spaCy/spacy/cli/ud_train.py

415 lines
15 KiB
Python
Raw Normal View History

'''Train for CONLL 2017 UD treebank evaluation. Takes .conllu files, writes
.conllu format for development data, allowing the official scorer to be used.
'''
from __future__ import unicode_literals
import plac
import tqdm
from pathlib import Path
import re
import sys
import json
import spacy
import spacy.util
from ..tokens import Token, Doc
from ..gold import GoldParse
Improve label management in parser and NER (#2108) This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly. Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable. We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense. To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort. Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training. To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make. Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths. This is a squash merge, as I made a lot of very small commits. Individual commit messages below. * Simplify label management for TransitionSystem and its subclasses * Fix serialization for new label handling format in parser * Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir * Set actions in transition system * Require thinc 6.11.1.dev4 * Fix error in parser init * Add unicode declaration * Fix unicode declaration * Update textcat test * Try to get model training on less memory * Print json loc for now * Try rapidjson to reduce memory use * Remove rapidjson requirement * Try rapidjson for reduced mem usage * Handle None heads when projectivising * Stream json docs * Fix train script * Handle projectivity in GoldParse * Fix projectivity handling * Add minibatch_by_words util from ud_train * Minibatch by number of words in spacy.cli.train * Move minibatch_by_words util to spacy.util * Fix label handling * More hacking at label management in parser * Fix encoding in msgpack serialization in GoldParse * Adjust batch sizes in parser training * Fix minibatch_by_words * Add merge_subtokens function to pipeline.pyx * Register merge_subtokens factory * Restore use of msgpack tmp directory * Use minibatch-by-words in train * Handle retokenization in scorer * Change back-off approach for missing labels. Use 'dep' label * Update NER for new label management * Set NER tags for over-segmented words * Fix label alignment in gold * Fix label back-off for infrequent labels * Fix int type in labels dict key * Fix int type in labels dict key * Update feature definition for 8 feature set * Update ud-train script for new label stuff * Fix json streamer * Print the line number if conll eval fails * Update children and sentence boundaries after deprojectivisation * Export set_children_from_heads from doc.pxd * Render parses during UD training * Remove print statement * Require thinc 6.11.1.dev6. Try adding wheel as install_requires * Set different dev version, to flush pip cache * Update thinc version * Update GoldCorpus docs * Remove print statements * Fix formatting and links [ci skip]
2018-03-19 04:58:08 +03:00
from ..util import compounding, minibatch_by_words
from ..syntax.nonproj import projectivize
from ..matcher import Matcher
Improve label management in parser and NER (#2108) This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly. Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable. We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense. To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort. Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training. To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make. Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths. This is a squash merge, as I made a lot of very small commits. Individual commit messages below. * Simplify label management for TransitionSystem and its subclasses * Fix serialization for new label handling format in parser * Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir * Set actions in transition system * Require thinc 6.11.1.dev4 * Fix error in parser init * Add unicode declaration * Fix unicode declaration * Update textcat test * Try to get model training on less memory * Print json loc for now * Try rapidjson to reduce memory use * Remove rapidjson requirement * Try rapidjson for reduced mem usage * Handle None heads when projectivising * Stream json docs * Fix train script * Handle projectivity in GoldParse * Fix projectivity handling * Add minibatch_by_words util from ud_train * Minibatch by number of words in spacy.cli.train * Move minibatch_by_words util to spacy.util * Fix label handling * More hacking at label management in parser * Fix encoding in msgpack serialization in GoldParse * Adjust batch sizes in parser training * Fix minibatch_by_words * Add merge_subtokens function to pipeline.pyx * Register merge_subtokens factory * Restore use of msgpack tmp directory * Use minibatch-by-words in train * Handle retokenization in scorer * Change back-off approach for missing labels. Use 'dep' label * Update NER for new label management * Set NER tags for over-segmented words * Fix label alignment in gold * Fix label back-off for infrequent labels * Fix int type in labels dict key * Fix int type in labels dict key * Update feature definition for 8 feature set * Update ud-train script for new label stuff * Fix json streamer * Print the line number if conll eval fails * Update children and sentence boundaries after deprojectivisation * Export set_children_from_heads from doc.pxd * Render parses during UD training * Remove print statement * Require thinc 6.11.1.dev6. Try adding wheel as install_requires * Set different dev version, to flush pip cache * Update thinc version * Update GoldCorpus docs * Remove print statements * Fix formatting and links [ci skip]
2018-03-19 04:58:08 +03:00
from .. import displacy
from collections import defaultdict, Counter
from timeit import default_timer as timer
import itertools
import random
import numpy.random
import cytoolz
from . import conll17_ud_eval
from .. import lang
from .. import lang
from ..lang import zh
from ..lang import ja
################
# Data reading #
################
space_re = re.compile('\s+')
def split_text(text):
return [space_re.sub(' ', par.strip()) for par in text.split('\n\n')]
def read_data(nlp, conllu_file, text_file, raw_text=True, oracle_segments=False,
max_doc_length=None, limit=None):
'''Read the CONLLU format into (Doc, GoldParse) tuples. If raw_text=True,
include Doc objects created using nlp.make_doc and then aligned against
the gold-standard sequences. If oracle_segments=True, include Doc objects
created from the gold-standard segments. At least one must be True.'''
