spaCy/spacy/language.py

1200 lines
46 KiB
Python
Raw Normal View History

# coding: utf8
from __future__ import absolute_import, unicode_literals
2017-05-25 04:10:54 +03:00
import random
2017-07-25 19:57:59 +03:00
import itertools
2020-04-28 14:37:37 +03:00
import warnings
from thinc.extra import load_nlp
import weakref
import functools
2017-10-27 22:07:59 +03:00
from collections import OrderedDict
from contextlib import contextmanager
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
from copy import copy, deepcopy
2017-10-27 22:07:59 +03:00
from thinc.neural import Model
import srsly
import multiprocessing as mp
from itertools import chain, cycle
2017-05-18 12:25:19 +03:00
from .tokenizer import Tokenizer
from .tokens.underscore import Underscore
from .vocab import Vocab
from .lemmatizer import Lemmatizer
from .lookups import Lookups
from .analysis import analyze_pipes, analyze_all_pipes, validate_attrs
from .compat import izip, basestring_, is_python2, class_types
from .gold import GoldParse
from .scorer import Scorer
2017-11-06 17:06:27 +03:00
from ._ml import link_vectors_to_models, create_default_optimizer
Reduce stored lexemes data, move feats to lookups (#5238) * Reduce stored lexemes data, move feats to lookups * Move non-derivable lexemes features (`norm / cluster / prob`) to `spacy-lookups-data` as lookups * Get/set `norm` in both lookups and `LexemeC`, serialize in lookups * Remove `cluster` and `prob` from `LexemesC`, get/set/serialize in lookups only * Remove serialization of lexemes data as `vocab/lexemes.bin` * Remove `SerializedLexemeC` * Remove `Lexeme.to_bytes/from_bytes` * Modify normalization exception loading: * Always create `Vocab.lookups` table `lexeme_norm` for normalization exceptions * Load base exceptions from `lang.norm_exceptions`, but load language-specific exceptions from lookups * Set `lex_attr_getter[NORM]` including new lookups table in `BaseDefaults.create_vocab()` and when deserializing `Vocab` * Remove all cached lexemes when deserializing vocab to override existing normalizations with the new normalizations (as a replacement for the previous step that replaced all lexemes data with the deserialized data) * Skip English normalization test Skip English normalization test because the data is now in `spacy-lookups-data`. * Remove norm exceptions Moved to spacy-lookups-data. * Move norm exceptions test to spacy-lookups-data * Load extra lookups from spacy-lookups-data lazily Load extra lookups (currently for cluster and prob) lazily from the entry point `lg_extra` as `Vocab.lookups_extra`. * Skip creating lexeme cache on load To improve model loading times, do not create the full lexeme cache when loading. The lexemes will be created on demand when processing. * Identify numeric values in Lexeme.set_attrs() With the removal of a special case for `PROB`, also identify `float` to avoid trying to convert it with the `StringStore`. * Skip lexeme cache init in from_bytes * Unskip and update lookups tests for python3.6+ * Update vocab pickle to include lookups_extra * Update vocab serialization tests Check strings rather than lexemes since lexemes aren't initialized automatically, account for addition of "_SP". * Re-skip lookups test because of python3.5 * Skip PROB/float values in Lexeme.set_attrs * Convert is_oov from lexeme flag to lex in vectors Instead of storing `is_oov` as a lexeme flag, `is_oov` reports whether the lexeme has a vector. Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
2020-05-19 16:59:14 +03:00
from .attrs import IS_STOP, LANG, NORM
from .lang.punctuation import TOKENIZER_PREFIXES, TOKENIZER_SUFFIXES
from .lang.punctuation import TOKENIZER_INFIXES
from .lang.tokenizer_exceptions import TOKEN_MATCH, URL_MATCH
Reduce stored lexemes data, move feats to lookups (#5238) * Reduce stored lexemes data, move feats to lookups * Move non-derivable lexemes features (`norm / cluster / prob`) to `spacy-lookups-data` as lookups * Get/set `norm` in both lookups and `LexemeC`, serialize in lookups * Remove `cluster` and `prob` from `LexemesC`, get/set/serialize in lookups only * Remove serialization of lexemes data as `vocab/lexemes.bin` * Remove `SerializedLexemeC` * Remove `Lexeme.to_bytes/from_bytes` * Modify normalization exception loading: * Always create `Vocab.lookups` table `lexeme_norm` for normalization exceptions * Load base exceptions from `lang.norm_exceptions`, but load language-specific exceptions from lookups * Set `lex_attr_getter[NORM]` including new lookups table in `BaseDefaults.create_vocab()` and when deserializing `Vocab` * Remove all cached lexemes when deserializing vocab to override existing normalizations with the new normalizations (as a replacement for the previous step that replaced all lexemes data with the deserialized data) * Skip English normalization test Skip English normalization test because the data is now in `spacy-lookups-data`. * Remove norm exceptions Moved to spacy-lookups-data. * Move norm exceptions test to spacy-lookups-data * Load extra lookups from spacy-lookups-data lazily Load extra lookups (currently for cluster and prob) lazily from the entry point `lg_extra` as `Vocab.lookups_extra`. * Skip creating lexeme cache on load To improve model loading times, do not create the full lexeme cache when loading. The lexemes will be created on demand when processing. * Identify numeric values in Lexeme.set_attrs() With the removal of a special case for `PROB`, also identify `float` to avoid trying to convert it with the `StringStore`. * Skip lexeme cache init in from_bytes * Unskip and update lookups tests for python3.6+ * Update vocab pickle to include lookups_extra * Update vocab serialization tests Check strings rather than lexemes since lexemes aren't initialized automatically, account for addition of "_SP". * Re-skip lookups test because of python3.5 * Skip PROB/float values in Lexeme.set_attrs * Convert is_oov from lexeme flag to lex in vectors Instead of storing `is_oov` as a lexeme flag, `is_oov` reports whether the lexeme has a vector. Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
2020-05-19 16:59:14 +03:00
from .lang.norm_exceptions import BASE_NORMS
2017-05-09 00:58:31 +03:00
from .lang.tag_map import TAG_MAP
from .tokens import Doc
from .lang.lex_attrs import LEX_ATTRS, is_stop
2020-04-28 14:37:37 +03:00
from .errors import Errors, Warnings
from .git_info import GIT_VERSION
from . import util
from . import about
ENABLE_PIPELINE_ANALYSIS = False
2016-09-24 21:26:17 +03:00
class BaseDefaults(object):
2016-10-18 17:18:25 +03:00
@classmethod
def create_lemmatizer(cls, nlp=None, lookups=None):
2019-09-08 19:08:09 +03:00
if lookups is None:
lookups = cls.create_lookups(nlp=nlp)
return Lemmatizer(lookups=lookups, is_base_form=cls.is_base_form)
@classmethod
def create_lookups(cls, nlp=None):
root = util.get_module_path(cls)
filenames = {name: root / filename for name, filename in cls.resources}
if LANG in cls.lex_attr_getters:
lang = cls.lex_attr_getters[LANG](None)
if lang in util.registry.lookups:
filenames.update(util.registry.lookups.get(lang))
lookups = Lookups()
for name, filename in filenames.items():
data = util.load_language_data(filename)
lookups.add_table(name, data)
return lookups
2016-10-18 17:18:25 +03:00
@classmethod
def create_vocab(cls, nlp=None):
lookups = cls.create_lookups(nlp)
lemmatizer = cls.create_lemmatizer(nlp, lookups=lookups)
lex_attr_getters = dict(cls.lex_attr_getters)
