def main():
parser = argparse.ArgumentParser(
description='NER with bi-directional RNN-CNN')
parser.add_argument('--config',
type=str,
help='config file',
required=True)
parser.add_argument('--num_epochs',
type=int,
default=100,
help='Number of training epochs')
parser.add_argument('--batch_size',
type=int,
default=16,
help='Number of sentences in each batch')
parser.add_argument('--loss_type',
choices=['sentence', 'token'],
default='sentence',
help='loss type (default: sentence)')
parser.add_argument('--optim',
choices=['sgd', 'adam'],
help='type of optimizer',
required=True)
parser.add_argument('--learning_rate',
type=float,
default=0.1,
help='Learning rate')
parser.add_argument('--lr_decay',
type=float,
default=0.999995,
help='Decay rate of learning rate')
parser.add_argument('--amsgrad', action='store_true', help='AMS Grad')
parser.add_argument('--grad_clip',
type=float,
default=0,
help='max norm for gradient clip (default 0: no clip')
parser.add_argument('--warmup_steps',
type=int,
default=0,
metavar='N',
help='number of steps to warm up (default: 0)')
parser.add_argument('--weight_decay',
type=float,
default=0.0,
help='weight for l2 norm decay')
parser.add_argument('--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
parser.add_argument('--embedding',
choices=['glove', 'senna', 'sskip', 'polyglot'],
help='Embedding for words',
required=True)
parser.add_argument('--embedding_dict', help='path for embedding dict')
parser.add_argument('--train',
help='path for training file.',
required=True)
parser.add_argument('--dev', help='path for dev file.', required=True)
parser.add_argument('--test', help='path for test file.', required=True)
parser.add_argument('--model_path',
help='path for saving model file.',
required=True)
args = parser.parse_args()
logger = get_logger("NER")
args.cuda = torch.cuda.is_available()
device = torch.device('cuda', 0) if args.cuda else torch.device('cpu')
train_path = args.train
dev_path = args.dev
test_path = args.test
num_epochs = args.num_epochs
batch_size = args.batch_size
optim = args.optim
learning_rate = args.learning_rate
lr_decay = args.lr_decay
amsgrad = args.amsgrad
warmup_steps = args.warmup_steps
weight_decay = args.weight_decay
grad_clip = args.grad_clip
loss_ty_token = args.loss_type == 'token'
unk_replace = args.unk_replace
model_path = args.model_path
model_name = os.path.join(model_path, 'model.pt')
embedding = args.embedding
embedding_path = args.embedding_dict
print(args)
embedd_dict, embedd_dim = utils.load_embedding_dict(
embedding, embedding_path)
logger.info("Creating Alphabets")
alphabet_path = os.path.join(model_path, 'alphabets')
word_alphabet, char_alphabet, pos_alphabet, chunk_alphabet, ner_alphabet = conll03_data.create_alphabets(
alphabet_path,
train_path,
data_paths=[dev_path, test_path],
embedd_dict=embedd_dict,
max_vocabulary_size=50000)
logger.info("Word Alphabet Size: %d" % word_alphabet.size())
logger.info("Character Alphabet Size: %d" % char_alphabet.size())
logger.info("POS Alphabet Size: %d" % pos_alphabet.size())
logger.info("Chunk Alphabet Size: %d" % chunk_alphabet.size())
logger.info("NER Alphabet Size: %d" % ner_alphabet.size())
logger.info("Reading Data")
data_train = conll03_data.read_bucketed_data(train_path, word_alphabet,
char_alphabet, pos_alphabet,
chunk_alphabet, ner_alphabet)
num_data = sum(data_train[1])
num_labels = ner_alphabet.size()
data_dev = conll03_data.read_data(dev_path, word_alphabet, char_alphabet,
pos_alphabet, chunk_alphabet,
ner_alphabet)
data_test = conll03_data.read_data(test_path, word_alphabet, char_alphabet,
pos_alphabet, chunk_alphabet,
ner_alphabet)
writer = CoNLL03Writer(word_alphabet, char_alphabet, pos_alphabet,
chunk_alphabet, ner_alphabet)
def construct_word_embedding_table():
scale = np.sqrt(3.0 / embedd_dim)
table = np.empty([word_alphabet.size(), embedd_dim], dtype=np.float32)
table[conll03_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, embedd_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in embedd_dict:
embedding = embedd_dict[word]
elif word.lower() in embedd_dict:
embedding = embedd_dict[word.lower()]
else:
embedding = np.random.uniform(
-scale, scale, [1, embedd_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('oov: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
logger.info("constructing network...")
