* Work on word vectors, and other stuff

spacy/en/__init__.py

@@ -1,5 +1,6 @@
 from __future__ import unicode_literals
 from os import path
+import re
 
 from .. import orth
 from ..vocab import Vocab
@@ -11,6 +12,9 @@ from .pos import POS_TAGS
 from .attrs import get_flags
 
 
+from ..util import read_lang_data
+
+
 def get_lex_props(string):
     return {
         'flags': get_flags(string),
@@ -64,11 +68,16 @@ class English(object):
         tag_names = list(POS_TAGS.keys())
         tag_names.sort()
         if data_dir is None:
-            self.tokenizer = Tokenizer(self.vocab, {}, None, None, None,
+            tok_rules = {}
-                                       POS_TAGS, tag_names)
+            prefix_re = None
+            suffix_re = None
+            infix_re = None
         else:
-            self.tokenizer = Tokenizer.from_dir(self.vocab, path.join(data_dir, 'tokenizer'),
+            tok_data_dir = path.join(data_dir, 'tokenizer')
-                                                POS_TAGS, tag_names)
+            tok_rules, prefix_re, suffix_re, infix_re = read_lang_data(tok_data_dir)
+        self.tokenizer = Tokenizer(self.vocab, tok_rules, re.compile(prefix_re),
+                                   re.compile(suffix_re), re.compile(infix_re),
+                                   POS_TAGS, tag_names)
         self.strings = self.vocab.strings
         self._tagger = None
         self._parser = None
@@ -100,11 +109,11 @@ class English(object):
         Returns:
             tokens (spacy.tokens.Tokens):
         """
-        tokens = self.tokenizer.tokenize(text)
+        tokens = self.tokenizer(text)
         if tag:
             self.tagger(tokens)
         if parse:
-            self.parser.parse(tokens)
+            self.parser(tokens)
         return tokens
 
     @property

spacy/en/attrs.pxd

@@ -1,18 +1,16 @@
-from ..typedefs cimport FLAG0, FLAG1, FLAG2, FLAG3, FLAG4, FLAG5, FLAG6, FLAG7
+from ..attrs cimport FLAG0, FLAG1, FLAG2, FLAG3, FLAG4, FLAG5, FLAG6, FLAG7
-from ..typedefs cimport FLAG8, FLAG9
+from ..attrs cimport FLAG8, FLAG9
-from ..typedefs cimport ID as _ID
+from ..attrs cimport SIC as _SIC
-from ..typedefs cimport SIC as _SIC
+from ..attrs cimport SHAPE as _SHAPE
-from ..typedefs cimport SHAPE as _SHAPE
+from ..attrs cimport NORM1 as _NORM1
-from ..typedefs cimport NORM1 as _NORM1
+from ..attrs cimport NORM2 as _NORM2
-from ..typedefs cimport NORM2 as _NORM2
+from ..attrs cimport CLUSTER as _CLUSTER
-from ..typedefs cimport CLUSTER as _CLUSTER
+from ..attrs cimport PREFIX as _PREFIX
-from ..typedefs cimport PREFIX as _PREFIX
+from ..attrs cimport SUFFIX as _SUFFIX
-from ..typedefs cimport SUFFIX as _SUFFIX
+from ..attrs cimport LEMMA as _LEMMA
-from ..typedefs cimport LEMMA as _LEMMA
+from ..attrs cimport POS as _POS
-from ..typedefs cimport POS as _POS
 
 
-# Work around the lack of global cpdef variables
 cpdef enum:
     IS_ALPHA = FLAG0
     IS_ASCII = FLAG1
@@ -25,7 +23,6 @@ cpdef enum:
     LIKE_URL = FLAG8
     LIKE_NUM = FLAG9
 
-    ID = _ID
     SIC = _SIC
     SHAPE = _SHAPE
     NORM1 = _NORM1

spacy/en/download.py

@@ -4,33 +4,52 @@ import tarfile
 import shutil
 import requests
 
-PARSER_URL = 'https://s3-us-west-1.amazonaws.com/media.spacynlp.com/en.tgz'
+PARSER_URL = 'http://s3-us-west-1.amazonaws.com/media.spacynlp.com/en.tgz'
 
