spaCy/spacy/vocab.pyx
Adriane Boyd c053f158c5
Add support for floret vectors (#8909)
* Add support for fasttext-bloom hash-only vectors

Overview:

* Extend `Vectors` to have two modes: `default` and `ngram`
  * `default` is the default mode and equivalent to the current
    `Vectors`
  * `ngram` supports the hash-only ngram tables from `fasttext-bloom`
* Extend `spacy.StaticVectors.v2` to handle both modes with no changes
  for `default` vectors
* Extend `spacy init vectors` to support ngram tables

The `ngram` mode **only** supports vector tables produced by this
fork of fastText, which adds an option to represent all vectors using
only the ngram buckets table and which uses the exact same ngram
generation algorithm and hash function (`MurmurHash3_x64_128`).
`fasttext-bloom` produces an additional `.hashvec` table, which can be
loaded by `spacy init vectors --fasttext-bloom-vectors`.

https://github.com/adrianeboyd/fastText/tree/feature/bloom

Implementation details:

* `Vectors` now includes the `StringStore` as `Vectors.strings` so that
  the API can stay consistent for both `default` (which can look up from
  `str` or `int`) and `ngram` (which requires `str` to calculate the
  ngrams).

* In ngram mode `Vectors` uses a default `Vectors` object as a cache
  since the ngram vectors lookups are relatively expensive.

  * The default cache size is the same size as the provided ngram vector
    table.

  * Once the cache is full, no more entries are added. The user is
    responsible for managing the cache in cases where the initial
    documents are not representative of the texts.

  * The cache can be resized by setting `Vectors.ngram_cache_size` or
    cleared with `vectors._ngram_cache.clear()`.

* The API ends up a bit split between methods for `default` and for
  `ngram`, so functions that only make sense for `default` or `ngram`
  include warnings with custom messages suggesting alternatives where
  possible.

* `Vocab.vectors` becomes a property so that the string stores can be
  synced when assigning vectors to a vocab.

* `Vectors` serializes its own config settings as `vectors.cfg`.

* The `Vectors` serialization methods have added support for `exclude`
  so that the `Vocab` can exclude the `Vectors` strings while serializing.

Removed:

* The `minn` and `maxn` options and related code from
  `Vocab.get_vector`, which does not work in a meaningful way for default
  vector tables.

* The unused `GlobalRegistry` in `Vectors`.

* Refactor to use reduce_mean

Refactor to use reduce_mean and remove the ngram vectors cache.

* Rename to floret

* Rename to floret in error messages

* Use --vectors-mode in CLI, vector init

* Fix vectors mode in init

* Remove unused var

* Minor API and docstrings adjustments

* Rename `--vectors-mode` to `--mode` in `init vectors` CLI
* Rename `Vectors.get_floret_vectors` to `Vectors.get_batch` and support
  both modes.
* Minor updates to Vectors docstrings.

