spaCy/spacy/vectors.pyx
Adriane Boyd fb0da3e097
Support custom token/lexeme attribute for vectors (#12625)
* Support custom token/lexeme attribute for vectors

* Fix imports

* Back off to ORTH without Vectors.attr

* Fallback if vectors.attr doesn't exist

* Update docs
2023-06-28 09:43:14 +02:00

713 lines
26 KiB
Cython

cimport numpy as np
from cython.operator cimport dereference as deref
from libc.stdint cimport uint32_t, uint64_t
from libcpp.set cimport set as cppset
from murmurhash.mrmr cimport hash128_x64
import functools
import warnings
from enum import Enum
from typing import cast
import numpy
import srsly
from thinc.api import Ops, get_array_module, get_current_ops
from thinc.backends import get_array_ops
from thinc.types import Floats2d
from .attrs cimport ORTH, attr_id_t
from .strings cimport StringStore
from . import util
from .attrs import IDS
from .errors import Errors, Warnings
from .strings import get_string_id
def unpickle_vectors(bytes_data):
return Vectors().from_bytes(bytes_data)
class Mode(str, Enum):
default = "default"
floret = "floret"
@classmethod
def values(cls):
return list(cls.__members__.keys())
cdef class Vectors:
"""Store, save and load word vectors.
Vectors data is kept in the vectors.data attribute, which should be an
instance of numpy.ndarray (for CPU vectors) or cupy.ndarray
(for GPU vectors).
In the default mode, `vectors.key2row` is a dictionary mapping word hashes
to rows in the vectors.data table. Multiple keys can be mapped to the same
vector, and not all of the rows in the table need to be assigned - so
len(list(vectors.keys())) may be greater or smaller than vectors.shape[0].
In floret mode, the floret settings (minn, maxn, etc.) are used to
calculate the vector from the rows corresponding to the key's ngrams.
DOCS: https://spacy.io/api/vectors
"""
cdef public object strings
cdef public object name
cdef readonly object mode
cdef public object data
cdef public object key2row
cdef cppset[int] _unset
cdef readonly uint32_t minn
cdef readonly uint32_t maxn
cdef readonly uint32_t hash_count
cdef readonly uint32_t hash_seed
cdef readonly unicode bow
cdef readonly unicode eow
cdef readonly attr_id_t attr
def __init__(self, *, strings=None, shape=None, data=None, keys=None, name=None, mode=Mode.default, minn=0, maxn=0, hash_count=1, hash_seed=0, bow="<", eow=">", attr="ORTH"):
"""Create a new vector store.
strings (StringStore): The string store.
shape (tuple): Size of the table, as (# entries, # columns)
data (numpy.ndarray or cupy.ndarray): The vector data.
keys (iterable): A sequence of keys, aligned with the data.
name (str): A name to identify the vectors table.
mode (str): Vectors mode: "default" or "floret" (default: "default").
minn (int): The floret char ngram minn (default: 0).
maxn (int): The floret char ngram maxn (default: 0).
hash_count (int): The floret hash count (1-4, default: 1).
hash_seed (int): The floret hash seed (default: 0).
bow (str): The floret BOW string (default: "<").
eow (str): The floret EOW string (default: ">").
attr (Union[int, str]): The token attribute for the vector keys
(default: "ORTH").
