spaCy/spacy/pipeline.pyx

1334 lines
50 KiB
Cython
Raw Normal View History

Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
# cython: infer_types=True
# cython: profile=True
# coding: utf8
from __future__ import unicode_literals
2017-05-07 19:04:24 +03:00
import numpy
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
cimport numpy as np
from collections import OrderedDict, defaultdict
import srsly
2017-10-27 21:29:08 +03:00
from thinc.api import chain
2018-12-01 16:41:24 +03:00
from thinc.v2v import Affine, Maxout, Softmax
from thinc.misc import LayerNorm
2017-10-27 21:29:08 +03:00
from thinc.t2v import Pooling, max_pool, mean_pool
from thinc.neural.util import to_categorical, copy_array
from thinc.neural._classes.difference import Siamese, CauchySimilarity
2017-05-08 15:53:45 +03:00
from .tokens.doc cimport Doc
from .syntax.nn_parser cimport Parser
from .syntax import nonproj
from .syntax.ner cimport BiluoPushDown
from .syntax.arc_eager cimport ArcEager
from .morphology cimport Morphology
from .vocab cimport Vocab
2017-05-22 13:17:44 +03:00
from .syntax import nonproj
from .matcher import Matcher
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
from .matcher import Matcher, PhraseMatcher
from .tokens.span import Span
2018-11-03 13:52:50 +03:00
from .attrs import POS, ID
from .parts_of_speech import X
2017-10-27 21:29:08 +03:00
from ._ml import Tok2Vec, build_text_classifier, build_tagger_model
from ._ml import build_simple_cnn_text_classifier
2017-11-03 22:20:26 +03:00
from ._ml import link_vectors_to_models, zero_init, flatten
from ._ml import create_default_optimizer
💫 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 ._ml import masked_language_model
from .errors import Errors, TempErrors
from .compat import basestring_
2017-10-27 21:29:08 +03:00
from . import util
2017-09-02 13:53:38 +03:00
class SentenceSegmenter(object):
2017-09-25 19:37:13 +03:00
"""A simple spaCy hook, to allow custom sentence boundary detection logic
2017-10-27 21:29:08 +03:00
(that doesn't require the dependency parse). To change the sentence
boundary detection strategy, pass a generator function `strategy` on
initialization, or assign a new strategy to the .strategy attribute.
2017-09-02 13:53:38 +03:00
Sentence detection strategies should be generators that take `Doc` objects
and yield `Span` objects for each sentence.
2017-09-25 19:37:13 +03:00
"""
name = 'sentencizer'
2017-09-02 13:53:38 +03:00
def __init__(self, vocab, strategy=None):
self.vocab = vocab
if strategy is None or strategy == 'on_punct':
strategy = self.split_on_punct
self.strategy = strategy
def __call__(self, doc):
doc.user_hooks['sents'] = self.strategy
return doc
2017-09-02 13:53:38 +03:00
@staticmethod
def split_on_punct(doc):
start = 0
seen_period = False
for i, word in enumerate(doc):
if seen_period and not word.is_punct:
2017-10-27 21:29:08 +03:00
yield doc[start:word.i]
2017-09-02 13:53:38 +03:00
start = word.i
seen_period = False
elif word.text in ['.', '!', '?']:
seen_period = True
if start < len(doc):
2017-10-27 21:29:08 +03:00
yield doc[start:len(doc)]
2017-09-02 13:53:38 +03:00
def merge_noun_chunks(doc):
"""Merge noun chunks into a single token.
doc (Doc): The Doc object.
RETURNS (Doc): The Doc object with merged noun chunks.
"""
if not doc.is_parsed:
return doc
spans = [(np.start_char, np.end_char, np.root.tag, np.root.dep)
for np in doc.noun_chunks]
for start, end, tag, dep in spans:
doc.merge(start, end, tag=tag, dep=dep)
return doc
def merge_entities(doc):
"""Merge entities into a single token.
doc (Doc): The Doc object.
RETURNS (Doc): The Doc object with merged noun entities.
"""
spans = [(e.start_char, e.end_char, e.root.tag, e.root.dep, e.label)
for e in doc.ents]
for start, end, tag, dep, ent_type in spans:
doc.merge(start, end, tag=tag, dep=dep, ent_type=ent_type)
return doc
def merge_subtokens(doc, label='subtok'):
merger = Matcher(doc.vocab)
merger.add('SUBTOK', None, [{'DEP': label, 'op': '+'}])
matches = merger(doc)
spans = [doc[start:end+1] for _, start, end in matches]
offsets = [(span.start_char, span.end_char) for span in spans]
for start_char, end_char in offsets:
doc.merge(start_char, end_char)
return doc
class EntityRuler(object):
name = 'entity_ruler'
def __init__(self, nlp, **cfg):
"""Initialise the entitiy ruler. If patterns are supplied here, they
need to be a list of dictionaries with a `"label"` and `"pattern"`
key. A pattern can either be a token pattern (list) or a phrase pattern
(string). For example: `{'label': 'ORG', 'pattern': 'Apple'}`.
nlp (Language): The shared nlp object to pass the vocab to the matchers
and process phrase patterns.
patterns (iterable): Optional patterns to load in.
overwrite_ents (bool): If existing entities are present, e.g. entities
added by the model, overwrite them by matches if necessary.
**cfg: Other config parameters. If pipeline component is loaded as part
of a model pipeline, this will include all keyword arguments passed
to `spacy.load`.
RETURNS (EntityRuler): The newly constructed object.
"""
self.nlp = nlp
self.overwrite = cfg.get('overwrite_ents', False)
self.token_patterns = defaultdict(list)
self.phrase_patterns = defaultdict(list)
self.matcher = Matcher(nlp.vocab)
self.phrase_matcher = PhraseMatcher(nlp.vocab)
patterns = cfg.get('patterns')
if patterns is not None:
self.add_patterns(patterns)
def __len__(self):
"""The number of all patterns added to the entity ruler."""
n_token_patterns = sum(len(p) for p in self.token_patterns.values())
n_phrase_patterns = sum(len(p) for p in self.phrase_patterns.values())
return n_token_patterns + n_phrase_patterns
def __contains__(self, label):
"""Whether a label is present in the patterns."""
return label in self.token_patterns or label in self.phrase_patterns
def __call__(self, doc):
"""Find matches in document and add them as entities.
doc (Doc): The Doc object in the pipeline.
RETURNS (Doc): The Doc with added entities, if available.
"""
matches = list(self.matcher(doc)) + list(self.phrase_matcher(doc))
matches = set([(m_id, start, end) for m_id, start, end in matches
if start != end])
get_sort_key = lambda m: (m[2] - m[1], m[1])
matches = sorted(matches, key=get_sort_key, reverse=True)
entities = list(doc.ents)
new_entities = []
seen_tokens = set()
for match_id, start, end in matches:
if any(t.ent_type for t in doc[start:end]) and not self.overwrite:
continue
# check for end - 1 here because boundaries are inclusive
if start not in seen_tokens and end - 1 not in seen_tokens:
new_entities.append(Span(doc, start, end, label=match_id))
entities = [e for e in entities
if not (e.start < end and e.end > start)]
seen_tokens.update(range(start, end))
doc.ents = entities + new_entities
return doc
@property
def labels(self):
"""All labels present in the match patterns.