* Update API docs for Vectors and init vectors CLI

* Update types for StaticVectors
2021-10-27 14:08:31 +02:00

552 lines
22 KiB
Cython

# cython: profile=True
from libc.string cimport memcpy
import srsly
from thinc.api import get_array_module, get_current_ops
import functools
from .lexeme cimport EMPTY_LEXEME, OOV_RANK
from .lexeme cimport Lexeme
from .typedefs cimport attr_t
from .tokens.token cimport Token
from .attrs cimport LANG, ORTH
from .compat import copy_reg
from .errors import Errors
from .attrs import intify_attrs, NORM, IS_STOP
from .vectors import Vectors, Mode as VectorsMode
from .util import registry
from .lookups import Lookups
from . import util
from .lang.norm_exceptions import BASE_NORMS
from .lang.lex_attrs import LEX_ATTRS, is_stop, get_lang
def create_vocab(lang, defaults, vectors_name=None):
# If the spacy-lookups-data package is installed, we pre-populate the lookups
# with lexeme data, if available
lex_attrs = {**LEX_ATTRS, **defaults.lex_attr_getters}
# This is messy, but it's the minimal working fix to Issue #639.
lex_attrs[IS_STOP] = functools.partial(is_stop, stops=defaults.stop_words)
# Ensure that getter can be pickled
lex_attrs[LANG] = functools.partial(get_lang, lang=lang)
lex_attrs[NORM] = util.add_lookups(
lex_attrs.get(NORM, LEX_ATTRS[NORM]),
BASE_NORMS,
)
return Vocab(
lex_attr_getters=lex_attrs,
writing_system=defaults.writing_system,
get_noun_chunks=defaults.syntax_iterators.get("noun_chunks"),
vectors_name=vectors_name,
)
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.
DOCS: https://spacy.io/api/vocab
"""
def __init__(self, lex_attr_getters=None, strings=tuple(), lookups=None,
oov_prob=-20., vectors_name=None, writing_system={},
get_noun_chunks=None, **deprecated_kwargs):
"""Create the vocabulary.
lex_attr_getters (dict): A dictionary mapping attribute IDs to
functions to compute them. Defaults to `None`.
strings (StringStore): StringStore that maps strings to integers, and
vice versa.
lookups (Lookups): Container for large lookup tables and dictionaries.
oov_prob (float): Default OOV probability.
vectors_name (str): Optional name to identify the vectors table.
get_noun_chunks (Optional[Callable[[Union[Doc, Span], Iterator[Tuple[int, int, int]]]]]):
A function that yields base noun phrases used for Doc.noun_chunks.
"""
lex_attr_getters = lex_attr_getters if lex_attr_getters is not None else {}
if lookups in (None, True, False):
lookups = Lookups()
self.cfg = {'oov_prob': oov_prob}
self.mem = Pool()
self._by_orth = PreshMap()
self.strings = StringStore()
self.length = 0
if strings:
for string in strings:
_ = self[string]
self.lex_attr_getters = lex_attr_getters
self.morphology = Morphology(self.strings)
self.vectors = Vectors(strings=self.strings, name=vectors_name)
self.lookups = lookups
self.writing_system = writing_system
self.get_noun_chunks = get_noun_chunks
property vectors:
def __get__(self):
return self._vectors
def __set__(self, vectors):
for s in vectors.strings:
self.strings.add(s)
self._vectors = vectors
self._vectors.strings = self.strings
@property
def lang(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.
The flag_getter function will be called over the words currently in the
vocab, and then applied to new words as they occur. You'll then be able
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`.
flag_getter (callable): A function `f(str) -> 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.
DOCS: https://spacy.io/api/vocab#add_flag
"""
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(Errors.E062)
elif flag_id >= 64 or flag_id < 1:
raise ValueError(Errors.E063.format(value=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, str string) except NULL:
"""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.
"""
if string == "":
return &EMPTY_LEXEME
cdef LexemeC* lex
cdef hash_t key = self.strings[string]
lex = <LexemeC*>self._by_orth.get(key)
cdef size_t addr
if lex != NULL:
assert lex.orth in self.strings
if lex.orth != key:
raise KeyError(Errors.E064.format(string=lex.orth,
orth=key, orth_id=string))
return lex
else:
return self._new_lexeme(mem, string)
cdef const LexemeC* get_by_orth(self, Pool mem, attr_t orth) except NULL:
"""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.
"""
if orth == 0:
return &EMPTY_LEXEME
cdef LexemeC* lex
lex = <LexemeC*>self._by_orth.get(orth)
if lex != NULL:
return lex
else:
return self._new_lexeme(mem, self.strings[orth])
cdef const LexemeC* _new_lexeme(self, Pool mem, str string) except NULL:
# I think this heuristic is bad, and the Vocab should always
# own the lexemes. It avoids weird bugs this way, as it's how the thing
# was originally supposed to work. The best solution to the growing
# memory use is to periodically reset the vocab, which is an action
# that should be up to the user to do (so we don't need to keep track
# of the doc ownership).
# TODO: Change the C API so that the mem isn't passed in here.
mem = self.mem
