gh-3243: propose a different caching approach for sentence

flair/data.py

@@ -547,25 +547,9 @@ def set_label(self, typename: str, value: str, score: float = 1.0):
 class Span(_PartOfSentence):
     """This class represents one textual span consisting of Tokens."""
 
-    def __new__(self, tokens: List[Token]):
-        # check if the span already exists. If so, return it
-        unlabeled_identifier = self._make_unlabeled_identifier(tokens)
-        if unlabeled_identifier in tokens[0].sentence._known_spans:
-            span = tokens[0].sentence._known_spans[unlabeled_identifier]
-            return span
-
-        # else make a new span
-        else:
-            span = super().__new__(self)
-            span.initialized = False
-            tokens[0].sentence._known_spans[unlabeled_identifier] = span
-            return span
-
     def __init__(self, tokens: List[Token]) -> None:
-        if not self.initialized:
-            super().__init__(tokens[0].sentence)
-            self.tokens = tokens
-            self.initialized: bool = True
+        super().__init__(tokens[0].sentence)
+        self.tokens = tokens
 
     @property
     def start_position(self) -> int:
@@ -606,26 +590,10 @@ def embedding(self):
 
 
 class Relation(_PartOfSentence):
-    def __new__(self, first: Span, second: Span):
-        # check if the relation already exists. If so, return it
-        unlabeled_identifier = self._make_unlabeled_identifier(first, second)
-        if unlabeled_identifier in first.sentence._known_spans:
-            span = first.sentence._known_spans[unlabeled_identifier]
-            return span
-
-        # else make a new relation
-        else:
-            span = super().__new__(self)
-            span.initialized = False
-            first.sentence._known_spans[unlabeled_identifier] = span
-            return span
-
     def __init__(self, first: Span, second: Span) -> None:
-        if not self.initialized:
-            super().__init__(sentence=first.sentence)
-            self.first: Span = first
-            self.second: Span = second
-            self.initialized: bool = True
+        super().__init__(sentence=first.sentence)
+        self.first: Span = first
+        self.second: Span = second
 
     def __repr__(self) -> str:
         return str(self)
@@ -692,7 +660,7 @@ def __init__(
         self.tokens: List[Token] = []
 
         # private field for all known spans
-        self._known_spans: Dict[str, _PartOfSentence] = {}
+        self._known_parts: Dict[str, _PartOfSentence] = {}
 
         self.language_code: Optional[str] = language_code
 
@@ -769,7 +737,7 @@ def get_relations(self, type: str) -> List[Relation]:
 
     def get_spans(self, type: str) -> List[Span]:
         spans: List[Span] = []
-        for potential_span in self._known_spans.values():
+        for potential_span in self._known_parts.values():
             if isinstance(potential_span, Span) and potential_span.has_label(type):
                 spans.append(potential_span)
         return sorted(spans)
@@ -949,8 +917,7 @@ def to_dict(self, tag_type: Optional[str] = None):
         return {"text": self.to_original_text(), "all labels": labels}
 
     def get_span(self, start: int, stop: int):
-        span_slice = slice(start, stop)
-        return self[span_slice]
+        return self[start:stop]
 
     @typing.overload
     def __getitem__(self, idx: int) -> Token:
@@ -960,9 +927,27 @@ def __getitem__(self, idx: int) -> Token:
     def __getitem__(self, s: slice) -> Span:
         ...
 
+    @typing.overload
+    def __getitem__(self, s: typing.Tuple[Span, Span]) -> Relation:
+        ...
+
     def __getitem__(self, subscript):
-        if isinstance(subscript, slice):
-            return Span(self.tokens[subscript])
+        if isinstance(subscript, tuple):
+            first, second = subscript
+            identifier = ""
+            if isinstance(first, Span) and isinstance(second, Span):
+                identifier = Relation._make_unlabeled_identifier(first, second)
+                if identifier not in self._known_parts:
+                    self._known_parts[identifier] = Relation(first, second)
+
+            return self._known_parts[identifier]
+        elif isinstance(subscript, slice):
+            identifier = Span._make_unlabeled_identifier(self.tokens[subscript])
+
+            if identifier not in self._known_parts:
+                self._known_parts[identifier] = Span(self.tokens[subscript])
+
+            return self._known_parts[identifier]
         else:
             return self.tokens[subscript]
 