RETURNS (set): The string labels.
"""
all_labels = set(self.token_patterns.keys())
all_labels.update(self.phrase_patterns.keys())
return all_labels
@property
def patterns(self):
"""Get all patterns that were added to the entity ruler.
RETURNS (list): The original patterns, one dictionary per pattern.
"""
all_patterns = []
for label, patterns in self.token_patterns.items():
for pattern in patterns:
all_patterns.append({'label': label, 'pattern': pattern})
for label, patterns in self.phrase_patterns.items():
for pattern in patterns:
all_patterns.append({'label': label, 'pattern': pattern.text})
return all_patterns
def add_patterns(self, patterns):
"""Add patterns to the entitiy ruler. A pattern can either be a token
pattern (list of dicts) or a phrase pattern (string). For example:
{'label': 'ORG', 'pattern': 'Apple'}
{'label': 'GPE', 'pattern': [{'lower': 'san'}, {'lower': 'francisco'}]}
patterns (list): The patterns to add.
"""
for entry in patterns:
label = entry['label']
pattern = entry['pattern']
if isinstance(pattern, basestring_):
self.phrase_patterns[label].append(self.nlp(pattern))
elif isinstance(pattern, list):
self.token_patterns[label].append(pattern)
else:
raise ValueError(Errors.E097.format(pattern=pattern))
for label, patterns in self.token_patterns.items():
self.matcher.add(label, None, *patterns)
for label, patterns in self.phrase_patterns.items():
self.phrase_matcher.add(label, None, *patterns)
def from_bytes(self, patterns_bytes, **kwargs):
"""Load the entity ruler from a bytestring.
patterns_bytes (bytes): The bytestring to load.
**kwargs: Other config paramters, mostly for consistency.
RETURNS (EntityRuler): The loaded entity ruler.
"""
patterns = srsly.msgpack_loads(patterns_bytes)
self.add_patterns(patterns)
return self
def to_bytes(self, **kwargs):
"""Serialize the entity ruler patterns to a bytestring.
RETURNS (bytes): The serialized patterns.
"""
return srsly.msgpack_dumps(self.patterns)
def from_disk(self, path, **kwargs):
"""Load the entity ruler from a file. Expects a file containing
newline-delimited JSON (JSONL) with one entry per line.
path (unicode / Path): The JSONL file to load.
**kwargs: Other config paramters, mostly for consistency.
RETURNS (EntityRuler): The loaded entity ruler.
"""
path = util.ensure_path(path)
path = path.with_suffix('.jsonl')
patterns = srsly.read_jsonl(path)
self.add_patterns(patterns)
return self
def to_disk(self, path, **kwargs):
"""Save the entity ruler patterns to a directory. The patterns will be
saved as newline-delimited JSON (JSONL).
path (unicode / Path): The JSONL file to load.
**kwargs: Other config paramters, mostly for consistency.
RETURNS (EntityRuler): The loaded entity ruler.
"""
path = util.ensure_path(path)
path = path.with_suffix('.jsonl')
srsly.write_jsonl(path, self.patterns)
2017-10-26 13:40:40 +03:00
class Pipe(object):
2017-10-27 21:29:08 +03:00
"""This class is not instantiated directly. Components inherit from it, and
it defines the interface that components should follow to function as
components in a spaCy analysis pipeline.
"""
name = None
@classmethod
def Model(cls, *shape, **kwargs):
2017-09-25 19:37:13 +03:00
"""Initialize a model for the pipe."""
raise NotImplementedError
def __init__(self, vocab, model=True, **cfg):
2017-09-25 19:37:13 +03:00
"""Create a new pipe instance."""
raise NotImplementedError
def __call__(self, doc):
2017-09-25 19:37:13 +03:00
"""Apply the pipe to one document. The document is
2017-09-25 17:20:49 +03:00
modified in-place, and returned.
2017-09-25 19:37:13 +03:00
2017-09-25 17:20:49 +03:00
Both __call__ and pipe should delegate to the `predict()`
and `set_annotations()` methods.
2017-09-25 19:37:13 +03:00
"""
self.require_model()
2017-11-03 13:20:05 +03:00
scores, tensors = self.predict([doc])
self.set_annotations([doc], scores, tensors=tensors)
return doc
def require_model(self):
"""Raise an error if the component's model is not initialized."""
if getattr(self, 'model', None) in (None, True, False):
raise ValueError(Errors.E109.format(name=self.name))
def pipe(self, stream, batch_size=128, n_threads=-1):
2017-09-25 19:37:13 +03:00
"""Apply the pipe to a stream of documents.
2017-09-25 17:20:49 +03:00
Both __call__ and pipe should delegate to the `predict()`
and `set_annotations()` methods.
2017-09-25 19:37:13 +03:00
"""
2018-12-03 04:19:12 +03:00
for docs in util.minibatch(stream, size=batch_size):
docs = list(docs)
2017-11-03 13:20:05 +03:00
scores, tensors = self.predict(docs)
self.set_annotations(docs, scores, tensor=tensors)
yield from docs
def predict(self, docs):
2017-09-25 19:37:13 +03:00
"""Apply the pipeline's model to a batch of docs, without
2017-09-25 17:20:49 +03:00
modifying them.
2017-09-25 19:37:13 +03:00
"""
self.require_model()
raise NotImplementedError
2017-11-03 13:20:05 +03:00
def set_annotations(self, docs, scores, tensors=None):
2017-09-25 19:37:13 +03:00
"""Modify a batch of documents, using pre-computed scores."""
raise NotImplementedError
2017-09-25 17:20:49 +03:00
def update(self, docs, golds, drop=0., sgd=None, losses=None):
2017-09-25 19:37:13 +03:00
"""Learn from a batch of documents and gold-standard information,
2017-09-25 17:20:49 +03:00
updating the pipe's model.
Delegates to predict() and get_loss().
2017-09-25 19:37:13 +03:00
"""
self.require_model()
raise NotImplementedError
💫 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):
pass
def get_loss(self, docs, golds, scores):
2017-09-25 19:37:13 +03:00
"""Find the loss and gradient of loss for the batch of
documents and their predicted scores."""
raise NotImplementedError
def add_label(self, label):
"""Add an output label, to be predicted by the model.
It's possible to extend pre-trained models with new labels,
but care should be taken to avoid the "catastrophic forgetting"
problem.
"""
raise NotImplementedError
def create_optimizer(self):
return create_default_optimizer(self.model.ops,
**self.cfg.get('optimizer', {}))
def begin_training(self, get_gold_tuples=lambda: [], pipeline=None, sgd=None,
**kwargs):
2017-09-25 19:37:13 +03:00
"""Initialize the pipe for training, using data exampes if available.
