Clean all caches

spacy/language.py

@@ -559,13 +559,14 @@ class Language(object):
                 nr_seen += 1
             elif len(old_refs) == 0:
                 old_refs, recent_refs = recent_refs, old_refs
-                self.vocab.strings._cleanup_stale_strings()
+                keys, strings = self.vocab.strings._cleanup_stale_strings()
+                self.vocab._reset_cache(keys, strings)
+                self.tokenizer._reset_cache(keys)
+                for string in original_strings_data:
+                    self.vocab.strings.add(string)
                 nr_seen = 0
         # We can't know which strings from the last batch have really expired.
-        # So we don't erase the strings — we just extend with the original
-        # content.
-        for string in original_strings_data:
-            self.vocab.strings.add(string)
+        # So we don't erase the strings.
 
     def to_disk(self, path, disable=tuple()):
         """Save the current state to a directory.  If a model is loaded, this

spacy/strings.pyx

@@ -250,15 +250,23 @@ cdef class StringStore:
             self.add(string)
 
     def _cleanup_stale_strings(self):
+        """
+        RETURNS (keys, strings): Dropped strings and keys that can be dropped from other places
+        """
         if self.hits.size() == 0:
             # If we don't have any hits, just skip cleanup
             return
 
         cdef vector[hash_t] tmp
+        dropped_strings = []
+        dropped_keys = []
         for i in range(self.keys.size()):
             key = self.keys[i]
             if self.hits.count(key) != 0:
                 tmp.push_back(key)
+            else:
+                dropped_keys.append(key)
+                dropped_strings.append(self[key])
 
         self.keys.swap(tmp)
         strings = list(self)
@@ -266,6 +274,8 @@ cdef class StringStore:
         # Here we have strings but hits to it should be reseted
         self.hits.clear()
 
+        return dropped_keys, dropped_strings
+
     cdef const Utf8Str* intern_unicode(self, unicode py_string):
         # 0 means missing, but we don't bother offsetting the index.
         cdef bytes byte_string = py_string.encode('utf8')

spacy/tokenizer.pyx

@@ -132,6 +132,11 @@ cdef class Tokenizer:
         for text in texts:
             yield self(text)
 
+    def _reset_cache(self, keys):
+        for k in keys:
+            del self._cache[k]
+            del self._specials[k]
+
     cdef int _try_cache(self, hash_t key, Doc tokens) except -1:
         cached = <_Cached*>self._cache.get(key)
         if cached == NULL:

spacy/vocab.pyx

@@ -465,6 +465,12 @@ cdef class Vocab:
             self._by_orth.set(lexeme.orth, lexeme)
             self.length += 1
 
+    def _reset_cache(self, keys, strings):
+        for k in keys:
+            del self._by_hash[k]
+
+        self._by_orth = PreshMap()
+
 
 def pickle_vocab(vocab):
     sstore = vocab.strings