def __init__(self, config, use_gpu):
# set up configurations
# get pretrained word vectors
self.pretrain = Pretrain(config['pretrain_path'])
# set up trainer
self.trainer = Trainer(pretrain=self.pretrain,
model_file=config['model_path'],
use_cuda=use_gpu)
self.build_final_config(config)
def _set_up_model(self, config, use_gpu):
# set up trainer
self._args = {
'charlm_forward_file': config['forward_charlm_path'],
'charlm_backward_file': config['backward_charlm_path']
}
self._pretrain = Pretrain(config['pretrain_path'])
self._trainer = Trainer(args=self._args,
pretrain=self.pretrain,
model_file=config['model_path'],
use_cuda=use_gpu)
def evaluate(args):
# file paths
system_pred_file = args['output_file']
gold_file = args['gold_file']
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_tagger.pt'.format(args['save_dir'], args['shorthand'])
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'],
args['save_name'])
# load pretrain
pretrain = Pretrain(pretrain_file)
# load model
print("Loading model from: {}".format(model_file))
use_cuda = args['cuda'] and not args['cpu']
trainer = Trainer(pretrain=pretrain,
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'
] or k == 'mode':
loaded_args[k] = args[k]
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
batch = DataLoader(args['eval_file'],
args['batch_size'],
loaded_args,
pretrain,
vocab=vocab,
evaluation=True)
if len(batch) > 0:
print("Start evaluation...")
preds = []
for i, b in enumerate(batch):
preds += trainer.predict(b)
else:
# skip eval if dev data does not exist
preds = []
# write to file and score
batch.conll.set(['upos', 'xpos', 'feats'], [y for x in preds for y in x])
batch.conll.write_conll(system_pred_file)
if gold_file is not None:
_, _, score = scorer.score(system_pred_file, gold_file)
print("Tagger score:")
print("{} {:.2f}".format(args['shorthand'], score * 100))
def evaluate(args):
model_file = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_lm.pt'.format(args['save_dir'], args['shorthand'])
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'],
args['shorthand'])
# load pretrain
pretrain = Pretrain(pretrain_file)
# load model
use_cuda = args['cuda'] and not args['cpu']
trainer = Trainer(pretrain=pretrain,
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'
] or k == 'mode':
loaded_args[k] = args[k]
# load data
print("Loading data with batch size {}...".format(args['eval_batch_size']))
batch = DataLoader(args['eval_file'],
args['eval_batch_size'],
loaded_args,
pretrain,
vocab=vocab,
evaluation=True)
loss = sum([trainer.update(b, eval=True) for b in batch]) / len(batch)
print('Test ppl = {:.6f}'.format(np.exp(loss)))
if args['output_file'] is not None:
preds = []
for b in batch:
preds += trainer.predict(b)
with open(args['output_file'], 'w') as fout:
for sent in preds:
fout.write(
' '.join([vocab['word'].id2unit(wid)
for wid in sent]) + '\n')
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 '{}/{}_tagger.pt'.format(args['save_dir'], args['shorthand'])
# load pretrained vectors
vec_file = args['wordvec_file']
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'],
args['save_name'])
pretrain = Pretrain(pretrain_file, vec_file, args['pretrain_max_vocab'])
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
train_batch = DataLoader(args['train_file'],
args['batch_size'],
args,
pretrain,
evaluation=False)
vocab = train_batch.vocab
dev_batch = DataLoader(args['eval_file'],
args['batch_size'],
args,
pretrain,
vocab=vocab,
evaluation=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 tagger...")