If no model has been initialized yet, the model is added."""
if self.model is True:
2017-09-25 17:20:49 +03:00
self.model = self.Model(**self.cfg)
2017-09-25 17:22:07 +03:00
link_vectors_to_models(self.vocab)
if sgd is None:
sgd = self.create_optimizer()
return sgd
def use_params(self, params):
2017-10-27 21:29:08 +03:00
"""Modify the pipe's model, to use the given parameter values."""
with self.model.use_params(params):
yield
def to_bytes(self, **exclude):
2017-09-25 19:37:13 +03:00
"""Serialize the pipe to a bytestring."""
serialize = OrderedDict()
serialize['cfg'] = lambda: srsly.json_dumps(self.cfg)
if self.model in (True, False, None):
serialize['model'] = lambda: self.model
else:
serialize['model'] = self.model.to_bytes
serialize['vocab'] = self.vocab.to_bytes
return util.to_bytes(serialize, exclude)
def from_bytes(self, bytes_data, **exclude):
2017-09-25 19:37:13 +03:00
"""Load the pipe from a bytestring."""
2017-09-02 16:17:20 +03:00
def load_model(b):
# TODO: Remove this once we don't have to handle previous models
if self.cfg.get('pretrained_dims') and 'pretrained_vectors' not in self.cfg:
2018-04-03 22:40:29 +03:00
self.cfg['pretrained_vectors'] = self.vocab.vectors.name
2017-09-02 16:17:20 +03:00
if self.model is True:
self.model = self.Model(**self.cfg)
self.model.from_bytes(b)
deserialize = OrderedDict((
('cfg', lambda b: self.cfg.update(srsly.json_loads(b))),
2017-09-26 14:45:14 +03:00
('vocab', lambda b: self.vocab.from_bytes(b)),
('model', load_model),
))
util.from_bytes(bytes_data, deserialize, exclude)
return self
def to_disk(self, path, **exclude):
2017-09-25 19:37:13 +03:00
"""Serialize the pipe to disk."""
serialize = OrderedDict()
serialize['cfg'] = lambda p: srsly.write_json(p, self.cfg)
serialize['vocab'] = lambda p: self.vocab.to_disk(p)
if self.model not in (None, True, False):
serialize['model'] = lambda p: p.open('wb').write(self.model.to_bytes())
util.to_disk(path, serialize, exclude)
def from_disk(self, path, **exclude):
2017-09-25 19:37:13 +03:00
"""Load the pipe from disk."""
2017-09-02 16:17:20 +03:00
def load_model(p):
# TODO: Remove this once we don't have to handle previous models
if self.cfg.get('pretrained_dims') and 'pretrained_vectors' not in self.cfg:
2018-04-03 22:40:29 +03:00
self.cfg['pretrained_vectors'] = self.vocab.vectors.name
2017-09-02 16:17:20 +03:00
if self.model is True:
self.model = self.Model(**self.cfg)
self.model.from_bytes(p.open('rb').read())
deserialize = OrderedDict((
2017-09-02 16:17:20 +03:00
('cfg', lambda p: self.cfg.update(_load_cfg(p))),
2017-07-23 01:33:43 +03:00
('vocab', lambda p: self.vocab.from_disk(p)),
('model', load_model),
))
util.from_disk(path, deserialize, exclude)
return self
2017-07-23 15:11:07 +03:00
def _load_cfg(path):
if path.exists():
return srsly.read_json(path)
2017-07-23 15:11:07 +03:00
else:
return {}
2017-10-26 13:40:40 +03:00
class Tensorizer(Pipe):
"""Pre-train position-sensitive vectors for tokens."""
2017-05-31 14:42:39 +03:00
name = 'tensorizer'
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
@classmethod
2018-11-03 13:52:50 +03:00
def Model(cls, output_size=300, **cfg):
2017-05-19 01:00:02 +03:00
"""Create a new statistical model for the class.
width (int): Output size of the model.
embed_size (int): Number of vectors in the embedding table.
**cfg: Config parameters.
RETURNS (Model): A `thinc.neural.Model` or similar instance.
"""
2018-11-30 19:40:11 +03:00
input_size = util.env_opt('token_vector_width', cfg.get('input_size', 96))
return zero_init(Affine(output_size, input_size, drop_factor=0.0))
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
def __init__(self, vocab, model=True, **cfg):
2017-05-19 01:00:02 +03:00
"""Construct a new statistical model. Weights are not allocated on
initialisation.
2017-10-27 21:29:08 +03:00
vocab (Vocab): A `Vocab` instance. The model must share the same
`Vocab` instance with the `Doc` objects it will process.
2017-05-19 01:00:02 +03:00
model (Model): A `Model` instance or `True` allocate one later.
**cfg: Config parameters.
EXAMPLE:
>>> from spacy.pipeline import TokenVectorEncoder
>>> tok2vec = TokenVectorEncoder(nlp.vocab)
>>> tok2vec.model = tok2vec.Model(128, 5000)
"""
self.vocab = vocab
2017-05-18 12:29:51 +03:00
self.model = model
2017-11-03 22:20:26 +03:00
self.input_models = []
2017-07-23 01:52:47 +03:00
self.cfg = dict(cfg)
2017-09-22 17:38:22 +03:00
self.cfg.setdefault('cnn_maxout_pieces', 3)
2017-05-28 16:11:58 +03:00
def __call__(self, doc):
2017-05-19 01:00:02 +03:00
"""Add context-sensitive vectors to a `Doc`, e.g. from a CNN or LSTM
model. Vectors are set to the `Doc.tensor` attribute.
docs (Doc or iterable): One or more documents to add vectors to.
RETURNS (dict or None): Intermediate computations.
"""
2017-05-28 16:11:58 +03:00
tokvecses = self.predict([doc])
self.set_annotations([doc], tokvecses)
return doc
2017-05-18 16:30:59 +03:00
def pipe(self, stream, batch_size=128, n_threads=-1):
2017-05-19 01:00:02 +03:00
"""Process `Doc` objects as a stream.
stream (iterator): A sequence of `Doc` objects to process.
batch_size (int): Number of `Doc` objects to group.
n_threads (int): Number of threads.
2017-05-21 21:46:23 +03:00
YIELDS (iterator): A sequence of `Doc` objects, in order of input.
2017-05-19 01:00:02 +03:00
"""
2018-12-03 04:19:12 +03:00
for docs in util.minibatch(stream, size=batch_size):
2017-05-22 01:52:01 +03:00
docs = list(docs)
2017-11-03 22:20:26 +03:00
tensors = self.predict(docs)
self.set_annotations(docs, tensors)
yield from docs
2017-05-18 12:29:51 +03:00
def predict(self, docs):
2017-05-19 01:00:02 +03:00
"""Return a single tensor for a batch of documents.
docs (iterable): A sequence of `Doc` objects.
2017-10-27 21:29:08 +03:00
RETURNS (object): Vector representations for each token in the docs.
2017-05-19 01:00:02 +03:00
"""
self.require_model()
2017-11-03 22:20:26 +03:00
inputs = self.model.ops.flatten([doc.tensor for doc in docs])
outputs = self.model(inputs)
return self.model.ops.unflatten(outputs, [len(d) for d in docs])
2017-11-03 22:20:26 +03:00
def set_annotations(self, docs, tensors):
2017-05-19 01:00:02 +03:00
"""Set the tensor attribute for a batch of documents.
docs (iterable): A sequence of `Doc` objects.
2017-11-03 22:20:26 +03:00
tensors (object): Vector representation for each token in the docs.
