bug fix

Author: voorhs

autointent/_ranker.py

@@ -3,6 +3,7 @@
 Can be used to rank retrieved sentences by meaning closeness to provided utterance.
 """
 
+import gc
 import itertools as it
 import json
 import logging
@@ -274,6 +275,7 @@ def load(cls, path: Path) -> "Ranker":
         return cls(**metadata, classifier_head=clf)
 
     def clear_ram(self) -> None:
-        self.cross_encoder.cpu()
+        self.cross_encoder.model.cpu()
         del self.cross_encoder
+        gc.collect()
         torch.cuda.empty_cache()

autointent/modules/scoring/_knn/knn.py

@@ -124,15 +124,15 @@ def get_embedder_name(self) -> str:
         """
         return self.embedder_name
 
-    def fit(self, utterances: list[str], labels: ListOfLabels) -> None:
+    def fit(self, utterances: list[str], labels: ListOfLabels, clear_cache: bool = False) -> None:
         """
         Fit the scorer by training or loading the vector index.
 
         :param utterances: List of training utterances.
         :param labels: List of labels corresponding to the utterances.
         :raises ValueError: If the vector index mismatches the provided utterances.
         """
-        if hasattr(self, "_vector_index"):
+        if hasattr(self, "_vector_index") and clear_cache:
             self.clear_cache()
 
         self._validate_task(labels)

autointent/modules/scoring/_knn/rerank_scorer.py

@@ -139,10 +139,11 @@ def fit(self, utterances: list[str], labels: ListOfLabels) -> None:
         )
         self._scorer.fit(utterances, labels)
 
-        super().fit(utterances, labels)
+        super().fit(utterances, labels, clear_cache=False)
 
     def clear_cache(self) -> None:
         self._scorer.clear_ram()
+        super().clear_cache()
 
     def _predict(self, utterances: list[str]) -> tuple[npt.NDArray[Any], list[list[str]]]:
         """