hyps = json.load(open(args.config, 'r'))
json.dump(hyps,
open(os.path.join(model_path, 'config.json'), 'w'),
indent=2)
dropout = hyps['dropout']
crf = hyps['crf']
bigram = hyps['bigram']
assert embedd_dim == hyps['embedd_dim']
char_dim = hyps['char_dim']
mode = hyps['rnn_mode']
hidden_size = hyps['hidden_size']
out_features = hyps['out_features']
num_layers = hyps['num_layers']
p_in = hyps['p_in']
p_out = hyps['p_out']
p_rnn = hyps['p_rnn']
activation = hyps['activation']
if dropout == 'std':
if crf:
network = BiRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
mode,
hidden_size,
out_features,
num_layers,
num_labels,
embedd_word=word_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
activation=activation)
else:
network = BiRecurrentConv(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
mode,
hidden_size,
out_features,
num_layers,
num_labels,
embedd_word=word_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
activation=activation)
elif dropout == 'variational':
if crf:
network = BiVarRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
mode,
hidden_size,
out_features,
num_layers,
num_labels,
embedd_word=word_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
activation=activation)
else:
network = BiVarRecurrentConv(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
mode,
hidden_size,
out_features,
num_layers,
num_labels,
embedd_word=word_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
activation=activation)
else:
raise ValueError('Unkown dropout type: {}'.format(dropout))
network = network.to(device)
optimizer, scheduler = get_optimizer(network.parameters(), optim,
learning_rate, lr_decay, amsgrad,
weight_decay, warmup_steps)
model = "{}-CNN{}".format(mode, "-CRF" if crf else "")
logger.info("Network: %s, num_layer=%d, hidden=%d, act=%s" %
(model, num_layers, hidden_size, activation))
logger.info(
"training: l2: %f, (#training data: %d, batch: %d, unk replace: %.2f)"
% (weight_decay, num_data, batch_size, unk_replace))
logger.info("dropout(in, out, rnn): %s(%.2f, %.2f, %s)" %
(dropout, p_in, p_out, p_rnn))
print('# of Parameters: %d' %
(sum([param.numel() for param in network.parameters()])))
best_f1 = 0.0
best_acc = 0.0
best_precision = 0.0
best_recall = 0.0
test_f1 = 0.0
test_acc = 0.0
test_precision = 0.0
test_recall = 0.0
best_epoch = 0
patient = 0
num_batches = num_data // batch_size + 1
result_path = os.path.join(model_path, 'tmp')
if not os.path.exists(result_path):
os.makedirs(result_path)
for epoch in range(1, num_epochs + 1):
start_time = time.time()
train_loss = 0.