# This is messy, but it's the minimal working fix to Issue #639.
lex_attr_getters[IS_STOP] = functools.partial(is_stop, stops=cls.stop_words)
vocab = Vocab(
lex_attr_getters=lex_attr_getters,
tag_map=cls.tag_map,
lemmatizer=lemmatizer,
lookups=lookups,
)
Reduce stored lexemes data, move feats to lookups (#5238) * Reduce stored lexemes data, move feats to lookups * Move non-derivable lexemes features (`norm / cluster / prob`) to `spacy-lookups-data` as lookups * Get/set `norm` in both lookups and `LexemeC`, serialize in lookups * Remove `cluster` and `prob` from `LexemesC`, get/set/serialize in lookups only * Remove serialization of lexemes data as `vocab/lexemes.bin` * Remove `SerializedLexemeC` * Remove `Lexeme.to_bytes/from_bytes` * Modify normalization exception loading: * Always create `Vocab.lookups` table `lexeme_norm` for normalization exceptions * Load base exceptions from `lang.norm_exceptions`, but load language-specific exceptions from lookups * Set `lex_attr_getter[NORM]` including new lookups table in `BaseDefaults.create_vocab()` and when deserializing `Vocab` * Remove all cached lexemes when deserializing vocab to override existing normalizations with the new normalizations (as a replacement for the previous step that replaced all lexemes data with the deserialized data) * Skip English normalization test Skip English normalization test because the data is now in `spacy-lookups-data`. * Remove norm exceptions Moved to spacy-lookups-data. * Move norm exceptions test to spacy-lookups-data * Load extra lookups from spacy-lookups-data lazily Load extra lookups (currently for cluster and prob) lazily from the entry point `lg_extra` as `Vocab.lookups_extra`. * Skip creating lexeme cache on load To improve model loading times, do not create the full lexeme cache when loading. The lexemes will be created on demand when processing. * Identify numeric values in Lexeme.set_attrs() With the removal of a special case for `PROB`, also identify `float` to avoid trying to convert it with the `StringStore`. * Skip lexeme cache init in from_bytes * Unskip and update lookups tests for python3.6+ * Update vocab pickle to include lookups_extra * Update vocab serialization tests Check strings rather than lexemes since lexemes aren't initialized automatically, account for addition of "_SP". * Re-skip lookups test because of python3.5 * Skip PROB/float values in Lexeme.set_attrs * Convert is_oov from lexeme flag to lex in vectors Instead of storing `is_oov` as a lexeme flag, `is_oov` reports whether the lexeme has a vector. Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
2020-05-19 16:59:14 +03:00
vocab.lex_attr_getters[NORM] = util.add_lookups(
2020-05-21 15:14:01 +03:00
vocab.lex_attr_getters.get(NORM, LEX_ATTRS[NORM]),
BASE_NORMS,
vocab.lookups.get_table("lexeme_norm"),
Reduce stored lexemes data, move feats to lookups (#5238) * Reduce stored lexemes data, move feats to lookups * Move non-derivable lexemes features (`norm / cluster / prob`) to `spacy-lookups-data` as lookups * Get/set `norm` in both lookups and `LexemeC`, serialize in lookups * Remove `cluster` and `prob` from `LexemesC`, get/set/serialize in lookups only * Remove serialization of lexemes data as `vocab/lexemes.bin` * Remove `SerializedLexemeC` * Remove `Lexeme.to_bytes/from_bytes` * Modify normalization exception loading: * Always create `Vocab.lookups` table `lexeme_norm` for normalization exceptions * Load base exceptions from `lang.norm_exceptions`, but load language-specific exceptions from lookups * Set `lex_attr_getter[NORM]` including new lookups table in `BaseDefaults.create_vocab()` and when deserializing `Vocab` * Remove all cached lexemes when deserializing vocab to override existing normalizations with the new normalizations (as a replacement for the previous step that replaced all lexemes data with the deserialized data) * Skip English normalization test Skip English normalization test because the data is now in `spacy-lookups-data`. * Remove norm exceptions Moved to spacy-lookups-data. * Move norm exceptions test to spacy-lookups-data * Load extra lookups from spacy-lookups-data lazily Load extra lookups (currently for cluster and prob) lazily from the entry point `lg_extra` as `Vocab.lookups_extra`. * Skip creating lexeme cache on load To improve model loading times, do not create the full lexeme cache when loading. The lexemes will be created on demand when processing. * Identify numeric values in Lexeme.set_attrs() With the removal of a special case for `PROB`, also identify `float` to avoid trying to convert it with the `StringStore`. * Skip lexeme cache init in from_bytes * Unskip and update lookups tests for python3.6+ * Update vocab pickle to include lookups_extra * Update vocab serialization tests Check strings rather than lexemes since lexemes aren't initialized automatically, account for addition of "_SP". * Re-skip lookups test because of python3.5 * Skip PROB/float values in Lexeme.set_attrs * Convert is_oov from lexeme flag to lex in vectors Instead of storing `is_oov` as a lexeme flag, `is_oov` reports whether the lexeme has a vector. Co-authored-by: Matthew Honnibal <honnibal+gh@gmail.com>
2020-05-19 16:59:14 +03:00
)
2017-03-15 17:24:40 +03:00
for tag_str, exc in cls.morph_rules.items():
for orth_str, attrs in exc.items():
vocab.morphology.add_special_case(tag_str, orth_str, attrs)
return vocab
2016-12-18 18:54:52 +03:00
2016-10-18 17:18:25 +03:00
@classmethod
def create_tokenizer(cls, nlp=None):
rules = cls.tokenizer_exceptions
token_match = cls.token_match
url_match = cls.url_match
prefix_search = (
util.compile_prefix_regex(cls.prefixes).search if cls.prefixes else None
)
suffix_search = (
util.compile_suffix_regex(cls.suffixes).search if cls.suffixes else None
)
infix_finditer = (
util.compile_infix_regex(cls.infixes).finditer if cls.infixes else None
)
2016-10-18 17:18:25 +03:00
vocab = nlp.vocab if nlp is not None else cls.create_vocab(nlp)
return Tokenizer(
vocab,
rules=rules,
prefix_search=prefix_search,
suffix_search=suffix_search,
infix_finditer=infix_finditer,
token_match=token_match,
url_match=url_match,
)
pipe_names = ["tagger", "parser", "ner"]
2017-05-09 00:58:31 +03:00
token_match = TOKEN_MATCH
url_match = URL_MATCH
2017-05-09 00:58:31 +03:00
prefixes = tuple(TOKENIZER_PREFIXES)
suffixes = tuple(TOKENIZER_SUFFIXES)
infixes = tuple(TOKENIZER_INFIXES)
tag_map = dict(TAG_MAP)
tokenizer_exceptions = {}
2016-09-24 21:26:17 +03:00
stop_words = set()
2017-03-15 17:24:40 +03:00
morph_rules = {}
is_base_form = None
2017-05-09 01:58:10 +03:00
lex_attr_getters = LEX_ATTRS
syntax_iterators = {}
resources = {}
writing_system = {"direction": "ltr", "has_case": True, "has_letters": True}
single_orth_variants = []
paired_orth_variants = []
2015-09-14 10:48:51 +03:00
class Language(object):
"""A text-processing pipeline. Usually you'll load this once per process,
and pass the instance around your application.