#if len(string) < 3 or self.length < 10000:
# mem = self.mem
cdef bint is_oov = mem is not self.mem
lex = <LexemeC*>mem.alloc(1, sizeof(LexemeC))
lex.orth = self.strings.add(string)
lex.length = len(string)
if self.vectors is not None:
lex.id = self.vectors.key2row.get(lex.orth, OOV_RANK)
else:
lex.id = OOV_RANK
if self.lex_attr_getters is not None:
for attr, func in self.lex_attr_getters.items():
value = func(string)
if isinstance(value, str):
value = self.strings.add(value)
if value is not None:
Lexeme.set_struct_attr(lex, attr, value)
if not is_oov:
self._add_lex_to_vocab(lex.orth, lex)
if lex == NULL:
raise ValueError(Errors.E085.format(string=string))
return lex
cdef int _add_lex_to_vocab(self, hash_t key, const LexemeC* lex) except -1:
self._by_orth.set(lex.orth, <void*>lex)
self.length += 1
def __contains__(self, key):
"""Check whether the string or int key has an entry in the vocabulary.
string (str): The ID string.
RETURNS (bool) Whether the string has an entry in the vocabulary.
DOCS: https://spacy.io/api/vocab#contains
"""
cdef hash_t int_key
if isinstance(key, bytes):
int_key = self.strings[key.decode("utf8")]
elif isinstance(key, str):
int_key = self.strings[key]
else:
int_key = key
lex = self._by_orth.get(int_key)
return lex is not NULL
def __iter__(self):
"""Iterate over the lexemes in the vocabulary.
YIELDS (Lexeme): An entry in the vocabulary.
DOCS: https://spacy.io/api/vocab#iter
"""
cdef attr_t key
cdef size_t addr
for key, addr in self._by_orth.items():
lex = Lexeme(self, key)
yield lex
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 str): 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"]
DOCS: https://spacy.io/api/vocab#getitem
"""
cdef attr_t orth
if isinstance(id_or_string, str):
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):
props = intify_attrs(props, strings_map=self.strings,
_do_deprecated=True)
token = &tokens[i]
# Set the special tokens up to have arbitrary attributes
lex = <LexemeC*>self.get_by_orth(self.mem, props[ORTH])
token.lex = lex
for attr_id, value in props.items():
Token.set_struct_attr(token, attr_id, value)
# NORM is the only one that overlaps between the two
# (which is maybe not great?)
if attr_id != NORM:
Lexeme.set_struct_attr(lex, attr_id, value)
return tokens
@property
def vectors_length(self):
return self.vectors.data.shape[1]
def reset_vectors(self, *, width=None, shape=None):
"""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.
"""
if width is not None and shape is not None:
raise ValueError(Errors.E065.format(width=width, shape=shape))
elif shape is not None:
self.vectors = Vectors(strings=self.strings, shape=shape)
else:
width = width if width is not None else self.vectors.data.shape[1]
self.vectors = Vectors(strings=self.strings, shape=(self.vectors.shape[0], width))
def prune_vectors(self, nr_row, batch_size=1024):
"""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.
DOCS: https://spacy.io/api/vocab#prune_vectors
"""
if self.vectors.mode != VectorsMode.default:
raise ValueError(Errors.E866)
ops = get_current_ops()
xp = get_array_module(self.vectors.data)
# Make sure all vectors are in the vocab
for orth in self.vectors:
self[orth]
# 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="uint64")
keys = xp.asarray([key for (prob, i, key) in priority], dtype="uint64")
keep = xp.ascontiguousarray(self.vectors.data[indices[:nr_row]])
toss = xp.ascontiguousarray(self.vectors.data[indices[nr_row:]])
self.vectors = Vectors(strings=self.strings, data=keep, keys=keys[:nr_row], name=self.vectors.name)
syn_keys, syn_rows, scores = self.vectors.most_similar(toss, batch_size=batch_size)
syn_keys = ops.to_numpy(syn_keys)
remap = {}
for i, key in enumerate(ops.to_numpy(keys[nr_row:])):
self.vectors.add(key, row=syn_rows[i][0])
word = self.strings[key]
synonym = self.strings[syn_keys[i][0]]
score = scores[i][0]
remap[word] = (synonym, score)
return remap
def get_vector(self, orth):
"""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.
orth (int / unicode): The hash value of a word, or its unicode string.
RETURNS (numpy.ndarray or cupy.ndarray): A word vector. Size
and shape determined by the `vocab.vectors` instance. Usually, a
numpy ndarray of shape (300,) and dtype float32.
DOCS: https://spacy.io/api/vocab#get_vector
"""
if isinstance(orth, str):
orth = self.strings.add(orth)
if self.has_vector(orth):
return self.vectors[orth]
xp = get_array_module(self.vectors.data)
vectors = xp.zeros((self.vectors_length,), dtype="f")
return vectors
def set_vector(self, orth, vector):
"""Set a vector for a word in the vocabulary. Words can be referenced
by string or int ID.