@@ -1108,11 +1093,11 @@ def remove_labels(self, typename: str):
             token.remove_labels(typename)
 
         # labels also need to be deleted at all known spans
-        for span in self._known_spans.values():
+        for span in self._known_parts.values():
             span.remove_labels(typename)
 
         # remove spans without labels
-        self._known_spans = {k: v for k, v in self._known_spans.items() if len(v.labels) > 0}
+        self._known_parts = {k: v for k, v in self._known_parts.items() if len(v.labels) > 0}
 
         # delete labels at object itself
         super().remove_labels(typename)

flair/datasets/sequence_labeling.py

@@ -26,7 +26,6 @@
     Corpus,
     FlairDataset,
     MultiCorpus,
-    Relation,
     Sentence,
     Token,
     get_spans_from_bio,
@@ -684,9 +683,7 @@ def _convert_lines_to_sentence(
                     tail_end = int(indices[3])
                     label = indices[4]
                     # head and tail span indices are 1-indexed and end index is inclusive
-                    relation = Relation(
-                        first=sentence[head_start - 1 : head_end], second=sentence[tail_start - 1 : tail_end]
-                    )
+                    relation = sentence[sentence[head_start - 1 : head_end], sentence[tail_start - 1 : tail_end]]
                     remapped = self._remap_label(label)
                     if remapped != "O":
                         relation.add_label(typename="relation", value=remapped)

flair/models/regexp_tagger.py

@@ -38,7 +38,7 @@ def get_token_span(self, span: Tuple[int, int]) -> Span:
         """
         span_start: int = self.__tokens_start_pos.index(span[0])
         span_end: int = self.__tokens_end_pos.index(span[1])
-        return Span(self.tokens[span_start : span_end + 1])
+        return self.sentence[span_start : span_end + 1]
 
 
 class RegexpTagger:

flair/models/relation_classifier_model.py

@@ -377,11 +377,9 @@ def _entity_pair_permutations(
         """
         valid_entities: List[_Entity] = list(self._valid_entities(sentence))
 
-        # Use a dictionary to find gold relation annotations for a given entity pair
-        relation_to_gold_label: Dict[str, str] = {
-            relation.unlabeled_identifier: relation.get_label(self.label_type, zero_tag_value=self.zero_tag_value).value
-            for relation in sentence.get_relations(self.label_type)
-        }
+        # ensure that all existing relations without label have the label set to zero_tag_value.
+        for relation in sentence.get_relations(self.label_type):
+            relation.set_label(self.label_type, relation.get_label(self.label_type, self.zero_tag_value).value)
 
         # Yield head and tail entity pairs from the cross product of all entities
         for head, tail in itertools.product(valid_entities, repeat=2):
@@ -398,9 +396,8 @@ def _entity_pair_permutations(
                 continue
 
             # Obtain gold label, if existing
-            original_relation: Relation = Relation(first=head.span, second=tail.span)
-            gold_label: Optional[str] = relation_to_gold_label.get(original_relation.unlabeled_identifier)
-
+            gold_relation = sentence[head.span, tail.span]
+            gold_label: Optional[str] = gold_relation.get_label(self.label_type, zero_tag_value=None).value
             yield head, tail, gold_label
 
     def _encode_sentence(
@@ -481,7 +478,7 @@ def _encode_sentence_for_inference(
                 tail=tail,
                 gold_label=gold_label if gold_label is not None else self.zero_tag_value,
             )
-            original_relation: Relation = Relation(first=head.span, second=tail.span)
+            original_relation: Relation = sentence[head.span, tail.span]
             yield masked_sentence, original_relation
 
     def _encode_sentence_for_training(self, sentence: Sentence) -> Iterator[EncodedSentence]:

flair/models/relation_extractor_model.py

@@ -79,7 +79,7 @@ def _get_data_points_from_sentence(self, sentence: Sentence) -> List[Relation]:
                 ):
                     continue
 