if not raw_text and not oracle_segments:
raise ValueError("At least one of raw_text or oracle_segments must be True")
paragraphs = split_text(text_file.read())
conllu = read_conllu(conllu_file)
# sd is spacy doc; cd is conllu doc
# cs is conllu sent, ct is conllu token
docs = []
golds = []
for doc_id, (text, cd) in enumerate(zip(paragraphs, conllu)):
sent_annots = []
for cs in cd:
sent = defaultdict(list)
for id_, word, lemma, pos, tag, morph, head, dep, _, space_after in cs:
if '.' in id_:
continue
if '-' in id_:
continue
id_ = int(id_)-1
head = int(head)-1 if head != '0' else id_
sent['words'].append(word)
sent['tags'].append(tag)
sent['heads'].append(head)
sent['deps'].append('ROOT' if dep == 'root' else dep)
sent['spaces'].append(space_after == '_')
sent['entities'] = ['-'] * len(sent['words'])
sent['heads'], sent['deps'] = projectivize(sent['heads'],
sent['deps'])
if oracle_segments:
docs.append(Doc(nlp.vocab, words=sent['words'], spaces=sent['spaces']))
golds.append(GoldParse(docs[-1], **sent))
sent_annots.append(sent)
if raw_text and max_doc_length and len(sent_annots) >= max_doc_length:
doc, gold = _make_gold(nlp, None, sent_annots)
sent_annots = []
docs.append(doc)
golds.append(gold)
if limit and len(docs) >= limit:
return docs, golds
if raw_text and sent_annots:
doc, gold = _make_gold(nlp, None, sent_annots)
docs.append(doc)
golds.append(gold)
if limit and len(docs) >= limit:
return docs, golds
return docs, golds
def read_conllu(file_):
docs = []
sent = []
doc = []
for line in file_:
if line.startswith('# newdoc'):
if doc:
docs.append(doc)
doc = []
elif line.startswith('#'):
continue
elif not line.strip():
if sent:
doc.append(sent)
sent = []
else:
sent.append(list(line.strip().split('\t')))
if len(sent[-1]) != 10:
print(repr(line))
raise ValueError
if sent:
doc.append(sent)
if doc:
docs.append(doc)
return docs
2018-05-07 16:52:47 +03:00
def _make_gold(nlp, text, sent_annots, drop_deps=0.0):
# Flatten the conll annotations, and adjust the head indices
flat = defaultdict(list)
2018-05-07 16:52:47 +03:00
sent_starts = []
for sent in sent_annots:
flat['heads'].extend(len(flat['words'])+head for head in sent['heads'])
for field in ['words', 'tags', 'deps', 'entities', 'spaces']:
flat[field].extend(sent[field])
2018-05-07 16:52:47 +03:00
sent_starts.append(True)
sent_starts.extend([False] * (len(sent['words'])-1))
# Construct text if necessary
assert len(flat['words']) == len(flat['spaces'])
if text is None:
text = ''.join(word+' '*space for word, space in zip(flat['words'], flat['spaces']))
doc = nlp.make_doc(text)
flat.pop('spaces')
gold = GoldParse(doc, **flat)
2018-05-07 16:52:47 +03:00
gold.sent_starts = sent_starts
for i in range(len(gold.heads)):
if random.random() < drop_deps:
gold.heads[i] = None
gold.labels[i] = None
return doc, gold
#############################
# Data transforms for spaCy #
#############################
def golds_to_gold_tuples(docs, golds):
'''Get out the annoying 'tuples' format used by begin_training, given the
GoldParse objects.'''