Defaults (class): Settings, data and factory methods for creating the `nlp`
object and processing pipeline.
lang (unicode): Two-letter language ID, i.e. ISO code.
DOCS: https://spacy.io/api/language
"""
2016-09-24 21:26:17 +03:00
Defaults = BaseDefaults
lang = None
2015-08-25 16:37:17 +03:00
2019-10-28 14:43:55 +03:00
factories = {"tokenizer": lambda nlp: nlp.Defaults.create_tokenizer(nlp)}
def __init__(
self, vocab=True, make_doc=True, max_length=10 ** 6, meta={}, **kwargs
):
"""Initialise a Language object.
vocab (Vocab): A `Vocab` object. If `True`, a vocab is created via
`Language.Defaults.create_vocab`.
make_doc (callable): A function that takes text and returns a `Doc`
object. Usually a `Tokenizer`.
meta (dict): Custom meta data for the Language class. Is written to by
models to add model meta data.
max_length (int) :
Maximum number of characters in a single text. The current v2 models
may run out memory on extremely long texts, due to large internal
allocations. You should segment these texts into meaningful units,
e.g. paragraphs, subsections etc, before passing them to spaCy.
Default maximum length is 1,000,000 characters (1mb). As a rule of
thumb, if all pipeline components are enabled, spaCy's default
models currently requires roughly 1GB of temporary memory per
100,000 characters in one text.
RETURNS (Language): The newly constructed object.
"""
user_factories = util.registry.factories.get_all()
self.factories.update(user_factories)
2017-07-23 01:50:18 +03:00
self._meta = dict(meta)
2017-10-25 12:57:43 +03:00
self._path = None
if vocab is True:
factory = self.Defaults.create_vocab
vocab = factory(self, **meta.get("vocab", {}))
if vocab.vectors.name is None:
vocab.vectors.name = meta.get("vectors", {}).get("name")
else:
if (self.lang and vocab.lang) and (self.lang != vocab.lang):
raise ValueError(Errors.E150.format(nlp=self.lang, vocab=vocab.lang))
self.vocab = vocab
if make_doc is True:
factory = self.Defaults.create_tokenizer
make_doc = factory(self, **meta.get("tokenizer", {}))
2017-05-29 16:40:45 +03:00
self.tokenizer = make_doc
self.pipeline = []
self.max_length = max_length
self._optimizer = None
2017-10-25 12:57:43 +03:00
@property
def path(self):
return self._path
2017-07-23 01:50:18 +03:00
@property
def meta(self):
if self.vocab.lang:
self._meta.setdefault("lang", self.vocab.lang)
else:
self._meta.setdefault("lang", self.lang)
self._meta.setdefault("name", "model")
self._meta.setdefault("version", "0.0.0")
self._meta.setdefault("spacy_version", ">={}".format(about.__version__))
self._meta.setdefault("description", "")
self._meta.setdefault("author", "")
self._meta.setdefault("email", "")
self._meta.setdefault("url", "")
self._meta.setdefault("license", "")
self._meta.setdefault("spacy_git_version", GIT_VERSION)
self._meta["vectors"] = {
"width": self.vocab.vectors_length,
"vectors": len(self.vocab.vectors),
"keys": self.vocab.vectors.n_keys,
"name": self.vocab.vectors.name,
}
self._meta["pipeline"] = self.pipe_names
self._meta["factories"] = self.pipe_factories
2019-09-12 12:34:25 +03:00
self._meta["labels"] = self.pipe_labels
2017-07-23 01:50:18 +03:00
return self._meta
@meta.setter
def meta(self, value):
self._meta = value
# Conveniences to access pipeline components
2019-03-15 18:23:17 +03:00
# Shouldn't be used anymore!
@property
def tensorizer(self):
return self.get_pipe("tensorizer")
@property
def tagger(self):
return self.get_pipe("tagger")
@property
def parser(self):
return self.get_pipe("parser")
@property
def entity(self):
return self.get_pipe("ner")
2019-03-21 19:33:25 +03:00
@property
def linker(self):
2019-03-22 15:55:10 +03:00
return self.get_pipe("entity_linker")
2019-03-21 19:33:25 +03:00
@property
def matcher(self):
return self.get_pipe("matcher")
@property
def pipe_names(self):
"""Get names of available pipeline components.
RETURNS (list): List of component name strings, in order.
"""
return [pipe_name for pipe_name, _ in self.pipeline]
@property
def pipe_factories(self):
"""Get the component factories for the available pipeline components.
RETURNS (dict): Factory names, keyed by component names.
"""
factories = {}
for pipe_name, pipe in self.pipeline:
factories[pipe_name] = getattr(pipe, "factory", pipe_name)
return factories
@property
def pipe_labels(self):
2019-09-12 14:03:38 +03:00
"""Get the labels set by the pipeline components, if available (if
the component exposes a labels property).
RETURNS (dict): Labels keyed by component name.
"""
labels = OrderedDict()
for name, pipe in self.pipeline:
if hasattr(pipe, "labels"):
labels[name] = list(pipe.labels)
return labels
def get_pipe(self, name):
"""Get a pipeline component for a given component name.
name (unicode): Name of pipeline component to get.
RETURNS (callable): The pipeline component.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#get_pipe
"""
for pipe_name, component in self.pipeline:
if pipe_name == name:
return component
raise KeyError(Errors.E001.format(name=name, opts=self.pipe_names))
def create_pipe(self, name, config=dict()):
"""Create a pipeline component from a factory.
name (unicode): Factory name to look up in `Language.factories`.
2017-10-07 02:04:50 +03:00
config (dict): Configuration parameters to initialise component.
RETURNS (callable): Pipeline component.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#create_pipe
"""
if name not in self.factories:
if name == "sbd":
raise KeyError(Errors.E108.format(name=name))
else:
raise KeyError(Errors.E002.format(name=name))
factory = self.factories[name]
return factory(self, **config)
def add_pipe(
self, component, name=None, before=None, after=None, first=None, last=None
):
"""Add a component to the processing pipeline. Valid components are
callables that take a `Doc` object, modify it and return it. Only one
of before/after/first/last can be set. Default behaviour is "last".
component (callable): The pipeline component.
name (unicode): Name of pipeline component. Overwrites existing
component.name attribute if available. If no name is set and
the component exposes no name attribute, component.__name__ is
used. An error is raised if a name already exists in the pipeline.
before (unicode): Component name to insert component directly before.
after (unicode): Component name to insert component directly after.
first (bool): Insert component first / not first in the pipeline.
