 
@@ -81,9 +83,11 @@ def test_regressor_evaluation(linear_regressor_model_uri, diabetes_dataset):
 def test_multi_classifier_evaluation(multiclass_logistic_regressor_model_uri, iris_dataset):
     with mlflow.start_run() as run:
         result = evaluate(
-            model=multiclass_logistic_regressor_model_uri,
+            multiclass_logistic_regressor_model_uri,
+            iris_dataset._constructor_args["data"],
             model_type="classifier",
-            dataset=iris_dataset,
+            targets=iris_dataset._constructor_args["targets"],
+            dataset_name=iris_dataset.name,
             evaluators="default",
         )
 
@@ -132,9 +136,11 @@ def test_multi_classifier_evaluation(multiclass_logistic_regressor_model_uri, ir
 def test_bin_classifier_evaluation(binary_logistic_regressor_model_uri, breast_cancer_dataset):
     with mlflow.start_run() as run:
         result = evaluate(
-            model=binary_logistic_regressor_model_uri,
+            binary_logistic_regressor_model_uri,
+            breast_cancer_dataset._constructor_args["data"],
             model_type="classifier",
-            dataset=breast_cancer_dataset,
+            targets=breast_cancer_dataset._constructor_args["targets"],
+            dataset_name=breast_cancer_dataset.name,
             evaluators="default",
         )
 
@@ -184,9 +190,11 @@ def test_bin_classifier_evaluation(binary_logistic_regressor_model_uri, breast_c
 def test_spark_regressor_model_evaluation(spark_linear_regressor_model_uri, diabetes_spark_dataset):
     with mlflow.start_run() as run:
         result = evaluate(
-            model=spark_linear_regressor_model_uri,
+            spark_linear_regressor_model_uri,
+            diabetes_spark_dataset._constructor_args["data"],
             model_type="regressor",
-            dataset=diabetes_spark_dataset,
+            targets=diabetes_spark_dataset._constructor_args["targets"],
+            dataset_name=diabetes_spark_dataset.name,
             evaluators="default",
             evaluator_config={"log_model_explainability": True},
         )
@@ -222,9 +230,11 @@ def test_spark_regressor_model_evaluation(spark_linear_regressor_model_uri, diab
 def test_svm_classifier_evaluation(svm_model_uri, breast_cancer_dataset):
     with mlflow.start_run() as run:
         result = evaluate(
-            model=svm_model_uri,
+            svm_model_uri,
+            breast_cancer_dataset._constructor_args["data"],
             model_type="classifier",
-            dataset=breast_cancer_dataset,
+            targets=breast_cancer_dataset._constructor_args["targets"],
+            dataset_name=breast_cancer_dataset.name,
             evaluators="default",
         )