def main(args):
logger.info('Args = {}'.format(args))
corpus = CorpusLoader().load_corpus(CORPUS_PATH[args.corpus])
tokenizer = TokenizerFactory().tokenizer(args.corpus)
logger.info('Loaded corpus: {}'.format(corpus))
model_dir = train_utils.model_dir(corpus.name, args.run_id)
os.makedirs(model_dir, exist_ok=True)
logger.info('Get sentences...')
docs = list(itertools.chain(corpus.train, corpus.dev))
sents, _ = tagging_utils.standoff_to_sents(docs, tokenizer, verbose=True)
logger.info('Compute features...')
feature_extractor, meta_sentence_filter = crf_util.FEATURE_EXTRACTOR[
args.feature_extractor]
X, y = crf_labeler.sents_to_features_and_labels(sents, feature_extractor)
logger.info('len(X) = {}'.format(len(X)))
logger.info('len(y) = {}'.format(len(y)))
crf = crf_labeler.SentenceFilterCRF(ignore_sentence=meta_sentence_filter,
ignored_label='O',
algorithm='lbfgs',
c1=0.1,
c2=0.1,
max_iterations=100,
all_possible_transitions=True)
logger.info('Start learing curve computation...')
with parallel_backend('multiprocessing'):
# scikit-learn changed the default multiprocessing to 'loky' in 0.21. It appears that this
# is not supported by sklearn_crfsuite. Therefore, we switch to the legacy 'multiprocessing'
# parallel backend.
plot_learning_curve(crf,
'CRF learning curve (sentences: N={})'.format(
len(X)),
X,
y,
out_dir=model_dir,
cv=5,
n_jobs=12)
logger.info('Done...')
def main(args):
logger.info('Args = {}'.format(args))
corpus = CorpusLoader().load_corpus(CORPUS_PATH[args.corpus])
tokenizer = TokenizerFactory().tokenizer(args.corpus)
logger.info('Loaded corpus: {}'.format(corpus))
model_dir = train_utils.model_dir(corpus.name, args.run_id)
os.makedirs(model_dir, exist_ok=True)
logger.info('Get sentences...')
train_sents, train_docs = tagging_utils.standoff_to_sents(corpus.train,
tokenizer,
verbose=True)
dev_sents, dev_docs = tagging_utils.standoff_to_sents(corpus.dev,
tokenizer,
verbose=True)
test_sents, test_docs = tagging_utils.standoff_to_sents(corpus.test,
tokenizer,
verbose=True)
logger.info('Compute features...')
feature_extractor, meta_sentence_filter = crf_util.FEATURE_EXTRACTOR[
args.feature_extractor]
X_train, y_train = crf_labeler.sents_to_features_and_labels(
train_sents, feature_extractor)
X_dev, y_dev = crf_labeler.sents_to_features_and_labels(
dev_sents, feature_extractor)
X_test, y_test = crf_labeler.sents_to_features_and_labels(
test_sents, feature_extractor)
logger.info('len(X_train) = {}'.format(len(X_train)))
logger.info('len(y_train) = {}'.format(len(y_train)))
logger.info('len(X_dev) = {}'.format(len(X_dev)))
logger.info('len(X_test) = {}'.format(len(X_test)))
X_train_combined = X_train + X_dev
y_train_combined = y_train + y_dev
train_indices = [-1] * len(X_train)
dev_indices = [0] * len(X_dev)
test_fold = train_indices + dev_indices
labels = list(set(label for sent in y_train_combined for label in sent))
labels.remove('O')
logger.info('Labels: {}'.format(labels))
f1_scorer = make_scorer(flat_f1_score, labels=labels, average='micro')
crf = crf_labeler.SentenceFilterCRF(ignore_sentence=meta_sentence_filter,
ignored_label='O',
algorithm='lbfgs',
max_iterations=100,
all_possible_transitions=True)
ps = PredefinedSplit(test_fold)
rs = RandomizedSearchCV(crf,
PARAM_SPACE,
cv=ps,
verbose=1,
n_jobs=args.n_jobs,
n_iter=args.n_iter,
scoring=f1_scorer,
return_train_score=True)
logger.info('Start RandomizedSearchCV... {}'.format(crf))
with parallel_backend('multiprocessing'):
rs.fit(X_train_combined, y_train_combined)
logger.info('best params: {}'.format(rs.best_params_))
logger.info('best CV score: {}'.format(rs.best_score_))
logger.info('model size: {:0.2f}M'.format(rs.best_estimator_.size_ /
1000000))
logger.info('Make predictions...')
crf = rs.best_estimator_
y_pred_train = crf.predict(X_train)
y_pred_dev = crf.predict(X_dev)
y_pred_test = crf.predict(X_test)
train_utils.save_predictions(
corpus_name=corpus.name,
run_id=args.run_id,
train=tagging_utils.sents_to_standoff(y_pred_train, train_docs),
dev=tagging_utils.sents_to_standoff(y_pred_dev, dev_docs),
test=tagging_utils.sents_to_standoff(y_pred_test, test_docs))
_save_model_aritfacts(rs, model_dir, y_test, y_pred_test)
test=tagging_utils.sents_to_standoff(y_pred_test, test_docs))
_save_model_aritfacts(rs, model_dir, y_test, y_pred_test)
def arg_parser():
parser = argparse.ArgumentParser()
parser.add_argument("corpus",
choices=CORPUS_PATH.keys(),
help="Corpus identifier.")