DOCS: https://spacy.io/api/vectors#init
"""
self.strings = strings
if self.strings is None:
self.strings = StringStore()
self.name = name
if mode not in Mode.values():
raise ValueError(
Errors.E202.format(
name="vectors",
mode=mode,
modes=str(Mode.values())
)
)
self.mode = Mode(mode).value
self.key2row = {}
self.minn = minn
self.maxn = maxn
self.hash_count = hash_count
self.hash_seed = hash_seed
self.bow = bow
self.eow = eow
if isinstance(attr, (int, long)):
self.attr = attr
else:
attr = attr.upper()
if attr == "TEXT":
attr = "ORTH"
self.attr = IDS.get(attr, ORTH)
if self.mode == Mode.default:
if data is None:
if shape is None:
shape = (0,0)
ops = get_current_ops()
data = ops.xp.zeros(shape, dtype="f")
self._unset = cppset[int]({i for i in range(data.shape[0])})
else:
self._unset = cppset[int]()
self.data = data
if keys is not None:
for i, key in enumerate(keys):
self.add(key, row=i)
elif self.mode == Mode.floret:
if maxn < minn:
raise ValueError(Errors.E863)
if hash_count < 1 or hash_count >= 5:
raise ValueError(Errors.E862)
if data is None:
raise ValueError(Errors.E864)
if keys is not None:
raise ValueError(Errors.E861)
self.data = data
self._unset = cppset[int]()
@property
def shape(self):
"""Get `(rows, dims)` tuples of number of rows and number of dimensions
in the vector table.
RETURNS (tuple): A `(rows, dims)` pair.
DOCS: https://spacy.io/api/vectors#shape
"""
return self.data.shape
@property
def size(self):
"""The vector size i,e. rows * dims.
RETURNS (int): The vector size.
DOCS: https://spacy.io/api/vectors#size
"""
return self.data.size
@property
def is_full(self):
"""Whether the vectors table is full.
RETURNS (bool): `True` if no slots are available for new keys.
DOCS: https://spacy.io/api/vectors#is_full
"""
if self.mode == Mode.floret:
return True
return self._unset.size() == 0
@property
def n_keys(self):
"""Get the number of keys in the table. Note that this is the number
of all keys, not just unique vectors.
RETURNS (int): The number of keys in the table for default vectors.
For floret vectors, return -1.
DOCS: https://spacy.io/api/vectors#n_keys
"""
if self.mode == Mode.floret:
return -1
return len(self.key2row)
def __reduce__(self):
return (unpickle_vectors, (self.to_bytes(),))
def __getitem__(self, key):
"""Get a vector by key. If the key is not found, a KeyError is raised.
key (str/int): The key to get the vector for.
RETURNS (ndarray): The vector for the key.
DOCS: https://spacy.io/api/vectors#getitem
"""
if self.mode == Mode.default:
i = self.key2row.get(get_string_id(key), None)
if i is None:
raise KeyError(Errors.E058.format(key=key))
else:
return self.data[i]
elif self.mode == Mode.floret:
return self.get_batch([key])[0]
raise KeyError(Errors.E058.format(key=key))
def __setitem__(self, key, vector):
"""Set a vector for the given key.
key (str/int): The key to set the vector for.
vector (ndarray): The vector to set.
DOCS: https://spacy.io/api/vectors#setitem
"""
if self.mode == Mode.floret:
warnings.warn(Warnings.W115.format(method="Vectors.__setitem__"))
return
key = get_string_id(key)
i = self.key2row[key]
self.data[i] = vector
if self._unset.count(i):
self._unset.erase(self._unset.find(i))
def __iter__(self):
"""Iterate over the keys in the table.
YIELDS (int): A key in the table.
DOCS: https://spacy.io/api/vectors#iter
"""
yield from self.key2row
def __len__(self):
"""Return the number of vectors in the table.
RETURNS (int): The number of vectors in the data.
DOCS: https://spacy.io/api/vectors#len
"""
return self.data.shape[0]
def __contains__(self, key):
"""Check whether a key has been mapped to a vector entry in the table.
key (int): The key to check.
RETURNS (bool): Whether the key has a vector entry.
DOCS: https://spacy.io/api/vectors#contains
"""
if self.mode == Mode.floret:
return True
else:
return key in self.key2row
def __eq__(self, other):
# Check for equality, with faster checks first
return (
self.shape == other.shape
and self.key2row == other.key2row
and self.to_bytes(exclude=["strings"])
== other.to_bytes(exclude=["strings"])
)
def resize(self, shape, inplace=False):
"""Resize the underlying vectors array. If inplace=True, the memory
is reallocated. This may cause other references to the data to become
invalid, so only use inplace=True if you're sure that's what you want.
