fix: handle empty features_in gracefully + initialize in __init__

datasieve/pipeline.py

@@ -23,6 +23,7 @@ def __init__(self, steps: List[Tuple] = [],
         self.fitparams: Dict[str, dict] = self._validate_fitparams(fitparams, steps)
         self.pandas_types: bool = False
         self.feature_list: list = []
+        self.features_in: list = []
         self.label_list: list = []
         self.step_strings: list = []
 
@@ -164,7 +165,14 @@ def _validate_arguments(self, X, y, sample_weight, fit=False, outlier_check=Fals
                 else:
                     self.label_list = [0]
         elif isinstance(X, pd.DataFrame) and not fit:
-            if list(X.columns) != list(self.features_in):
+            # If features_in is empty (pipeline loaded from cached state or
+            # never fit'd with DataFrame), populate from first call as a
+            # graceful fallback. Otherwise, validate.
+            if not list(self.features_in):
+                self.features_in = X.columns
+                self.feature_list = X.columns
+                self.pandas_types = True
+            elif list(X.columns) != list(self.features_in):
                 raise Exception(f"Pipeline expected {self.features_in} but got {X.columns}.")
         elif not isinstance(X, pd.DataFrame) and not fit and self.pandas_types:
             X = pd.DataFrame(X, columns=self.features_in)

tests/test_pipeline.py

@@ -103,3 +103,37 @@ def test_getitem(dummy_array_without_nans, dummy_array2_without_nans):
     di_values = pipeline["di"].di_values
 
     assert di_values.shape[0] == 100
+
+
+def test_pipeline_features_in_initialized_empty(dummy_df_without_nans):
+    """
+    Pipeline.features_in must be initialized as empty list (not undefined).
+    Regression test for issue #13167 (freqtrade): predict before fit raised
+    AttributeError because features_in did not exist.
+    """
+    pipeline = Pipeline([])
+    # Should not raise AttributeError
+    assert pipeline.features_in == []
+    assert pipeline.feature_list == []
+
+
+def test_pipeline_transform_without_prior_fit_graceful(dummy_df_without_nans):
+    """
+    If pipeline.transform() is called and features_in is empty (e.g. pipeline
+    loaded from cached state where attribute didn't persist), the validation
+    should gracefully populate from the first call instead of raising.
+
+    Regression test for issue #13167 (freqtrade).
+    """
+    import datasieve.transforms as ts
+    pipeline = Pipeline([("detect_constants", ts.VarianceThreshold(threshold=0))])
+
+    df = dummy_df_without_nans.copy()
+    Xdf, ydf = extract_features_and_labels(df)
+
+    # Simulate "pipeline loaded fresh" state (features_in empty)
+    assert pipeline.features_in == []
+
+    # transform should populate features_in instead of raising
+    Xdf_t, ydf_t, _ = pipeline.transform(Xdf, ydf)
+    assert list(pipeline.features_in) == list(Xdf.columns)