num_insts = 0
num_words = 0
num_back = 0
network.train()
lr = scheduler.get_lr()[0]
print('Epoch %d (%s, lr=%.6f, lr decay=%.6f, amsgrad=%s, l2=%.1e): ' %
(epoch, optim, lr, lr_decay, amsgrad, weight_decay))
if args.cuda:
torch.cuda.empty_cache()
gc.collect()
for step, data in enumerate(
iterate_data(data_train,
batch_size,
bucketed=True,
unk_replace=unk_replace,
shuffle=True)):
optimizer.zero_grad()
words = data['WORD'].to(device)
chars = data['CHAR'].to(device)
labels = data['NER'].to(device)
masks = data['MASK'].to(device)
nbatch = words.size(0)
nwords = masks.sum().item()
loss_total = network.loss(words, chars, labels, mask=masks).sum()
if loss_ty_token:
loss = loss_total.div(nwords)
else:
loss = loss_total.div(nbatch)
loss.backward()
if grad_clip > 0:
clip_grad_norm_(network.parameters(), grad_clip)
optimizer.step()
scheduler.step()
with torch.no_grad():
num_insts += nbatch
num_words += nwords
train_loss += loss_total.item()
# update log
if step % 100 == 0:
torch.cuda.empty_cache()
sys.stdout.write("\b" * num_back)
sys.stdout.write(" " * num_back)
sys.stdout.write("\b" * num_back)
curr_lr = scheduler.get_lr()[0]
log_info = '[%d/%d (%.0f%%) lr=%.6f] loss: %.4f (%.4f)' % (
step, num_batches, 100. * step / num_batches, curr_lr,
train_loss / num_insts, train_loss / num_words)
sys.stdout.write(log_info)
sys.stdout.flush()
num_back = len(log_info)
sys.stdout.write("\b" * num_back)
sys.stdout.write(" " * num_back)
sys.stdout.write("\b" * num_back)
print('total: %d (%d), loss: %.4f (%.4f), time: %.2fs' %
(num_insts, num_words, train_loss / num_insts,
train_loss / num_words, time.time() - start_time))
print('-' * 100)
# evaluate performance on dev data
with torch.no_grad():
outfile = os.path.join(result_path, 'pred_dev%d' % epoch)
scorefile = os.path.join(result_path, "score_dev%d" % epoch)
acc, precision, recall, f1 = eval(data_dev, network, writer,
outfile, scorefile, device)
print(
'Dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%%'
% (acc, precision, recall, f1))
if best_f1 < f1:
torch.save(network.state_dict(), model_name)
best_f1 = f1
best_acc = acc
best_precision = precision
best_recall = recall
best_epoch = epoch
# evaluate on test data when better performance detected
outfile = os.path.join(result_path, 'pred_test%d' % epoch)
scorefile = os.path.join(result_path, "score_test%d" % epoch)
test_acc, test_precision, test_recall, test_f1 = eval(
data_test, network, writer, outfile, scorefile, device)
print(
'test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%%'
% (test_acc, test_precision, test_recall, test_f1))
patient = 0
else:
patient += 1
print('-' * 100)
print(
"Best dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d (%d))"
% (best_acc, best_precision, best_recall, best_f1, best_epoch,
patient))
print(
"Best test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d (%d))"
% (test_acc, test_precision, test_recall, test_f1, best_epoch,
patient))
print('=' * 100)
if patient > 4:
logger.info('reset optimizer momentums')
scheduler.reset_state()
patient = 0
def main():
parser = argparse.ArgumentParser(
description='Tuning with bi-directional RNN-CNN-CRF')
parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU'],
help='architecture of rnn',
required=True)
parser.add_argument('--cuda', action='store_true', help='using GPU')
parser.add_argument('--num_epochs',
type=int,
default=100,
help='Number of training epochs')
parser.add_argument('--batch_size',
type=int,
default=16,
help='Number of sentences in each batch')
parser.add_argument('--hidden_size',
type=int,
default=128,
help='Number of hidden units in RNN')
parser.add_argument('--tag_space',
type=int,
default=0,
help='Dimension of tag space')
parser.add_argument('--num_layers',
type=int,
default=1,
help='Number of layers of RNN')
parser.add_argument('--num_filters',
type=int,
default=30,
help='Number of filters in CNN')
parser.add_argument('--char_dim',
type=int,
default=30,
help='Dimension of Character embeddings')
parser.add_argument('--learning_rate',
type=float,
default=0.015,
help='Learning rate')
parser.add_argument('--decay_rate',
type=float,
default=0.1,
help='Decay rate of learning rate')
parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for regularization')
parser.add_argument('--dropout',
choices=['std', 'variational'],