-                relation = Relation(span_1, span_2)
+                relation = sentence[span_1, span_2]
                 if self.training and self.train_on_gold_pairs_only and relation.get_label(self.label_type).value == "O":
                     continue
                 entity_pairs.append(relation)

flair/models/tars_model.py

@@ -404,10 +404,9 @@ def _get_tars_formatted_sentence(self, label, sentence):
 
         for entity_label in sentence.get_labels(self.label_type):
             if entity_label.value == label:
-                new_span = Span(
-                    [tars_sentence.get_token(token.idx + label_length) for token in entity_label.data_point]
-                )
-                new_span.add_label(self.static_label_type, value="entity")
+                start_pos = entity_label.data_point[0].idx + label_length - 1
+                end_pos = entity_label.data_point[-1].idx + label_length
+                tars_sentence[start_pos:end_pos].add_label(self.static_label_type, value="entity")
         tars_sentence.copy_context_from_sentence(sentence)
         return tars_sentence
 
@@ -588,9 +587,10 @@ def predict(
                             # only add if all tokens have no label
                             if tag_this:
                                 # make and add a corresponding predicted span
-                                predicted_span = Span(
-                                    [sentence.get_token(token.idx - label_length) for token in label.data_point]
-                                )
+                                start_pos = label.data_point.data_point[0].idx - label_length - 1
+                                end_pos = label.data_point.data_point[-1].idx - label_length
+
+                                predicted_span = sentence[start_pos:end_pos]
                                 predicted_span.add_label(label_name, value=label.value, score=label.score)
 
                                 # set indices so that no token can be tagged twice

tests/test_labels.py

@@ -189,9 +189,9 @@ def test_relation_tags():
     sentence = Sentence("Humboldt Universität zu Berlin is located in Berlin .")
 
     # create two relation label
-    Relation(sentence[0:4], sentence[7:8]).add_label("rel", "located in")
-    Relation(sentence[0:2], sentence[3:4]).add_label("rel", "university of")
-    Relation(sentence[0:2], sentence[3:4]).add_label("syntactic", "apposition")
+    sentence[sentence[0:4], sentence[7:8]].add_label("rel", "located in")
+    sentence[sentence[0:2], sentence[3:4]].add_label("rel", "university of")
+    sentence[sentence[0:2], sentence[3:4]].add_label("syntactic", "apposition")
 
     # there should be two relation labels
     labels: List[Label] = sentence.get_labels("rel")

tests/test_sentence.py

@@ -1,3 +1,6 @@
+import copy
+import pickle
+
 from flair.data import Sentence
 
 
@@ -73,3 +76,37 @@ def test_start_end_position_pretokenized() -> None:
         (10, 18),
         (19, 20),
     ]
+
+
+def test_spans_support_deepcopy() -> None:
+    sentence = Sentence(["I", "live", "in", "Vienna", "."])
+    sentence[3:4].add_label("ner", "LOC")
+
+    _ = copy.deepcopy(sentence)
+
+
+def test_spans_support_pickle() -> None:
+    sentence = Sentence(["I", "live", "in", "Vienna", "."])
+    sentence[3:4].add_label("ner", "LOC")
+
+    pickle_data = pickle.dumps(sentence)
+    _ = pickle.loads(pickle_data)
+
+
+def test_relations_support_deepcopy() -> None:
+    sentence = Sentence(["Vienna", "is", "the", "capital", "of", "Austria"])
+    sentence[0:1].add_label("ner", "LOC")
+    sentence[5:6].add_label("ner", "LOC")
+    sentence[sentence[0:1], sentence[5:6]].add_label("rel", "capital")
+
+    _ = copy.deepcopy(sentence)
+
+
+def test_relations_support_pickle() -> None:
+    sentence = Sentence(["Vienna", "is", "the", "capital", "of", "Austria"])
+    sentence[0:1].add_label("ner", "LOC")
+    sentence[5:6].add_label("ner", "LOC")
+    sentence[sentence[0:1], sentence[5:6]].add_label("rel", "capital")
+
+    pickle_data = pickle.dumps(sentence)
+    _ = pickle.loads(pickle_data)