feat: added eps for zeros (#35)

* feat: added eps for zeros

* Add embedding normalization

---------

Co-authored-by: Egor Sergeenko <[email protected]>

Author: Darinochka

autointent/context/embedder.py

@@ -71,4 +71,6 @@ def embed(self, utterances: list[str]) -> npt.NDArray[np.float32]:
         )
         if self.max_length is not None:
             self.embedding_model.max_seq_length = self.max_length
-        return self.embedding_model.encode(utterances, convert_to_numpy=True, batch_size=self.batch_size)  # type: ignore[return-value]
+        return self.embedding_model.encode(
+            utterances, convert_to_numpy=True, batch_size=self.batch_size, normalize_embeddings=True,
+        )  # type: ignore[return-value]

autointent/metrics/scoring.py

@@ -24,7 +24,7 @@ def __call__(self, labels: LABELS_VALUE_TYPE, scores: SCORES_VALUE_TYPE) -> floa
         ...
 
 
-def scoring_log_likelihood(labels: LABELS_VALUE_TYPE, scores: SCORES_VALUE_TYPE) -> float:
+def scoring_log_likelihood(labels: LABELS_VALUE_TYPE, scores: SCORES_VALUE_TYPE, eps: float = 1e-10) -> float:
     """
     supports multiclass and multilabel
 
@@ -45,9 +45,10 @@ def scoring_log_likelihood(labels: LABELS_VALUE_TYPE, scores: SCORES_VALUE_TYPE)
     where `s[i,c]` is a predicted score of `i`th utterance having ground truth label `c`
     """
     labels_array, scores_array = transform(labels, scores)
+    scores_array[scores_array == 0] = eps
 
     if np.any((scores_array <= 0) | (scores_array > 1)):
-        msg = "One or more scores are not from [0,1]. It is incompatible with `scoring_log_likelihood` metric"
+        msg = "One or more scores are not from (0,1]. It is incompatible with `scoring_log_likelihood` metric"
         logger.error(msg)
         raise ValueError(msg)