help='type of dropout',
required=True)
parser.add_argument('--p_rnn',
nargs=2,
type=float,
required=True,
help='dropout rate for RNN')
parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input embeddings')
parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
parser.add_argument('--bigram',
action='store_true',
help='bi-gram parameter for CRF')
parser.add_argument('--schedule',
type=int,
help='schedule for learning rate decay')
parser.add_argument('--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
parser.add_argument('--embedding',
choices=['glove', 'senna', 'sskip', 'polyglot'],
help='Embedding for words',
required=True)
parser.add_argument('--embedding_dict', help='path for embedding dict')
parser.add_argument(
'--train') # "data/POS-penn/wsj/split1/wsj1.train.original"
parser.add_argument(
'--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original"
parser.add_argument(
'--test') # "data/POS-penn/wsj/split1/wsj1.test.original"
args = parser.parse_args()
logger = get_logger("NERCRF")
mode = args.mode
train_path = args.train
dev_path = args.dev
test_path = args.test
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
num_filters = args.num_filters
learning_rate = args.learning_rate
momentum = 0.9
decay_rate = args.decay_rate
gamma = args.gamma
schedule = args.schedule
p_rnn = tuple(args.p_rnn)
p_in = args.p_in
p_out = args.p_out
unk_replace = args.unk_replace
bigram = args.bigram
embedding = args.embedding
embedding_path = args.embedding_dict
embedd_dict, embedd_dim = utils.load_embedding_dict(
embedding, embedding_path)
logger.info("Creating Alphabets")
word_alphabet, char_alphabet, pos_alphabet, \
chunk_alphabet, ner_alphabet = conll03_data.create_alphabets("data/alphabets/ner_crf/", train_path, data_paths=[dev_path, test_path],
embedd_dict=embedd_dict, max_vocabulary_size=50000)
logger.info("Word Alphabet Size: %d" % word_alphabet.size())
logger.info("Character Alphabet Size: %d" % char_alphabet.size())
logger.info("POS Alphabet Size: %d" % pos_alphabet.size())
logger.info("Chunk Alphabet Size: %d" % chunk_alphabet.size())
logger.info("NER Alphabet Size: %d" % ner_alphabet.size())
logger.info("Reading Data")
device = torch.device('cuda') if args.cuda else torch.device('cpu')
data_train = conll03_data.read_data_to_tensor(train_path,
word_alphabet,
char_alphabet,
pos_alphabet,
chunk_alphabet,
ner_alphabet,
device=device)
num_data = sum(data_train[1])
num_labels = ner_alphabet.size()
data_dev = conll03_data.read_data_to_tensor(dev_path,
word_alphabet,
char_alphabet,
pos_alphabet,
chunk_alphabet,
ner_alphabet,
device=device)
data_test = conll03_data.read_data_to_tensor(test_path,
word_alphabet,
char_alphabet,
pos_alphabet,
chunk_alphabet,
ner_alphabet,
device=device)
writer = CoNLL03Writer(word_alphabet, char_alphabet, pos_alphabet,
chunk_alphabet, ner_alphabet)
def construct_word_embedding_table():
scale = np.sqrt(3.0 / embedd_dim)
table = np.empty([word_alphabet.size(), embedd_dim], dtype=np.float32)
table[conll03_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, embedd_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in embedd_dict:
embedding = embedd_dict[word]
elif word.lower() in embedd_dict:
embedding = embedd_dict[word.lower()]
else:
embedding = np.random.uniform(
-scale, scale, [1, embedd_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('oov: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
logger.info("constructing network...")
char_dim = args.char_dim
window = 3
num_layers = args.num_layers
tag_space = args.tag_space
initializer = nn.init.xavier_uniform_
if args.dropout == 'std':
network = BiRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
num_filters,
window,
mode,
hidden_size,
num_layers,
num_labels,
tag_space=tag_space,
embedd_word=word_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
initializer=initializer)
else:
network = BiVarRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
num_filters,
window,
mode,
hidden_size,
num_layers,
num_labels,
tag_space=tag_space,
embedd_word=word_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
initializer=initializer)
network = network.to(device)
lr = learning_rate
optim = SGD(network.parameters(),
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
logger.info(