If the number of vectors is reduced, keys mapped to rows that have been
deleted are removed. These removed items are returned as a list of
`(key, row)` tuples.
shape (tuple): A `(rows, dims)` tuple.
inplace (bool): Reallocate the memory.
RETURNS (list): The removed items as a list of `(key, row)` tuples.
DOCS: https://spacy.io/api/vectors#resize
"""
if self.mode == Mode.floret:
warnings.warn(Warnings.W115.format(method="Vectors.resize"))
return -1
xp = get_array_module(self.data)
if inplace:
if shape[1] != self.data.shape[1]:
raise ValueError(Errors.E193.format(new_dim=shape[1], curr_dim=self.data.shape[1]))
if xp == numpy:
self.data.resize(shape, refcheck=False)
else:
raise ValueError(Errors.E192)
else:
resized_array = xp.zeros(shape, dtype=self.data.dtype)
copy_shape = (min(shape[0], self.data.shape[0]), min(shape[1], self.data.shape[1]))
resized_array[:copy_shape[0], :copy_shape[1]] = self.data[:copy_shape[0], :copy_shape[1]]
self.data = resized_array
self._sync_unset()
removed_items = []
for key, row in self.key2row.copy().items():
if row >= shape[0]:
self.key2row.pop(key)
removed_items.append((key, row))
return removed_items
def keys(self):
"""RETURNS (iterable): A sequence of keys in the table."""
return self.key2row.keys()
def values(self):
"""Iterate over vectors that have been assigned to at least one key.
Note that some vectors may be unassigned, so the number of vectors
returned may be less than the length of the vectors table.
YIELDS (ndarray): A vector in the table.
DOCS: https://spacy.io/api/vectors#values
"""
for row, vector in enumerate(range(self.data.shape[0])):
if not self._unset.count(row):
yield vector
def items(self):
"""Iterate over `(key, vector)` pairs.
YIELDS (tuple): A key/vector pair.
DOCS: https://spacy.io/api/vectors#items
"""
for key, row in self.key2row.items():
yield key, self.data[row]
def find(self, *, key=None, keys=None, row=None, rows=None):
"""Look up one or more keys by row, or vice versa.
key (Union[int, str]): Find the row that the given key points to.
Returns int, -1 if missing.
keys (Iterable[Union[int, str]]): Find rows that the keys point to.
Returns ndarray.
row (int): Find the first key that points to the row.
Returns int.
rows (Iterable[int]): Find the keys that point to the rows.
Returns ndarray.
RETURNS: The requested key, keys, row or rows.
"""
if self.mode == Mode.floret:
raise ValueError(
Errors.E858.format(
mode=self.mode,
alternative="Use Vectors[key] instead.",
)
)
if sum(arg is None for arg in (key, keys, row, rows)) != 3:
bad_kwargs = {"key": key, "keys": keys, "row": row, "rows": rows}
raise ValueError(Errors.E059.format(kwargs=bad_kwargs))
xp = get_array_module(self.data)
if key is not None:
key = get_string_id(key)
return self.key2row.get(int(key), -1)
elif keys is not None:
keys = [get_string_id(key) for key in keys]
rows = [self.key2row.get(int(key), -1) for key in keys]
return xp.asarray(rows, dtype="i")
else:
row2key = {row: key for key, row in self.key2row.items()}
if row is not None:
return row2key[row]
else:
results = [row2key[row] for row in rows]
return xp.asarray(results, dtype="uint64")
def _get_ngram_hashes(self, unicode s):
"""Calculate up to 4 32-bit hash values with MurmurHash3_x64_128 using
the floret hash settings.
key (str): The string key.
RETURNS: A list of the integer hashes.