last (bool): Insert component last / not last in the pipeline.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#add_pipe
"""
if not hasattr(component, "__call__"):
msg = Errors.E003.format(component=repr(component), name=name)
if isinstance(component, basestring_) and component in self.factories:
msg += Errors.E004.format(component=component)
raise ValueError(msg)
if name is None:
name = util.get_component_name(component)
if name in self.pipe_names:
raise ValueError(Errors.E007.format(name=name, opts=self.pipe_names))
if sum([bool(before), bool(after), bool(first), bool(last)]) >= 2:
raise ValueError(Errors.E006)
pipe_index = 0
pipe = (name, component)
if last or not any([first, before, after]):
pipe_index = len(self.pipeline)
self.pipeline.append(pipe)
elif first:
self.pipeline.insert(0, pipe)
elif before and before in self.pipe_names:
pipe_index = self.pipe_names.index(before)
self.pipeline.insert(self.pipe_names.index(before), pipe)
elif after and after in self.pipe_names:
pipe_index = self.pipe_names.index(after) + 1
self.pipeline.insert(self.pipe_names.index(after) + 1, pipe)
else:
raise ValueError(
Errors.E001.format(name=before or after, opts=self.pipe_names)
)
if ENABLE_PIPELINE_ANALYSIS:
analyze_pipes(self.pipeline, name, component, pipe_index)
2017-10-17 12:20:07 +03:00
def has_pipe(self, name):
"""Check if a component name is present in the pipeline. Equivalent to
`name in nlp.pipe_names`.
name (unicode): Name of the component.
RETURNS (bool): Whether a component of the name exists in the pipeline.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#has_pipe
2017-10-17 12:20:07 +03:00
"""
return name in self.pipe_names
def replace_pipe(self, name, component):
"""Replace a component in the pipeline.
name (unicode): Name of the component to replace.
component (callable): Pipeline component.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#replace_pipe
"""
if name not in self.pipe_names:
raise ValueError(Errors.E001.format(name=name, opts=self.pipe_names))
if not hasattr(component, "__call__"):
msg = Errors.E003.format(component=repr(component), name=name)
if isinstance(component, basestring_) and component in self.factories:
msg += Errors.E135.format(name=name)
raise ValueError(msg)
self.pipeline[self.pipe_names.index(name)] = (name, component)
if ENABLE_PIPELINE_ANALYSIS:
analyze_all_pipes(self.pipeline)
def rename_pipe(self, old_name, new_name):
"""Rename a pipeline component.
old_name (unicode): Name of the component to rename.
new_name (unicode): New name of the component.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#rename_pipe
"""
if old_name not in self.pipe_names:
raise ValueError(Errors.E001.format(name=old_name, opts=self.pipe_names))
if new_name in self.pipe_names:
raise ValueError(Errors.E007.format(name=new_name, opts=self.pipe_names))
i = self.pipe_names.index(old_name)
self.pipeline[i] = (new_name, self.pipeline[i][1])
def remove_pipe(self, name):
"""Remove a component from the pipeline.
name (unicode): Name of the component to remove.
2017-10-07 02:04:50 +03:00
RETURNS (tuple): A `(name, component)` tuple of the removed component.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#remove_pipe
"""
if name not in self.pipe_names:
raise ValueError(Errors.E001.format(name=name, opts=self.pipe_names))
removed = self.pipeline.pop(self.pipe_names.index(name))
if ENABLE_PIPELINE_ANALYSIS:
analyze_all_pipes(self.pipeline)
return removed
def __call__(self, text, disable=[], component_cfg=None):
"""Apply the pipeline to some text. The text can span multiple sentences,
and can contain arbitrary whitespace. Alignment into the original string
2015-08-25 16:37:17 +03:00
is preserved.
2016-12-18 18:54:52 +03:00
text (unicode): The text to be processed.
disable (list): Names of the pipeline components to disable.
component_cfg (dict): An optional dictionary with extra keyword arguments
for specific components.
RETURNS (Doc): A container for accessing the annotations.
2016-11-01 14:25:36 +03:00
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#call
2015-08-25 16:37:17 +03:00
"""
doc = self.make_doc(text)
if component_cfg is None:
component_cfg = {}
for name, proc in self.pipeline:
if name in disable:
continue
if not hasattr(proc, "__call__"):
raise ValueError(Errors.E003.format(component=type(proc), name=name))
doc = proc(doc, **component_cfg.get(name, {}))
if doc is None:
raise ValueError(Errors.E005.format(name=name))
return doc
2015-08-25 16:37:17 +03:00
2017-10-25 14:46:41 +03:00
def disable_pipes(self, *names):
"""Disable one or more pipeline components. If used as a context
manager, the pipeline will be restored to the initial state at the end
of the block. Otherwise, a DisabledPipes object is returned, that has
a `.restore()` method you can use to undo your changes.
2017-10-25 14:46:41 +03:00
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#disable_pipes
"""
if len(names) == 1 and isinstance(names[0], (list, tuple)):
names = names[0] # support list of names instead of spread
2017-10-25 14:46:41 +03:00
return DisabledPipes(self, *names)
2017-05-29 16:40:45 +03:00
def make_doc(self, text):
if len(text) > self.max_length:
raise ValueError(
Errors.E088.format(length=len(text), max_length=self.max_length)
)
2017-05-29 16:40:45 +03:00
return self.tokenizer(text)
def _format_docs_and_golds(self, docs, golds):
"""Format golds and docs before update models."""
expected_keys = ("words", "tags", "heads", "deps", "entities", "cats", "links")
gold_objs = []
doc_objs = []
for doc, gold in zip(docs, golds):
if isinstance(doc, basestring_):
doc = self.make_doc(doc)
if not isinstance(gold, GoldParse):
unexpected = [k for k in gold if k not in expected_keys]
if unexpected:
err = Errors.E151.format(unexp=unexpected, exp=expected_keys)
raise ValueError(err)
gold = GoldParse(doc, **gold)
doc_objs.append(doc)
gold_objs.append(gold)
return doc_objs, gold_objs
def update(self, docs, golds, drop=0.0, sgd=None, losses=None, component_cfg=None):
"""Update the models in the pipeline.
docs (iterable): A batch of `Doc` objects.
golds (iterable): A batch of `GoldParse` objects.
drop (float): The dropout rate.
sgd (callable): An optimizer.
2019-05-24 15:06:26 +03:00
losses (dict): Dictionary to update with the loss, keyed by component.
component_cfg (dict): Config parameters for specific pipeline
components, keyed by component name.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#update
"""
2017-08-01 23:10:17 +03:00
if len(docs) != len(golds):
raise IndexError(Errors.E009.format(n_docs=len(docs), n_golds=len(golds)))
2017-08-01 23:10:17 +03:00
if len(docs) == 0:
return
if sgd is None:
if self._optimizer is None:
self._optimizer = create_default_optimizer(Model.ops)
sgd = self._optimizer
# Allow dict of args to GoldParse, instead of GoldParse objects.
docs, golds = self._format_docs_and_golds(docs, golds)
2017-05-25 04:10:54 +03:00
grads = {}
2017-05-25 04:10:54 +03:00
def get_grads(W, dW, key=None):
grads[key] = (W, dW)
get_grads.alpha = sgd.alpha
get_grads.b1 = sgd.b1
get_grads.b2 = sgd.b2
pipes = list(self.pipeline)
random.shuffle(pipes)
if component_cfg is None:
component_cfg = {}
for name, proc in pipes:
if not hasattr(proc, "update"):
2017-05-22 02:43:31 +03:00
continue
grads = {}
kwargs = component_cfg.get(name, {})
kwargs.setdefault("drop", drop)
proc.update(docs, golds, sgd=get_grads, losses=losses, **kwargs)
for key, (W, dW) in grads.items():
sgd(W, dW, key=key)
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
def rehearse(self, docs, sgd=None, losses=None, config=None):
"""Make a "rehearsal" update to the models in the pipeline, to prevent
forgetting. Rehearsal updates run an initial copy of the model over some
data, and update the model so its current predictions are more like the
2019-10-02 11:37:39 +03:00
initial ones. This is useful for keeping a pretrained model on-track,
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
even if you're updating it with a smaller set of examples.
docs (iterable): A batch of `Doc` objects.
drop (float): The dropout rate.