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}'
if args['adapt_eval_interval']:
args['eval_interval'] = utils.get_adaptive_eval_interval(
dev_batch.num_examples, 2000, args['eval_interval'])
print("Evaluating the model every {} steps...".format(
args['eval_interval']))
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_batch.conll.set(['upos', 'xpos', 'feats'],
[y for x in dev_preds for y in x])
dev_batch.conll.write_conll(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 _set_up_model(self, config, use_gpu):
self._pretrain = Pretrain(config['pretrain_path'])
self._trainer = Trainer(pretrain=self.pretrain,
model_file=config['model_path'],
use_cuda=use_gpu)
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 '{}/{}_lm.pt'.format(args['save_dir'], args['shorthand'])
# load pretrained vectors
vec_file = 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)
# load data
print("Loading data with batch size {}...".format(args['batch_size']))
train_batch = DataLoader(args['train_file'],
args['batch_size'],
args,
pretrain,
evaluation=False)
vocab = train_batch.vocab
# train_dev_batch = DataLoader(args['train_file'], args['batch_size'], args, pretrain, vocab=vocab, evaluation=True)
dev_batch = DataLoader(args['eval_file'],
args['eval_batch_size'],
args,
pretrain,
vocab=vocab,
evaluation=True)
# 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 language model...")
trainer = Trainer(args=args,
vocab=vocab,
pretrain=pretrain,
use_cuda=args['cuda'])
print()
print('Parameters:')
n_param = 0
for p_name, p in trainer.model.named_parameters():
if p.requires_grad == True:
n_param += np.prod(list(p.size()))
print('\t{:10} {}'.format(p_name, p.size()))
print('\tTotal paramamters: {}'.format(n_param))
global_step = 0
max_steps = args['max_steps']
dev_score_history = []
current_lr = args['lr']
global_start_time = time.time()
format_str = '{}: step {}/{}, loss = {:.6f} ({:.3f} sec/batch), ppl = {:.6f}, lr: {:.6f}'
last_best_step = 0
log_loss = 0
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
log_loss += loss
train_loss += loss
if global_step % args['log_step'] == 0:
duration = time.time() - start_time
log_loss /= args['log_step']
print(
format_str.format(
datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
global_step, max_steps, log_loss, duration,
np.exp(log_loss), current_lr))
log_loss = 0
if global_step % args['eval_interval'] == 0:
# eval on dev
print("Evaluating on dev set...")
dev_loss = 0
for batch in dev_batch:
dev_loss += trainer.update(batch, eval=True)
dev_loss /= len(dev_batch)
train_loss = train_loss / args[
'eval_interval'] # avg loss per batch
print("step {}: train_ppl = {:.6f}, dev_ppl = {:.6f}".format(
global_step, np.exp(train_loss), np.exp(dev_loss)))
train_loss = 0
# save best model
if len(dev_score_history
) == 0 or dev_loss < min(dev_score_history):
last_best_step = global_step
trainer.save(model_file)
print("new best model saved.")
dev_score_history.append(dev_loss)
print()
if global_step - last_best_step >= args['max_steps_before_stop']:
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_ppl, best_eval = np.exp(
min(dev_score_history)), np.argmin(dev_score_history) + 1
print("Best dev ppl = {:.2f}, at iteration = {}".format(
best_ppl, best_eval * args['eval_interval']))
def _set_up_model(self, config, use_gpu):
# get pretrained word vectors
self._pretrain = Pretrain(config['pretrain_path'])
# set up trainer
self._trainer = Trainer(pretrain=self.pretrain, model_file=config['model_path'], use_cuda=use_gpu)
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
vec_file = args['wordvec_file']
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'],
args['save_name'])
pretrain = Pretrain(pretrain_file, vec_file, args['pretrain_max_vocab'])
"""
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:
print(
"CharLM Shorthand is required for loading pretrained CharLM model..."