-DEST_DIR = path.join(path.dirname(__file__), 'data', 'deps')
+DEP_VECTORS_URL = 'http://u.cs.biu.ac.il/~yogo/data/syntemb/deps.words.bz2'
+
+DEST_DIR = path.join(path.dirname(__file__), 'data')
 
 def download_file(url):
     local_filename = url.split('/')[-1]
+    return path.join(DEST_DIR, local_filename)
     # NOTE the stream=True parameter
     r = requests.get(url, stream=True)
+    print "Download %s" % url
+    i = 0
     with open(local_filename, 'wb') as f:
         for chunk in r.iter_content(chunk_size=1024): 
             if chunk: # filter out keep-alive new chunks
                 f.write(chunk)
                 f.flush()
+        print i
+        i += 1
     return local_filename
 
-def main():
+def install_parser_model(url, dest_dir):
-    if not os.path.exists(DEST_DIR):
+    if not os.path.exists(dest_dir):
-        os.mkdir(DEST_DIR)
+        os.mkdir(dest_dir)
-    assert not path.exists(path.join(DEST_DIR, 'en'))
+    assert not path.exists(path.join(dest_dir, 'en'))
 
-
+    filename = download_file(url)
-    filename = download_file(URL)
     t = tarfile.open(filename, mode=":gz")
-    t.extractall(DEST_DIR)
+    t.extractall(dest_dir)
-    shutil.move(path.join(DEST_DIR, 'en', 'deps', 'model'), DEST_DIR)
+    shutil.move(path.join(dest_dir, 'en', 'deps', 'model'), dest_dir)
-    shutil.move(path.join(DEST_DIR, 'en', 'deps', 'config.json'), DEST_DIR)
+    shutil.move(path.join(dest_dir, 'en', 'deps', 'config.json'), dest_dir)
-    shutil.rmtree(path.join(DEST_DIR, 'en'))
+    shutil.rmtree(path.join(dest_dir, 'en'))
+
+
+def install_dep_vectors(url, dest_dir):
+    if not os.path.exists(dest_dir):
+        os.mkdir(dest_dir)
+    
+    filename = download_file(url)
+    shutil.move(filename, path.join(dest_dir, 'vec.bz2'))
+
+
+def main():
+    #install_parser_model(PARSER_URL, path.join(DEST_DIR, 'deps'))
+    install_dep_vectors(DEP_VECTORS_URL, path.join(DEST_DIR, 'vocab'))
 
 
 if __name__ == '__main__':

spacy/en/pos.pyx

@@ -247,11 +247,12 @@ cdef class EnPosTagger:
         cdef atom_t[N_CONTEXT_FIELDS] context
         cdef const weight_t* scores
         for i in range(tokens.length):
-            if tokens.data[i].fine_pos == 0:
+            if tokens.data[i].pos == 0:
                 fill_context(context, i, tokens.data)
                 scores = self.model.score(context)
-                tokens.data[i].fine_pos = arg_max(scores, self.model.n_classes)
+                tokens.data[i].tag = arg_max(scores, self.model.n_classes)
                 self.set_morph(i, tokens.data)
+        tokens.pos_scheme = self.tag_map
 
     def train(self, Tokens tokens, object golds):
         cdef int i
@@ -263,13 +264,13 @@ cdef class EnPosTagger:
             scores = self.model.score(context)
             guess = arg_max(scores, self.model.n_classes)
             self.model.update(context, guess, golds[i], guess != golds[i])
-            tokens.data[i].fine_pos = guess
+            tokens.data[i].tag = guess
             self.set_morph(i, tokens.data)
             correct += guess == golds[i]
         return correct
 
     cdef int set_morph(self, const int i, TokenC* tokens) except -1:
-        cdef const PosTag* tag = &self.tags[tokens[i].fine_pos]
+        cdef const PosTag* tag = &self.tags[tokens[i].tag]
         tokens[i].pos = tag.pos
         cached = <_CachedMorph*>self._morph_cache.get(tag.id, tokens[i].lex.sic)
         if cached is NULL:

spacy/lexeme.pxd

@@ -3,39 +3,70 @@ from .typedefs cimport ID, SIC, NORM1, NORM2, SHAPE, PREFIX, SUFFIX, LENGTH, CLU
 from .structs cimport LexemeC
 from .strings cimport StringStore
 
+from numpy cimport ndarray
+
+
 
 cdef LexemeC EMPTY_LEXEME
 
 
-cdef int set_lex_struct_props(LexemeC* lex, dict props, StringStore strings) except -1
+cdef int set_lex_struct_props(LexemeC* lex, dict props, StringStore strings,
+                              const float* empty_vec) except -1
  
 cdef class Lexeme:
-    cdef const float* vec
+    cdef readonly ndarray vec
 
     cdef readonly flags_t flags
     cdef readonly attr_t id
     cdef readonly attr_t length
 
     cdef readonly attr_t sic
-    cdef readonly unicode norm1
+    cdef readonly attr_t norm1
-    cdef readonly unicode norm2
+    cdef readonly attr_t norm2
-    cdef readonly unicode shape
+    cdef readonly attr_t shape
-    cdef readonly unicode prefix
+    cdef readonly attr_t prefix
-    cdef readonly unicode suffix
+    cdef readonly attr_t suffix
 
-    cdef readonly attr_t sic_id
+    cdef readonly unicode sic_
-    cdef readonly attr_t norm1_id
+    cdef readonly unicode norm1_
-    cdef readonly attr_t norm2_id
+    cdef readonly unicode norm2_
-    cdef readonly attr_t shape_id
+    cdef readonly unicode shape_
-    cdef readonly attr_t prefix_id
+    cdef readonly unicode prefix_
-    cdef readonly attr_t suffix_id
+    cdef readonly unicode suffix_
 
     cdef readonly attr_t cluster
     cdef readonly float prob
     cdef readonly float sentiment
 
+    # Workaround for an apparent bug in the way the decorator is handled ---
+    # TODO: post bug report / patch to Cython.
+    @staticmethod
+    cdef inline Lexeme from_ptr(const LexemeC* ptr, StringStore strings):
+        cdef Lexeme py = Lexeme.__new__(Lexeme, 300)
+        for i in range(300):
+            py.vec[i] = ptr.vec[i]
+        py.flags = ptr.flags
+        py.id = ptr.id
+        py.length = ptr.length
 
-cdef Lexeme Lexeme_cinit(const LexemeC* c, StringStore strings)
+        py.sic = ptr.sic
+        py.norm1 = ptr.norm1
+        py.norm2 = ptr.norm2
+        py.shape = ptr.shape
+        py.prefix = ptr.prefix
+        py.suffix = ptr.suffix
+
+        py.sic_ = strings[ptr.sic]
+        py.norm1_ = strings[ptr.norm1]
+        py.norm2_ = strings[ptr.norm2]
+        py.shape_ = strings[ptr.shape]
+        py.prefix_ = strings[ptr.prefix]
+        py.suffix_ = strings[ptr.suffix]
+
+        py.cluster = ptr.cluster
+        py.prob = ptr.prob
+        py.sentiment = ptr.sentiment
+        return py
 
 
 cdef inline bint check_flag(const LexemeC* lexeme, attr_id_t flag_id) nogil:

spacy/lexeme.pyx

@@ -7,13 +7,14 @@ from libc.string cimport memset
 
 from .orth cimport word_shape
 from .typedefs cimport attr_t
+import numpy
 
 
 memset(&EMPTY_LEXEME, 0, sizeof(LexemeC))
 
 
-cdef int set_lex_struct_props(LexemeC* lex, dict props, StringStore string_store) except -1:
+cdef int set_lex_struct_props(LexemeC* lex, dict props, StringStore string_store,
-
+                              const float* empty_vec) except -1:
     lex.length = props['length']
     lex.sic = string_store[props['sic']]
     lex.norm1 = string_store[props['norm1']] 
@@ -27,39 +28,10 @@ cdef int set_lex_struct_props(LexemeC* lex, dict props, StringStore string_store
     lex.sentiment = props['sentiment']
 
     lex.flags = props['flags']
+    lex.vec = empty_vec
 
 
 cdef class Lexeme:
     """A dummy docstring"""
-    def __init__(self):
+    def __cinit__(self, int vec_size):
-        pass
+        self.vec = numpy.ndarray(shape=(vec_size,), dtype=numpy.float32)
-        
-
-cdef Lexeme Lexeme_cinit(const LexemeC* c, StringStore strings):
-    cdef Lexeme py = Lexeme.__new__(Lexeme)
-
-    py.vec = c.vec
-
-    py.flags = c.flags
-    py.id = c.id
-    py.length = c.length
-
-    py.sic = c.sic
-    py.norm1 = strings[c.norm1]
-    py.norm2 = strings[c.norm2]
-    py.shape = strings[c.shape]
-    py.prefix = strings[c.prefix]
-    py.suffix = strings[c.suffix]
-
-    py.sic_id = c.sic
-    py.norm1_id = c.norm1
-    py.norm2_id = c.norm2
-    py.shape_id = c.shape
-    py.prefix_id = c.prefix
-    py.suffix_id = c.suffix
-    
-    py.cluster = c.cluster
-
-    py.prob = c.prob
-    py.sentiment = c.sentiment
-    return py