"Network: %s, num_layer=%d, hidden=%d, filter=%d, tag_space=%d, crf=%s"
% (mode, num_layers, hidden_size, num_filters, tag_space,
'bigram' if bigram else 'unigram'))
logger.info(
"training: l2: %f, (#training data: %d, batch: %d, unk replace: %.2f)"
% (gamma, num_data, batch_size, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
num_batches = num_data / batch_size + 1
dev_f1 = 0.0
dev_acc = 0.0
dev_precision = 0.0
dev_recall = 0.0
test_f1 = 0.0
test_acc = 0.0
test_precision = 0.0
test_recall = 0.0
best_epoch = 0
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s(%s), learning rate=%.4f, decay rate=%.4f (schedule=%d)): '
% (epoch, mode, args.dropout, lr, decay_rate, schedule))
train_err = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
for batch in range(1, num_batches + 1):
word, char, _, _, labels, masks, lengths = conll03_data.get_batch_tensor(
data_train, batch_size, unk_replace=unk_replace)
optim.zero_grad()
loss = network.loss(word, char, labels, mask=masks)
loss.backward()
optim.step()
with torch.no_grad():
num_inst = word.size(0)
train_err += loss * num_inst
train_total += num_inst
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
# update log
if batch % 100 == 0:
sys.stdout.write("\b" * num_back)
sys.stdout.write(" " * num_back)
sys.stdout.write("\b" * num_back)
log_info = 'train: %d/%d loss: %.4f, time left (estimated): %.2fs' % (
batch, num_batches, train_err / train_total, time_left)
sys.stdout.write(log_info)
sys.stdout.flush()
num_back = len(log_info)
sys.stdout.write("\b" * num_back)
sys.stdout.write(" " * num_back)
sys.stdout.write("\b" * num_back)
print('train: %d loss: %.4f, time: %.2fs' %
(num_batches, train_err / train_total, time.time() - start_time))
# evaluate performance on dev data
with torch.no_grad():
network.eval()
tmp_filename = 'tmp/%s_dev%d' % (str(uid), epoch)
writer.start(tmp_filename)
for batch in conll03_data.iterate_batch_tensor(
data_dev, batch_size):
word, char, pos, chunk, labels, masks, lengths = batch
preds, _ = network.decode(
word,
char,
target=labels,
mask=masks,
leading_symbolic=conll03_data.NUM_SYMBOLIC_TAGS)
writer.write(word.cpu().numpy(),
pos.cpu().numpy(),
chunk.cpu().numpy(),
preds.cpu().numpy(),
labels.cpu().numpy(),
lengths.cpu().numpy())
writer.close()
acc, precision, recall, f1 = evaluate(tmp_filename)
print(
'dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%%'
% (acc, precision, recall, f1))
if dev_f1 < f1:
dev_f1 = f1
dev_acc = acc
dev_precision = precision
dev_recall = recall
best_epoch = epoch
# evaluate on test data when better performance detected
tmp_filename = 'tmp/%s_test%d' % (str(uid), epoch)
writer.start(tmp_filename)
for batch in conll03_data.iterate_batch_tensor(
data_test, batch_size):
word, char, pos, chunk, labels, masks, lengths = batch
preds, _ = network.decode(
word,
char,
target=labels,
mask=masks,
leading_symbolic=conll03_data.NUM_SYMBOLIC_TAGS)
writer.write(word.cpu().numpy(),
pos.cpu().numpy(),
chunk.cpu().numpy(),
preds.cpu().numpy(),
labels.cpu().numpy(),
lengths.cpu().numpy())
writer.close()
test_acc, test_precision, test_recall, test_f1 = evaluate(
tmp_filename)
print(
"best dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)"
% (dev_acc, dev_precision, dev_recall, dev_f1, best_epoch))
print(
"best test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)"
% (test_acc, test_precision, test_recall, test_f1, best_epoch))
if epoch % schedule == 0:
lr = learning_rate / (1.0 + epoch * decay_rate)
optim = SGD(network.parameters(),
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
def main():
# Arguments parser
parser = argparse.ArgumentParser(
description='Tuning with DNN Model for NER')
# Model Hyperparameters
parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU'],
help='architecture of rnn',
default='LSTM')
parser.add_argument('--encoder_mode',
choices=['cnn', 'lstm'],
help='Encoder type for sentence encoding',
default='lstm')
parser.add_argument('--char_method',
choices=['cnn', 'lstm'],
help='Method to create character-level embeddings',
required=True)
parser.add_argument(
'--hidden_size',
type=int,
default=128,
help='Number of hidden units in RNN for sentence level')
parser.add_argument('--char_hidden_size',
type=int,
default=30,
help='Output character-level embeddings size')
parser.add_argument('--char_dim',
type=int,
