Avoid pickling Doc inputs passed to Language.pipe() (#10864)

* `Language.pipe()`: Serialize `Doc` objects to bytes when using multiprocessing to avoid pickling overhead

* `Doc.to_dict()`: Serialize `_context` attribute (keeping in line with `(un)pickle_doc()`

* Correct type annotations

* Fix typo

* `Doc`: Do not serialize `_context`

* `Language.pipe`: Send context objects to child processes, Simplify `as_tuples` handling

* Fix type annotation

* `Language.pipe`: Simplify `as_tuple` multiprocessor handling

* Cleanup code, fix typos

* MyPy fixes

* Move doc preparation function into `_multiprocessing_pipe`
Whitespace changes

* Remove superfluous comma

* Rename `prepare_doc` to `prepare_input`

* Update spacy/errors.py

* Undo renaming for error

Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>

spacy/errors.py

@@ -931,6 +931,7 @@ class Errors(metaclass=ErrorsWithCodes):
              "could not be aligned to token boundaries.")
     E1040 = ("Doc.from_json requires all tokens to have the same attributes. "
              "Some tokens do not contain annotation for: {partial_attrs}")
+    E1041 = ("Expected a string, Doc, or bytes as input, but got: {type}")
 
 
 # Deprecated model shortcuts, only used in errors and warnings

spacy/language.py

@@ -1090,16 +1090,21 @@ class Language:
             )
         return self.tokenizer(text)
 
-    def _ensure_doc(self, doc_like: Union[str, Doc]) -> Doc:
+    def _ensure_doc(self, doc_like: Union[str, Doc, bytes]) -> Doc:
-        """Create a Doc if need be, or raise an error if the input is not a Doc or a string."""
+        """Create a Doc if need be, or raise an error if the input is not
+        a Doc, string, or a byte array (generated by Doc.to_bytes())."""
         if isinstance(doc_like, Doc):
             return doc_like
         if isinstance(doc_like, str):
             return self.make_doc(doc_like)
-        raise ValueError(Errors.E866.format(type=type(doc_like)))
+        if isinstance(doc_like, bytes):
+            return Doc(self.vocab).from_bytes(doc_like)
+        raise ValueError(Errors.E1041.format(type=type(doc_like)))
 
-    def _ensure_doc_with_context(self, doc_like: Union[str, Doc], context: Any) -> Doc:
+    def _ensure_doc_with_context(
-        """Create a Doc if need be and add as_tuples context, or raise an error if the input is not a Doc or a string."""
+        self, doc_like: Union[str, Doc, bytes], context: _AnyContext
+    ) -> Doc:
+        """Call _ensure_doc to generate a Doc and set its context object."""
         doc = self._ensure_doc(doc_like)
         doc._context = context
         return doc
@@ -1519,7 +1524,6 @@ class Language:
 
         DOCS: https://spacy.io/api/language#pipe
         """
-        # Handle texts with context as tuples
         if as_tuples:
             texts = cast(Iterable[Tuple[Union[str, Doc], _AnyContext]], texts)
             docs_with_contexts = (
@@ -1597,8 +1601,21 @@ class Language:
         n_process: int,
         batch_size: int,
     ) -> Iterator[Doc]:
+        def prepare_input(
+            texts: Iterable[Union[str, Doc]]
+        ) -> Iterable[Tuple[Union[str, bytes], _AnyContext]]:
+            # Serialize Doc inputs to bytes to avoid incurring pickling
+            # overhead when they are passed to child processes. Also yield
+            # any context objects they might have separately (as they are not serialized).
+            for doc_like in texts:
+                if isinstance(doc_like, Doc):
+                    yield (doc_like.to_bytes(), cast(_AnyContext, doc_like._context))
+                else:
+                    yield (doc_like, cast(_AnyContext, None))
+
+        serialized_texts_with_ctx = prepare_input(texts)  # type: ignore
         # raw_texts is used later to stop iteration.
-        texts, raw_texts = itertools.tee(texts)
+        texts, raw_texts = itertools.tee(serialized_texts_with_ctx)  # type: ignore
         # for sending texts to worker
         texts_q: List[mp.Queue] = [mp.Queue() for _ in range(n_process)]
         # for receiving byte-encoded docs from worker
@@ -1618,7 +1635,13 @@ class Language:
         procs = [
             mp.Process(
                 target=_apply_pipes,
-                args=(self._ensure_doc, pipes, rch, sch, Underscore.get_state()),
+                args=(
+                    self._ensure_doc_with_context,
+                    pipes,
+                    rch,
+                    sch,
+                    Underscore.get_state(),
+                ),
             )
             for rch, sch in zip(texts_q, bytedocs_send_ch)
         ]
@@ -1631,12 +1654,12 @@ class Language:
             recv.recv() for recv in cycle(bytedocs_recv_ch)
         )
         try:
-            for i, (_, (byte_doc, byte_context, byte_error)) in enumerate(
+            for i, (_, (byte_doc, context, byte_error)) in enumerate(
                 zip(raw_texts, byte_tuples), 1
             ):
                 if byte_doc is not None:
                     doc = Doc(self.vocab).from_bytes(byte_doc)
-                    doc._context = byte_context
+                    doc._context = context
                     yield doc
                 elif byte_error is not None:
                     error = srsly.msgpack_loads(byte_error)
@@ -2163,7 +2186,7 @@ def _copy_examples(examples: Iterable[Example]) -> List[Example]:
 
