def evaluate(args):
# file paths
model_file = os.path.join(args['save_dir'], args['save_name']) if args['save_name'] is not None \
else '{}/{}_nertagger.pt'.format(args['save_dir'], args['shorthand'])
loaded_args, trainer, vocab = load_model(args, model_file)
# load data
logger.info("Loading data with batch size {}...".format(args['batch_size']))
doc = Document(json.load(open(args['eval_file'])))
batch = DataLoader(doc, args['batch_size'], loaded_args, vocab=vocab, evaluation=True)
utils.warn_missing_tags([i for i in trainer.vocab['tag']], batch.tags, "eval_file")
logger.info("Start evaluation...")
preds = []
for i, b in enumerate(batch):
preds += trainer.predict(b)
gold_tags = batch.tags
_, _, score = scorer.score_by_entity(preds, gold_tags)
_, _, _, confusion = scorer.score_by_token(preds, gold_tags)
logger.info("NER tagger score:")
logger.info("{} {:.2f}".format(args['shorthand'], score*100))
logger.info("NER token confusion matrix:\n{}".format(format_confusion(confusion)))
def test_ner_scorer():
pred_sequences = [['O', 'S-LOC', 'O', 'O', 'B-PER', 'E-PER'],
['O', 'S-MISC', 'O', 'E-ORG', 'O', 'B-PER', 'I-PER', 'E-PER']]
gold_sequences = [['O', 'B-LOC', 'E-LOC', 'O', 'B-PER', 'E-PER'],
['O', 'S-MISC', 'B-ORG', 'E-ORG', 'O', 'B-PER', 'E-PER', 'S-LOC']]
token_p, token_r, token_f = score_by_token(pred_sequences, gold_sequences)
assert pytest.approx(token_p, abs=0.00001) == 0.625
assert pytest.approx(token_r, abs=0.00001) == 0.5
assert pytest.approx(token_f, abs=0.00001) == 0.55555
entity_p, entity_r, entity_f = score_by_entity(pred_sequences, gold_sequences)
assert pytest.approx(entity_p, abs=0.00001) == 0.4
assert pytest.approx(entity_r, abs=0.00001) == 0.33333
assert pytest.approx(entity_f, abs=0.00001) == 0.36363
def evaluate(args):
# file paths
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_nertagger.pt'.format(args['save_dir'], args['shorthand'])
# load model
use_cuda = args['cuda'] and not args['cpu']
trainer = Trainer(model_file=model_file, use_cuda=use_cuda)
loaded_args, vocab = trainer.args, trainer.vocab
# load config
for k in args:
if k.endswith('_dir') or k.endswith('_file') or k in [
'shorthand', 'mode', 'scheme'
]:
loaded_args[k] = args[k]
# load data
logger.info("Loading data with batch size {}...".format(
args['batch_size']))
doc = Document(json.load(open(args['eval_file'])))
batch = DataLoader(doc,
args['batch_size'],
loaded_args,
vocab=vocab,
evaluation=True)
logger.info("Start evaluation...")
preds = []
for i, b in enumerate(batch):
preds += trainer.predict(b)
gold_tags = batch.tags
_, _, score = scorer.score_by_entity(preds, gold_tags)
logger.info("NER tagger score:")
logger.info("{} {:.2f}".format(args['shorthand'], score * 100))
def train(args):
utils.ensure_dir(args['save_dir'])
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_nertagger.pt'.format(args['save_dir'], args['shorthand'])
# load pretrained vectors
if len(args['wordvec_file']) == 0:
vec_file = utils.get_wordvec_file(args['wordvec_dir'],
args['shorthand'])
else:
vec_file = args['wordvec_file']
# do not save pretrained embeddings individually
pretrain = Pretrain(None,
vec_file,
args['pretrain_max_vocab'],
save_to_file=False)
if args['charlm']:
if args['charlm_shorthand'] is None:
logger.info(
"CharLM Shorthand is required for loading pretrained CharLM model..."