"""
# MurmurHash3_x64_128 returns an array of 2 uint64_t values.
cdef uint64_t[2] out
chars = s.encode("utf8")
cdef char* utf8_string = chars
hash128_x64(utf8_string, len(chars), self.hash_seed, &out)
rows = [
out[0] & 0xffffffffu,
out[0] >> 32,
out[1] & 0xffffffffu,
out[1] >> 32,
]
return rows[:min(self.hash_count, 4)]
def _get_ngrams(self, unicode key):
"""Get all padded ngram strings using the ngram settings.
key (str): The string key.
RETURNS: A list of the ngram strings for the padded key.
"""
key = self.bow + key + self.eow
ngrams = [key] + [
key[start:start+ngram_size]
for ngram_size in range(self.minn, self.maxn + 1)
for start in range(0, len(key) - ngram_size + 1)
]
return ngrams
def get_batch(self, keys):
"""Get the vectors for the provided keys efficiently as a batch.
keys (Iterable[Union[int, str]]): The keys.
RETURNS: The requested vectors from the vector table.
"""
ops = get_array_ops(self.data)
if self.mode == Mode.default:
rows = self.find(keys=keys)
vecs = self.data[rows]
elif self.mode == Mode.floret:
keys = [self.strings.as_string(key) for key in keys]
if sum(len(key) for key in keys) == 0:
return ops.xp.zeros((len(keys), self.data.shape[1]))
unique_keys = tuple(set(keys))
row_index = {key: i for i, key in enumerate(unique_keys)}
rows = [row_index[key] for key in keys]
indices = []
lengths = []
for key in unique_keys:
if key == "":
ngram_rows = []
else:
ngram_rows = [
h % self.data.shape[0]
for ngram in self._get_ngrams(key)
for h in self._get_ngram_hashes(ngram)
]
indices.extend(ngram_rows)
lengths.append(len(ngram_rows))
indices = ops.asarray(indices, dtype="int32")
lengths = ops.asarray(lengths, dtype="int32")
vecs = ops.reduce_mean(cast(Floats2d, self.data[indices]), lengths)
vecs = vecs[rows]
return ops.as_contig(vecs)
def add(self, key, *, vector=None, row=None):
"""Add a key to the table. Keys can be mapped to an existing vector
by setting `row`, or a new vector can be added.
key (int): The key to add.
vector (ndarray / None): A vector to add for the key.
row (int / None): The row number of a vector to map the key to.
RETURNS (int): The row the vector was added to.
DOCS: https://spacy.io/api/vectors#add
"""
if self.mode == Mode.floret:
warnings.warn(Warnings.W115.format(method="Vectors.add"))
return -1
# use int for all keys and rows in key2row for more efficient access
# and serialization
key = int(get_string_id(key))
if row is not None:
row = int(row)
if row is None and key in self.key2row:
row = self.key2row[key]
elif row is None:
if self.is_full:
raise ValueError(Errors.E060.format(rows=self.data.shape[0],
cols=self.data.shape[1]))
row = deref(self._unset.begin())
if row < self.data.shape[0]:
self.key2row[key] = row
else:
raise ValueError(Errors.E197.format(row=row, key=key))
if vector is not None:
xp = get_array_module(self.data)
vector = xp.asarray(vector)
self.data[row] = vector
if self._unset.count(row):
self._unset.erase(self._unset.find(row))
return row
def most_similar(self, queries, *, batch_size=1024, n=1, sort=True):
"""For each of the given vectors, find the n most similar entries
to it, by cosine.
Queries are by vector. Results are returned as a `(keys, best_rows,
scores)` tuple. If `queries` is large, the calculations are performed in
chunks, to avoid consuming too much memory. You can set the `batch_size`
to control the size/space trade-off during the calculations.
queries (ndarray): An array with one or more vectors.
batch_size (int): The batch size to use.
n (int): The number of entries to return for each query.
sort (bool): Whether to sort the n entries returned by score.