 
 def _apply_pipes(
-    ensure_doc: Callable[[Union[str, Doc]], Doc],
+    ensure_doc: Callable[[Union[str, Doc, bytes], _AnyContext], Doc],
     pipes: Iterable[Callable[..., Iterator[Doc]]],
     receiver,
     sender,
@@ -2184,17 +2207,19 @@ def _apply_pipes(
     Underscore.load_state(underscore_state)
     while True:
         try:
-            texts = receiver.get()
+            texts_with_ctx = receiver.get()
-            docs = (ensure_doc(text) for text in texts)
+            docs = (
+                ensure_doc(doc_like, context) for doc_like, context in texts_with_ctx
+            )
             for pipe in pipes:
                 docs = pipe(docs)  # type: ignore[arg-type, assignment]
             # Connection does not accept unpickable objects, so send list.
             byte_docs = [(doc.to_bytes(), doc._context, None) for doc in docs]
-            padding = [(None, None, None)] * (len(texts) - len(byte_docs))
+            padding = [(None, None, None)] * (len(texts_with_ctx) - len(byte_docs))
             sender.send(byte_docs + padding)  # type: ignore[operator]
         except Exception:
             error_msg = [(None, None, srsly.msgpack_dumps(traceback.format_exc()))]
-            padding = [(None, None, None)] * (len(texts) - 1)
+            padding = [(None, None, None)] * (len(texts_with_ctx) - 1)
             sender.send(error_msg + padding)
 
 

spacy/tests/doc/test_pickle_doc.py

@@ -5,11 +5,9 @@ from spacy.compat import pickle
 def test_pickle_single_doc():
     nlp = Language()
     doc = nlp("pickle roundtrip")
-    doc._context = 3
     data = pickle.dumps(doc, 1)
     doc2 = pickle.loads(data)
     assert doc2.text == "pickle roundtrip"
-    assert doc2._context == 3
 
 
 def test_list_of_docs_pickles_efficiently():

spacy/tokens/doc.pyx

@@ -1880,18 +1880,17 @@ cdef int [:,:] _get_lca_matrix(Doc doc, int start, int end):
 def pickle_doc(doc):
     bytes_data = doc.to_bytes(exclude=["vocab", "user_data", "user_hooks"])
     hooks_and_data = (doc.user_data, doc.user_hooks, doc.user_span_hooks,
-                      doc.user_token_hooks, doc._context)
+                      doc.user_token_hooks)
     return (unpickle_doc, (doc.vocab, srsly.pickle_dumps(hooks_and_data), bytes_data))
 
 
 def unpickle_doc(vocab, hooks_and_data, bytes_data):
-    user_data, doc_hooks, span_hooks, token_hooks, _context = srsly.pickle_loads(hooks_and_data)
+    user_data, doc_hooks, span_hooks, token_hooks = srsly.pickle_loads(hooks_and_data)
 
     doc = Doc(vocab, user_data=user_data).from_bytes(bytes_data, exclude=["user_data"])
     doc.user_hooks.update(doc_hooks)
     doc.user_span_hooks.update(span_hooks)
     doc.user_token_hooks.update(token_hooks)
-    doc._context = _context
     return doc