take care of global vectors in multiprocessing (#5081)

* restore load_nlp.VECTORS in the child process

* add unit test

* fix test

* remove unnecessary import

* add utf8 encoding

* import unicode_literals

spacy/_ml.py

@@ -296,8 +296,7 @@ def link_vectors_to_models(vocab):
     key = (ops.device, vectors.name)
     if key in thinc.extra.load_nlp.VECTORS:
         if thinc.extra.load_nlp.VECTORS[key].shape != data.shape:
-            # This is a hack to avoid the problem in #3853. Maybe we should
+            # This is a hack to avoid the problem in #3853.
-            # print a warning as well?
             old_name = vectors.name
             new_name = vectors.name + "_%d" % data.shape[0]
             user_warning(Warnings.W019.format(old=old_name, new=new_name))

spacy/language.py

@@ -3,6 +3,9 @@ from __future__ import absolute_import, unicode_literals
 
 import random
 import itertools
+
+from thinc.extra import load_nlp
+
 from spacy.util import minibatch
 import weakref
 import functools
@@ -856,7 +859,7 @@ class Language(object):
         procs = [
             mp.Process(
                 target=_apply_pipes,
-                args=(self.make_doc, pipes, rch, sch, Underscore.get_state()),
+                args=(self.make_doc, pipes, rch, sch, Underscore.get_state(), load_nlp.VECTORS),
             )
             for rch, sch in zip(texts_q, bytedocs_send_ch)
         ]
@@ -1112,7 +1115,7 @@ def _pipe(docs, proc, kwargs):
         yield doc
 
 
-def _apply_pipes(make_doc, pipes, receiver, sender, underscore_state):
+def _apply_pipes(make_doc, pipes, receiver, sender, underscore_state, vectors):
     """Worker for Language.pipe
 
     receiver (multiprocessing.Connection): Pipe to receive text. Usually
@@ -1120,8 +1123,10 @@ def _apply_pipes(make_doc, pipes, receiver, sender, underscore_state):
     sender (multiprocessing.Connection): Pipe to send doc. Usually created by
         `multiprocessing.Pipe()`
     underscore_state (tuple): The data in the Underscore class of the parent
+    vectors (dict): The global vectors data, copied from the parent
     """
     Underscore.load_state(underscore_state)
+    load_nlp.VECTORS = vectors
     while True:
         texts = receiver.get()
         docs = (make_doc(text) for text in texts)

spacy/tests/regression/test_issue4725.py

@@ -0,0 +1,26 @@
+# coding: utf8
+from __future__ import unicode_literals
+
+import numpy
+
+from spacy.lang.en import English
+from spacy.vocab import Vocab
+
+
+def test_issue4725():
+    # ensures that this runs correctly and doesn't hang or crash because of the global vectors
+    vocab = Vocab(vectors_name="test_vocab_add_vector")
+    data = numpy.ndarray((5, 3), dtype="f")
+    data[0] = 1.0
+    data[1] = 2.0
+    vocab.set_vector("cat", data[0])
+    vocab.set_vector("dog", data[1])
+
+    nlp = English(vocab=vocab)
+    ner = nlp.create_pipe("ner")
+    nlp.add_pipe(ner)
+    nlp.begin_training()
+    docs = ["Kurt is in London."] * 10
+    for _ in nlp.pipe(docs, batch_size=2, n_process=2):
+        pass
+

spacy/tests/regression/test_issue4849.py

@@ -3,7 +3,6 @@ from __future__ import unicode_literals
 
 from spacy.lang.en import English
 from spacy.pipeline import EntityRuler
-from spacy.tokens.underscore import Underscore
 
 
 def test_issue4849():

spacy/tests/regression/test_issue4903.py

@@ -1,10 +1,8 @@
 # coding: utf8
 from __future__ import unicode_literals
 
-import spacy
 from spacy.lang.en import English
 from spacy.tokens import Span, Doc
-from spacy.tokens.underscore import Underscore
 
 
 class CustomPipe: