spaCy/spacy/vocab.pyx

507 lines
20 KiB
Cython
Raw Normal View History

# coding: utf8
# cython: profile=True
from __future__ import unicode_literals
2017-08-22 17:46:35 +00:00
import numpy
import dill
2017-05-29 11:04:40 +00:00
from collections import OrderedDict
2017-10-30 16:59:43 +00:00
from thinc.neural.util import get_array_module
from .lexeme cimport EMPTY_LEXEME
from .lexeme cimport Lexeme
from .strings cimport hash_string
2015-01-31 05:38:58 +00:00
from .typedefs cimport attr_t
from .tokens.token cimport Token
2017-10-27 17:45:19 +00:00
from .attrs cimport PROB, LANG, ORTH, TAG
from .structs cimport SerializedLexemeC
2017-10-27 17:45:19 +00:00
from .compat import copy_reg, basestring_
from .lemmatizer import Lemmatizer
from .attrs import intify_attrs
2017-08-19 16:50:16 +00:00
from .vectors import Vectors
2017-09-22 14:38:22 +00:00
from ._ml import link_vectors_to_models
2017-10-27 19:07:59 +00:00
from . import util
2014-12-24 06:42:00 +00:00
cdef class Vocab:
"""A look-up table that allows you to access `Lexeme` objects. The `Vocab`
instance also provides access to the `StringStore`, and owns underlying
C-data that is shared between `Doc` objects.
"""
def __init__(self, lex_attr_getters=None, tag_map=None, lemmatizer=None,
strings=tuple(), oov_prob=-20., **deprecated_kwargs):
"""Create the vocabulary.
2017-10-27 17:45:19 +00:00
lex_attr_getters (dict): A dictionary mapping attribute IDs to
functions to compute them. Defaults to `None`.
tag_map (dict): Dictionary mapping fine-grained tags to coarse-grained
parts-of-speech, and optionally morphological attributes.
lemmatizer (object): A lemmatizer. Defaults to `None`.
strings (StringStore): StringStore that maps strings to integers, and
vice versa.
2017-10-27 17:45:19 +00:00
RETURNS (Vocab): The newly constructed object.
"""
lex_attr_getters = lex_attr_getters if lex_attr_getters is not None else {}
tag_map = tag_map if tag_map is not None else {}
if lemmatizer in (None, True, False):
lemmatizer = Lemmatizer({}, {}, {})
self.cfg = {'oov_prob': oov_prob}
self.mem = Pool()
self._by_hash = PreshMap()
self._by_orth = PreshMap()
self.strings = StringStore()
2017-06-02 08:57:25 +00:00
self.length = 0
if strings:
for string in strings:
_ = self[string]
self.lex_attr_getters = lex_attr_getters
self.morphology = Morphology(self.strings, tag_map, lemmatizer)
2017-10-31 17:25:08 +00:00
self.vectors = Vectors()
2016-12-21 17:04:41 +00:00
property lang:
def __get__(self):
langfunc = None
if self.lex_attr_getters:
langfunc = self.lex_attr_getters.get(LANG, None)
return langfunc('_') if langfunc else ''
def __len__(self):
"""The current number of lexemes stored.
RETURNS (int): The current number of lexemes stored.
"""
return self.length
def add_flag(self, flag_getter, int flag_id=-1):
"""Set a new boolean flag to words in the vocabulary.
2016-12-21 17:04:41 +00:00
The flag_getter function will be called over the words currently in the
2016-11-01 11:25:36 +00:00
vocab, and then applied to new words as they occur. You'll then be able
2017-10-27 17:45:19 +00:00
to access the flag value on each token using token.check_flag(flag_id).
See also: `Lexeme.set_flag`, `Lexeme.check_flag`, `Token.set_flag`,
`Token.check_flag`.
2017-10-27 17:45:19 +00:00
flag_getter (callable): A function `f(unicode) -> bool`, to get the
flag value.
flag_id (int): An integer between 1 and 63 (inclusive), specifying
the bit at which the flag will be stored. If -1, the lowest
available bit will be chosen.
RETURNS (int): The integer ID by which the flag value can be checked.
EXAMPLE:
2017-10-27 17:45:19 +00:00
>>> my_product_getter = lambda text: text in ['spaCy', 'dislaCy']
>>> MY_PRODUCT = nlp.vocab.add_flag(my_product_getter)
>>> doc = nlp(u'I like spaCy')
>>> assert doc[2].check_flag(MY_PRODUCT) == True
"""
if flag_id == -1:
for bit in range(1, 64):
if bit not in self.lex_attr_getters:
flag_id = bit
break
else:
raise ValueError(
2017-10-27 17:45:19 +00:00
"Cannot find empty bit for new lexical flag. All bits "
"between 0 and 63 are occupied. You can replace one by "
"specifying the flag_id explicitly, e.g. "
"`nlp.vocab.add_flag(your_func, flag_id=IS_ALPHA`.")
elif flag_id >= 64 or flag_id < 1:
raise ValueError(
"Invalid value for flag_id: %d. Flag IDs must be between "
"1 and 63 (inclusive)" % flag_id)
for lex in self:
lex.set_flag(flag_id, flag_getter(lex.orth_))
self.lex_attr_getters[flag_id] = flag_getter
return flag_id
cdef const LexemeC* get(self, Pool mem, unicode string) except NULL:
2017-10-27 17:45:19 +00:00
"""Get a pointer to a `LexemeC` from the lexicon, creating a new
`Lexeme` if necessary using memory acquired from the given pool. If the
pool is the lexicon's own memory, the lexeme is saved in the lexicon.
"""
2015-07-25 22:18:30 +00:00
if string == u'':
return &EMPTY_LEXEME
cdef LexemeC* lex
cdef hash_t key = hash_string(string)
lex = <LexemeC*>self._by_hash.get(key)
cdef size_t addr
if lex != NULL:
if lex.orth != self.strings[string]:
raise LookupError.mismatched_strings(
lex.orth, self.strings[string], string)
return lex
else:
2015-08-22 20:04:34 +00:00
return self._new_lexeme(mem, string)
cdef const LexemeC* get_by_orth(self, Pool mem, attr_t orth) except NULL:
2017-10-27 17:45:19 +00:00
"""Get a pointer to a `LexemeC` from the lexicon, creating a new
`Lexeme` if necessary using memory acquired from the given pool. If the
pool is the lexicon's own memory, the lexeme is saved in the lexicon.
"""
2015-07-26 17:26:41 +00:00
if orth == 0:
2015-07-26 16:39:27 +00:00
return &EMPTY_LEXEME
cdef LexemeC* lex
lex = <LexemeC*>self._by_orth.get(orth)
if lex != NULL:
return lex
2015-08-22 20:04:34 +00:00
else:
return self._new_lexeme(mem, self.strings[orth])
2015-08-22 20:04:34 +00:00
cdef const LexemeC* _new_lexeme(self, Pool mem, unicode string) except NULL:
cdef hash_t key
if len(string) < 3 or self.length < 10000:
mem = self.mem
cdef bint is_oov = mem is not self.mem
lex = <LexemeC*>mem.alloc(sizeof(LexemeC), 1)
lex.orth = self.strings.add(string)
2015-08-26 17:21:46 +00:00
lex.length = len(string)
lex.id = self.vectors.key2row.get(lex.orth, 0)
if self.lex_attr_getters is not None:
for attr, func in self.lex_attr_getters.items():
value = func(string)
if isinstance(value, unicode):
value = self.strings.add(value)
2015-08-26 17:21:46 +00:00
if attr == PROB:
lex.prob = value
elif value is not None:
2015-08-26 17:21:46 +00:00
Lexeme.set_struct_attr(lex, attr, value)
if not is_oov:
key = hash_string(string)
2015-08-22 20:04:34 +00:00
self._add_lex_to_vocab(key, lex)
assert lex != NULL, string
return lex
cdef int _add_lex_to_vocab(self, hash_t key, const LexemeC* lex) except -1:
self._by_hash.set(key, <void*>lex)
self._by_orth.set(lex.orth, <void*>lex)
self.length += 1
def __contains__(self, unicode string):
"""Check whether the string has an entry in the vocabulary.
2016-11-01 11:25:36 +00:00
string (unicode): The ID string.
RETURNS (bool) Whether the string has an entry in the vocabulary.
"""
key = hash_string(string)
lex = self._by_hash.get(key)
2017-01-11 09:18:22 +00:00
return lex is not NULL
def __iter__(self):
"""Iterate over the lexemes in the vocabulary.
2016-11-01 11:25:36 +00:00
YIELDS (Lexeme): An entry in the vocabulary.
"""
2017-10-31 01:00:26 +00:00
cdef attr_t key
cdef size_t addr
2017-10-31 01:00:26 +00:00
for key, addr in self._by_orth.items():
lex = Lexeme(self, key)
yield lex
2017-10-27 17:45:19 +00:00
def __getitem__(self, id_or_string):
"""Retrieve a lexeme, given an int ID or a unicode string. If a
previously unseen unicode string is given, a new lexeme is created and
stored.
id_or_string (int or unicode): The integer ID of a word, or its unicode
string. If `int >= Lexicon.size`, `IndexError` is raised. If
`id_or_string` is neither an int nor a unicode string, `ValueError`
is raised.
RETURNS (Lexeme): The lexeme indicated by the given ID.
EXAMPLE:
>>> apple = nlp.vocab.strings['apple']
>>> assert nlp.vocab[apple] == nlp.vocab[u'apple']
"""
cdef attr_t orth
2017-10-31 01:00:26 +00:00
if isinstance(id_or_string, unicode):
orth = self.strings.add(id_or_string)
else:
orth = id_or_string
return Lexeme(self, orth)
cdef const TokenC* make_fused_token(self, substrings) except NULL:
cdef int i
tokens = <TokenC*>self.mem.alloc(len(substrings) + 1, sizeof(TokenC))
for i, props in enumerate(substrings):
2017-10-27 17:45:19 +00:00
props = intify_attrs(props, strings_map=self.strings,
_do_deprecated=True)
token = &tokens[i]
# Set the special tokens up to have arbitrary attributes
2017-10-27 17:45:19 +00:00
lex = <LexemeC*>self.get_by_orth(self.mem, props[ORTH])
token.lex = lex
2017-10-27 17:45:19 +00:00
if TAG in props:
self.morphology.assign_tag(token, props[TAG])
for attr_id, value in props.items():
Token.set_struct_attr(token, attr_id, value)
Lexeme.set_struct_attr(lex, attr_id, value)
return tokens
2016-12-21 17:04:41 +00:00
2017-05-30 21:34:50 +00:00
@property
def vectors_length(self):
2017-08-22 17:46:35 +00:00
return self.vectors.data.shape[1]
2017-05-30 21:34:50 +00:00
2017-10-31 17:25:08 +00:00
def reset_vectors(self, *, width=None, shape=None):
2017-05-30 21:34:50 +00:00
"""Drop the current vector table. Because all vectors must be the same
width, you have to call this to change the size of the vectors.
"""
2017-10-31 17:25:08 +00:00
if width is not None and shape is not None:
raise ValueError("Only one of width and shape can be specified")
elif shape is not None:
self.vectors = Vectors(shape=shape)
else:
width = width if width is not None else self.vectors.data.shape[1]
self.vectors = Vectors(shape=(self.vectors.shape[0], width))
2017-10-31 01:00:26 +00:00
2017-10-31 17:25:08 +00:00
def prune_vectors(self, nr_row, batch_size=1024):
2017-10-31 01:00:26 +00:00
"""Reduce the current vector table to `nr_row` unique entries. Words
mapped to the discarded vectors will be remapped to the closest vector
among those remaining.