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
sgd (callable): An optimizer.
RETURNS (dict): Results from the update.
EXAMPLE:
>>> raw_text_batches = minibatch(raw_texts)
>>> for labelled_batch in minibatch(zip(train_docs, train_golds)):
2018-12-18 15:48:10 +03:00
>>> docs, golds = zip(*train_docs)
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
>>> nlp.update(docs, golds)
>>> raw_batch = [nlp.make_doc(text) for text in next(raw_text_batches)]
>>> nlp.rehearse(raw_batch)
"""
2019-03-15 18:23:17 +03:00
# TODO: document
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
if len(docs) == 0:
return
if sgd is None:
if self._optimizer is None:
self._optimizer = create_default_optimizer(Model.ops)
sgd = self._optimizer
docs = list(docs)
for i, doc in enumerate(docs):
if isinstance(doc, basestring_):
docs[i] = self.make_doc(doc)
pipes = list(self.pipeline)
random.shuffle(pipes)
if config is None:
config = {}
grads = {}
def get_grads(W, dW, key=None):
grads[key] = (W, dW)
get_grads.alpha = sgd.alpha
get_grads.b1 = sgd.b1
get_grads.b2 = sgd.b2
for name, proc in pipes:
if not hasattr(proc, "rehearse"):
continue
grads = {}
proc.rehearse(docs, sgd=get_grads, losses=losses, **config.get(name, {}))
for key, (W, dW) in grads.items():
sgd(W, dW, key=key)
return losses
2017-05-21 17:07:06 +03:00
def preprocess_gold(self, docs_golds):
"""Can be called before training to pre-process gold data. By default,
it handles nonprojectivity and adds missing tags to the tag map.
docs_golds (iterable): Tuples of `Doc` and `GoldParse` objects.
YIELDS (tuple): Tuples of preprocessed `Doc` and `GoldParse` objects.
"""
for name, proc in self.pipeline:
if hasattr(proc, "preprocess_gold"):
2017-05-21 17:07:06 +03:00
docs_golds = proc.preprocess_gold(docs_golds)
for doc, gold in docs_golds:
yield doc, gold
def begin_training(self, get_gold_tuples=None, sgd=None, component_cfg=None, **cfg):
"""Allocate models, pre-process training data and acquire a trainer and
optimizer. Used as a contextmanager.
get_gold_tuples (function): Function returning gold data
component_cfg (dict): Config parameters for specific components.
**cfg: Config parameters.
2019-03-15 18:23:17 +03:00
RETURNS: An optimizer.
DOCS: https://spacy.io/api/language#begin_training
"""
if get_gold_tuples is None:
get_gold_tuples = lambda: []
# Populate vocab
else:
2017-09-21 03:15:20 +03:00
for _, annots_brackets in get_gold_tuples():
_ = annots_brackets.pop()
2017-09-21 03:15:20 +03:00
for annots, _ in annots_brackets:
for word in annots[1]:
2018-11-30 19:43:08 +03:00
_ = self.vocab[word] # noqa: F841
if cfg.get("device", -1) >= 0:
2018-11-30 19:43:08 +03:00
util.use_gpu(cfg["device"])
2017-09-19 02:04:16 +03:00
if self.vocab.vectors.data.shape[1] >= 1:
self.vocab.vectors.data = Model.ops.asarray(self.vocab.vectors.data)
2017-09-23 04:11:52 +03:00
link_vectors_to_models(self.vocab)
if self.vocab.vectors.data.shape[1]:
cfg["pretrained_vectors"] = self.vocab.vectors.name
2020-03-25 14:28:12 +03:00
cfg["pretrained_dims"] = self.vocab.vectors.data.shape[1]
if sgd is None:
sgd = create_default_optimizer(Model.ops)
self._optimizer = sgd
if component_cfg is None:
component_cfg = {}
for name, proc in self.pipeline:
if hasattr(proc, "begin_training"):
kwargs = component_cfg.get(name, {})
kwargs.update(cfg)
proc.begin_training(
get_gold_tuples,
pipeline=self.pipeline,
sgd=self._optimizer,
**kwargs
)
return self._optimizer
2017-05-21 17:07:06 +03:00
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
def resume_training(self, sgd=None, **cfg):
2019-10-02 11:37:39 +03:00
"""Continue training a pretrained model.
2018-12-18 15:48:10 +03:00
💫 Better support for semi-supervised learning (#3035) The new spacy pretrain command implemented BERT/ULMFit/etc-like transfer learning, using our Language Modelling with Approximate Outputs version of BERT's cloze task. Pretraining is convenient, but in some ways it's a bit of a strange solution. All we're doing is initialising the weights. At the same time, we're putting a lot of work into our optimisation so that it's less sensitive to initial conditions, and more likely to find good optima. I discuss this a bit in the pseudo-rehearsal blog post: https://explosion.ai/blog/pseudo-rehearsal-catastrophic-forgetting Support semi-supervised learning in spacy train One obvious way to improve these pretraining methods is to do multi-task learning, instead of just transfer learning. This has been shown to work very well: https://arxiv.org/pdf/1809.08370.pdf . This patch makes it easy to do this sort of thing. Add a new argument to spacy train, --raw-text. This takes a jsonl file with unlabelled data that can be used in arbitrary ways to do semi-supervised learning. Add a new method to the Language class and to pipeline components, .rehearse(). This is like .update(), but doesn't expect GoldParse objects. It takes a batch of Doc objects, and performs an update on some semi-supervised objective. Move the BERT-LMAO objective out from spacy/cli/pretrain.py into spacy/_ml.py, so we can create a new pipeline component, ClozeMultitask. This can be specified as a parser or NER multitask in the spacy train command. Example usage: python -m spacy train en ./tmp ~/data/en-core-web/train/nw.json ~/data/en-core-web/dev/nw.json --pipeline parser --raw-textt ~/data/unlabelled/reddit-100k.jsonl --vectors en_vectors_web_lg --parser-multitasks cloze Implement rehearsal methods for pipeline components The new --raw-text argument and nlp.rehearse() method also gives us a good place to implement the the idea in the pseudo-rehearsal blog post in the parser. This works as follows: Add a new nlp.resume_training() method. This allocates copies of pre-trained models in the pipeline, setting things up for the rehearsal updates. It also returns an optimizer object. This also greatly reduces confusion around the nlp.begin_training() method, which randomises the weights, making it not suitable for adding new labels or otherwise fine-tuning a pre-trained model. Implement rehearsal updates on the Parser class, making it available for the dependency parser and NER. During rehearsal, the initial model is used to supervise the model being trained. The current model is asked to match the predictions of the initial model on some data. This minimises catastrophic forgetting, by keeping the model's predictions close to the original. See the blog post for details. Implement rehearsal updates for tagger Implement rehearsal updates for text categoriz
2018-12-10 18:25:33 +03:00
Create and return an optimizer, and initialize "rehearsal" for any pipeline
component that has a .rehearse() method. Rehearsal is used to prevent
models from "forgetting" their initialised "knowledge". To perform
rehearsal, collect samples of text you want the models to retain performance
on, and call nlp.rehearse() with a batch of Doc objects.