tuples = []
for doc, gold in zip(docs, golds):
text = doc.text
ids, words, tags, heads, labels, iob = zip(*gold.orig_annot)
sents = [((ids, words, tags, heads, labels, iob), [])]
tuples.append((text, sents))
return tuples
##############
# Evaluation #
##############
def evaluate(nlp, text_loc, gold_loc, sys_loc, limit=None):
if text_loc.parts[-1].endswith('.conllu'):
docs = []
with text_loc.open() as file_:
for conllu_doc in read_conllu(file_):
for conllu_sent in conllu_doc:
words = [line[1] for line in conllu_sent]
docs.append(Doc(nlp.vocab, words=words))
for name, component in nlp.pipeline:
docs = list(component.pipe(docs))
else:
with text_loc.open('r', encoding='utf8') as text_file:
texts = split_text(text_file.read())
docs = list(nlp.pipe(texts))
with sys_loc.open('w', encoding='utf8') as out_file:
write_conllu(docs, out_file)
with gold_loc.open('r', encoding='utf8') as gold_file:
gold_ud = conll17_ud_eval.load_conllu(gold_file)
with sys_loc.open('r', encoding='utf8') as sys_file:
sys_ud = conll17_ud_eval.load_conllu(sys_file)
scores = conll17_ud_eval.evaluate(gold_ud, sys_ud)
Improve label management in parser and NER (#2108) This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly. Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable. We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense. To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort. Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training. To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make. Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths. This is a squash merge, as I made a lot of very small commits. Individual commit messages below. * Simplify label management for TransitionSystem and its subclasses * Fix serialization for new label handling format in parser * Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir * Set actions in transition system * Require thinc 6.11.1.dev4 * Fix error in parser init * Add unicode declaration * Fix unicode declaration * Update textcat test * Try to get model training on less memory * Print json loc for now * Try rapidjson to reduce memory use * Remove rapidjson requirement * Try rapidjson for reduced mem usage * Handle None heads when projectivising * Stream json docs * Fix train script * Handle projectivity in GoldParse * Fix projectivity handling * Add minibatch_by_words util from ud_train * Minibatch by number of words in spacy.cli.train * Move minibatch_by_words util to spacy.util * Fix label handling * More hacking at label management in parser * Fix encoding in msgpack serialization in GoldParse * Adjust batch sizes in parser training * Fix minibatch_by_words * Add merge_subtokens function to pipeline.pyx * Register merge_subtokens factory * Restore use of msgpack tmp directory * Use minibatch-by-words in train * Handle retokenization in scorer * Change back-off approach for missing labels. Use 'dep' label * Update NER for new label management * Set NER tags for over-segmented words * Fix label alignment in gold * Fix label back-off for infrequent labels * Fix int type in labels dict key * Fix int type in labels dict key * Update feature definition for 8 feature set * Update ud-train script for new label stuff * Fix json streamer * Print the line number if conll eval fails * Update children and sentence boundaries after deprojectivisation * Export set_children_from_heads from doc.pxd * Render parses during UD training * Remove print statement * Require thinc 6.11.1.dev6. Try adding wheel as install_requires * Set different dev version, to flush pip cache * Update thinc version * Update GoldCorpus docs * Remove print statements * Fix formatting and links [ci skip]
2018-03-19 04:58:08 +03:00
return docs, scores
def write_conllu(docs, file_):
merger = Matcher(docs[0].vocab)
merger.add('SUBTOK', None, [{'DEP': 'subtok', 'op': '+'}])
for i, doc in enumerate(docs):
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)
file_.write("# newdoc id = {i}\n".format(i=i))
for j, sent in enumerate(doc.sents):
file_.write("# sent_id = {i}.{j}\n".format(i=i, j=j))