2017-05-19 01:00:02 +03:00
"""
2017-11-03 22:20:26 +03:00
for doc, tensor in zip(docs, tensors):
if tensor.shape[0] != len(doc):
raise ValueError(Errors.E076.format(rows=tensor.shape[0], words=len(doc)))
2017-11-03 22:20:26 +03:00
doc.tensor = tensor
2017-05-25 04:09:51 +03:00
def update(self, docs, golds, state=None, drop=0., sgd=None, losses=None):
2017-05-19 01:00:02 +03:00
"""Update the model.
docs (iterable): A batch of `Doc` objects.
golds (iterable): A batch of `GoldParse` objects.
drop (float): The droput rate.
sgd (callable): An optimizer.
2017-05-19 01:00:02 +03:00
RETURNS (dict): Results from the update.
"""
self.require_model()
if isinstance(docs, Doc):
docs = [docs]
2017-11-03 22:20:26 +03:00
inputs = []
bp_inputs = []
for tok2vec in self.input_models:
tensor, bp_tensor = tok2vec.begin_update(docs, drop=drop)
inputs.append(tensor)
bp_inputs.append(bp_tensor)
inputs = self.model.ops.xp.hstack(inputs)
scores, bp_scores = self.model.begin_update(inputs, drop=drop)
loss, d_scores = self.get_loss(docs, golds, scores)
d_inputs = bp_scores(d_scores, sgd=sgd)
d_inputs = self.model.ops.xp.split(d_inputs, len(self.input_models), axis=1)
2017-11-05 14:25:10 +03:00
for d_input, bp_input in zip(d_inputs, bp_inputs):
2017-11-03 22:20:26 +03:00
bp_input(d_input, sgd=sgd)
if losses is not None:
losses.setdefault(self.name, 0.)
losses[self.name] += loss
return loss
def get_loss(self, docs, golds, prediction):
2018-11-03 13:52:50 +03:00
ids = self.model.ops.flatten([doc.to_array(ID).ravel() for doc in docs])
target = self.vocab.vectors.data[ids]
d_scores = (prediction - target) / prediction.shape[0]
2017-11-03 22:20:26 +03:00
loss = (d_scores**2).sum()
return loss, d_scores
def begin_training(self, gold_tuples=lambda: [], pipeline=None, sgd=None,
**kwargs):
"""Allocate models, pre-process training data and acquire an
2017-05-19 01:00:02 +03:00
optimizer.
gold_tuples (iterable): Gold-standard training data.
pipeline (list): The pipeline the model is part of.
"""
2018-11-03 01:51:37 +03:00
if pipeline is not None:
for name, model in pipeline:
if getattr(model, 'tok2vec', None):
self.input_models.append(model.tok2vec)
2017-05-18 12:29:51 +03:00
if self.model is True:
self.model = self.Model(**self.cfg)
link_vectors_to_models(self.vocab)
if sgd is None:
sgd = self.create_optimizer()
return sgd
2017-05-18 12:29:51 +03:00
2017-05-29 02:37:57 +03:00
2017-10-26 13:40:40 +03:00
class Tagger(Pipe):
2017-06-01 18:37:53 +03:00
name = 'tagger'
2017-10-27 21:29:08 +03:00
2017-07-23 01:52:47 +03:00
def __init__(self, vocab, model=True, **cfg):
self.vocab = vocab
self.model = model
💫 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
self._rehearsal_model = None
2017-11-08 14:10:49 +03:00
self.cfg = OrderedDict(sorted(cfg.items()))
self.cfg.setdefault('cnn_maxout_pieces', 2)
@property
def labels(self):
return self.vocab.morphology.tag_names
2017-11-03 22:20:26 +03:00
@property
def tok2vec(self):
if self.model in (None, True, False):
return None
else:
return chain(self.model.tok2vec, flatten)
def __call__(self, doc):
2017-11-03 13:20:05 +03:00
tags, tokvecs = self.predict([doc])
self.set_annotations([doc], tags, tensors=tokvecs)
2017-05-28 16:11:58 +03:00
return doc
def pipe(self, stream, batch_size=128, n_threads=-1):
2018-12-03 04:19:12 +03:00
for docs in util.minibatch(stream, size=batch_size):
docs = list(docs)
2017-11-03 13:20:05 +03:00
tag_ids, tokvecs = self.predict(docs)
self.set_annotations(docs, tag_ids, tensors=tokvecs)
yield from docs
def predict(self, docs):
self.require_model()
2018-06-29 14:44:25 +03:00
if not any(len(doc) for doc in docs):
# Handle case where there are no tokens in any docs.
2018-06-29 16:13:45 +03:00
n_labels = len(self.labels)
2018-06-29 17:05:40 +03:00
guesses = [self.model.ops.allocate((0, n_labels)) for doc in docs]
tokvecs = self.model.ops.allocate((0, self.model.tok2vec.nO))
return guesses, tokvecs
2017-11-03 13:20:05 +03:00
tokvecs = self.model.tok2vec(docs)
scores = self.model.softmax(tokvecs)
2017-11-03 15:29:36 +03:00
guesses = []
for doc_scores in scores:
doc_guesses = doc_scores.argmax(axis=1)
if not isinstance(doc_guesses, numpy.ndarray):
doc_guesses = doc_guesses.get()
guesses.append(doc_guesses)
2017-11-03 13:20:05 +03:00
return guesses, tokvecs
2017-11-03 13:20:05 +03:00
def set_annotations(self, docs, batch_tag_ids, tensors=None):
if isinstance(docs, Doc):
docs = [docs]
cdef Doc doc
cdef int idx = 0
2017-05-18 12:29:51 +03:00
cdef Vocab vocab = self.vocab
2017-05-08 15:53:45 +03:00
for i, doc in enumerate(docs):
2017-05-21 17:05:34 +03:00
doc_tag_ids = batch_tag_ids[i]
if hasattr(doc_tag_ids, 'get'):
doc_tag_ids = doc_tag_ids.get()
2017-05-18 12:29:51 +03:00
for j, tag_id in enumerate(doc_tag_ids):
2017-06-04 23:52:42 +03:00
# Don't clobber preset POS tags
if doc.c[j].tag == 0 and doc.c[j].pos == 0:
2017-11-06 14:36:05 +03:00
# Don't clobber preset lemmas
lemma = doc.c[j].lemma
vocab.morphology.assign_tag_id(&doc.c[j], tag_id)
2017-11-06 18:56:19 +03:00
if lemma != 0 and lemma != doc.c[j].lex.orth:
2017-11-06 14:36:05 +03:00
doc.c[j].lemma = lemma
2017-05-08 15:53:45 +03:00
idx += 1
2018-06-29 20:21:38 +03:00
if tensors is not None and len(tensors):
2017-11-05 17:34:40 +03:00
if isinstance(doc.tensor, numpy.ndarray) \
and not isinstance(tensors[i], numpy.ndarray):
doc.extend_tensor(tensors[i].get())
else:
doc.extend_tensor(tensors[i])
doc.is_tagged = True
2017-05-08 15:53:45 +03:00
def update(self, docs, golds, drop=0., sgd=None, losses=None):
self.require_model()
2017-08-20 15:42:23 +03:00
if losses is not None and self.name not in losses:
losses[self.name] = 0.
tag_scores, bp_tag_scores = self.model.begin_update(docs, drop=drop)
loss, d_tag_scores = self.get_loss(docs, golds, tag_scores)
2017-09-23 03:58:06 +03:00
bp_tag_scores(d_tag_scores, sgd=sgd)
2017-05-18 12:29:51 +03:00
2017-08-20 15:42:23 +03:00
if losses is not None:
losses[self.name] += loss
💫 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, drop=0., sgd=None, losses=None):
"""Perform a 'rehearsal' update, where we try to match the output of
an initial model.