spacy/orth.pyx

@@ -137,8 +137,42 @@ cpdef unicode word_shape(unicode string):
     return ''.join(shape)
 
 
+cpdef unicode norm1(unicode string, lower_pc=0.0, upper_pc=0.0, title_pc=0.0):
+    """Apply level 1 normalization:
+
+    * Case is canonicalized, using frequency statistics
+    * Unicode mapped to ascii, via unidecode
+    * Regional spelling variations are normalized
+    """
+    pass
+
+
 cpdef bytes asciied(unicode string):
     cdef str stripped = unidecode(string)
     if not stripped:
         return b'???'
     return stripped.encode('ascii')
+
+
+# Exceptions --- do not convert these
+_uk_us_except = set([
+    'our',
+    'ours',
+    'four',
+    'fours',
+    'your',
+    'yours',
+    'hour',
+    'hours',
+    'course',
+    'rise',
+])
+def uk_to_usa(unicode string):
+    if not string.islower():
+        return string
+    if string in _uk_us_except:
+        return string
+    our = re.compile(r'ours?$')
+    string = our.sub('or', string)
+
+    return string

spacy/structs.pxd

@@ -44,12 +44,12 @@ cdef struct TokenC:
     const LexemeC* lex
     Morphology morph
     univ_tag_t pos
-    int fine_pos
+    int tag
     int idx
     int lemma
     int sense
     int head
-    int dep_tag
+    int dep
     uint32_t l_kids
     uint32_t r_kids
 

spacy/syntax/_parse_features.pyx

@@ -28,7 +28,7 @@ cdef inline void fill_token(atom_t* context, const TokenC* token) nogil:
     else:
         context[0] = token.lex.sic
         context[1] = token.lemma
-        context[2] = token.fine_pos
+        context[2] = token.tag
         context[3] = token.lex.cluster
         # We've read in the string little-endian, so now we can take & (2**n)-1
         # to get the first n bits of the cluster.
@@ -44,7 +44,7 @@ cdef inline void fill_token(atom_t* context, const TokenC* token) nogil:
         # the source that are set to 1.
         context[4] = token.lex.cluster & 63
         context[5] = token.lex.cluster & 15
-        context[6] = token.dep_tag if has_head(token) else 0
+        context[6] = token.dep if has_head(token) else 0
 
 
 cdef int fill_context(atom_t* context, State* state) except -1:

spacy/syntax/_state.pyx

@@ -12,7 +12,7 @@ DEF NON_MONOTONIC = True
 cdef int add_dep(State *s, int head, int child, int label) except -1:
     cdef int dist = head - child
     s.sent[child].head = dist
-    s.sent[child].dep_tag = label
+    s.sent[child].dep = label
     # Keep a bit-vector tracking child dependencies.  If a word has a child at
     # offset i from it, set that bit (tracking left and right separately)
     if child > head:
@@ -38,7 +38,7 @@ cdef int push_stack(State *s) except -1:
     if at_eol(s):
         while s.stack_len != 0:
             if not has_head(get_s0(s)):
-                get_s0(s).dep_tag = 0
+                get_s0(s).dep = 0
             pop_stack(s)
 