file_.write("# text = {text}\n".format(text=sent.text))
for k, token in enumerate(sent):
file_.write(token._.get_conllu_lines(k) + '\n')
file_.write('\n')
def print_progress(itn, losses, ud_scores):
fields = {
'dep_loss': losses.get('parser', 0.0),
'tag_loss': losses.get('tagger', 0.0),
'words': ud_scores['Words'].f1 * 100,
'sents': ud_scores['Sentences'].f1 * 100,
'tags': ud_scores['XPOS'].f1 * 100,
'uas': ud_scores['UAS'].f1 * 100,
'las': ud_scores['LAS'].f1 * 100,
}
header = ['Epoch', 'Loss', 'LAS', 'UAS', 'TAG', 'SENT', 'WORD']
if itn == 0:
print('\t'.join(header))
tpl = '\t'.join((
'{:d}',
'{dep_loss:.1f}',
'{las:.1f}',
'{uas:.1f}',
'{tags:.1f}',
'{sents:.1f}',
'{words:.1f}',
))
print(tpl.format(itn, **fields))
#def get_sent_conllu(sent, sent_id):
# lines = ["# sent_id = {sent_id}".format(sent_id=sent_id)]
def get_token_conllu(token, i):
if token._.begins_fused:
n = 1
while token.nbor(n)._.inside_fused:
n += 1
id_ = '%d-%d' % (i, i+n)
lines = [id_, token.text, '_', '_', '_', '_', '_', '_', '_', '_']
else:
lines = []
if token.head.i == token.i:
head = 0
else:
head = i + (token.head.i - token.i) + 1
fields = [str(i+1), token.text, token.lemma_, token.pos_, token.tag_, '_',
str(head), token.dep_.lower(), '_', '_']
lines.append('\t'.join(fields))
return '\n'.join(lines)
Token.set_extension('get_conllu_lines', method=get_token_conllu)
Token.set_extension('begins_fused', default=False)
Token.set_extension('inside_fused', default=False)
##################
# Initialization #
##################
2018-04-29 16:49:32 +03:00
def load_nlp(corpus, config, vectors=None):
lang = corpus.split('_')[0]
nlp = spacy.blank(lang)
if config.vectors:
2018-04-29 16:49:32 +03:00
if not vectors:
raise ValueError("config asks for vectors, but no vectors "
"directory set on command line (use -v)")
if (Path(vectors) / corpus).exists():
nlp.vocab.from_disk(Path(vectors) / corpus / 'vocab')
nlp.meta['treebank'] = corpus
return nlp
2018-04-29 16:49:32 +03:00
2018-03-27 12:53:35 +03:00
def initialize_pipeline(nlp, docs, golds, config, device):
nlp.add_pipe(nlp.create_pipe('tagger'))
nlp.add_pipe(nlp.create_pipe('parser'))
if config.multitask_tag:
nlp.parser.add_multitask_objective('tag')
if config.multitask_sent:
nlp.parser.add_multitask_objective('sent_start')
for gold in golds:
for tag in gold.tags:
if tag is not None:
nlp.tagger.add_label(tag)
return nlp.begin_training(
lambda: golds_to_gold_tuples(docs, golds), device=device,
2018-08-27 03:12:39 +03:00
subword_features=config.subword_features, conv_depth=config.conv_depth)
########################
# Command line helpers #
########################
class Config(object):
2018-03-11 02:59:39 +03:00
def __init__(self, vectors=None, max_doc_length=10, multitask_tag=True,
2018-04-29 16:49:32 +03:00
multitask_sent=True, multitask_dep=True, multitask_vectors=False,
nr_epoch=30, batch_size=1000, dropout=0.2,
conv_depth=4, subword_features=True):
2018-03-11 03:26:45 +03:00
for key, value in locals().items():
2018-03-11 02:59:39 +03:00
setattr(self, key, value)
2018-04-29 16:49:32 +03:00
@classmethod
def load(cls, loc):
with Path(loc).open('r', encoding='utf8') as file_:
cfg = json.load(file_)
return cls(**cfg)
class Dataset(object):
def __init__(self, path, section):
self.path = path
self.section = section
self.conllu = None
self.text = None
for file_path in self.path.iterdir():
name = file_path.parts[-1]
if section in name and name.endswith('conllu'):
self.conllu = file_path
elif section in name and name.endswith('txt'):
self.text = file_path
if self.conllu is None:
msg = "Could not find .txt file in {path} for {section}"
raise IOError(msg.format(section=section, path=path))
if self.text is None:
msg = "Could not find .txt file in {path} for {section}"
self.lang = self.conllu.parts[-1].split('-')[0].split('_')[0]
class TreebankPaths(object):
def __init__(self, ud_path, treebank, **cfg):
self.train = Dataset(ud_path / treebank, 'train')
self.dev = Dataset(ud_path / treebank, 'dev')
self.lang = self.train.lang
@plac.annotations(
ud_dir=("Path to Universal Dependencies corpus", "positional", None, Path),
