Fix memory issues in Language.evaluate (#6386)

* Fix memory issues in Language.evaluate

Reset annotation in predicted docs before evaluating and store all data
in `examples`.

* Minor refactor to docs generator init

* Fix generator expression

* Fix final generator check

* Refactor pipeline loop

* Handle examples generator in Language.evaluate

* Add test with generator

* Use make_doc

spacy/language.py

@@ -1290,6 +1290,7 @@ class Language:
 
         DOCS: https://nightly.spacy.io/api/language#evaluate
         """
+        examples = list(examples)
         validate_examples(examples, "Language.evaluate")
         if batch_size is None:
             batch_size = self.batch_size
@@ -1301,27 +1302,21 @@ class Language:
             kwargs = dict(scorer_cfg)
             kwargs.setdefault("nlp", self)
             scorer = Scorer(**kwargs)
-        texts = [eg.reference.text for eg in examples]
+        # reset annotation in predicted docs and time tokenization
-        docs = [eg.predicted for eg in examples]
         start_time = timer()
-        # tokenize the texts only for timing purposes
+        for eg in examples:
-        if not hasattr(self.tokenizer, "pipe"):
+            eg.predicted = self.make_doc(eg.reference.text)
-            _ = [self.tokenizer(text) for text in texts]  # noqa: F841
+        # apply all pipeline components
-        else:
-            _ = list(self.tokenizer.pipe(texts))  # noqa: F841
         for name, pipe in self.pipeline:
             kwargs = component_cfg.get(name, {})
             kwargs.setdefault("batch_size", batch_size)
-            docs = _pipe(docs, pipe, kwargs)
+            for doc, eg in zip(
-        # iterate over the final generator
+                _pipe((eg.predicted for eg in examples), pipe, kwargs), examples
-        if len(self.pipeline):
+            ):
-            docs = list(docs)
+                eg.predicted = doc
         end_time = timer()
-        for i, (doc, eg) in enumerate(zip(docs, examples)):
-            util.logger.debug(doc)
-            eg.predicted = doc
         results = scorer.score(examples)
-        n_words = sum(len(doc) for doc in docs)
+        n_words = sum(len(eg.predicted) for eg in examples)
         results["speed"] = n_words / (end_time - start_time)
         return results
 

spacy/tests/test_language.py

@@ -53,7 +53,12 @@ def test_language_evaluate(nlp):
     annots = {"doc_annotation": {"cats": {"POSITIVE": 1.0, "NEGATIVE": 0.0}}}
     doc = Doc(nlp.vocab, words=text.split(" "))
     example = Example.from_dict(doc, annots)
-    nlp.evaluate([example])
+    scores = nlp.evaluate([example])
+    assert scores["speed"] > 0
+
+    # test with generator
+    scores = nlp.evaluate(eg for eg in [example])
+    assert scores["speed"] > 0
 
     # Not allowed to call with just one Example
     with pytest.raises(TypeError):

spacy/training/loop.py

@@ -249,9 +249,8 @@ def create_evaluation_callback(
     weights = {key: value for key, value in weights.items() if value is not None}
 
     def evaluate() -> Tuple[float, Dict[str, float]]:
-        dev_examples = list(dev_corpus(nlp))
         try:
-            scores = nlp.evaluate(dev_examples)
+            scores = nlp.evaluate(dev_corpus(nlp))
         except KeyError as e:
             raise KeyError(Errors.E900.format(pipeline=nlp.pipe_names)) from e
         # Calculate a weighted sum based on score_weights for the main score.