"""
if self._rehearsal_model is None:
return
guesses, backprop = self.model.begin_update(docs, drop=drop)
target = self._rehearsal_model(docs)
gradient = guesses - target
backprop(gradient, sgd=sgd)
if losses is not None:
losses.setdefault(self.name, 0.0)
losses[self.name] += (gradient**2).sum()
def get_loss(self, docs, golds, scores):
scores = self.model.ops.flatten(scores)
tag_index = {tag: i for i, tag in enumerate(self.labels)}
2017-05-18 12:29:51 +03:00
cdef int idx = 0
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
correct = numpy.zeros((scores.shape[0],), dtype='i')
guesses = scores.argmax(axis=1)
known_labels = numpy.ones((scores.shape[0], 1), dtype='f')
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
for gold in golds:
for tag in gold.tags:
if tag is None:
correct[idx] = guesses[idx]
elif tag in tag_index:
correct[idx] = tag_index[tag]
else:
correct[idx] = 0
known_labels[idx] = 0.
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
idx += 1
2017-05-18 16:30:59 +03:00
correct = self.model.ops.xp.array(correct, dtype='i')
Update draft of parser neural network model Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU. Outline of the model: We first predict context-sensitive vectors for each word in the input: (embed_lower | embed_prefix | embed_suffix | embed_shape) >> Maxout(token_width) >> convolution ** 4 This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features. To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a representation that's one affine transform from this informative lexical information. This is obviously good for the parser (which backprops to the convolutions too). The parser model makes a state vector by concatenating the vector representations for its context tokens. Current results suggest few context tokens works well. Maybe this is a bug. The current context tokens: * S0, S1, S2: Top three words on the stack * B0, B1: First two words of the buffer * S0L1, S0L2: Leftmost and second leftmost children of S0 * S0R1, S0R2: Rightmost and second rightmost children of S0 * S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0 This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately, there's a way to structure the computation to save some expense (and make it more GPU friendly). The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN -- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model is so big.) This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity. The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier. We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle in CUDA to train. Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to be 0 cost. This is defined as: (exp(score) / Z) - (exp(score) / gZ) Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly, but so far this isn't working well. Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit greatly from the pre-computation trick.
2017-05-13 00:09:15 +03:00
d_scores = scores - to_categorical(correct, nb_classes=scores.shape[1])
2018-09-13 15:14:38 +03:00
d_scores *= self.model.ops.asarray(known_labels)
2017-05-18 12:29:51 +03:00
loss = (d_scores**2).sum()
d_scores = self.model.ops.unflatten(d_scores, [len(d) for d in docs])
2017-05-18 16:30:59 +03:00
return float(loss), d_scores
def begin_training(self, get_gold_tuples=lambda: [], pipeline=None, sgd=None,
**kwargs):
2017-05-18 16:30:59 +03:00
orig_tag_map = dict(self.vocab.morphology.tag_map)
2017-11-08 14:10:49 +03:00
new_tag_map = OrderedDict()
for raw_text, annots_brackets in get_gold_tuples():
for annots, brackets in annots_brackets:
ids, words, tags, heads, deps, ents = annots
for tag in tags:
2017-05-18 16:30:59 +03:00
if tag in orig_tag_map:
new_tag_map[tag] = orig_tag_map[tag]
else:
new_tag_map[tag] = {POS: X}
cdef Vocab vocab = self.vocab
2017-06-01 11:04:36 +03:00
if new_tag_map:
vocab.morphology = Morphology(vocab.strings, new_tag_map,
2017-06-05 00:34:32 +03:00
vocab.morphology.lemmatizer,
exc=vocab.morphology.exc)
self.cfg['pretrained_vectors'] = kwargs.get('pretrained_vectors')
2017-05-29 21:23:47 +03:00
if self.model is True:
2018-12-18 02:08:31 +03:00
for hp in ['token_vector_width', 'conv_depth']:
if hp in kwargs:
self.cfg[hp] = kwargs[hp]
2017-09-21 21:07:26 +03:00
self.model = self.Model(self.vocab.morphology.n_tags, **self.cfg)
link_vectors_to_models(self.vocab)
if sgd is None:
sgd = self.create_optimizer()
return sgd
2017-05-29 21:23:47 +03:00
@classmethod
2017-09-23 03:58:06 +03:00
def Model(cls, n_tags, **cfg):
if cfg.get('pretrained_dims') and not cfg.get('pretrained_vectors'):
2018-04-03 22:40:29 +03:00
raise ValueError(TempErrors.T008)
2017-09-23 03:58:06 +03:00
return build_tagger_model(n_tags, **cfg)
def add_label(self, label, values=None):
if label in self.labels:
return 0
if self.model not in (True, False, None):