"""
if cfg.get("device", -1) >= 0:
util.use_gpu(cfg["device"])
if self.vocab.vectors.data.shape[1] >= 1:
self.vocab.vectors.data = Model.ops.asarray(self.vocab.vectors.data)
link_vectors_to_models(self.vocab)
if self.vocab.vectors.data.shape[1]:
cfg["pretrained_vectors"] = self.vocab.vectors.name
if sgd is None:
sgd = create_default_optimizer(Model.ops)
self._optimizer = sgd
for name, proc in self.pipeline:
if hasattr(proc, "_rehearsal_model"):
proc._rehearsal_model = deepcopy(proc.model)
return self._optimizer
def evaluate(
self, docs_golds, verbose=False, batch_size=256, scorer=None, component_cfg=None
):
2019-05-24 15:06:36 +03:00
"""Evaluate a model's pipeline components.
docs_golds (iterable): Tuples of `Doc` and `GoldParse` objects.
verbose (bool): Print debugging information.
batch_size (int): Batch size to use.
scorer (Scorer): Optional `Scorer` to use. If not passed in, a new one
will be created.
component_cfg (dict): An optional dictionary with extra keyword
arguments for specific components.
RETURNS (Scorer): The scorer containing the evaluation results.
DOCS: https://spacy.io/api/language#evaluate
"""
if scorer is None:
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
scorer = Scorer(pipeline=self.pipeline)
2019-03-15 17:20:09 +03:00
if component_cfg is None:
component_cfg = {}
2017-08-18 23:26:12 +03:00
docs, golds = zip(*docs_golds)
docs = [
self.make_doc(doc) if isinstance(doc, basestring_) else doc for doc in docs
]
2017-08-18 23:26:12 +03:00
golds = list(golds)
for name, pipe in self.pipeline:
kwargs = component_cfg.get(name, {})
kwargs.setdefault("batch_size", batch_size)
if not hasattr(pipe, "pipe"):
docs = _pipe(docs, pipe, kwargs)
2017-08-18 23:26:12 +03:00
else:
docs = pipe.pipe(docs, **kwargs)
2017-08-18 23:26:12 +03:00
for doc, gold in zip(docs, golds):
if not isinstance(gold, GoldParse):
gold = GoldParse(doc, **gold)
if verbose:
print(doc)
kwargs = component_cfg.get("scorer", {})
kwargs.setdefault("verbose", verbose)
scorer.score(doc, gold, **kwargs)
2017-05-21 17:07:06 +03:00
return scorer
2017-05-18 12:25:19 +03:00
@contextmanager
def use_params(self, params, **cfg):
"""Replace weights of models in the pipeline with those provided in the
params dictionary. Can be used as a contextmanager, in which case,
models go back to their original weights after the block.
params (dict): A dictionary of parameters keyed by model ID.
**cfg: Config parameters.
EXAMPLE:
>>> with nlp.use_params(optimizer.averages):
>>> nlp.to_disk('/tmp/checkpoint')
"""
contexts = [
pipe.use_params(params)
for name, pipe in self.pipeline
if hasattr(pipe, "use_params")
]
2017-05-18 16:30:59 +03:00
# TODO: Having trouble with contextlib
# Workaround: these aren't actually context managers atm.
for context in contexts:
try:
next(context)
except StopIteration:
pass
2017-05-18 12:25:19 +03:00
yield
for context in contexts:
try:
2017-05-18 16:30:59 +03:00
next(context)
2017-05-18 12:25:19 +03:00
except StopIteration:
pass
def pipe(
self,
texts,
as_tuples=False,
n_threads=-1,
batch_size=1000,
disable=[],
cleanup=False,
component_cfg=None,
n_process=1,
):
"""Process texts as a stream, and yield `Doc` objects in order.
texts (iterable): A sequence of texts to process.
2019-03-15 18:23:17 +03:00
as_tuples (bool): If set to True, inputs should be a sequence of
(text, context) tuples. Output will then be a sequence of
(doc, context) tuples. Defaults to False.
batch_size (int): The number of texts to buffer.
disable (list): Names of the pipeline components to disable.
2019-03-15 18:23:17 +03:00
cleanup (bool): If True, unneeded strings are freed to control memory
use. Experimental.
component_cfg (dict): An optional dictionary with extra keyword
arguments for specific components.
2019-10-18 12:33:38 +03:00
n_process (int): Number of processors to process texts, only supported
in Python3. If -1, set `multiprocessing.cpu_count()`.
YIELDS (Doc): Documents in the order of the original text.
2019-03-15 18:23:17 +03:00
DOCS: https://spacy.io/api/language#pipe
"""
if is_python2 and n_process != 1:
2020-04-28 14:37:37 +03:00
warnings.warn(Warnings.W023)
n_process = 1
if n_threads != -1:
2020-04-28 14:37:37 +03:00
warnings.warn(Warnings.W016, DeprecationWarning)
if n_process == -1:
n_process = mp.cpu_count()
if as_tuples:
2017-07-25 19:57:59 +03:00
text_context1, text_context2 = itertools.tee(texts)
texts = (tc[0] for tc in text_context1)
contexts = (tc[1] for tc in text_context2)
docs = self.pipe(
texts,
batch_size=batch_size,
disable=disable,
n_process=n_process,
component_cfg=component_cfg,
)
2017-07-25 19:57:59 +03:00
for doc, context in izip(docs, contexts):
yield (doc, context)
return
if component_cfg is None:
component_cfg = {}
pipes = (
[]
) # contains functools.partial objects to easily create multiprocess worker.