default=30,
help='Dimension of Character embeddings')
parser.add_argument('--tag_space',
type=int,
default=0,
help='Dimension of tag space')
parser.add_argument('--num_layers',
type=int,
default=1,
help='Number of layers of RNN')
parser.add_argument('--dropout',
choices=['std', 'weight_drop'],
help='Dropout method',
default='weight_drop')
parser.add_argument('--p_em',
type=float,
default=0.33,
help='dropout rate for input embeddings')
parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input of RNN model')
parser.add_argument('--p_rnn',
nargs=2,
type=float,
required=True,
help='dropout rate for RNN')
parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
parser.add_argument('--bigram',
action='store_true',
help='bi-gram parameter for CRF')
# Data loading and storing params
parser.add_argument('--embedding_dict', help='path for embedding dict')
parser.add_argument('--dataset_name',
type=str,
default='alexa',
help='Which dataset to use')
parser.add_argument('--train',
type=str,
required=True,
help='Path of train set')
parser.add_argument('--dev',
type=str,
required=True,
help='Path of dev set')
parser.add_argument('--test',
type=str,
required=True,
help='Path of test set')
parser.add_argument('--results_folder',
type=str,
default='results',
help='The folder to store results')
parser.add_argument('--tmp_folder',
type=str,
default='tmp',
help='The folder to store tmp files')
parser.add_argument('--alphabets_folder',
type=str,
default='data/alphabets',
help='The folder to store alphabets files')
parser.add_argument('--result_file_name',
type=str,
default='hyperparameters_tuning',
help='File name to store some results')
parser.add_argument('--result_file_path',
type=str,
default='results/hyperparameters_tuning',
help='File name to store some results')
# Training parameters
parser.add_argument('--cuda',
action='store_true',
help='whether using GPU')
parser.add_argument('--num_epochs',
type=int,
default=100,
help='Number of training epochs')
parser.add_argument('--batch_size',
type=int,
default=16,
help='Number of sentences in each batch')
parser.add_argument('--learning_rate',
type=float,
default=0.001,
help='Base learning rate')
parser.add_argument('--decay_rate',
type=float,
default=0.95,
help='Decay rate of learning rate')
parser.add_argument('--schedule',
type=int,
default=3,
help='schedule for learning rate decay')
parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for l2 regularization')
parser.add_argument('--max_norm',
type=float,
default=1.,
help='Max norm for gradients')
parser.add_argument('--gpu_id',
type=int,
nargs='+',
required=True,
help='which gpu to use for training')
# Misc
parser.add_argument('--embedding',
choices=['glove', 'senna', 'alexa'],
help='Embedding for words',
required=True)
parser.add_argument('--restore',
action='store_true',
help='whether restore from stored parameters')
parser.add_argument('--save_checkpoint',
type=str,
default='',
help='the path to save the model')
parser.add_argument('--o_tag',
type=str,
default='O',
help='The default tag for outside tag')
parser.add_argument('--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
parser.add_argument('--evaluate_raw_format',
action='store_true',
help='The tagging format for evaluation')
args = parser.parse_args()
logger = get_logger("NERCRF")
# rename the parameters
mode = args.mode
encoder_mode = args.encoder_mode
train_path = args.train
dev_path = args.dev
test_path = args.test
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
char_hidden_size = args.char_hidden_size
char_method = args.char_method
learning_rate = args.learning_rate
momentum = 0.9
decay_rate = args.decay_rate
gamma = args.gamma
max_norm = args.max_norm
schedule = args.schedule
dropout = args.dropout
p_em = args.p_em
p_rnn = tuple(args.p_rnn)
p_in = args.p_in
p_out = args.p_out
unk_replace = args.unk_replace
bigram = args.bigram
embedding = args.embedding
embedding_path = args.embedding_dict
dataset_name = args.dataset_name
result_file_name = args.result_file_name
evaluate_raw_format = args.evaluate_raw_format
o_tag = args.o_tag
restore = args.restore
save_checkpoint = args.save_checkpoint
gpu_id = args.gpu_id
results_folder = args.results_folder
tmp_folder = args.tmp_folder
alphabets_folder = args.alphabets_folder
use_elmo = False
p_em_vec = 0.