 

spacy/syntax/arc_eager.pyx

@@ -123,7 +123,7 @@ cdef class TransitionSystem:
         if t.move == SHIFT:
             # Set the dep label, in case we need it after we reduce
             if NON_MONOTONIC:
-                get_s0(s).dep_tag = t.label
+                get_s0(s).dep = t.label
             push_stack(s)
         elif t.move == LEFT:
             add_dep(s, s.i, s.stack[0], t.label)
@@ -132,7 +132,7 @@ cdef class TransitionSystem:
             add_dep(s, s.stack[0], s.i, t.label)
             push_stack(s)
         elif t.move == REDUCE:
-            add_dep(s, s.stack[-1], s.stack[0], get_s0(s).dep_tag)
+            add_dep(s, s.stack[-1], s.stack[0], get_s0(s).dep)
             pop_stack(s)
         else:
             raise Exception(t.move)

spacy/syntax/parser.pxd

@@ -9,5 +9,3 @@ cdef class GreedyParser:
     cdef object cfg
     cdef readonly Model model
     cdef TransitionSystem moves
-
-    cpdef int parse(self, Tokens tokens) except -1

spacy/syntax/parser.pyx

@@ -65,7 +65,7 @@ cdef class GreedyParser:
         hasty_templ, full_templ = get_templates(self.cfg.features)
         self.model = Model(self.moves.n_moves, full_templ, model_dir)
 
-    cpdef int parse(self, Tokens tokens) except -1:
+    def __call__(self, Tokens tokens):
         cdef:
             Transition guess
             uint64_t state_key

spacy/tokenizer.pxd

@@ -28,8 +28,6 @@ cdef class Tokenizer:
     cdef object _infix_re
 
     cpdef Tokens tokens_from_list(self, list strings)
-    cpdef Tokens tokenize(self, unicode text)
-
 
     cdef int _try_cache(self, int idx, hash_t key, Tokens tokens) except -1
     cdef int _tokenize(self, Tokens tokens, UniStr* span, int start, int end) except -1

spacy/tokenizer.pyx

@@ -31,18 +31,6 @@ cdef class Tokenizer:
         self.vocab = vocab
         self._load_special_tokenization(rules, pos_tags, tag_names)
 
-    @classmethod
-    def from_dir(cls, Vocab vocab, object data_dir, object pos_tags, object tag_names):
-        if not path.exists(data_dir):
-            raise IOError("Directory %s not found -- cannot load Tokenizer." % data_dir)
-        if not path.isdir(data_dir):
-            raise IOError("Path %s is a file, not a dir -- cannot load Tokenizer." % data_dir)
- 
-        assert path.exists(data_dir) and path.isdir(data_dir)
-        rules, prefix_re, suffix_re, infix_re = util.read_lang_data(data_dir)
-        return cls(vocab, rules, re.compile(prefix_re), re.compile(suffix_re),
-                   re.compile(infix_re), pos_tags, tag_names)
-
     cpdef Tokens tokens_from_list(self, list strings):
         cdef int length = sum([len(s) for s in strings])
         cdef Tokens tokens = Tokens(self.vocab, length)
@@ -57,7 +45,7 @@ cdef class Tokenizer:
             idx += len(py_string) + 1
         return tokens
 
-    cpdef Tokens tokenize(self, unicode string):
+    def __call__(self, unicode string):
         """Tokenize a string.
 
         The tokenization rules are defined in three places:
@@ -257,7 +245,7 @@ cdef class Tokenizer:
                     tokens[i].lemma = self.vocab.strings[lemma]
                 if 'pos' in props:
                     # TODO: Clean up this mess...
-                    tokens[i].fine_pos = tag_names.index(props['pos'])
+                    tokens[i].tag = tag_names.index(props['pos'])
                     tokens[i].pos = tag_map[props['pos']][0]
                     # These are defaults, which can be over-ridden by the
                     # token-specific props.

spacy/tokens.pyx

@@ -198,9 +198,8 @@ cdef class Token:
         self.sentiment = t.lex.sentiment
         self.flags = t.lex.flags
         self.lemma = t.lemma
-        self.pos = t.pos
+        self.tag = t.tag
-        self.fine_pos = t.fine_pos
+        self.dep = t.dep
-        self.dep_tag = t.dep_tag
 
     def __unicode__(self):
         cdef const TokenC* t = &self._seq.data[self.i]
@@ -220,6 +219,12 @@ cdef class Token:
         """
         return self._seq.data[self.i].lex.length
 
+    def check_flag(self, attr_id_t flag):
+        return False
+
+    def is_pos(self, univ_tag_t pos):
+        return False
+
     property head:
         """The token predicted by the parser to be the head of the current token."""
         def __get__(self):
@@ -267,16 +272,10 @@ cdef class Token:
             cdef unicode py_ustr = self._seq.vocab.strings[t.lemma]
             return py_ustr
 