# Here's how the model resizing will work, once the
# neuron-to-tag mapping is no longer controlled by
# the Morphology class, which sorts the tag names.
# The sorting makes adding labels difficult.
# smaller = self.model._layers[-1]
# larger = Softmax(len(self.labels)+1, smaller.nI)
# copy_array(larger.W[:smaller.nO], smaller.W)
# copy_array(larger.b[:smaller.nO], smaller.b)
# self.model._layers[-1] = larger
raise ValueError(TempErrors.T003)
tag_map = dict(self.vocab.morphology.tag_map)
if values is None:
values = {POS: "X"}
tag_map[label] = values
self.vocab.morphology = Morphology(
self.vocab.strings, tag_map=tag_map,
lemmatizer=self.vocab.morphology.lemmatizer,
exc=self.vocab.morphology.exc)
return 1
2017-05-18 16:30:59 +03:00
def use_params(self, params):
with self.model.use_params(params):
yield
2017-05-29 11:14:20 +03:00
def to_bytes(self, **exclude):
serialize = OrderedDict()
if self.model in (None, True, False):
serialize['model'] = lambda: self.model
else:
serialize['model'] = self.model.to_bytes
serialize['vocab'] = self.vocab.to_bytes
serialize['cfg'] = lambda: srsly.json_dumps(self.cfg)
2017-11-08 15:08:48 +03:00
tag_map = OrderedDict(sorted(self.vocab.morphology.tag_map.items()))
serialize['tag_map'] = lambda: srsly.msgpack_dumps(tag_map)
2017-05-29 11:14:20 +03:00
return util.to_bytes(serialize, exclude)
def from_bytes(self, bytes_data, **exclude):
2017-05-29 21:23:47 +03:00
def load_model(b):
# TODO: Remove this once we don't have to handle previous models
if self.cfg.get('pretrained_dims') and 'pretrained_vectors' not in self.cfg:
2018-04-03 22:40:29 +03:00
self.cfg['pretrained_vectors'] = self.vocab.vectors.name
2017-05-29 21:23:47 +03:00
if self.model is True:
2017-10-27 21:29:08 +03:00
token_vector_width = util.env_opt(
'token_vector_width',
self.cfg.get('token_vector_width', 96))
2017-10-27 21:29:08 +03:00
self.model = self.Model(self.vocab.morphology.n_tags,
**self.cfg)
self.model.from_bytes(b)
def load_tag_map(b):
tag_map = srsly.msgpack_loads(b)
self.vocab.morphology = Morphology(
self.vocab.strings, tag_map=tag_map,
2017-06-05 00:34:32 +03:00
lemmatizer=self.vocab.morphology.lemmatizer,
exc=self.vocab.morphology.exc)
2017-05-30 01:53:06 +03:00
deserialize = OrderedDict((
('vocab', lambda b: self.vocab.from_bytes(b)),
('tag_map', load_tag_map),
('cfg', lambda b: self.cfg.update(srsly.json_loads(b))),
2017-05-30 01:53:06 +03:00
('model', lambda b: load_model(b)),
))
2017-05-29 21:23:47 +03:00
util.from_bytes(bytes_data, deserialize, exclude)
2017-05-29 11:14:20 +03:00
return self
def to_disk(self, path, **exclude):
2017-11-08 15:08:48 +03:00
tag_map = OrderedDict(sorted(self.vocab.morphology.tag_map.items()))
serialize = OrderedDict((
('vocab', lambda p: self.vocab.to_disk(p)),
('tag_map', lambda p: srsly.write_msgpack(p, tag_map)),
('model', lambda p: p.open('wb').write(self.model.to_bytes())),
('cfg', lambda p: srsly.write_json(p, self.cfg))
))
util.to_disk(path, serialize, exclude)
def from_disk(self, path, **exclude):
def load_model(p):
# TODO: Remove this once we don't have to handle previous models
if self.cfg.get('pretrained_dims') and 'pretrained_vectors' not in self.cfg:
self.cfg['pretrained_vectors'] = self.vocab.vectors.name
if self.model is True:
2017-09-21 21:07:26 +03:00
self.model = self.Model(self.vocab.morphology.n_tags, **self.cfg)
with p.open('rb') as file_:
self.model.from_bytes(file_.read())
def load_tag_map(p):
tag_map = srsly.read_msgpack(p)
self.vocab.morphology = Morphology(
self.vocab.strings, tag_map=tag_map,
2017-06-05 00:34:32 +03:00
lemmatizer=self.vocab.morphology.lemmatizer,
exc=self.vocab.morphology.exc)
deserialize = OrderedDict((
('cfg', lambda p: self.cfg.update(_load_cfg(p))),
('vocab', lambda p: self.vocab.from_disk(p)),
('tag_map', load_tag_map),
('model', load_model),
))
util.from_disk(path, deserialize, exclude)
return self
2017-05-29 11:14:20 +03:00
class MultitaskObjective(Tagger):
2017-10-27 21:29:08 +03:00
"""Experimental: Assist training of a parser or tagger, by training a
side-objective.
"""
2017-05-22 01:52:30 +03:00
name = 'nn_labeller'
2017-10-27 21:29:08 +03:00
def __init__(self, vocab, model=True, target='dep_tag_offset', **cfg):
2017-05-22 01:52:30 +03:00
self.vocab = vocab
self.model = model
if target == 'dep':
self.make_label = self.make_dep
elif target == 'tag':
self.make_label = self.make_tag
elif target == 'ent':
self.make_label = self.make_ent
elif target == 'dep_tag_offset':
self.make_label = self.make_dep_tag_offset
elif target == 'ent_tag':
self.make_label = self.make_ent_tag
elif target == 'sent_start':
self.make_label = self.make_sent_start
elif hasattr(target, '__call__'):
self.make_label = target
else:
raise ValueError(Errors.E016)
2017-07-23 01:52:47 +03:00
self.cfg = dict(cfg)
self.cfg.setdefault('cnn_maxout_pieces', 2)
2017-07-23 01:52:47 +03:00
@property
def labels(self):
return self.cfg.setdefault('labels', {})
2017-07-23 01:52:47 +03:00
@labels.setter
def labels(self, value):
self.cfg['labels'] = value
2017-05-22 01:52:30 +03:00
2017-11-03 13:20:05 +03:00
def set_annotations(self, docs, dep_ids, tensors=None):
2017-05-22 01:52:30 +03:00
pass
def begin_training(self, get_gold_tuples=lambda: [], pipeline=None, tok2vec=None,
sgd=None, **kwargs):
gold_tuples = nonproj.preprocess_training_data(get_gold_tuples())
2017-05-22 01:52:30 +03:00
for raw_text, annots_brackets in gold_tuples:
for annots, brackets in annots_brackets:
ids, words, tags, heads, deps, ents = annots
for i in range(len(ids)):
label = self.make_label(i, words, tags, heads, deps, ents)
if label is not None and label not in self.labels:
self.labels[label] = len(self.labels)
2017-05-29 21:23:47 +03:00
if self.model is True:
2017-09-27 19:43:58 +03:00
token_vector_width = util.env_opt('token_vector_width')
2018-01-21 21:21:34 +03:00
self.model = self.Model(len(self.labels), tok2vec=tok2vec)
link_vectors_to_models(self.vocab)
if sgd is None:
sgd = self.create_optimizer()
return sgd
2017-05-29 21:23:47 +03:00
@classmethod
def Model(cls, n_tags, tok2vec=None, **cfg):
token_vector_width = util.env_opt('token_vector_width', 96)
2018-12-01 16:41:24 +03:00
softmax = Softmax(n_tags, token_vector_width*2)
2018-01-21 21:21:34 +03:00
model = chain(
tok2vec,
2018-12-01 16:41:24 +03:00
LayerNorm(Maxout(token_vector_width*2, token_vector_width, pieces=3)),
2018-01-21 21:21:34 +03:00
softmax
)
model.tok2vec = tok2vec
model.softmax = softmax
return model
def predict(self, docs):
self.require_model()
2018-01-21 21:21:34 +03:00
tokvecs = self.model.tok2vec(docs)
scores = self.model.softmax(tokvecs)
return tokvecs, scores
2017-05-22 01:52:30 +03:00
def get_loss(self, docs, golds, scores):
if len(docs) != len(golds):
raise ValueError(Errors.E077.format(value='loss', n_docs=len(docs),
n_golds=len(golds)))
2017-05-22 01:52:30 +03:00
cdef int idx = 0
correct = numpy.zeros((scores.shape[0],), dtype='i')
guesses = scores.argmax(axis=1)
2018-02-17 20:41:18 +03:00
for i, gold in enumerate(golds):
for j in range(len(docs[i])):
# Handes alignment for tokenization differences
label = self.make_label(j, gold.words, gold.tags,
2018-02-17 20:41:18 +03:00
gold.heads, gold.labels, gold.ents)
if label is None or label not in self.labels:
2017-05-22 01:52:30 +03:00
correct[idx] = guesses[idx]
else:
correct[idx] = self.labels[label]
2017-05-22 01:52:30 +03:00
idx += 1
correct = self.model.ops.xp.array(correct, dtype='i')
d_scores = scores - to_categorical(correct, nb_classes=scores.shape[1])
loss = (d_scores**2).sum()
return float(loss), d_scores
@staticmethod
def make_dep(i, words, tags, heads, deps, ents):
if deps[i] is None or heads[i] is None:
return None
return deps[i]
@staticmethod
def make_tag(i, words, tags, heads, deps, ents):
return tags[i]
@staticmethod
def make_ent(i, words, tags, heads, deps, ents):
if ents is None:
return None
return ents[i]
@staticmethod
def make_dep_tag_offset(i, words, tags, heads, deps, ents):
if deps[i] is None or heads[i] is None:
return None
offset = heads[i] - i
offset = min(offset, 2)
offset = max(offset, -2)
return '%s-%s:%d' % (deps[i], tags[i], offset)
@staticmethod
def make_ent_tag(i, words, tags, heads, deps, ents):
if ents is None or ents[i] is None:
return None
else:
return '%s-%s' % (tags[i], ents[i])
@staticmethod
def make_sent_start(target, words, tags, heads, deps, ents, cache=True, _cache={}):
'''A multi-task objective for representing sentence boundaries,
using BILU scheme. (O is impossible)
The implementation of this method uses an internal cache that relies
on the identity of the heads array, to avoid requiring a new piece
of gold data. You can pass cache=False if you know the cache will
do the wrong thing.
