spaCy/spacy/training/initialize.py
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

338 lines
13 KiB
Python

from typing import Union, Dict, Optional, Any, IO, TYPE_CHECKING
from thinc.api import Config, fix_random_seed, set_gpu_allocator
from thinc.api import ConfigValidationError
from pathlib import Path
import srsly
import numpy
import tarfile
import gzip
import zipfile
import tqdm
from itertools import islice
import warnings
from .pretrain import get_tok2vec_ref
from ..lookups import Lookups
from ..vectors import Vectors, Mode as VectorsMode
from ..errors import Errors, Warnings
from ..schemas import ConfigSchemaTraining
from ..util import registry, load_model_from_config, resolve_dot_names, logger
from ..util import load_model, ensure_path, get_sourced_components
from ..util import OOV_RANK, DEFAULT_OOV_PROB
if TYPE_CHECKING:
from ..language import Language # noqa: F401
def init_nlp(config: Config, *, use_gpu: int = -1) -> "Language":
raw_config = config
config = raw_config.interpolate()
if "seed" not in config["training"]:
raise ValueError(Errors.E1015.format(value="[training] seed"))
if "gpu_allocator" not in config["training"]:
raise ValueError(Errors.E1015.format(value="[training] gpu_allocator"))
if config["training"]["seed"] is not None:
fix_random_seed(config["training"]["seed"])
allocator = config["training"]["gpu_allocator"]
if use_gpu >= 0 and allocator:
set_gpu_allocator(allocator)
# Use original config here before it's resolved to functions
sourced = get_sourced_components(config)
nlp = load_model_from_config(raw_config, auto_fill=True)
logger.info("Set up nlp object from config")
config = nlp.config.interpolate()
# Resolve all training-relevant sections using the filled nlp config
T = registry.resolve(config["training"], schema=ConfigSchemaTraining)
dot_names = [T["train_corpus"], T["dev_corpus"]]
if not isinstance(T["train_corpus"], str):
raise ConfigValidationError(
desc=Errors.E897.format(
field="training.train_corpus", type=type(T["train_corpus"])
)
)
if not isinstance(T["dev_corpus"], str):
raise ConfigValidationError(
desc=Errors.E897.format(
field="training.dev_corpus", type=type(T["dev_corpus"])
)
)
train_corpus, dev_corpus = resolve_dot_names(config, dot_names)
optimizer = T["optimizer"]
# Components that shouldn't be updated during training
frozen_components = T["frozen_components"]
# Sourced components that require resume_training
resume_components = [p for p in sourced if p not in frozen_components]
logger.info(f"Pipeline: {nlp.pipe_names}")
if resume_components:
with nlp.select_pipes(enable=resume_components):
logger.info(f"Resuming training for: {resume_components}")
nlp.resume_training(sgd=optimizer)
# Make sure that listeners are defined before initializing further
nlp._link_components()
with nlp.select_pipes(disable=[*frozen_components, *resume_components]):
if T["max_epochs"] == -1:
sample_size = 100
logger.debug(
f"Due to streamed train corpus, using only first {sample_size} "
f"examples for initialization. If necessary, provide all labels "
f"in [initialize]. More info: https://spacy.io/api/cli#init_labels"
)
nlp.initialize(
lambda: islice(train_corpus(nlp), sample_size), sgd=optimizer
)
else:
nlp.initialize(lambda: train_corpus(nlp), sgd=optimizer)
logger.info(f"Initialized pipeline components: {nlp.pipe_names}")
# Detect components with listeners that are not frozen consistently
for name, proc in nlp.pipeline:
for listener in getattr(
proc, "listening_components", []
): # e.g. tok2vec/transformer
# Don't warn about components not in the pipeline
if listener not in nlp.pipe_names:
continue
if listener in frozen_components and name not in frozen_components:
logger.warning(Warnings.W087.format(name=name, listener=listener))
# We always check this regardless, in case user freezes tok2vec
if listener not in frozen_components and name in frozen_components:
if name not in T["annotating_components"]:
logger.warning(Warnings.W086.format(name=name, listener=listener))
return nlp
def init_vocab(
nlp: "Language",
*,
data: Optional[Path] = None,
lookups: Optional[Lookups] = None,
vectors: Optional[str] = None,
) -> None:
if lookups:
nlp.vocab.lookups = lookups