corpus=("UD corpus to train and evaluate on, e.g. en, es_ancora, etc",
"positional", None, str),
parses_dir=("Directory to write the development parses", "positional", None, Path),
2018-03-11 03:26:45 +03:00
config=("Path to json formatted config file", "positional"),
2018-03-27 12:53:35 +03:00
limit=("Size limit", "option", "n", int),
2018-04-29 16:49:32 +03:00
use_gpu=("Use GPU", "option", "g", int),
use_oracle_segments=("Use oracle segments", "flag", "G", int),
2018-04-29 16:49:32 +03:00
vectors_dir=("Path to directory with pre-trained vectors, named e.g. en/",
"option", "v", Path),
)
def main(ud_dir, parses_dir, config, corpus, limit=0, use_gpu=-1, vectors_dir=None,
use_oracle_segments=False):
2018-03-27 12:53:35 +03:00
spacy.util.fix_random_seed()
lang.zh.Chinese.Defaults.use_jieba = False
lang.ja.Japanese.Defaults.use_janome = False
2018-03-11 03:26:45 +03:00
config = Config.load(config)
paths = TreebankPaths(ud_dir, corpus)
if not (parses_dir / corpus).exists():
(parses_dir / corpus).mkdir()
print("Train and evaluate", corpus, "using lang", paths.lang)
2018-04-29 16:49:32 +03:00
nlp = load_nlp(paths.lang, config, vectors=vectors_dir)
docs, golds = read_data(nlp, paths.train.conllu.open(), paths.train.text.open(),
max_doc_length=3, limit=limit)
2018-03-27 12:53:35 +03:00
optimizer = initialize_pipeline(nlp, docs, golds, config, use_gpu)
2018-03-27 12:53:35 +03:00
batch_sizes = compounding(config.batch_size//10, config.batch_size, 1.001)
2018-05-16 21:11:59 +03:00
beam_prob = compounding(0.2, 0.8, 1.001)
for i in range(config.nr_epoch):
docs, golds = read_data(nlp, paths.train.conllu.open(), paths.train.text.open(),
max_doc_length=config.max_doc_length, limit=limit,
oracle_segments=use_oracle_segments,
raw_text=not use_oracle_segments)
Improve label management in parser and NER (#2108) This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly. Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable. We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense. To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort. Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training. To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make. Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths. This is a squash merge, as I made a lot of very small commits. Individual commit messages below. * Simplify label management for TransitionSystem and its subclasses * Fix serialization for new label handling format in parser * Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir * Set actions in transition system * Require thinc 6.11.1.dev4 * Fix error in parser init * Add unicode declaration * Fix unicode declaration * Update textcat test * Try to get model training on less memory * Print json loc for now * Try rapidjson to reduce memory use * Remove rapidjson requirement * Try rapidjson for reduced mem usage * Handle None heads when projectivising * Stream json docs * Fix train script * Handle projectivity in GoldParse * Fix projectivity handling * Add minibatch_by_words util from ud_train * Minibatch by number of words in spacy.cli.train * Move minibatch_by_words util to spacy.util * Fix label handling * More hacking at label management in parser * Fix encoding in msgpack serialization in GoldParse * Adjust batch sizes in parser training * Fix minibatch_by_words * Add merge_subtokens function to pipeline.pyx * Register merge_subtokens factory * Restore use of msgpack tmp directory * Use minibatch-by-words in train * Handle retokenization in scorer * Change back-off approach for missing labels. Use 'dep' label * Update NER for new label management * Set NER tags for over-segmented words * Fix label alignment in gold * Fix label back-off for infrequent labels * Fix int type in labels dict key * Fix int type in labels dict key * Update feature definition for 8 feature set * Update ud-train script for new label stuff * Fix json streamer * Print the line number if conll eval fails * Update children and sentence boundaries after deprojectivisation * Export set_children_from_heads from doc.pxd * Render parses during UD training * Remove print statement * Require thinc 6.11.1.dev6. Try adding wheel as install_requires * Set different dev version, to flush pip cache * Update thinc version * Update GoldCorpus docs * Remove print statements * Fix formatting and links [ci skip]