for name, proc in self.pipeline:
if name in disable:
continue
kwargs = component_cfg.get(name, {})
# Allow component_cfg to overwrite the top-level kwargs.
kwargs.setdefault("batch_size", batch_size)
if hasattr(proc, "pipe"):
f = functools.partial(proc.pipe, **kwargs)
else:
2017-05-22 02:43:31 +03:00
# Apply the function, but yield the doc
f = functools.partial(_pipe, proc=proc, kwargs=kwargs)
pipes.append(f)
if n_process != 1:
docs = self._multiprocessing_pipe(texts, pipes, n_process, batch_size)
else:
# if n_process == 1, no processes are forked.
docs = (self.make_doc(text) for text in texts)
for pipe in pipes:
docs = pipe(docs)
# Track weakrefs of "recent" documents, so that we can see when they
# expire from memory. When they do, we know we don't need old strings.
# This way, we avoid maintaining an unbounded growth in string entries
# in the string store.
recent_refs = weakref.WeakSet()
old_refs = weakref.WeakSet()
# Keep track of the original string data, so that if we flush old strings,
# we can recover the original ones. However, we only want to do this if we're
# really adding strings, to save up-front costs.
original_strings_data = None
nr_seen = 0
for doc in docs:
yield doc
if cleanup:
recent_refs.add(doc)
if nr_seen < 10000:
old_refs.add(doc)
nr_seen += 1
elif len(old_refs) == 0:
old_refs, recent_refs = recent_refs, old_refs
if original_strings_data is None:
original_strings_data = list(self.vocab.strings)
else:
keys, strings = self.vocab.strings._cleanup_stale_strings(
original_strings_data
)
self.vocab._reset_cache(keys, strings)
self.tokenizer._reset_cache(keys)
nr_seen = 0
def _multiprocessing_pipe(self, texts, pipes, n_process, batch_size):
# raw_texts is used later to stop iteration.
texts, raw_texts = itertools.tee(texts)
# for sending texts to worker
texts_q = [mp.Queue() for _ in range(n_process)]
# for receiving byte-encoded docs from worker
bytedocs_recv_ch, bytedocs_send_ch = zip(
*[mp.Pipe(False) for _ in range(n_process)]
)
batch_texts = util.minibatch(texts, batch_size)
# Sender sends texts to the workers.
# This is necessary to properly handle infinite length of texts.
# (In this case, all data cannot be sent to the workers at once)
sender = _Sender(batch_texts, texts_q, chunk_size=n_process)
# send twice to make process busy
sender.send()
sender.send()
procs = [
mp.Process(
target=_apply_pipes,
2020-03-25 14:28:12 +03:00
args=(
self.make_doc,
pipes,
rch,
sch,
Underscore.get_state(),
load_nlp.VECTORS,
),
)
for rch, sch in zip(texts_q, bytedocs_send_ch)
]
for proc in procs:
proc.start()
# Cycle channels not to break the order of docs.
# The received object is a batch of byte-encoded docs, so flatten them with chain.from_iterable.
byte_docs = chain.from_iterable(recv.recv() for recv in cycle(bytedocs_recv_ch))
docs = (Doc(self.vocab).from_bytes(byte_doc) for byte_doc in byte_docs)
try:
for i, (_, doc) in enumerate(zip(raw_texts, docs), 1):
yield doc
if i % batch_size == 0:
# tell `sender` that one batch was consumed.
sender.step()
finally:
for proc in procs:
proc.terminate()
def to_disk(self, path, exclude=tuple(), disable=None):
"""Save the current state to a directory. If a model is loaded, this
will include the model.
2017-04-17 02:40:26 +03:00
path (unicode or Path): Path to a directory, which will be created if
it doesn't exist.
exclude (list): Names of components or serialization fields to exclude.
DOCS: https://spacy.io/api/language#to_disk
"""
if disable is not None:
2020-04-28 14:37:37 +03:00
warnings.warn(Warnings.W014, DeprecationWarning)
exclude = disable
path = util.ensure_path(path)
serializers = OrderedDict()
serializers["tokenizer"] = lambda p: self.tokenizer.to_disk(
p, exclude=["vocab"]
)
serializers["meta.json"] = lambda p: srsly.write_json(p, self.meta)
for name, proc in self.pipeline:
if not hasattr(proc, "name"):
2017-05-31 14:42:39 +03:00
continue
if name in exclude:
2017-05-31 14:42:39 +03:00
continue
if not hasattr(proc, "to_disk"):
2017-05-31 14:42:39 +03:00
continue
serializers[name] = lambda p, proc=proc: proc.to_disk(p, exclude=["vocab"])
serializers["vocab"] = lambda p: self.vocab.to_disk(p)
util.to_disk(path, serializers, exclude)
2017-05-31 14:42:39 +03:00
def from_disk(self, path, exclude=tuple(), disable=None):
"""Loads state from a directory. Modifies the object in place and
returns it. If the saved `Language` object contains a model, the
model will be loaded.
path (unicode or Path): A path to a directory.
exclude (list): Names of components or serialization fields to exclude.
RETURNS (Language): The modified `Language` object.
DOCS: https://spacy.io/api/language#from_disk
"""
def deserialize_meta(path):
if path.exists():
data = srsly.read_json(path)
self.meta.update(data)
# self.meta always overrides meta["vectors"] with the metadata
# from self.vocab.vectors, so set the name directly
self.vocab.vectors.name = data.get("vectors", {}).get("name")
def deserialize_vocab(path):
if path.exists():
self.vocab.from_disk(path)
_fix_pretrained_vectors_name(self)
if disable is not None:
2020-04-28 14:37:37 +03:00
warnings.warn(Warnings.W014, DeprecationWarning)
exclude = disable
path = util.ensure_path(path)
deserializers = OrderedDict()
deserializers["meta.json"] = deserialize_meta
deserializers["vocab"] = deserialize_vocab
deserializers["tokenizer"] = lambda p: self.tokenizer.from_disk(
p, exclude=["vocab"]
)
for name, proc in self.pipeline:
if name in exclude:
2017-05-31 14:42:39 +03:00
continue
if not hasattr(proc, "from_disk"):
2017-05-31 14:42:39 +03:00
continue
deserializers[name] = lambda p, proc=proc: proc.from_disk(
p, exclude=["vocab"]
)
if not (path / "vocab").exists() and "vocab" not in exclude:
# Convert to list here in case exclude is (default) tuple
exclude = list(exclude) + ["vocab"]
2017-06-01 15:38:35 +03:00
util.from_disk(path, deserializers, exclude)
2017-10-25 12:57:43 +03:00
self._path = path
2017-05-31 14:42:39 +03:00
return self
def to_bytes(self, exclude=tuple(), disable=None, **kwargs):
"""Serialize the current state to a binary string.
2016-12-18 18:54:52 +03:00
exclude (list): Names of components or serialization fields to exclude.
RETURNS (bytes): The serialized form of the `Language` object.
DOCS: https://spacy.io/api/language#to_bytes
"""
if disable is not None:
2020-04-28 14:37:37 +03:00
warnings.warn(Warnings.W014, DeprecationWarning)
exclude = disable
serializers = OrderedDict()
serializers["vocab"] = lambda: self.vocab.to_bytes()
serializers["tokenizer"] = lambda: self.tokenizer.to_bytes(exclude=["vocab"])
2020-05-21 15:14:01 +03:00
serializers["meta.json"] = lambda: srsly.json_dumps(
OrderedDict(sorted(self.meta.items()))
)
for name, proc in self.pipeline:
if name in exclude:
continue
if not hasattr(proc, "to_bytes"):
continue
serializers[name] = lambda proc=proc: proc.to_bytes(exclude=["vocab"])
exclude = util.get_serialization_exclude(serializers, exclude, kwargs)
return util.to_bytes(serializers, exclude)
def from_bytes(self, bytes_data, exclude=tuple(), disable=None, **kwargs):
"""Load state from a binary string.
bytes_data (bytes): The data to load from.
exclude (list): Names of components or serialization fields to exclude.