parser.add_argument("run_id", help="Run identifier")
parser.add_argument("feature_extractor",
choices=crf_util.FEATURE_EXTRACTOR.keys(),
help="Feature extractor.")
parser.add_argument("--n_iter",
help="Number of random search trials",
default=1,
type=int)
parser.add_argument("--n_jobs",
help="Number of concurrent jobs",
default=1,
type=int)
return parser.parse_args()
if __name__ == '__main__':
ARGS = arg_parser()
ARGS.run_id = 'crf_' + ARGS.run_id
logger.add(
join(train_utils.model_dir(ARGS.corpus, ARGS.run_id), 'training.log'))
main(ARGS)
plot_learning_curve(crf,
'CRF learning curve (sentences: N={})'.format(
len(X)),
X,
y,
out_dir=model_dir,
cv=5,
n_jobs=12)
logger.info('Done...')
def arg_parser():
parser = argparse.ArgumentParser()
parser.add_argument("corpus",
choices=CORPUS_PATH.keys(),
help="Corpus identifier.")
parser.add_argument("run_id", help="Run identifier")
parser.add_argument("feature_extractor",
choices=crf_util.FEATURE_EXTRACTOR.keys(),
help="Feature extractor.")
return parser.parse_args()
if __name__ == '__main__':
ARGS = arg_parser()
ARGS.run_id = 'crf_' + ARGS.run_id
logger.add(
join(train_utils.model_dir(ARGS.corpus, ARGS.run_id),
'learning-curve.log'))
main(ARGS)
def main(args):
logger.info('Args = {}'.format(args))
corpus = CorpusLoader().load_corpus(CORPUS_PATH[args.corpus])
tokenizer = TokenizerFactory().tokenizer(args.corpus)
logger.info('Loaded corpus: {}'.format(corpus))
model_dir = train_utils.model_dir(corpus.name, args.run_id)
os.makedirs(model_dir, exist_ok=True)
logger.info('Get sentences...')
train_sents, train_docs = flair_utils.standoff_to_flair_sents(corpus.train,
tokenizer,
verbose=True)
dev_sents, dev_docs = flair_utils.standoff_to_flair_sents(corpus.dev,
tokenizer,
verbose=True)
test_sents, test_docs = flair_utils.standoff_to_flair_sents(corpus.test,
tokenizer,
verbose=True)
flair_corpus = flair_utils.FilteredCorpus(train=train_sents,
dev=dev_sents,
test=test_sents,
ignore_sentence=_ignore_sentence)
logger.info(flair_corpus)
if args.model_file:
logger.info('Load existing model from {}'.format(args.model_file))
tagger = SequenceTagger.load(args.model_file)
else:
logger.info('Train model...')
tagger = get_model(
flair_corpus,
corpus_name=args.corpus,
embedding_lang=args.embedding_lang,
pooled_contextual_embeddings=args.pooled_contextual_embeddings,
contextual_forward_path=args.contextual_forward_path,
contextual_backward_path=args.contextual_backward_path)
if args.fine_tune or not args.model_file:
trainer = ModelTrainer(tagger, flair_corpus)
trainer.train(join(model_dir, 'flair'),
max_epochs=150,
monitor_train=False,
train_with_dev=args.train_with_dev)
if not args.train_with_dev:
# Model performance is judged by dev data, so we also pick the best performing model
# according to the dev score to make our final predictions.
tagger = SequenceTagger.load(
join(model_dir, 'flair', 'best-model.pt'))
else:
# Training is stopped if train loss converges - here, we do not have a "best model" and
# use the final model to make predictions.
pass
logger.info('Make predictions...')
make_predictions(tagger, flair_corpus)
train_utils.save_predictions(
corpus_name=corpus.name,
run_id=args.run_id,
train=flair_utils.flair_sents_to_standoff(train_sents, train_docs),
dev=flair_utils.flair_sents_to_standoff(dev_sents, dev_docs),
test=flair_utils.flair_sents_to_standoff(test_sents, test_docs))
evaluator.token_level_blind().to_csv(
join(model_dir, 'scores_token_blind.csv'))
logger.info('Done.')
def arg_parser():
parser = argparse.ArgumentParser()
parser.add_argument("corpus",
choices=CORPUS_PATH.keys(),
help="Corpus identifier.")
parser.add_argument("run_id", help="Run identifier")
parser.add_argument("--train_sample_frac",
help="Fraction of the training data to use.",
type=float,
default=0.1)
parser.add_argument("--random_seed",
help="Seed for the training set sampler.",
type=int,
default=42)
return parser.parse_args()
if __name__ == '__main__':
ARGS = arg_parser()
ARGS.run_id = 'bilstmcrf_{}_frac_{}_seed_{}'.format(
ARGS.run_id, ARGS.train_sample_frac, ARGS.random_seed)
MODEL_DIR = train_utils.model_dir(ARGS.corpus + '-subsets', ARGS.run_id)
os.makedirs(MODEL_DIR, exist_ok=True)
logger.add(join(MODEL_DIR, 'training.log'))
main(ARGS, MODEL_DIR)