-    property pos_:
+    property tag_:
         def __get__(self):
-            return self._seq.vocab.strings[self.pos]
+            return self._seq.tag_names[self.tag]
 
-    property fine_pos_:
+    property dep_:
         def __get__(self):
-            return self._seq.vocab.strings[self.fine_pos]
+            return self._seq.dep_names[self.dep]
-
-    property dep_tag_:
-        def __get__(self):
-            return self._seq.vocab.strings[self.dep_tag]
-
-

spacy/vocab.pyx

@@ -1,16 +1,20 @@
 from libc.stdio cimport fopen, fclose, fread, fwrite, FILE
 from libc.string cimport memset
+from libc.stdint cimport int32_t
 
+import bz2
 from os import path
 import codecs
 
 from .lexeme cimport EMPTY_LEXEME
 from .lexeme cimport set_lex_struct_props
-from .lexeme cimport Lexeme_cinit
+from .lexeme cimport Lexeme
 from .strings cimport slice_unicode
 from .strings cimport hash_string
 from .orth cimport word_shape
 
+from cymem.cymem cimport Address
+
 
 DEF MAX_VEC_SIZE = 100000
 
@@ -34,12 +38,15 @@ cdef class Vocab:
         if data_dir is not None:
             if not path.exists(data_dir):
                 raise IOError("Directory %s not found -- cannot load Vocab." % data_dir)
+        assert EMPTY_LEXEME.vec != NULL
         if data_dir is not None:
             if not path.isdir(data_dir):
                 raise IOError("Path %s is a file, not a dir -- cannot load Vocab." % data_dir)
             self.strings.load(path.join(data_dir, 'strings.txt'))
             self.load_lexemes(path.join(data_dir, 'lexemes.bin'))
-            #self.load_vectors(path.join(data_dir, 'deps.words'))
+            self.load_vectors(path.join(data_dir, 'vec.bin'))
+        for i in range(self.lexemes.size()):
+            assert self.lexemes[i].vec != NULL, repr(self.strings[self.lexemes[i].sic])
 
     def __len__(self):
         """The current number of lexemes stored."""
@@ -52,13 +59,15 @@ cdef class Vocab:
         cdef LexemeC* lex
         lex = <LexemeC*>self._map.get(c_str.key)
         if lex != NULL:
+            assert lex.vec != NULL
             return lex
         if c_str.n < 3:
             mem = self.mem
         cdef unicode py_str = c_str.chars[:c_str.n]
         lex = <LexemeC*>mem.alloc(sizeof(LexemeC), 1)
         props = self.lexeme_props_getter(py_str)
-        set_lex_struct_props(lex, props, self.strings)
+        set_lex_struct_props(lex, props, self.strings, EMPTY_VEC)
+        assert lex.vec != NULL
         if mem is self.mem:
             lex.id = self.lexemes.size()
             self._add_lex_to_vocab(c_str.key, lex)
@@ -98,7 +107,7 @@ cdef class Vocab:
             lexeme = self.get(self.mem, &c_str)
         else:
             raise ValueError("Vocab unable to map type: %s. Maps unicode --> int or int --> unicode" % str(type(id_or_string)))
-        return Lexeme_cinit(lexeme, self.strings)
+        return Lexeme.from_ptr(lexeme, self.strings)
 
     def __setitem__(self, unicode py_str, dict props):
         cdef UniStr c_str
@@ -109,7 +118,8 @@ cdef class Vocab:
             lex = <LexemeC*>self.mem.alloc(sizeof(LexemeC), 1)
             lex.id = self.lexemes.size()
             self._add_lex_to_vocab(c_str.key, lex)
-        set_lex_struct_props(lex, props, self.strings)
+        set_lex_struct_props(lex, props, self.strings, EMPTY_VEC)
+        assert lex.vec != NULL
         assert lex.sic < 1000000
 
     def dump(self, loc):
@@ -147,8 +157,9 @@ cdef class Vocab:
             if st != 1:
                 break
             lexeme = <LexemeC*>self.mem.alloc(sizeof(LexemeC), 1)
-            lexeme.vec = EMPTY_VEC
+            # Copies data from the file into the lexeme
             st = fread(lexeme, sizeof(LexemeC), 1, fp)
+            lexeme.vec = EMPTY_VEC
             if st != 1:
                 break
             self._map.set(key, lexeme)
@@ -159,27 +170,96 @@ cdef class Vocab:
         fclose(fp)
     
