def test_fasttext_txt():
"""
Test searching for fasttext and txt files
"""
with tempfile.TemporaryDirectory(dir=f'{TEST_WORKING_DIR}/out') as temp_dir:
# make a fake directory for English word vectors
fasttext_dir = os.path.join(temp_dir, 'fasttext', 'English')
os.makedirs(fasttext_dir)
# make a fake English word vector file
fake_file = os.path.join(fasttext_dir, 'en.vectors.txt')
fout = open(fake_file, 'w')
fout.close()
# get_wordvec_file should now find this fake file
filename = utils.get_wordvec_file(wordvec_dir=temp_dir, shorthand='en_foo')
assert filename == fake_file
def load_pretrain(args):
pretrain = None
if args['pretrain']:
if args['wordvec_pretrain_file']:
pretrain_file = args['wordvec_pretrain_file']
else:
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'],
args['shorthand'])
if os.path.exists(pretrain_file):
vec_file = None
else:
vec_file = args['wordvec_file'] if args[
'wordvec_file'] else utils.get_wordvec_file(
args['wordvec_dir'], args['shorthand'])
pretrain = Pretrain(pretrain_file, vec_file,
args['pretrain_max_vocab'])
return pretrain
def load_pretrain(args):
if args.wordvec_pretrain_file:
pretrain_file = args.wordvec_pretrain_file
elif args.wordvec_type:
pretrain_file = '{}/{}.{}.pretrain.pt'.format(
args.save_dir, args.shorthand, args.wordvec_type.name.lower())
else:
raise Exception(
"TODO: need to get the wv type back from get_wordvec_file")
logger.info("Looking for pretrained vectors in {}".format(pretrain_file))
if os.path.exists(pretrain_file):
vec_file = None
elif args.wordvec_raw_file:
vec_file = args.wordvec_raw_file
logger.info("Pretrain not found. Looking in {}".format(vec_file))
else:
vec_file = utils.get_wordvec_file(args.wordvec_dir, args.shorthand,
args.wordvec_type.name.lower())
logger.info("Pretrain not found. Looking in {}".format(vec_file))
pretrain = Pretrain(pretrain_file, vec_file, args.pretrain_max_vocab)
logger.info("Embedding shape: %s" % str(pretrain.emb.shape))
return pretrain
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 = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_parser.pt'.format(args['save_dir'], args['shorthand'])
# load pretrained vectors if needed
pretrain = None
if args['pretrain']:
vec_file = args['wordvec_file'] if args['wordvec_file'] else utils.get_wordvec_file(args['wordvec_dir'], args['shorthand'])
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'], args['shorthand'])
pretrain = Pretrain(pretrain_file, vec_file, args['pretrain_max_vocab'])
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
train_doc = Document(CoNLL.conll2dict(input_file=args['train_file']))
train_batch = DataLoader(train_doc, args['batch_size'], args, pretrain, evaluation=False)
vocab = train_batch.vocab
dev_doc = Document(CoNLL.conll2dict(input_file=args['eval_file']))
dev_batch = DataLoader(dev_doc, args['batch_size'], args, pretrain, vocab=vocab, evaluation=True, sort_during_eval=True)
# pred and gold path
system_pred_file = args['output_file']
gold_file = args['gold_file']
# skip training if the language does not have training or dev data
if len(train_batch) == 0 or len(dev_batch) == 0:
print("Skip training because no data available...")
sys.exit(0)
print("Training parser...")
trainer = Trainer(args=args, vocab=vocab, pretrain=pretrain, use_cuda=args['cuda'])
global_step = 0
max_steps = args['max_steps']
dev_score_history = []
best_dev_preds = []
current_lr = args['lr']
global_start_time = time.time()
format_str = '{}: step {}/{}, loss = {:.6f} ({:.3f} sec/batch), lr: {:.6f}'
using_amsgrad = False
last_best_step = 0
# start training
train_loss = 0
while True:
do_break = 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
print(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
print("Evaluating on dev set...")
dev_preds = []
for batch in dev_batch:
preds = trainer.predict(batch)
dev_preds += preds
dev_preds = utils.unsort(dev_preds, dev_batch.data_orig_idx)
dev_batch.doc.set([HEAD, DEPREL], [y for x in dev_preds for y in x])
CoNLL.dict2conll(dev_batch.doc.to_dict(), system_pred_file)
_, _, dev_score = scorer.score(system_pred_file, gold_file)
train_loss = train_loss / args['eval_interval'] # avg loss per batch
print("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):
last_best_step = global_step
trainer.save(model_file)
print("new best model saved.")
best_dev_preds = dev_preds
dev_score_history += [dev_score]
print("")
if global_step - last_best_step >= args['max_steps_before_stop']:
if not using_amsgrad:
print("Switching to AMSGrad")
last_best_step = global_step
using_amsgrad = True
trainer.optimizer = optim.Adam(trainer.model.parameters(), amsgrad=True, lr=args['lr'], betas=(.9, args['beta2']), eps=1e-6)
else:
do_break = True
break
if global_step >= args['max_steps']:
do_break = True
break
if do_break: break
train_batch.reshuffle()
print("Training ended with {} steps.".format(global_step))
best_f, best_eval = max(dev_score_history)*100, np.argmax(dev_score_history)+1
print("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']))