logger.info(f"Added vocab lookups: {', '.join(lookups.tables)}")
data_path = ensure_path(data)
if data_path is not None:
lex_attrs = srsly.read_jsonl(data_path)
for lexeme in nlp.vocab:
lexeme.rank = OOV_RANK
for attrs in lex_attrs:
if "settings" in attrs:
continue
lexeme = nlp.vocab[attrs["orth"]]
lexeme.set_attrs(**attrs)
if len(nlp.vocab):
oov_prob = min(lex.prob for lex in nlp.vocab) - 1
else:
oov_prob = DEFAULT_OOV_PROB
nlp.vocab.cfg.update({"oov_prob": oov_prob})
logger.info(f"Added {len(nlp.vocab)} lexical entries to the vocab")
logger.info("Created vocabulary")
if vectors is not None:
load_vectors_into_model(nlp, vectors)
logger.info(f"Added vectors: {vectors}")
# warn if source model vectors are not identical
sourced_vectors_hashes = nlp.meta.pop("_sourced_vectors_hashes", {})
vectors_hash = hash(nlp.vocab.vectors.to_bytes())
for sourced_component, sourced_vectors_hash in sourced_vectors_hashes.items():
if vectors_hash != sourced_vectors_hash:
warnings.warn(Warnings.W113.format(name=sourced_component))
logger.info("Finished initializing nlp object")
def load_vectors_into_model(
nlp: "Language", name: Union[str, Path], *, add_strings: bool = True
) -> None:
"""Load word vectors from an installed model or path into a model instance."""
try:
# Load with the same vocab, which automatically adds the vectors to
# the current nlp object. Exclude lookups so they are not modified.
exclude = ["lookups"]
if not add_strings:
exclude.append("strings")
vectors_nlp = load_model(name, vocab=nlp.vocab, exclude=exclude)
except ConfigValidationError as e:
title = f"Config validation error for vectors {name}"
desc = (
"This typically means that there's a problem in the config.cfg included "
"with the packaged vectors. Make sure that the vectors package you're "
"loading is compatible with the current version of spaCy."
)
err = ConfigValidationError.from_error(e, title=title, desc=desc)
raise err from None
if (
len(vectors_nlp.vocab.vectors.keys()) == 0
and vectors_nlp.vocab.vectors.mode != VectorsMode.floret
) or (
vectors_nlp.vocab.vectors.data.shape[0] == 0
and vectors_nlp.vocab.vectors.mode == VectorsMode.floret
):
logger.warning(Warnings.W112.format(name=name))
for lex in nlp.vocab:
lex.rank = nlp.vocab.vectors.key2row.get(lex.orth, OOV_RANK) # type: ignore[attr-defined]
def init_tok2vec(
nlp: "Language", pretrain_config: Dict[str, Any], init_config: Dict[str, Any]
) -> bool:
# Load pretrained tok2vec weights - cf. CLI command 'pretrain'
P = pretrain_config
I = init_config
weights_data = None
init_tok2vec = ensure_path(I["init_tok2vec"])
if init_tok2vec is not None:
if not init_tok2vec.exists():
err = f"can't find pretrained tok2vec: {init_tok2vec}"
errors = [{"loc": ["initialize", "init_tok2vec"], "msg": err}]
raise ConfigValidationError(config=nlp.config, errors=errors)
with init_tok2vec.open("rb") as file_:
weights_data = file_.read()
if weights_data is not None:
layer = get_tok2vec_ref(nlp, P)
layer.from_bytes(weights_data)
logger.info(f"Loaded pretrained weights from {init_tok2vec}")
return True
return False
def convert_vectors(
nlp: "Language",
vectors_loc: Optional[Path],
*,
truncate: int,
prune: int,
name: Optional[str] = None,
mode: str = VectorsMode.default,
) -> None:
vectors_loc = ensure_path(vectors_loc)
if vectors_loc and vectors_loc.parts[-1].endswith(".npz"):
nlp.vocab.vectors = Vectors(
strings=nlp.vocab.strings, data=numpy.load(vectors_loc.open("rb"))
)
for lex in nlp.vocab:
if lex.rank and lex.rank != OOV_RANK:
nlp.vocab.vectors.add(lex.orth, row=lex.rank) # type: ignore[attr-defined]
else:
if vectors_loc:
logger.info(f"Reading vectors from {vectors_loc}")
vectors_data, vector_keys, floret_settings = read_vectors(
vectors_loc,
truncate,
mode=mode,
)
logger.info(f"Loaded vectors from {vectors_loc}")
else:
vectors_data, vector_keys = (None, None)
if vector_keys is not None and mode != VectorsMode.floret:
for word in vector_keys:
if word not in nlp.vocab:
nlp.vocab[word]
if vectors_data is not None:
if mode == VectorsMode.floret:
nlp.vocab.vectors = Vectors(
strings=nlp.vocab.strings,
data=vectors_data,
**floret_settings,
)
else:
nlp.vocab.vectors = Vectors(
strings=nlp.vocab.strings, data=vectors_data, keys=vector_keys
)
if name is None:
# TODO: Is this correct? Does this matter?
nlp.vocab.vectors.name = f"{nlp.meta['lang']}_{nlp.meta['name']}.vectors"
else:
nlp.vocab.vectors.name = name
nlp.meta["vectors"]["name"] = nlp.vocab.vectors.name
if prune >= 1 and mode != VectorsMode.floret:
nlp.vocab.prune_vectors(prune)
def read_vectors(
vectors_loc: Path, truncate_vectors: int, *, mode: str = VectorsMode.default
):
f = ensure_shape(vectors_loc)
header_parts = next(f).split()
shape = tuple(int(size) for size in header_parts[:2])
floret_settings = {}
if mode == VectorsMode.floret:
if len(header_parts) != 8:
raise ValueError(
"Invalid header for floret vectors. "
"Expected: bucket dim minn maxn hash_count hash_seed BOW EOW"
)
floret_settings = {
"mode": "floret",
"minn": int(header_parts[2]),
"maxn": int(header_parts[3]),
"hash_count": int(header_parts[4]),
"hash_seed": int(header_parts[5]),
"bow": header_parts[6],
"eow": header_parts[7],
}
if truncate_vectors >= 1:
raise ValueError(Errors.E860)
else:
assert len(header_parts) == 2
if truncate_vectors >= 1:
shape = (truncate_vectors, shape[1])
vectors_data = numpy.zeros(shape=shape, dtype="f")
vectors_keys = []
for i, line in enumerate(tqdm.tqdm(f)):
line = line.rstrip()
pieces = line.rsplit(" ", vectors_data.shape[1])
word = pieces.pop(0)
if len(pieces) != vectors_data.shape[1]:
raise ValueError(Errors.E094.format(line_num=i, loc=vectors_loc))
vectors_data[i] = numpy.asarray(pieces, dtype="f")
vectors_keys.append(word)
if i == truncate_vectors - 1:
break
return vectors_data, vectors_keys, floret_settings
def open_file(loc: Union[str, Path]) -> IO:
"""Handle .gz, .tar.gz or unzipped files"""
loc = ensure_path(loc)
if tarfile.is_tarfile(str(loc)):
return tarfile.open(str(loc), "r:gz") # type: ignore[return-value]
elif loc.parts[-1].endswith("gz"):
return (line.decode("utf8") for line in gzip.open(str(loc), "r")) # type: ignore[return-value]
elif loc.parts[-1].endswith("zip"):
zip_file = zipfile.ZipFile(str(loc))
names = zip_file.namelist()
file_ = zip_file.open(names[0])
return (line.decode("utf8") for line in file_) # type: ignore[return-value]
else:
return loc.open("r", encoding="utf8")
def ensure_shape(vectors_loc):
"""Ensure that the first line of the data is the vectors shape.
If it's not, we read in the data and output the shape as the first result,
so that the reader doesn't have to deal with the problem.
"""
lines = open_file(vectors_loc)
first_line = next(lines)
try:
shape = tuple(int(size) for size in first_line.split()[:2])
except ValueError:
shape = None
if shape is not None:
# All good, give the data
yield first_line
yield from lines
else:
# Figure out the shape, make it the first value, and then give the
# rest of the data.
width = len(first_line.split()) - 1
length = 1
for _ in lines:
length += 1
yield f"{length} {width}"
# Reading the lines in again from file. This to avoid having to
# store all the results in a list in memory
lines2 = open_file(vectors_loc)
yield from lines2