)
sys.exit(0)
logger.info('Use pretrained contextualized char embedding')
args['charlm_forward_file'] = '{}/{}_forward_charlm.pt'.format(
args['charlm_save_dir'], args['charlm_shorthand'])
args['charlm_backward_file'] = '{}/{}_backward_charlm.pt'.format(
args['charlm_save_dir'], args['charlm_shorthand'])
# load data
logger.info("Loading data with batch size {}...".format(
args['batch_size']))
train_doc = Document(json.load(open(args['train_file'])))
train_batch = DataLoader(train_doc,
args['batch_size'],
args,
pretrain,
evaluation=False)
vocab = train_batch.vocab
dev_doc = Document(json.load(open(args['eval_file'])))
dev_batch = DataLoader(dev_doc,
args['batch_size'],
args,
pretrain,
vocab=vocab,
evaluation=True)
dev_gold_tags = dev_batch.tags
# skip training if the language does not have training or dev data
if len(train_batch) == 0 or len(dev_batch) == 0:
logger.info("Skip training because no data available...")
sys.exit(0)
logger.info("Training tagger...")
trainer = Trainer(args=args,
vocab=vocab,
pretrain=pretrain,
use_cuda=args['cuda'])
logger.info(trainer.model)
global_step = 0
max_steps = args['max_steps']
dev_score_history = []
best_dev_preds = []
current_lr = trainer.optimizer.param_groups[0]['lr']
global_start_time = time.time()
format_str = '{}: step {}/{}, loss = {:.6f} ({:.3f} sec/batch), lr: {:.6f}'
# LR scheduling
if args['lr_decay'] > 0:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(trainer.optimizer, mode='max', factor=args['lr_decay'], \
patience=args['patience'], verbose=True, min_lr=args['min_lr'])
else:
scheduler = None
# start training
train_loss = 0
while True:
should_stop = False
for i, batch in enumerate(train_batch):
start_time = time.time()
global_step += 1
loss = trainer.update(batch, eval=False) # update step
train_loss += loss
if global_step % args['log_step'] == 0:
duration = time.time() - start_time
logger.info(format_str.format(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), global_step,\
max_steps, loss, duration, current_lr))
if global_step % args['eval_interval'] == 0:
# eval on dev
logger.info("Evaluating on dev set...")
dev_preds = []
for batch in dev_batch:
preds = trainer.predict(batch)
dev_preds += preds
_, _, dev_score = scorer.score_by_entity(
dev_preds, dev_gold_tags)
train_loss = train_loss / args[
'eval_interval'] # avg loss per batch
logger.info(
"step {}: train_loss = {:.6f}, dev_score = {:.4f}".format(
global_step, train_loss, dev_score))
train_loss = 0
# save best model
if len(dev_score_history
) == 0 or dev_score > max(dev_score_history):
trainer.save(model_file)
logger.info("New best model saved.")
best_dev_preds = dev_preds
dev_score_history += [dev_score]
logger.info("")
# lr schedule
if scheduler is not None:
scheduler.step(dev_score)
# check stopping
current_lr = trainer.optimizer.param_groups[0]['lr']
if global_step >= args['max_steps'] or current_lr <= args['min_lr']:
should_stop = True
break
if should_stop:
break
train_batch.reshuffle()
logger.info("Training ended with {} steps.".format(global_step))
best_f, best_eval = max(dev_score_history) * 100, np.argmax(
dev_score_history) + 1
logger.info("Best dev F1 = {:.2f}, at iteration = {}".format(
best_f, best_eval * args['eval_interval']))
def train(args):
utils.ensure_dir(args['save_dir'])
model_file = os.path.join(args['save_dir'], args['save_name']) if args['save_name'] is not None \
else '{}/{}_nertagger.pt'.format(args['save_dir'], args['shorthand'])
pretrain = None
vocab = None
trainer = None
if args['finetune'] and args['finetune_load_name']:
logger.warning('Finetune is ON. Using model from "{}"'.format(args['finetune_load_name']))
_, trainer, vocab = load_model(args, args['finetune_load_name'])
elif args['finetune'] and os.path.exists(model_file):
logger.warning('Finetune is ON. Using model from "{}"'.format(model_file))
_, trainer, vocab = load_model(args, model_file)
else:
if args['finetune']:
raise FileNotFoundError('Finetune is set to true but model file is not found: {}'.format(model_file))
# load pretrained vectors
if args['wordvec_pretrain_file']:
pretrain_file = args['wordvec_pretrain_file']
pretrain = Pretrain(pretrain_file, None, args['pretrain_max_vocab'], save_to_file=False)
else:
if len(args['wordvec_file']) == 0:
vec_file = utils.get_wordvec_file(args['wordvec_dir'], args['shorthand'])
else:
vec_file = args['wordvec_file']
# do not save pretrained embeddings individually
pretrain = Pretrain(None, vec_file, args['pretrain_max_vocab'], save_to_file=False)
if pretrain is not None:
word_emb_dim = pretrain.emb.shape[1]
if args['word_emb_dim'] and args['word_emb_dim'] != word_emb_dim:
logger.warning("Embedding file has a dimension of {}. Model will be built with that size instead of {}".format(word_emb_dim, args['word_emb_dim']))
args['word_emb_dim'] = word_emb_dim
if args['charlm']:
if args['charlm_shorthand'] is None:
raise ValueError("CharLM Shorthand is required for loading pretrained CharLM model...")