'''
assert len(words) == len(heads)
assert target < len(words), (target, len(words))
if cache:
if id(heads) in _cache:
return _cache[id(heads)][target]
else:
for key in list(_cache.keys()):
_cache.pop(key)
sent_tags = ['I-SENT'] * len(words)
_cache[id(heads)] = sent_tags
else:
sent_tags = ['I-SENT'] * len(words)
def _find_root(child):
seen = set([child])
while child is not None and heads[child] != child:
seen.add(child)
child = heads[child]
return child
sentences = {}
for i in range(len(words)):
root = _find_root(i)
if root is None:
sent_tags[i] = None
else:
sentences.setdefault(root, []).append(i)
for root, span in sorted(sentences.items()):
if len(span) == 1:
sent_tags[span[0]] = 'U-SENT'
else:
sent_tags[span[0]] = 'B-SENT'
sent_tags[span[-1]] = 'L-SENT'
return sent_tags[target]
💫 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
class ClozeMultitask(Pipe):
@classmethod
def Model(cls, vocab, tok2vec, **cfg):
output_size = vocab.vectors.data.shape[1]
output_layer = chain(
LayerNorm(Maxout(output_size, tok2vec.nO, pieces=3)),
zero_init(Affine(output_size, output_size, drop_factor=0.0))
)
model = chain(tok2vec, output_layer)
model = masked_language_model(vocab, model)
model.tok2vec = tok2vec
model.output_layer = output_layer
return model
def __init__(self, vocab, model=True, **cfg):
self.vocab = vocab
self.model = model
self.cfg = cfg
def set_annotations(self, docs, dep_ids, tensors=None):
pass
def begin_training(self, get_gold_tuples=lambda: [], pipeline=None,
tok2vec=None, sgd=None, **kwargs):
link_vectors_to_models(self.vocab)
if self.model is True:
self.model = self.Model(self.vocab, tok2vec)
X = self.model.ops.allocate((5, self.model.tok2vec.nO))
self.model.output_layer.begin_training(X)
if sgd is None:
sgd = self.create_optimizer()
return sgd
def predict(self, docs):
self.require_model()
💫 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
tokvecs = self.model.tok2vec(docs)
vectors = self.model.output_layer(tokvecs)
return tokvecs, vectors
def get_loss(self, docs, vectors, prediction):
# The simplest way to implement this would be to vstack the
# token.vector values, but that's a bit inefficient, especially on GPU.
# Instead we fetch the index into the vectors table for each of our tokens,
# and look them up all at once. This prevents data copying.
ids = self.model.ops.flatten([doc.to_array(ID).ravel() for doc in docs])
target = vectors[ids]
gradient = (prediction - target) / prediction.shape[0]
loss = (gradient**2).sum()
return float(loss), gradient
def update(self, docs, golds, drop=0., sgd=None, losses=None):
pass
def rehearse(self, docs, drop=0., sgd=None, losses=None):
self.require_model()
💫 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 losses is not None and self.name not in losses:
losses[self.name] = 0.
predictions, bp_predictions = self.model.begin_update(docs, drop=drop)
loss, d_predictions = self.get_loss(docs, self.vocab.vectors.data, predictions)
bp_predictions(d_predictions, sgd=sgd)
if losses is not None:
losses[self.name] += loss
2017-10-26 13:40:40 +03:00
class SimilarityHook(Pipe):
"""
2017-10-27 21:29:08 +03:00
Experimental: A pipeline component to install a hook for supervised
similarity into `Doc` objects. Requires a `Tensorizer` to pre-process
documents. The similarity model can be any object obeying the Thinc `Model`
interface. By default, the model concatenates the elementwise mean and
elementwise max of the two tensors, and compares them using the
Cauchy-like similarity function from Chen (2013):
2017-10-27 21:29:08 +03:00
>>> similarity = 1. / (1. + (W * (vec1-vec2)**2).sum())
Where W is a vector of dimension weights, initialized to 1.
"""
name = 'similarity'
2017-10-27 21:29:08 +03:00
2017-07-23 01:52:47 +03:00
def __init__(self, vocab, model=True, **cfg):
self.vocab = vocab
self.model = model
2017-07-23 01:52:47 +03:00
self.cfg = dict(cfg)
@classmethod
def Model(cls, length):
return Siamese(Pooling(max_pool, mean_pool), CauchySimilarity(length))
def __call__(self, doc):
2017-09-25 19:37:13 +03:00
"""Install similarity hook"""
doc.user_hooks['similarity'] = self.predict
return doc
def pipe(self, docs, **kwargs):
for doc in docs:
yield self(doc)
def predict(self, doc1, doc2):
self.require_model()
return self.model.predict([(doc1, doc2)])
def update(self, doc1_doc2, golds, sgd=None, drop=0.):
self.require_model()
sims, bp_sims = self.model.begin_update(doc1_doc2, drop=drop)
def begin_training(self, _=tuple(), pipeline=None, sgd=None, **kwargs):
2017-10-27 21:29:08 +03:00
"""Allocate model, using width from tensorizer in pipeline.
gold_tuples (iterable): Gold-standard training data.
pipeline (list): The pipeline the model is part of.