For example, suppose the original table had vectors for the words:
['sat', 'cat', 'feline', 'reclined']. If we prune the vector table to,
two rows, we would discard the vectors for 'feline' and 'reclined'.
These words would then be remapped to the closest remaining vector
-- so "feline" would have the same vector as "cat", and "reclined"
would have the same vector as "sat".
The similarities are judged by cosine. The original vectors may
be large, so the cosines are calculated in minibatches, to reduce
memory usage.
nr_row (int): The number of rows to keep in the vector table.
batch_size (int): Batch of vectors for calculating the similarities.
Larger batch sizes might be faster, while temporarily requiring
more memory.
RETURNS (dict): A dictionary keyed by removed words mapped to
`(string, score)` tuples, where `string` is the entry the removed
word was mapped to, and `score` the similarity score between the
two words.
"""
xp = get_array_module(self.vectors.data)
2017-10-31 17:25:08 +00:00
# Make prob negative so it sorts by rank ascending
# (key2row contains the rank)
priority = [(-lex.prob, self.vectors.key2row[lex.orth], lex.orth)
for lex in self if lex.orth in self.vectors.key2row]
priority.sort()
indices = xp.asarray([i for (prob, i, key) in priority], dtype='i')
keys = xp.asarray([key for (prob, i, key) in priority], dtype='uint64')
2017-10-31 23:34:55 +00:00
2017-10-31 17:25:08 +00:00
keep = xp.ascontiguousarray(self.vectors.data[indices[:nr_row]])
toss = xp.ascontiguousarray(self.vectors.data[indices[nr_row:]])
self.vectors = Vectors(data=keep, keys=keys)
2017-10-31 23:34:55 +00:00
syn_keys, syn_rows, scores = self.vectors.most_similar(toss)
2017-10-31 17:25:08 +00:00
2017-10-31 10:40:46 +00:00
remap = {}
2017-10-31 17:25:08 +00:00
for i, key in enumerate(keys[nr_row:]):
self.vectors.add(key, row=syn_rows[i])
word = self.strings[key]
synonym = self.strings[syn_keys[i]]
score = scores[i]
remap[word] = (synonym, score)
2017-10-31 01:00:26 +00:00
link_vectors_to_models(self)
return remap
2017-05-30 21:34:50 +00:00
2017-05-28 09:45:32 +00:00
def get_vector(self, orth):
2017-10-27 17:45:19 +00:00
"""Retrieve a vector for a word in the vocabulary. Words can be looked
up by string or int ID. If no vectors data is loaded, ValueError is
raised.
2017-05-28 09:45:32 +00:00
2017-10-27 17:45:19 +00:00
RETURNS (numpy.ndarray): A word vector. Size
and shape determined by the `vocab.vectors` instance. Usually, a
numpy ndarray of shape (300,) and dtype float32.
2017-05-28 09:45:32 +00:00
"""
2017-08-19 17:52:25 +00:00
if isinstance(orth, basestring_):
orth = self.strings.add(orth)
2017-08-22 17:46:35 +00:00
if orth in self.vectors.key2row:
return self.vectors[orth]
else:
return numpy.zeros((self.vectors_length,), dtype='f')
2017-05-28 09:45:32 +00:00
2017-05-30 21:34:50 +00:00
def set_vector(self, orth, vector):
2017-10-27 17:45:19 +00:00
"""Set a vector for a word in the vocabulary. Words can be referenced
by string or int ID.
2017-05-30 21:34:50 +00:00
"""
2017-10-31 17:25:08 +00:00
if isinstance(orth, basestring_):
orth = self.strings.add(orth)
if self.vectors.is_full and orth not in self.vectors:
new_rows = max(100, int(self.vectors.shape[0]*1.3))
if self.vectors.shape[1] == 0:
width = vector.size
else:
width = self.vectors.shape[1]
self.vectors.resize((new_rows, width))
self.vectors.add(orth, vector=vector)
2017-08-19 18:35:33 +00:00
self.vectors.add(orth, vector=vector)
2017-05-30 21:34:50 +00:00
2017-05-28 09:45:32 +00:00
def has_vector(self, orth):
2017-10-27 17:45:19 +00:00
"""Check whether a word has a vector. Returns False if no vectors have
been loaded. Words can be looked up by string or int ID."""