RETURNS (tuple): The most similar entries as a `(keys, best_rows, scores)`
tuple.
"""
if self.mode == Mode.floret:
raise ValueError(Errors.E858.format(
mode=self.mode,
alternative="",
))
xp = get_array_module(self.data)
filled = sorted(list({row for row in self.key2row.values()}))
if len(filled) < n:
raise ValueError(Errors.E198.format(n=n, n_rows=len(filled)))
filled = xp.asarray(filled)
norms = xp.linalg.norm(self.data[filled], axis=1, keepdims=True)
norms[norms == 0] = 1
vectors = self.data[filled] / norms
best_rows = xp.zeros((queries.shape[0], n), dtype='i')
scores = xp.zeros((queries.shape[0], n), dtype='f')
# Work in batches, to avoid memory problems.
for i in range(0, queries.shape[0], batch_size):
batch = queries[i : i+batch_size]
batch_norms = xp.linalg.norm(batch, axis=1, keepdims=True)
batch_norms[batch_norms == 0] = 1
batch /= batch_norms
# batch e.g. (1024, 300)
# vectors e.g. (10000, 300)
# sims e.g. (1024, 10000)
sims = xp.dot(batch, vectors.T)
best_rows[i:i+batch_size] = xp.argpartition(sims, -n, axis=1)[:,-n:]
scores[i:i+batch_size] = xp.partition(sims, -n, axis=1)[:,-n:]
if sort and n >= 2:
sorted_index = xp.arange(scores.shape[0])[:,None][i:i+batch_size],xp.argsort(scores[i:i+batch_size], axis=1)[:,::-1]
scores[i:i+batch_size] = scores[sorted_index]
best_rows[i:i+batch_size] = best_rows[sorted_index]
for i, j in numpy.ndindex(best_rows.shape):
best_rows[i, j] = filled[best_rows[i, j]]
# Round values really close to 1 or -1
scores = xp.around(scores, decimals=4, out=scores)
# Account for numerical error we want to return in range -1, 1
scores = xp.clip(scores, a_min=-1, a_max=1, out=scores)
row2key = {row: key for key, row in self.key2row.items()}
numpy_rows = get_current_ops().to_numpy(best_rows)
keys = xp.asarray(
[[row2key[row] for row in numpy_rows[i] if row in row2key]
for i in range(len(queries)) ], dtype="uint64")
return (keys, best_rows, scores)
def to_ops(self, ops: Ops):
self.data = ops.asarray(self.data)
def _get_cfg(self):
if self.mode == Mode.default:
return {
"mode": Mode(self.mode).value,
}
elif self.mode == Mode.floret:
return {
"mode": Mode(self.mode).value,
"minn": self.minn,
"maxn": self.maxn,
"hash_count": self.hash_count,
"hash_seed": self.hash_seed,
"bow": self.bow,
"eow": self.eow,
"attr": self.attr,
}
def _set_cfg(self, cfg):
self.mode = Mode(cfg.get("mode", Mode.default)).value
self.minn = cfg.get("minn", 0)
self.maxn = cfg.get("maxn", 0)
self.hash_count = cfg.get("hash_count", 0)
self.hash_seed = cfg.get("hash_seed", 0)
self.bow = cfg.get("bow", "<")
self.eow = cfg.get("eow", ">")
self.attr = cfg.get("attr", ORTH)
def to_disk(self, path, *, exclude=tuple()):
"""Save the current state to a directory.
path (str / Path): A path to a directory, which will be created if
it doesn't exists.