result_file_path = args.result_file_path
score_file = "%s/score_gpu_%s" % (tmp_folder, '-'.join(map(str, gpu_id)))
if not os.path.exists(results_folder):
os.makedirs(results_folder)
if not os.path.exists(tmp_folder):
os.makedirs(tmp_folder)
if not os.path.exists(alphabets_folder):
os.makedirs(alphabets_folder)
embedd_dict, embedd_dim = utils.load_embedding_dict(
embedding, embedding_path)
logger.info("Creating Alphabets")
word_alphabet, char_alphabet, ner_alphabet = conll03_data.create_alphabets(
"{}/{}/".format(alphabets_folder, dataset_name),
train_path,
data_paths=[dev_path, test_path],
embedd_dict=embedd_dict,
max_vocabulary_size=50000)
logger.info("Word Alphabet Size: %d" % word_alphabet.size())
logger.info("Character Alphabet Size: %d" % char_alphabet.size())
logger.info("NER Alphabet Size: %d" % ner_alphabet.size())
logger.info("Reading Data")
device = torch.device('cuda') if args.cuda else torch.device('cpu')
print(device)
data_train = conll03_data.read_data_to_tensor(train_path,
word_alphabet,
char_alphabet,
ner_alphabet,
device=device)
num_data = sum(data_train[1])
num_labels = ner_alphabet.size()
data_dev = conll03_data.read_data_to_tensor(dev_path,
word_alphabet,
char_alphabet,
ner_alphabet,
device=device)
data_test = conll03_data.read_data_to_tensor(test_path,
word_alphabet,
char_alphabet,
ner_alphabet,
device=device)
writer = CoNLL03Writer(word_alphabet, char_alphabet, ner_alphabet)
def construct_word_embedding_table():
scale = np.sqrt(3.0 / embedd_dim)
table = np.empty([word_alphabet.size(), embedd_dim], dtype=np.float32)
table[conll03_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, embedd_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in embedd_dict:
embedding = embedd_dict[word]
elif word.lower() in embedd_dict:
embedding = embedd_dict[word.lower()]
else:
embedding = np.random.uniform(
-scale, scale, [1, embedd_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('oov: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
logger.info("constructing network...")
char_dim = args.char_dim
window = 3
num_layers = args.num_layers
tag_space = args.tag_space
initializer = nn.init.xavier_uniform_
if args.dropout == 'std':
network = BiRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
char_hidden_size,
window,
mode,
encoder_mode,
hidden_size,
num_layers,
num_labels,
tag_space=tag_space,
embedd_word=word_table,
use_elmo=use_elmo,
p_em_vec=p_em_vec,
p_em=p_em,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
initializer=initializer)
elif args.dropout == 'var':
network = BiVarRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
char_hidden_size,
window,
mode,
encoder_mode,
hidden_size,
num_layers,
num_labels,
tag_space=tag_space,
embedd_word=word_table,
use_elmo=use_elmo,
p_em_vec=p_em_vec,
p_em=p_em,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
initializer=initializer)
else:
network = BiWeightDropRecurrentConvCRF(embedd_dim,
word_alphabet.size(),
char_dim,
char_alphabet.size(),
char_hidden_size,
window,
mode,
encoder_mode,
hidden_size,
num_layers,
num_labels,
tag_space=tag_space,
embedd_word=word_table,
p_em=p_em,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
bigram=bigram,
initializer=initializer)
network = network.to(device)
lr = learning_rate
optim = SGD(network.parameters(),
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
# optim = Adam(network.parameters(), lr=lr, weight_decay=gamma, amsgrad=True)
nn.utils.clip_grad_norm_(network.parameters(), max_norm)
logger.info("Network: %s, encoder_mode=%s, num_layer=%d, hidden=%d, char_hidden_size=%d, char_method=%s, tag_space=%d, crf=%s" % \
(mode, encoder_mode, num_layers, hidden_size, char_hidden_size, char_method, tag_space, 'bigram' if bigram else 'unigram'))
logger.info(
"training: l2: %f, (#training data: %d, batch: %d, unk replace: %.2f)"
% (gamma, num_data, batch_size, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
num_batches = num_data // batch_size + 1
dev_f1 = 0.0
dev_acc = 0.0
dev_precision = 0.0
dev_recall = 0.0
test_f1 = 0.0
test_acc = 0.0
test_precision = 0.0
test_recall = 0.0
best_epoch = 0
