Only keep models starting with specified model prefix (SNB_)

Thus, the baseline model (starting with (B_)) is taken account of, but
is not used (so that regressions are not entailed).

Author: popescu-v

khiops/sklearn/estimators.py

@@ -284,6 +284,19 @@ def _get_main_dictionary(self):
         self._assert_is_fitted()
         return self.model_.get_dictionary(self.model_main_dictionary_name_)
 
+    def _read_model_from_dictionary_file(self, model_dictionary_file_path):
+        """Removes dictionaries that do not have the model prefix in their name
+
+        This function is necessary for the regression case because Khiops generates a
+        baseline model which has to be removed for sklearn predictor.
+        """
+        model = kh.read_dictionary_file(model_dictionary_file_path)
+        assert self._khiops_model_prefix is not None
+        for dictionary_name in [kdic.name for kdic in model.dictionaries]:
+            if not dictionary_name.startswith(self._khiops_model_prefix):
+                model.remove_dictionary(dictionary_name)
+        return model
+
     def export_report_file(self, report_file_path):
         """Exports the model report to a JSON file
 
@@ -309,7 +322,7 @@ def export_dictionary_file(self, dictionary_file_path):
 
     def _import_model(self, kdic_path):
         """Sets model instance attribute by importing model from ``.kdic``"""
-        self.model_ = kh.read_dictionary_file(kdic_path)
+        self.model_ = self._read_model_from_dictionary_file(kdic_path)
 
     def _get_output_dir(self, fallback_dir):
         if self.output_dir:
@@ -808,7 +821,7 @@ def _fit_train_model(self, ds, computation_dir, **kwargs):
 
         # Update the `model_` attribute of the coclustering estimator to the
         # new coclustering model
-        self.model_ = kh.read_dictionary_file(
+        self.model_ = self._read_model_from_dictionary_file(
             fs.get_child_path(
                 output_dir, f"{self.model_main_dictionary_name_}_deployed.kdic"
             )
@@ -1021,7 +1034,7 @@ def _simplify(
 
             # Set the `model_` attribute of the new coclustering estimator to
             # the new coclustering model
-            simplified_cc.model_ = kh.read_dictionary_file(
+            simplified_cc.model_ = self._read_model_from_dictionary_file(
                 fs.get_child_path(
                     output_dir, f"{self.model_main_dictionary_name_}_deployed.kdic"
                 )
@@ -1205,6 +1218,7 @@ def __init__(
         self.construction_rules = construction_rules
         self._original_target_dtype = None
         self._predicted_target_meta_data_tag = None
+        self._khiops_baseline_model_prefix = None
 
     def __sklearn_tags__(self):
         # If we don't implement this trivial method it's not found by the sklearn. This
@@ -1295,7 +1309,7 @@ def _fit_train_model(self, ds, computation_dir, **kwargs):
             return
 
         # Save the model domain object and report
-        self.model_ = kh.read_dictionary_file(model_kdic_file_path)
+        self.model_ = self._read_model_from_dictionary_file(model_kdic_file_path)
         self.model_report_ = kh.read_analysis_results_file(report_file_path)
 
     @abstractmethod
@@ -1384,15 +1398,24 @@ def _fit_training_post_process(self, ds):
             self.model_main_dictionary_name_ = self.model_.dictionaries[0].name
         else:
             for dictionary in self.model_.dictionaries:
-                assert dictionary.name.startswith(self._khiops_model_prefix), (
+
+                # The baseline model is mandatory for regression;
+                # absent for classification and encoding
+                assert dictionary.name.startswith(
+                    self._khiops_model_prefix
+                ) or dictionary.name.startswith(self._khiops_baseline_model_prefix), (
                     f"Dictionary '{dictionary.name}' "
-                    f"does not have prefix '{self._khiops_model_prefix}'"
-                )
-                initial_dictionary_name = dictionary.name.replace(
-                    self._khiops_model_prefix, "", 1
+                    f"does not have prefix '{self._khiops_model_prefix}' "
+                    f"or '{self._khiops_baseline_model_prefix}'."
                 )
-                if initial_dictionary_name == ds.main_table.name:
-                    self.model_main_dictionary_name_ = dictionary.name
+
+                # Skip baseline model
+                if dictionary.name.startswith(self._khiops_model_prefix):
+                    initial_dictionary_name = dictionary.name.replace(
+                        self._khiops_model_prefix, "", 1
+                    )
+                    if initial_dictionary_name == ds.main_table.name:
+                        self.model_main_dictionary_name_ = dictionary.name
         if self.model_main_dictionary_name_ is None:
             raise ValueError("No model dictionary after Khiops call")
 
@@ -2185,6 +2208,7 @@ def __init__(
             auto_sort=auto_sort,
         )
         self._khiops_model_prefix = "SNB_"
+        self._khiops_baseline_model_prefix = "B_"
         self._predicted_target_meta_data_tag = "Mean"
         self._predicted_target_name_prefix = "M"
         self._original_target_dtype = np.float64
@@ -2286,6 +2310,9 @@ def predict(self, X):
             - str (a path for the file containing the array) if X is a dataset spec
               containing file-path tables.
         """
+        assert (
+            self._khiops_baseline_model_prefix is not None
+        ), "Baseline model prefix is not set (mandatory for regression)"
         # Call the parent's method
         y_pred = super().predict(X)