DOCS: https://spacy.io/api/vectors#to_disk
"""
xp = get_array_module(self.data)
if xp is numpy:
save_array = lambda arr, file_: xp.save(file_, arr, allow_pickle=False)
else:
save_array = lambda arr, file_: xp.save(file_, arr)
def save_vectors(path):
# the source of numpy.save indicates that the file object is closed after use.
# but it seems that somehow this does not happen, as ResourceWarnings are raised here.
# in order to not rely on this, wrap in context manager.
ops = get_current_ops()
with path.open("wb") as _file:
save_array(ops.to_numpy(self.data, byte_order="<"), _file)
serializers = {
"strings": lambda p: self.strings.to_disk(p.with_suffix(".json")),
"vectors": lambda p: save_vectors(p),
"key2row": lambda p: srsly.write_msgpack(p, self.key2row),
"vectors.cfg": lambda p: srsly.write_json(p, self._get_cfg()),
}
return util.to_disk(path, serializers, exclude)
def from_disk(self, path, *, exclude=tuple()):
"""Loads state from a directory. Modifies the object in place and
returns it.
path (str / Path): Directory path, string or Path-like object.
RETURNS (Vectors): The modified object.
DOCS: https://spacy.io/api/vectors#from_disk
"""
def load_key2row(path):
if path.exists():
self.key2row = srsly.read_msgpack(path)
for key, row in self.key2row.items():
if self._unset.count(row):
self._unset.erase(self._unset.find(row))
def load_keys(path):
if path.exists():
keys = numpy.load(str(path))
for i, key in enumerate(keys):
self.add(key, row=i)
def load_vectors(path):
ops = get_current_ops()
if path.exists():
self.data = ops.xp.load(str(path))
self.to_ops(ops)
def load_settings(path):
if path.exists():
self._set_cfg(srsly.read_json(path))
serializers = {
"strings": lambda p: self.strings.from_disk(p.with_suffix(".json")),
"vectors": load_vectors,
"keys": load_keys,
"key2row": load_key2row,
"vectors.cfg": load_settings,
}
util.from_disk(path, serializers, exclude)
self._sync_unset()
return self
def to_bytes(self, *, exclude=tuple()):
"""Serialize the current state to a binary string.
exclude (list): String names of serialization fields to exclude.
RETURNS (bytes): The serialized form of the `Vectors` object.
DOCS: https://spacy.io/api/vectors#to_bytes
"""
def serialize_weights():
if hasattr(self.data, "to_bytes"):
return self.data.to_bytes()
else:
ops = get_current_ops()
return srsly.msgpack_dumps(ops.to_numpy(self.data, byte_order="<"))
serializers = {
"strings": lambda: self.strings.to_bytes(),
"key2row": lambda: srsly.msgpack_dumps(self.key2row),
"vectors": serialize_weights,
"vectors.cfg": lambda: srsly.json_dumps(self._get_cfg()),
}
return util.to_bytes(serializers, exclude)
def from_bytes(self, data, *, exclude=tuple()):
"""Load state from a binary string.
data (bytes): The data to load from.
exclude (list): String names of serialization fields to exclude.
RETURNS (Vectors): The `Vectors` object.
DOCS: https://spacy.io/api/vectors#from_bytes
"""
def deserialize_weights(b):
if hasattr(self.data, "from_bytes"):
self.data.from_bytes()
else:
xp = get_array_module(self.data)
self.data = xp.asarray(srsly.msgpack_loads(b))
ops = get_current_ops()
self.to_ops(ops)
deserializers = {
"strings": lambda b: self.strings.from_bytes(b),
"key2row": lambda b: self.key2row.update(srsly.msgpack_loads(b)),
"vectors": deserialize_weights,
"vectors.cfg": lambda b: self._set_cfg(srsly.json_loads(b))
}
util.from_bytes(data, deserializers, exclude)
self._sync_unset()
return self
def clear(self):
"""Clear all entries in the vector table.
DOCS: https://spacy.io/api/vectors#clear
"""
if self.mode == Mode.floret:
raise ValueError(Errors.E859)
self.key2row = {}
self._sync_unset()
def _sync_unset(self):
filled = {row for row in self.key2row.values()}
self._unset = cppset[int]({row for row in range(self.data.shape[0]) if row not in filled})