best_test_f1 = 0.0
best_test_acc = 0.0
best_test_precision = 0.0
best_test_recall = 0.0
best_test_epoch = 0.0
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s(%s), learning rate=%.4f, decay rate=%.4f (schedule=%d)): '
% (epoch, mode, args.dropout, lr, decay_rate, schedule))
train_err = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
for batch in range(1, num_batches + 1):
_, word, char, labels, masks, lengths = conll03_data.get_batch_tensor(
data_train, batch_size, unk_replace=unk_replace)
optim.zero_grad()
loss = network.loss(_, word, char, labels, mask=masks)
loss.backward()
optim.step()
with torch.no_grad():
num_inst = word.size(0)
train_err += loss * num_inst
train_total += num_inst
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
# update log
if batch % 20 == 0:
sys.stdout.write("\b" * num_back)
sys.stdout.write(" " * num_back)
sys.stdout.write("\b" * num_back)
log_info = 'train: %d/%d loss: %.4f, time left (estimated): %.2fs' % (
batch, num_batches, train_err / train_total, time_left)
sys.stdout.write(log_info)
sys.stdout.flush()
num_back = len(log_info)
sys.stdout.write("\b" * num_back)
sys.stdout.write(" " * num_back)
sys.stdout.write("\b" * num_back)
print('train: %d loss: %.4f, time: %.2fs' %
(num_batches, train_err / train_total, time.time() - start_time))
# evaluate performance on dev data
with torch.no_grad():
network.eval()
tmp_filename = '%s/gpu_%s_dev' % (tmp_folder, '-'.join(
map(str, gpu_id)))
writer.start(tmp_filename)
for batch in conll03_data.iterate_batch_tensor(
data_dev, batch_size):
_, word, char, labels, masks, lengths = batch
preds, _ = network.decode(
_,
word,
char,
target=labels,
mask=masks,
leading_symbolic=conll03_data.NUM_SYMBOLIC_TAGS)
writer.write(word.cpu().numpy(),
preds.cpu().numpy(),
labels.cpu().numpy(),
lengths.cpu().numpy())
writer.close()
acc, precision, recall, f1 = evaluate(tmp_filename, score_file,
evaluate_raw_format, o_tag)
print(
'dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%%'
% (acc, precision, recall, f1))
if dev_f1 < f1:
dev_f1 = f1
dev_acc = acc
dev_precision = precision
dev_recall = recall
best_epoch = epoch
# evaluate on test data when better performance detected
tmp_filename = '%s/gpu_%s_test' % (tmp_folder, '-'.join(
map(str, gpu_id)))
writer.start(tmp_filename)
for batch in conll03_data.iterate_batch_tensor(
data_test, batch_size):
_, word, char, labels, masks, lengths = batch
preds, _ = network.decode(
_,
word,
char,
target=labels,
mask=masks,
leading_symbolic=conll03_data.NUM_SYMBOLIC_TAGS)
writer.write(word.cpu().numpy(),
preds.cpu().numpy(),
labels.cpu().numpy(),
lengths.cpu().numpy())
writer.close()
test_acc, test_precision, test_recall, test_f1 = evaluate(
tmp_filename, score_file, evaluate_raw_format, o_tag)
if best_test_f1 < test_f1:
best_test_acc, best_test_precision, best_test_recall, best_test_f1 = test_acc, test_precision, test_recall, test_f1
best_test_epoch = epoch
print(
"best dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)"
% (dev_acc, dev_precision, dev_recall, dev_f1, best_epoch))
print(
"best test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)"
% (test_acc, test_precision, test_recall, test_f1, best_epoch))
print(
"overall best test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)"
% (best_test_acc, best_test_precision, best_test_recall,
best_test_f1, best_test_epoch))
if epoch % schedule == 0:
lr = learning_rate / (1.0 + epoch * decay_rate)
optim = SGD(network.parameters(),
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
with open(result_file_path, 'a') as ofile:
ofile.write(
"best dev acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)\n"
% (dev_acc, dev_precision, dev_recall, dev_f1, best_epoch))
ofile.write(
"best test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)\n"
% (test_acc, test_precision, test_recall, test_f1, best_epoch))
ofile.write(
"overall best test acc: %.2f%%, precision: %.2f%%, recall: %.2f%%, F1: %.2f%% (epoch: %d)\n\n"
% (best_test_acc, best_test_precision, best_test_recall,
best_test_f1, best_test_epoch))
print('Training finished!')