Merge pull request #1424 from explosion/feature/streaming-data-memory-growth

💫 Fix streaming data memory growth (!!)

spacy/language.py

@@ -8,6 +8,7 @@ import random
 import ujson
 from collections import OrderedDict
 import itertools
+import weakref
 
 from .tokenizer import Tokenizer
 from .vocab import Vocab
@@ -510,8 +511,33 @@ class Language(object):
             else:
                 # Apply the function, but yield the doc
                 docs = _pipe(proc, docs)
+        # Track weakrefs of "recent" documents, so that we can see when they
+        # expire from memory. When they do, we know we don't need old strings.
+        # This way, we avoid maintaining an unbounded growth in string entries
+        # in the string store.
+        recent_refs = weakref.WeakSet()
+        old_refs = weakref.WeakSet()
+        original_strings_data = self.vocab.strings.to_bytes()
+        StringStore = self.vocab.strings.__class__
+        recent_strings = StringStore().from_bytes(original_strings_data)
+        nr_seen = 0
         for doc in docs:
             yield doc
+            for word in doc:
+                recent_strings.add(word.text)
+            recent_refs.add(doc)
+            if nr_seen < 10000:
+                old_refs.add(doc)
+                nr_seen += 1
+            elif len(old_refs) == 0:
+                # All the docs in the 'old' set have expired, so the only
+                # difference between the backup strings and the current
+                # string-store should be obsolete. We therefore swap out the
+                # old strings data.
+                old_refs, recent_refs = recent_refs, old_refs
+                self.vocab.strings._reset_and_load(recent_strings)
+                recent_strings = StringStore().from_bytes(original_strings_data)
+                nr_seen = 0
 
     def to_disk(self, path, disable=tuple()):
         """Save the current state to a directory.  If a model is loaded, this

spacy/strings.pxd

@@ -21,11 +21,9 @@ ctypedef union Utf8Str:
 
 cdef class StringStore:
     cdef Pool mem
-    cdef bint is_frozen
 
     cdef vector[hash_t] keys
     cdef public PreshMap _map
-    cdef public PreshMap _oov
 
     cdef const Utf8Str* intern_unicode(self, unicode py_string)
     cdef const Utf8Str* _intern_utf8(self, char* utf8_string, int length)

spacy/strings.pyx

@@ -86,8 +86,6 @@ cdef class StringStore:
         """
         self.mem = Pool()
         self._map = PreshMap()
-        self._oov = PreshMap()
-        self.is_frozen = freeze
         if strings is not None:
             for string in strings:
                 self.add(string)
@@ -243,21 +241,12 @@ cdef class StringStore:
             self.add(word)
         return self
 
-    def set_frozen(self, bint is_frozen):
+    def _reset_and_load(self, strings):
-        # TODO
-        self.is_frozen = is_frozen
-
-    def flush_oov(self):
-        self._oov = PreshMap()
-
-    def _reset_and_load(self, strings, freeze=False):
         self.mem = Pool()
         self._map = PreshMap()
-        self._oov = PreshMap()
         self.keys.clear()
         for string in strings:
             self.add(string)
-        self.is_frozen = freeze
 
     cdef const Utf8Str* intern_unicode(self, unicode py_string):
         # 0 means missing, but we don't bother offsetting the index.
@@ -272,18 +261,6 @@ cdef class StringStore:
         cdef Utf8Str* value = <Utf8Str*>self._map.get(key)
         if value is not NULL:
             return value
-        value = <Utf8Str*>self._oov.get(key)
-        if value is not NULL:
-            return value
-        if self.is_frozen:
-            # OOV store uses 32 bit hashes. Pretty ugly :(
-            key32 = hash32_utf8(utf8_string, length)
-            # Important: Make the OOV store own the memory. That way it's trivial
-            # to flush them all.
-            value = _allocate(self._oov.mem, <unsigned char*>utf8_string, length)
-            self._oov.set(key32, value)
-            return NULL
-
         value = _allocate(self.mem, <unsigned char*>utf8_string, length)
         self._map.set(key, value)
         self.keys.push_back(key)

spacy/tests/parser/test_parse_navigate.py

@@ -57,9 +57,9 @@ def test_parser_parse_navigate_consistency(en_tokenizer, text, heads):
     doc = get_doc(tokens.vocab, [t.text for t in tokens], heads=heads)
     for head in doc:
         for child in head.lefts:
-            assert child.head is head
+            assert child.head == head
         for child in head.rights:
-            assert child.head is head
+            assert child.head == head
 
 
 def test_parser_parse_navigate_child_consistency(en_tokenizer, text, heads):

spacy/tokens/doc.pxd

@@ -54,6 +54,8 @@ cdef class Doc:
 
     cdef public object noun_chunks_iterator
 
+    cdef object __weakref__
+
     cdef int push_back(self, LexemeOrToken lex_or_tok, bint has_space) except -1
 
     cpdef np.ndarray to_array(self, object features)

spacy/tokens/doc.pyx

@@ -140,7 +140,6 @@ cdef class Doc:
         self.user_span_hooks = {}
         self.tensor = numpy.zeros((0,), dtype='float32')
         self.user_data = {}
-        self._py_tokens = []
         self._vector = None
         self.noun_chunks_iterator = _get_chunker(self.vocab.lang)
         cdef unicode orth
@@ -209,10 +208,7 @@ cdef class Doc:
         if i < 0:
             i = self.length + i
         bounds_check(i, self.length, PADDING)
-        if self._py_tokens[i] is not None:
+        return Token.cinit(self.vocab, &self.c[i], i, self)
-            return self._py_tokens[i]
-        else:
-            return Token.cinit(self.vocab, &self.c[i], i, self)
 
     def __iter__(self):
         """Iterate over `Token`  objects, from which the annotations can be
@@ -226,10 +222,7 @@ cdef class Doc:
         """
         cdef int i
         for i in range(self.length):
-            if self._py_tokens[i] is not None:
+            yield Token.cinit(self.vocab, &self.c[i], i, self)
-                yield self._py_tokens[i]
-            else:
-                yield Token.cinit(self.vocab, &self.c[i], i, self)
 
     def __len__(self):
         """The number of tokens in the document.
@@ -535,7 +528,6 @@ cdef class Doc:
         self.length += 1
         # Set morphological attributes, e.g. by lemma, if possible
         self.vocab.morphology.assign_untagged(t)
-        self._py_tokens.append(None)
         return t.idx + t.lex.length + t.spacy
 
     @cython.boundscheck(False)
@@ -841,7 +833,6 @@ cdef class Doc:
         # Set the left/right children, left/right edges
         set_children_from_heads(self.c, self.length)
         # Clear the cached Python objects
-        self._py_tokens = [None] * self.length
         # Return the merged Python object
         return self[start]
 

spacy/tokens/token.pxd

@@ -19,10 +19,7 @@ cdef class Token:
         if offset < 0 or offset >= doc.length:
             msg = "Attempt to access token at %d, max length %d"
             raise IndexError(msg % (offset, doc.length))
-        if doc._py_tokens[offset] != None:
-            return doc._py_tokens[offset]
         cdef Token self = Token.__new__(Token, vocab, doc, offset)
-        doc._py_tokens[offset] = self
         return self
 
     #cdef inline TokenC struct_from_attrs(Vocab vocab, attrs):