logger.info('Using pretrained contextualized char embedding')
if not args['charlm_forward_file']:
args['charlm_forward_file'] = '{}/{}_forward_charlm.pt'.format(args['charlm_save_dir'], args['charlm_shorthand'])
if not args['charlm_backward_file']:
args['charlm_backward_file'] = '{}/{}_backward_charlm.pt'.format(args['charlm_save_dir'], args['charlm_shorthand'])
# load data
logger.info("Loading data with batch size {}...".format(args['batch_size']))
train_doc = Document(json.load(open(args['train_file'])))
train_batch = DataLoader(train_doc, args['batch_size'], args, pretrain, vocab=vocab, evaluation=False)
vocab = train_batch.vocab
dev_doc = Document(json.load(open(args['eval_file'])))
dev_batch = DataLoader(dev_doc, args['batch_size'], args, pretrain, vocab=vocab, evaluation=True)
dev_gold_tags = dev_batch.tags
if args['finetune']:
utils.warn_missing_tags([i for i in trainer.vocab['tag']], train_batch.tags, "training set")
utils.warn_missing_tags(train_batch.tags, dev_batch.tags, "dev set")
# skip training if the language does not have training or dev data
if len(train_batch) == 0 or len(dev_batch) == 0:
logger.info("Skip training because no data available...")
sys.exit(0)
logger.info("Training tagger...")
if trainer is None: # init if model was not loaded previously from file
trainer = Trainer(args=args, vocab=vocab, pretrain=pretrain, use_cuda=args['cuda'],
train_classifier_only=args['train_classifier_only'])
logger.info(trainer.model)
global_step = 0
max_steps = args['max_steps']
dev_score_history = []
best_dev_preds = []
current_lr = trainer.optimizer.param_groups[0]['lr']
global_start_time = time.time()
format_str = '{}: step {}/{}, loss = {:.6f} ({:.3f} sec/batch), lr: {:.6f}'
# LR scheduling
if args['lr_decay'] > 0:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(trainer.optimizer, mode='max', factor=args['lr_decay'], \
patience=args['patience'], verbose=True, min_lr=args['min_lr'])
else:
scheduler = None
# start training
train_loss = 0
while True:
should_stop = False
for i, batch in enumerate(train_batch):
start_time = time.time()
global_step += 1
loss = trainer.update(batch, eval=False) # update step
train_loss += loss
if global_step % args['log_step'] == 0:
duration = time.time() - start_time
logger.info(format_str.format(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), global_step,\
max_steps, loss, duration, current_lr))
if global_step % args['eval_interval'] == 0:
# eval on dev
logger.info("Evaluating on dev set...")
dev_preds = []
for batch in dev_batch:
preds = trainer.predict(batch)
dev_preds += preds
_, _, dev_score = scorer.score_by_entity(dev_preds, dev_gold_tags)
train_loss = train_loss / args['eval_interval'] # avg loss per batch
logger.info("step {}: train_loss = {:.6f}, dev_score = {:.4f}".format(global_step, train_loss, dev_score))
train_loss = 0
# save best model
if len(dev_score_history) == 0 or dev_score > max(dev_score_history):
trainer.save(model_file)
logger.info("New best model saved.")
best_dev_preds = dev_preds
dev_score_history += [dev_score]
logger.info("")
# lr schedule
if scheduler is not None:
scheduler.step(dev_score)
# check stopping
current_lr = trainer.optimizer.param_groups[0]['lr']
if global_step >= args['max_steps'] or current_lr <= args['min_lr']:
should_stop = True
break
if should_stop:
break
train_batch.reshuffle()
logger.info("Training ended with {} steps.".format(global_step))
if len(dev_score_history) > 0:
best_f, best_eval = max(dev_score_history)*100, np.argmax(dev_score_history)+1
logger.info("Best dev F1 = {:.2f}, at iteration = {}".format(best_f, best_eval * args['eval_interval']))