Merge pull request #3621 from pycaret/fix_3606

Force np.float64 for CBLOF

pycaret/containers/models/anomaly.py

@@ -181,7 +181,7 @@ class CBLOFAnomalyContainer(AnomalyContainer):
     def __init__(self, experiment):
         get_logger()
         np.random.seed(experiment.seed)
-        from pyod.models.cblof import CBLOF
+        from pycaret.internal.patches.pyod import CBLOFForceToDouble
 
         args = {
             "random_state": experiment.seed,
@@ -194,7 +194,7 @@ def __init__(self, experiment):
         super().__init__(
             id="cluster",
             name="Clustering-Based Local Outlier",
-            class_def=CBLOF,
+            class_def=CBLOFForceToDouble,
             args=args,
             tune_grid=tune_grid,
             tune_distribution=tune_distributions,

pycaret/internal/patches/pyod.py

@@ -0,0 +1,30 @@
+import numpy as np
+import pandas as pd
+from pyod.models.cblof import CBLOF
+
+
+def convert_to_fp64(X):
+    if isinstance(X, pd.DataFrame):
+        X = X.astype(
+            {
+                col: np.float64
+                for col in X.columns
+                if X.dtypes[col] in (np.float32, np.float16)
+            }
+        )
+    elif X.dtype == np.float32:
+        X = X.astype(np.float64)
+    return X
+
+
+# Fixes https://github.com/pycaret/pycaret/issues/3606
+class CBLOFForceToDouble(CBLOF):
+    """CBLOF with forced float32 -> float64 conversion"""
+
+    def fit(self, X, y=None):
+        X = convert_to_fp64(X)
+        return super().fit(X, y)
+
+    def decision_function(self, X):
+        X = convert_to_fp64(X)
+        return super().decision_function(X)

pycaret/internal/pycaret_experiment/unsupervised_experiment.py

@@ -1099,10 +1099,10 @@ def _create_model(
                             )
                             model_fit_start = time.time()
                             pipeline_with_model.fit(data_X, **fit_kwargs)
-                        except Exception:
+                        except Exception as e:
                             raise RuntimeError(
                                 "Could not form valid cluster separation. Try a different dataset or model."
-                            )
+                            ) from e
                 else:
                     pipeline_with_model.fit(data_X, **fit_kwargs)
             model_fit_end = time.time()

tests/test_anomaly.py

@@ -30,6 +30,8 @@ def test_anomaly(data):
     # create model
     iforest = pycaret.anomaly.create_model("iforest", experiment_custom_tags={"tag": 1})
     knn = pycaret.anomaly.create_model("knn", experiment_custom_tags={"tag": 1})
+    # https://github.com/pycaret/pycaret/issues/3606
+    cluster = pycaret.anomaly.create_model("cluster", experiment_custom_tags={"tag": 1})
 
     # Plot model
     pycaret.anomaly.plot_model(iforest)
@@ -38,14 +40,18 @@ def test_anomaly(data):
     # assign model
     iforest_results = pycaret.anomaly.assign_model(iforest)
     knn_results = pycaret.anomaly.assign_model(knn)
+    cluster_results = pycaret.anomaly.assign_model(cluster)
     assert isinstance(iforest_results, pd.DataFrame)
     assert isinstance(knn_results, pd.DataFrame)
+    assert isinstance(cluster_results, pd.DataFrame)
 
     # predict model
     iforest_predictions = pycaret.anomaly.predict_model(model=iforest, data=data)
     knn_predictions = pycaret.anomaly.predict_model(model=knn, data=data)
+    cluster_predictions = pycaret.anomaly.predict_model(model=cluster, data=data)
     assert isinstance(iforest_predictions, pd.DataFrame)
     assert isinstance(knn_predictions, pd.DataFrame)
+    assert isinstance(cluster_predictions, pd.DataFrame)
 
     # get config
     X = pycaret.anomaly.get_config("X")