2017-08-19 17:52:25 +00:00
if isinstance(orth, basestring_):
orth = self.strings.add(orth)
return orth in self.vectors
2017-05-28 09:45:32 +00:00
2017-08-18 18:46:56 +00:00
def to_disk(self, path, **exclude):
"""Save the current state to a directory.
path (unicode or Path): A path to a directory, which will be created if
2017-10-27 17:45:19 +00:00
it doesn't exist. Paths may be either strings or Path-like objects.
"""
path = util.ensure_path(path)
if not path.exists():
path.mkdir()
2017-05-28 21:34:12 +00:00
self.strings.to_disk(path / 'strings.json')
with (path / 'lexemes.bin').open('wb') as file_:
file_.write(self.lexemes_to_bytes())
2017-08-18 18:46:56 +00:00
if self.vectors is not None:
2017-08-19 19:27:35 +00:00
self.vectors.to_disk(path)
2017-08-18 18:46:56 +00:00
def from_disk(self, path, **exclude):
"""Loads state from a directory. Modifies the object in place and
returns it.
path (unicode or Path): A path to a directory. Paths may be either
strings or `Path`-like objects.
RETURNS (Vocab): The modified `Vocab` object.
"""
path = util.ensure_path(path)
2017-05-28 21:34:12 +00:00
self.strings.from_disk(path / 'strings.json')
with (path / 'lexemes.bin').open('rb') as file_:
self.lexemes_from_bytes(file_.read())
2017-08-18 18:46:56 +00:00
if self.vectors is not None:
2017-08-19 19:27:35 +00:00
self.vectors.from_disk(path, exclude='strings.json')
link_vectors_to_models(self)
2017-05-28 21:34:12 +00:00
return self
2016-11-01 11:25:36 +00:00
def to_bytes(self, **exclude):
"""Serialize the current state to a binary string.
**exclude: Named attributes to prevent from being serialized.
RETURNS (bytes): The serialized form of the `Vocab` object.
"""
2017-08-18 18:46:56 +00:00
def deserialize_vectors():
if self.vectors is None:
return None
else:
return self.vectors.to_bytes()
2017-05-29 22:52:36 +00:00
getters = OrderedDict((
('strings', lambda: self.strings.to_bytes()),
('lexemes', lambda: self.lexemes_to_bytes()),
2017-08-18 18:46:56 +00:00
('vectors', deserialize_vectors)
2017-05-29 22:52:36 +00:00
))
2017-05-29 08:14:20 +00:00
return util.to_bytes(getters, exclude)
2017-05-21 12:18:46 +00:00
def from_bytes(self, bytes_data, **exclude):
"""Load state from a binary string.
bytes_data (bytes): The data to load from.
**exclude: Named attributes to prevent from being loaded.
RETURNS (Vocab): The `Vocab` object.