"""
if self.model is True:
self.model = self.Model(pipeline[0].model.nO)
2017-09-22 17:38:22 +03:00
link_vectors_to_models(self.vocab)
if sgd is None:
sgd = self.create_optimizer()
return sgd
2017-10-26 13:40:40 +03:00
class TextCategorizer(Pipe):
2017-07-22 02:14:07 +03:00
name = 'textcat'
@classmethod
2018-04-29 16:48:53 +03:00
def Model(cls, nr_class, **cfg):
embed_size = util.env_opt("embed_size", 2000)
if "token_vector_width" in cfg:
token_vector_width = cfg["token_vector_width"]
else:
token_vector_width = util.env_opt("token_vector_width", 96)
tok2vec = Tok2Vec(token_vector_width, embed_size, **cfg)
return build_simple_cnn_text_classifier(tok2vec, nr_class, **cfg)
2018-11-03 01:51:37 +03:00
@property
def tok2vec(self):
if self.model in (None, True, False):
return None
else:
return self.model.tok2vec
2018-11-03 01:51:37 +03:00
def __init__(self, vocab, model=True, **cfg):
self.vocab = vocab
self.model = model
💫 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
self._rehearsal_model = None
2017-07-23 01:52:47 +03:00
self.cfg = dict(cfg)
2017-07-23 01:33:43 +03:00
@property
def labels(self):
return self.cfg.setdefault('labels', [])
2017-07-23 01:33:43 +03:00
@labels.setter
def labels(self, value):
self.cfg['labels'] = value
def __call__(self, doc):
2017-11-03 13:20:05 +03:00
scores, tensors = self.predict([doc])
self.set_annotations([doc], scores, tensors=tensors)
return doc
def pipe(self, stream, batch_size=128, n_threads=-1):
2018-12-03 04:19:12 +03:00
for docs in util.minibatch(stream, size=batch_size):
docs = list(docs)
2017-11-03 13:20:05 +03:00
scores, tensors = self.predict(docs)
self.set_annotations(docs, scores, tensors=tensors)
yield from docs
def predict(self, docs):
self.require_model()
scores = self.model(docs)
scores = self.model.ops.asarray(scores)
2017-11-05 14:25:10 +03:00
tensors = [doc.tensor for doc in docs]
return scores, tensors
2017-11-03 13:20:05 +03:00
def set_annotations(self, docs, scores, tensors=None):
for i, doc in enumerate(docs):
for j, label in enumerate(self.labels):
doc.cats[label] = float(scores[i, j])
def update(self, docs, golds, state=None, drop=0., sgd=None, losses=None):
scores, bp_scores = self.model.begin_update(docs, drop=drop)
loss, d_scores = self.get_loss(docs, golds, scores)
bp_scores(d_scores, sgd=sgd)
if losses is not None:
losses.setdefault(self.name, 0.0)
losses[self.name] += loss
💫 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, drop=0., sgd=None, losses=None):
if self._rehearsal_model is None:
return
scores, bp_scores = self.model.begin_update(docs, drop=drop)
target = self._rehearsal_model(docs)
gradient = scores - target
bp_scores(gradient, sgd=sgd)
if losses is not None:
losses.setdefault(self.name, 0.0)
losses[self.name] += (gradient**2).sum()
def get_loss(self, docs, golds, scores):
truths = numpy.zeros((len(golds), len(self.labels)), dtype='f')
not_missing = numpy.ones((len(golds), len(self.labels)), dtype='f')
for i, gold in enumerate(golds):
for j, label in enumerate(self.labels):
if label in gold.cats:
truths[i, j] = gold.cats[label]
else:
not_missing[i, j] = 0.
truths = self.model.ops.asarray(truths)
not_missing = self.model.ops.asarray(not_missing)
d_scores = (scores-truths) / scores.shape[0]
d_scores *= not_missing
mean_square_error = ((scores-truths)**2).sum(axis=1).mean()
return float(mean_square_error), d_scores
def add_label(self, label):
if label in self.labels:
return 0
2017-11-01 19:06:43 +03:00
if self.model not in (None, True, False):
# This functionality was available previously, but was broken.
# The problem is that we resize the last layer, but the last layer
# is actually just an ensemble. We're not resizing the child layers
# -- a huge problem.
raise ValueError(
"Cannot currently add labels to pre-trained text classifier. "
"Add labels before training begins. This functionality was "
"available in previous versions, but had significant bugs that "
"let to poor performance")
#smaller = self.model._layers[-1]
#larger = Affine(len(self.labels)+1, smaller.nI)
#copy_array(larger.W[:smaller.nO], smaller.W)
#copy_array(larger.b[:smaller.nO], smaller.b)
#self.model._layers[-1] = larger
self.labels.append(label)
return 1
def begin_training(self, get_gold_tuples=lambda: [], pipeline=None, sgd=None,
**kwargs):
2017-09-02 15:56:30 +03:00
if pipeline and getattr(pipeline[0], 'name', None) == 'tensorizer':
token_vector_width = pipeline[0].model.nO
else:
token_vector_width = 64
if self.model is True:
self.cfg['pretrained_vectors'] = kwargs.get('pretrained_vectors')
2018-04-29 16:48:53 +03:00
self.model = self.Model(len(self.labels), **self.cfg)
2017-09-22 17:38:22 +03:00
link_vectors_to_models(self.vocab)
if sgd is None:
sgd = self.create_optimizer()
return sgd
cdef class DependencyParser(Parser):
name = 'parser'
TransitionSystem = ArcEager
@property
def postprocesses(self):
return [nonproj.deprojectivize]
def add_multitask_objective(self, target):
💫 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 target == 'cloze':
cloze = ClozeMultitask(self.vocab)
self._multitasks.append(cloze)
else:
labeller = MultitaskObjective(self.vocab, target=target)
self._multitasks.append(labeller)
def init_multitask_objectives(self, get_gold_tuples, pipeline, sgd=None, **cfg):
for labeller in self._multitasks:
2018-09-13 15:08:55 +03:00
tok2vec = self.model.tok2vec
labeller.begin_training(get_gold_tuples, pipeline=pipeline,
tok2vec=tok2vec, sgd=sgd)
def __reduce__(self):
2017-10-27 21:29:08 +03:00
return (DependencyParser, (self.vocab, self.moves, self.model),
None, None)
cdef class EntityRecognizer(Parser):
2017-05-31 14:42:39 +03:00
name = 'ner'
TransitionSystem = BiluoPushDown
nr_feature = 6
def add_multitask_objective(self, target):
💫 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 target == 'cloze':
cloze = ClozeMultitask(self.vocab)
self._multitasks.append(cloze)
else:
labeller = MultitaskObjective(self.vocab, target=target)
self._multitasks.append(labeller)
def init_multitask_objectives(self, get_gold_tuples, pipeline, sgd=None, **cfg):
for labeller in self._multitasks:
2018-09-13 15:08:55 +03:00
tok2vec = self.model.tok2vec
labeller.begin_training(get_gold_tuples, pipeline=pipeline,
2017-10-27 21:29:08 +03:00
tok2vec=tok2vec)
def __reduce__(self):
2017-10-27 21:29:08 +03:00
return (EntityRecognizer, (self.vocab, self.moves, self.model),
None, None)
2018-11-18 02:06:26 +03:00
@property
def labels(self):
# Get the labels from the model by looking at the available moves, e.g.
# B-PERSON, I-PERSON, L-PERSON, U-PERSON
return [move.split('-')[1] for move in self.move_names
if move[0] in ('B', 'I', 'L', 'U')]
2017-03-15 17:27:41 +03:00
2018-11-18 02:06:13 +03:00
__all__ = ['Tagger', 'DependencyParser', 'EntityRecognizer', 'Tensorizer', 'TextCategorizer']