2018-03-19 04:58:08 +03:00
Xs = list(zip(docs, golds))
random.shuffle(Xs)
batches = minibatch_by_words(Xs, size=batch_sizes)
losses = {}
n_train_words = sum(len(doc) for doc in docs)
with tqdm.tqdm(total=n_train_words, leave=False) as pbar:
for batch in batches:
batch_docs, batch_gold = zip(*batch)
pbar.update(sum(len(doc) for doc in batch_docs))
2018-05-16 21:11:59 +03:00
nlp.parser.cfg['beam_update_prob'] = next(beam_prob)
nlp.update(batch_docs, batch_gold, sgd=optimizer,
drop=config.dropout, losses=losses)
out_path = parses_dir / corpus / 'epoch-{i}.conllu'.format(i=i)
with nlp.use_params(optimizer.averages):
if use_oracle_segments:
parsed_docs, scores = evaluate(nlp, paths.dev.conllu,
paths.dev.conllu, out_path)
else:
parsed_docs, scores = evaluate(nlp, paths.dev.text,
paths.dev.conllu, out_path)
print_progress(i, losses, scores)
Improve label management in parser and NER (#2108) This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly. Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable. We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense. To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort. Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training. To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make. Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths. This is a squash merge, as I made a lot of very small commits. Individual commit messages below. * Simplify label management for TransitionSystem and its subclasses * Fix serialization for new label handling format in parser * Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir * Set actions in transition system * Require thinc 6.11.1.dev4 * Fix error in parser init * Add unicode declaration * Fix unicode declaration * Update textcat test * Try to get model training on less memory * Print json loc for now * Try rapidjson to reduce memory use * Remove rapidjson requirement * Try rapidjson for reduced mem usage * Handle None heads when projectivising * Stream json docs * Fix train script * Handle projectivity in GoldParse * Fix projectivity handling * Add minibatch_by_words util from ud_train * Minibatch by number of words in spacy.cli.train * Move minibatch_by_words util to spacy.util * Fix label handling * More hacking at label management in parser * Fix encoding in msgpack serialization in GoldParse * Adjust batch sizes in parser training * Fix minibatch_by_words * Add merge_subtokens function to pipeline.pyx * Register merge_subtokens factory * Restore use of msgpack tmp directory * Use minibatch-by-words in train * Handle retokenization in scorer * Change back-off approach for missing labels. Use 'dep' label * Update NER for new label management * Set NER tags for over-segmented words * Fix label alignment in gold * Fix label back-off for infrequent labels * Fix int type in labels dict key * Fix int type in labels dict key * Update feature definition for 8 feature set * Update ud-train script for new label stuff * Fix json streamer * Print the line number if conll eval fails * Update children and sentence boundaries after deprojectivisation * Export set_children_from_heads from doc.pxd * Render parses during UD training * Remove print statement * Require thinc 6.11.1.dev6. Try adding wheel as install_requires * Set different dev version, to flush pip cache * Update thinc version * Update GoldCorpus docs * Remove print statements * Fix formatting and links [ci skip]
2018-03-19 04:58:08 +03:00
_render_parses(i, parsed_docs[:50])
def _render_parses(i, to_render):
to_render[0].user_data['title'] = "Batch %d" % i
with Path('/tmp/parses.html').open('w') as file_:
html = displacy.render(to_render[:5], style='dep', page=True)
file_.write(html)
if __name__ == '__main__':
plac.call(main)