     def load_vectors(self, loc):
-        cdef int i
+        file_ = _CFile(loc, 'rb')
-        cdef unicode line
+        cdef int32_t word_len
-        cdef unicode word
+        cdef int32_t vec_len
-        cdef unicode val_str
-        cdef hash_t key
-        cdef LexemeC* lex
         cdef float* vec
+        cdef Address mem
+        cdef id_t string_id
+        cdef bytes py_word
+        cdef vector[float*] vectors
+        cdef int i
+        while True:
+            try:
+                file_.read(&word_len, sizeof(word_len), 1)
+            except IOError:
+                break
+            file_.read(&vec_len, sizeof(vec_len), 1)
             
-        with codecs.open(loc, 'r', 'utf8') as file_:
+            mem = Address(word_len, sizeof(char))
-            for line in file_:
+            chars = <char*>mem.ptr
-                pieces = line.split()
+            vec = <float*>self.mem.alloc(vec_len, sizeof(float))
-                word = pieces.pop(0)
+
-                if len(pieces) >= MAX_VEC_SIZE:
+            file_.read(chars, sizeof(char), word_len)
-                    sizes = (len(pieces), MAX_VEC_SIZE)
+            file_.read(vec, sizeof(float), vec_len)
-                    msg = ("Your vector is %d elements."
+
-                           "The compile-time limit is %d elements." % sizes)
+            string_id = self.strings[chars[:word_len]]
-                    raise ValueError(msg)
+            while string_id >= vectors.size():
-                key = hash_string(word)
+                vectors.push_back(EMPTY_VEC)
-                lex = <LexemeC*>self._map.get(key)
+            assert vec != NULL
-                if lex is not NULL:
+            vectors[string_id] = vec
-                    vec = <float*>self.mem.alloc(len(pieces), sizeof(float))
+        cdef LexemeC* lex
-                    for i, val_str in enumerate(pieces):
+        for i in range(self.lexemes.size()):
-                        vec[i] = float(val_str)
+            # Cast away the const, cos we can modify our lexemes
-                    lex.vec = vec
+            lex = <LexemeC*>self.lexemes[i]
+            if lex.sic < vectors.size():
+                lex.vec = vectors[lex.sic]
+            else:
+                lex.vec = EMPTY_VEC
+            assert lex.vec != NULL
+
+
+def write_binary_vectors(in_loc, out_loc):
+    cdef _CFile out_file = _CFile(out_loc, 'wb')
+    cdef Address mem
+    cdef int32_t word_len
+    cdef int32_t vec_len
+    cdef char* chars
+    with bz2.BZ2File(in_loc, 'r') as file_:
+        for line in file_:
+            pieces = line.split()
+            word = pieces.pop(0)
+            mem = Address(len(pieces), sizeof(float))
+            vec = <float*>mem.ptr
+            for i, val_str in enumerate(pieces):
+                vec[i] = float(val_str)
+
+            word_len = len(word)
+            vec_len = len(pieces)
+
+            out_file.write(sizeof(word_len), 1, &word_len)
+            out_file.write(sizeof(vec_len), 1, &vec_len)
+
+            chars = <char*>word
+            out_file.write(sizeof(char), len(word), chars)
+            out_file.write(sizeof(float), vec_len, vec)
+
+
+cdef class _CFile:
+    cdef FILE* fp
+    def __init__(self, loc, mode):
+        cdef bytes bytes_loc = loc.encode('utf8') if type(loc) == unicode else loc
+        self.fp = fopen(<char*>bytes_loc, mode)
+        if self.fp == NULL:
+            raise IOError
+
+    def __dealloc__(self):
+        fclose(self.fp)
+
+    def close(self):
+        fclose(self.fp)
+
+    cdef int read(self, void* dest, size_t elem_size, size_t n) except -1:
+        st = fread(dest, elem_size, n, self.fp)
+        if st != n:
+            raise IOError
+
+    cdef int write(self, size_t elem_size, size_t n, void* data) except -1:
+        st = fwrite(data, elem_size, n, self.fp)
+        if st != n:
+            raise IOError
+
+    cdef int write_unicode(self, unicode value):
+        cdef bytes py_bytes = value.encode('utf8')
+        cdef char* chars = <char*>py_bytes
+        self.write(sizeof(char), len(py_bytes), chars)