)
sys.exit(0)
print('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
print("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:
print("Skip training because no data available...")
sys.exit(0)
print("Training tagger...")
trainer = Trainer(args=args,
vocab=vocab,
pretrain=pretrain,
use_cuda=args['cuda'])
print(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
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_score = scorer.score_by_entity(
dev_preds, dev_gold_tags)
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):
trainer.save(model_file)
print("New best model saved.")
best_dev_preds = dev_preds
dev_score_history += [dev_score]
print("")
# 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()
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 = args['save_dir'] + '/' + args['save_name'] if args['save_name'] is not None \
else '{}/{}_parser.pt'.format(args['save_dir'], args['shorthand'])
pretrain_file = '{}/{}.pretrain.pt'.format(args['save_dir'],
args['shorthand'])
vec_file = utils.get_wordvec_file(args['wordvec_dir'], args['shorthand'])
pretrain = Pretrain(pretrain_file, vec_file)
use_cuda = args['cuda'] and not args['cpu']
lm_train_batch = LMDataLoader(args['lm_file'],
args['lm_batch_size'],
args,
pretrain,
vocab=None,
evaluation=False,
cutoff=args['vocab_cutoff'])
vocab = lm_train_batch.vocab
dp_train_batch = DPDataLoader(args['train_file'],
args['batch_size'],
args,
pretrain,
vocab=None,
evaluation=False,
cutoff=args['vocab_cutoff'])
vocab['deprel'] = dp_train_batch.vocab['deprel']
dp_train_batch = DPDataLoader(args['train_file'],
args['batch_size'],
args,
pretrain,
vocab=vocab,
evaluation=False,
cutoff=args['vocab_cutoff'])
train_dev_batch = DPDataLoader(args['train_file'],
args['batch_size'],
args,
pretrain,
vocab=vocab,
evaluation=True)
dev_batch = DPDataLoader(args['eval_file'],
args['batch_size'],
args,
pretrain,
vocab=vocab,
evaluation=True)
lm_train_iter = iter(lm_train_batch)
dp_train_iter = iter(dp_train_batch)
# pred and gold path
system_pred_file = args['output_file']
gold_file = args['gold_file']
print("Training parser...")
trainer = Trainer(args=args,
vocab=vocab,
pretrain=pretrain,
use_cuda=args['cuda'],
weight_decay=args['wdecay'])
print()
print('Parameters that require grad:')
for p_name, p in trainer.model.named_parameters():
if p.requires_grad == True:
print('\t{:10} {}'.format(p_name, p.size()))
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), dp_loss = {:.4f}, ppl = {:.2f}, lr: {:.6f}'
using_amsgrad = False
last_best_step = 0
# start training
log_loss = np.zeros(3)
train_loss = np.zeros(3)
while True:
do_break = False
try:
lm_batch = next(lm_train_iter)
except StopIteration:
lm_train_iter = iter(lm_train_batch)
lm_batch = next(lm_train_iter)
try:
dp_batch = next(dp_train_iter)
except StopIteration:
dp_train_iter = iter(dp_train_batch)
dp_batch = next(dp_train_iter)
start_time = time.time()
global_step += 1
dp_loss, lm_loss, loss = trainer.update(dp_batch, lm_batch,
eval=False) # update step
log_loss += np.array([lm_loss, dp_loss, loss])
train_loss += np.array([lm_loss, dp_loss, loss])
if global_step % args['log_step'] == 0:
duration = time.time() - start_time
log_loss = log_loss / args['log_step']
print(
format_str.format(datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
global_step, max_steps,
log_loss[2], duration, log_loss[1],
np.exp(log_loss[0]), current_lr))
log_loss[:] = 0
if global_step % args['eval_interval'] == 0:
# eval on train
train_preds = []
for batch in train_dev_batch:
preds = trainer.predict(batch)
train_preds += preds
train_dev_batch.conll.set(['head', 'deprel'],
[y for x in train_preds for y in x])
train_dev_batch.conll.write_conll(system_pred_file)
_, _, train_score = scorer.score(system_pred_file,
args['train_file'])
# eval on dev
print("Evaluating on dev set...")
dev_preds = []
for batch in dev_batch:
preds = trainer.predict(batch)
dev_preds += preds
dev_batch.conll.set(['head', 'deprel'],
[y for x in dev_preds for y in x])
dev_batch.conll.write_conll(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_score = {:.4f}, dev_score = {:.4f}".format(
global_step, train_score, 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']))