RETURNS (Language): The `Language` object.
DOCS: https://spacy.io/api/language#from_bytes
"""
def deserialize_meta(b):
data = srsly.json_loads(b)
self.meta.update(data)
# self.meta always overrides meta["vectors"] with the metadata
# from self.vocab.vectors, so set the name directly
self.vocab.vectors.name = data.get("vectors", {}).get("name")
def deserialize_vocab(b):
self.vocab.from_bytes(b)
_fix_pretrained_vectors_name(self)
if disable is not None:
2020-04-28 14:37:37 +03:00
warnings.warn(Warnings.W014, DeprecationWarning)
exclude = disable
deserializers = OrderedDict()
deserializers["meta.json"] = deserialize_meta
deserializers["vocab"] = deserialize_vocab
deserializers["tokenizer"] = lambda b: self.tokenizer.from_bytes(
b, exclude=["vocab"]
)
for name, proc in self.pipeline:
if name in exclude:
continue
if not hasattr(proc, "from_bytes"):
continue
deserializers[name] = lambda b, proc=proc: proc.from_bytes(
b, exclude=["vocab"]
)
exclude = util.get_serialization_exclude(deserializers, exclude, kwargs)
util.from_bytes(bytes_data, deserializers, exclude)
return self
2017-05-22 02:43:31 +03:00
class component(object):
"""Decorator for pipeline components. Can decorate both function components
and class components and will automatically register components in the
Language.factories. If the component is a class and needs access to the
nlp object or config parameters, it can expose a from_nlp classmethod
that takes the nlp object and **cfg arguments and returns the initialized
component.
"""
# NB: This decorator needs to live here, because it needs to write to
# Language.factories. All other solutions would cause circular import.
def __init__(self, name=None, assigns=tuple(), requires=tuple(), retokenizes=False):
"""Decorate a pipeline component.
name (unicode): Default component and factory name.
assigns (list): Attributes assigned by component, e.g. `["token.pos"]`.
requires (list): Attributes required by component, e.g. `["token.dep"]`.
retokenizes (bool): Whether the component changes the tokenization.
"""
self.name = name
self.assigns = validate_attrs(assigns)
self.requires = validate_attrs(requires)
self.retokenizes = retokenizes
def __call__(self, *args, **kwargs):
obj = args[0]
args = args[1:]
factory_name = self.name or util.get_component_name(obj)
obj.name = factory_name
obj.factory = factory_name
obj.assigns = self.assigns
obj.requires = self.requires
obj.retokenizes = self.retokenizes
def factory(nlp, **cfg):
if hasattr(obj, "from_nlp"):
return obj.from_nlp(nlp, **cfg)
elif isinstance(obj, class_types):
return obj()
return obj
Language.factories[obj.factory] = factory
return obj
def _fix_pretrained_vectors_name(nlp):
# TODO: Replace this once we handle vectors consistently as static
# data
if "vectors" in nlp.meta and "name" in nlp.meta["vectors"]:
nlp.vocab.vectors.name = nlp.meta["vectors"]["name"]
elif not nlp.vocab.vectors.size:
nlp.vocab.vectors.name = None
elif "name" in nlp.meta and "lang" in nlp.meta:
vectors_name = "%s_%s.vectors" % (nlp.meta["lang"], nlp.meta["name"])
nlp.vocab.vectors.name = vectors_name
else:
2018-04-03 22:40:29 +03:00
raise ValueError(Errors.E092)
2018-04-04 02:31:25 +03:00
if nlp.vocab.vectors.size != 0:
link_vectors_to_models(nlp.vocab, skip_rank=True)
for name, proc in nlp.pipeline:
if not hasattr(proc, "cfg"):
continue
proc.cfg.setdefault("deprecation_fixes", {})
proc.cfg["deprecation_fixes"]["vectors_name"] = nlp.vocab.vectors.name
2017-10-25 14:46:41 +03:00
class DisabledPipes(list):
"""Manager for temporary pipeline disabling."""
2017-10-25 14:46:41 +03:00
def __init__(self, nlp, *names):
self.nlp = nlp
self.names = names
# Important! Not deep copy -- we just want the container (but we also
# want to support people providing arbitrarily typed nlp.pipeline
# objects.)
2017-10-27 22:07:59 +03:00
self.original_pipeline = copy(nlp.pipeline)
2017-10-25 14:46:41 +03:00
list.__init__(self)
self.extend(nlp.remove_pipe(name) for name in names)
def __enter__(self):
2017-10-25 15:56:16 +03:00
return self
2017-10-25 14:46:41 +03:00
def __exit__(self, *args):
self.restore()
def restore(self):
"""Restore the pipeline to its state when DisabledPipes was created."""
2017-10-25 14:46:41 +03:00
current, self.nlp.pipeline = self.nlp.pipeline, self.original_pipeline
unexpected = [name for name, pipe in current if not self.nlp.has_pipe(name)]
2017-10-25 14:46:41 +03:00
if unexpected:
# Don't change the pipeline if we're raising an error.
self.nlp.pipeline = current
raise ValueError(Errors.E008.format(names=unexpected))
2017-10-25 14:46:41 +03:00
self[:] = []
def _pipe(docs, proc, kwargs):
2019-03-12 15:32:56 +03:00
# We added some args for pipe that __call__ doesn't expect.
kwargs = dict(kwargs)
for arg in ["n_threads", "batch_size"]:
if arg in kwargs:
kwargs.pop(arg)
2017-05-22 02:43:31 +03:00
for doc in docs:
doc = proc(doc, **kwargs)
2017-05-22 02:43:31 +03:00
yield doc
def _apply_pipes(make_doc, pipes, receiver, sender, underscore_state, vectors):
"""Worker for Language.pipe
2019-10-18 12:33:38 +03:00
receiver (multiprocessing.Connection): Pipe to receive text. Usually
created by `multiprocessing.Pipe()`
sender (multiprocessing.Connection): Pipe to send doc. Usually created by
`multiprocessing.Pipe()`
2020-02-12 14:06:27 +03:00
underscore_state (tuple): The data in the Underscore class of the parent
vectors (dict): The global vectors data, copied from the parent
"""
Underscore.load_state(underscore_state)
load_nlp.VECTORS = vectors
while True:
texts = receiver.get()
docs = (make_doc(text) for text in texts)
for pipe in pipes:
docs = pipe(docs)
# Connection does not accept unpickable objects, so send list.
sender.send([doc.to_bytes() for doc in docs])
class _Sender:
"""Util for sending data to multiprocessing workers in Language.pipe"""
def __init__(self, data, queues, chunk_size):
self.data = iter(data)
self.queues = iter(cycle(queues))
self.chunk_size = chunk_size
self.count = 0
def send(self):
"""Send chunk_size items from self.data to channels."""
for item, q in itertools.islice(
zip(self.data, cycle(self.queues)), self.chunk_size
):
# cycle channels so that distribute the texts evenly
q.put(item)
def step(self):
2019-10-10 18:08:39 +03:00
"""Tell sender that comsumed one item.
Data is sent to the workers after every chunk_size calls."""
self.count += 1
if self.count >= self.chunk_size:
self.count = 0
self.send()