"""
2017-08-18 18:46:56 +00:00
def serialize_vectors(b):
if self.vectors is None:
return None
else:
return self.vectors.from_bytes(b)
2017-05-29 11:04:40 +00:00
setters = OrderedDict((
('strings', lambda b: self.strings.from_bytes(b)),
2017-05-29 22:52:36 +00:00
('lexemes', lambda b: self.lexemes_from_bytes(b)),
2017-08-18 18:46:56 +00:00
('vectors', lambda b: serialize_vectors(b))
2017-05-29 11:04:40 +00:00
))
2017-06-02 08:56:40 +00:00
util.from_bytes(bytes_data, setters, exclude)
return self
2017-05-28 21:34:12 +00:00
def lexemes_to_bytes(self):
cdef hash_t key
cdef size_t addr
2017-03-11 18:43:09 +00:00
cdef LexemeC* lexeme = NULL
cdef SerializedLexemeC lex_data
cdef int size = 0
2015-07-27 08:58:15 +00:00
for key, addr in self._by_hash.items():
if addr == 0:
continue
size += sizeof(lex_data.data)
byte_string = b'\0' * size
byte_ptr = <unsigned char*>byte_string
cdef int j
cdef int i = 0
for key, addr in self._by_hash.items():
if addr == 0:
continue
lexeme = <LexemeC*>addr
lex_data = Lexeme.c_to_bytes(lexeme)
for j in range(sizeof(lex_data.data)):
byte_ptr[i] = lex_data.data[j]
i += 1
return byte_string
2016-11-01 11:25:36 +00:00
def lexemes_from_bytes(self, bytes bytes_data):
"""Load the binary vocabulary data from the given string."""
cdef LexemeC* lexeme
2017-03-07 19:25:12 +00:00
cdef hash_t key
cdef unicode py_str
cdef int i = 0
cdef int j = 0
cdef SerializedLexemeC lex_data
chunk_size = sizeof(lex_data.data)
cdef void* ptr
cdef unsigned char* bytes_ptr = bytes_data
for i in range(0, len(bytes_data), chunk_size):
lexeme = <LexemeC*>self.mem.alloc(1, sizeof(LexemeC))
for j in range(sizeof(lex_data.data)):
lex_data.data[j] = bytes_ptr[i+j]
Lexeme.c_from_bytes(lexeme, lex_data)
2017-03-07 19:25:12 +00:00
ptr = self.strings._map.get(lexeme.orth)
if ptr == NULL:
continue
2017-03-07 19:25:12 +00:00
py_str = self.strings[lexeme.orth]
assert self.strings[py_str] == lexeme.orth, (py_str, lexeme.orth)
2017-03-07 19:25:12 +00:00
key = hash_string(py_str)
self._by_hash.set(key, lexeme)
self._by_orth.set(lexeme.orth, lexeme)
self.length += 1
2017-03-07 19:25:12 +00:00
def pickle_vocab(vocab):
sstore = vocab.strings
morph = vocab.morphology
length = vocab.length
data_dir = vocab.data_dir
lex_attr_getters = dill.dumps(vocab.lex_attr_getters)
lexemes_data = vocab.lexemes_to_bytes()
2017-03-07 19:25:12 +00:00
return (unpickle_vocab,
2017-10-27 17:45:19 +00:00
(sstore, morph, data_dir, lex_attr_getters, lexemes_data, length))
2017-03-07 19:25:12 +00:00
2017-05-09 15:28:50 +00:00
def unpickle_vocab(sstore, morphology, data_dir,
2017-10-27 17:45:19 +00:00
lex_attr_getters, bytes lexemes_data, int length):
2017-03-07 19:25:12 +00:00
cdef Vocab vocab = Vocab()
vocab.length = length
vocab.strings = sstore
vocab.morphology = morphology
vocab.data_dir = data_dir
vocab.lex_attr_getters = dill.loads(lex_attr_getters)
vocab.lexemes_from_bytes(lexemes_data)
2017-03-07 19:25:12 +00:00
vocab.length = length
2017-09-22 14:38:22 +00:00
link_vectors_to_models(vocab)
2017-03-07 19:25:12 +00:00
return vocab
copy_reg.pickle(Vocab, pickle_vocab, unpickle_vocab)
class LookupError(Exception):
@classmethod
def mismatched_strings(cls, id_, id_string, original_string):
return cls(
2017-10-27 17:45:19 +00:00
"Error fetching a Lexeme from the Vocab. When looking up a "
"string, the lexeme returned had an orth ID that did not match "
"the query string. This means that the cached lexeme structs are "
"mismatched to the string encoding table. The mismatched:\n"
"Query string: {}\n"
"Orth cached: {}\n"
"Orth ID: {}".format(repr(original_string), repr(id_string), id_))