orth (int / str): The word.
vector (numpy.ndarray or cupy.nadarry[ndim=1, dtype='float32']): The vector to set.
DOCS: https://spacy.io/api/vocab#set_vector
"""
if isinstance(orth, str):
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))
lex = self[orth] # Add word to vocab if necessary
row = self.vectors.add(orth, vector=vector)
if row >= 0:
lex.rank = row
def has_vector(self, orth):
"""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.
orth (int / str): The word.
RETURNS (bool): Whether the word has a vector.
DOCS: https://spacy.io/api/vocab#has_vector
"""
if isinstance(orth, str):
orth = self.strings.add(orth)
return orth in self.vectors
property lookups:
def __get__(self):
return self._lookups
def __set__(self, lookups):
self._lookups = lookups
if lookups.has_table("lexeme_norm"):
self.lex_attr_getters[NORM] = util.add_lookups(
self.lex_attr_getters.get(NORM, LEX_ATTRS[NORM]),
self.lookups.get_table("lexeme_norm"),
)
def to_disk(self, path, *, exclude=tuple()):
"""Save the current state to a directory.
path (str or Path): A path to a directory, which will be created if
it doesn't exist.
exclude (Iterable[str]): String names of serialization fields to exclude.
DOCS: https://spacy.io/api/vocab#to_disk
"""
path = util.ensure_path(path)
if not path.exists():
path.mkdir()
setters = ["strings", "vectors"]
if "strings" not in exclude:
self.strings.to_disk(path / "strings.json")
if "vectors" not in "exclude":
self.vectors.to_disk(path, exclude=["strings"])
if "lookups" not in "exclude":
self.lookups.to_disk(path)
def from_disk(self, path, *, exclude=tuple()):
"""Loads state from a directory. Modifies the object in place and
returns it.
path (str or Path): A path to a directory.
exclude (Iterable[str]): String names of serialization fields to exclude.
RETURNS (Vocab): The modified `Vocab` object.
DOCS: https://spacy.io/api/vocab#to_disk
"""
path = util.ensure_path(path)
getters = ["strings", "vectors"]
if "strings" not in exclude:
self.strings.from_disk(path / "strings.json") # TODO: add exclude?
if "vectors" not in exclude:
if self.vectors is not None:
self.vectors.from_disk(path, exclude=["strings"])
if "lookups" not in exclude:
self.lookups.from_disk(path)
if "lexeme_norm" in self.lookups:
self.lex_attr_getters[NORM] = util.add_lookups(
self.lex_attr_getters.get(NORM, LEX_ATTRS[NORM]), self.lookups.get_table("lexeme_norm")
)
self.length = 0
self._by_orth = PreshMap()
return self
def to_bytes(self, *, exclude=tuple()):
"""Serialize the current state to a binary string.
exclude (Iterable[str]): String names of serialization fields to exclude.
RETURNS (bytes): The serialized form of the `Vocab` object.
DOCS: https://spacy.io/api/vocab#to_bytes
"""
def deserialize_vectors():
if self.vectors is None:
return None
else:
return self.vectors.to_bytes(exclude=["strings"])
getters = {
"strings": lambda: self.strings.to_bytes(),
"vectors": deserialize_vectors,
"lookups": lambda: self.lookups.to_bytes(),
}
return util.to_bytes(getters, exclude)
def from_bytes(self, bytes_data, *, exclude=tuple()):
"""Load state from a binary string.
bytes_data (bytes): The data to load from.
exclude (Iterable[str]): String names of serialization fields to exclude.
RETURNS (Vocab): The `Vocab` object.
DOCS: https://spacy.io/api/vocab#from_bytes
"""
def serialize_vectors(b):
if self.vectors is None:
return None
else:
return self.vectors.from_bytes(b, exclude=["strings"])
setters = {
"strings": lambda b: self.strings.from_bytes(b),
"vectors": lambda b: serialize_vectors(b),
"lookups": lambda b: self.lookups.from_bytes(b),
}
util.from_bytes(bytes_data, setters, exclude)
if "lexeme_norm" in self.lookups:
self.lex_attr_getters[NORM] = util.add_lookups(
self.lex_attr_getters.get(NORM, LEX_ATTRS[NORM]), self.lookups.get_table("lexeme_norm")
)
self.length = 0
self._by_orth = PreshMap()
return self
def _reset_cache(self, keys, strings):
# I'm not sure this made sense. Disable it for now.
raise NotImplementedError
def pickle_vocab(vocab):
sstore = vocab.strings
vectors = vocab.vectors
morph = vocab.morphology
_unused_object = vocab._unused_object
lex_attr_getters = srsly.pickle_dumps(vocab.lex_attr_getters)
lookups = vocab.lookups
get_noun_chunks = vocab.get_noun_chunks
return (unpickle_vocab,
(sstore, vectors, morph, _unused_object, lex_attr_getters, lookups, get_noun_chunks))
def unpickle_vocab(sstore, vectors, morphology, _unused_object,
lex_attr_getters, lookups, get_noun_chunks):
cdef Vocab vocab = Vocab()
vocab.vectors = vectors
vocab.strings = sstore
vocab.morphology = morphology
vocab._unused_object = _unused_object
vocab.lex_attr_getters = srsly.pickle_loads(lex_attr_getters)
vocab.lookups = lookups
vocab.get_noun_chunks = get_noun_chunks
return vocab
copy_reg.pickle(Vocab, pickle_vocab, unpickle_vocab)