def decode():
# Prepare NLC data.
global reverse_vocab, vocab, lm
if FLAGS.lmfile is not None:
print("Loading Language model from %s" % FLAGS.lmfile)
lm = kenlm.LanguageModel(FLAGS.lmfile)
print("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(), FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, reverse_vocab = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
print("Vocabulary size: %d" % vocab_size)
with tf.Session() as sess:
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
while True:
sent = raw_input("Enter a sentence: ")
output_sent = fix_sent(model, sess, sent)
print("Candidate: ", output_sent)
def decode():
# Prepare NLC data.
global reverse_vocab, vocab, lm
if FLAGS.lmfile is not None:
print("Loading Language model from %s" % FLAGS.lmfile)
lm = kenlm.LanguageModel(FLAGS.lmfile)
print("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(), FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, reverse_vocab = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
print("Vocabulary size: %d" % vocab_size)
config = tf.ConfigProto(
device_count={'GPU': 0}
)
with tf.Session(config=config) as sess:
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
while True:
sent = input("Enter a sentence: ")
output_sent = fix_sent(model, sess, sent)
print("Candidate: ", output_sent)
def minimize_language(args):
tokenizer = util.get_tokenizer(args.tokenizer_name)
if not args.splits:
minimize_partition(args.filename, "conll", args, tokenizer)
else:
minimize_partition('dev', 'v4_gold_conll', args, tokenizer)
minimize_partition('test', 'v4_gold_conll', args, tokenizer)
minimize_partition('train', 'v4_gold_conll', args, tokenizer)
def __init__(self, config, language='english'):
self.config = config
self.language = language
self.max_seg_len = config['max_segment_len']
self.max_training_seg = config['max_training_sentences']
self.data_dir = config['data_dir']
self.tokenizer = util.get_tokenizer(config['bert_tokenizer_name'])
self.tensor_samples, self.stored_info = None, None # For dataset samples; lazy loading
def tokenize(sent, vocab, depth=FLAGS.num_layers):
align = pow(2, depth - 1)
token_ids = nlc_data.sentence_to_token_ids(sent, vocab, get_tokenizer(FLAGS))
ones = [1] * len(token_ids)
pad = (align - len(token_ids)) % align
token_ids += [nlc_data.PAD_ID] * pad
ones += [0] * pad
source = np.array(token_ids).reshape([-1, 1])
mask = np.array(ones).reshape([-1, 1])
return source, mask
def tokenize(sent, vocab, depth=FLAGS.num_layers): align = pow(2, depth - 1) token_ids = nlc_data.sentence_to_token_ids(sent, vocab, get_tokenizer(FLAGS)) ones = [1] * len(token_ids) pad = (align - len(token_ids)) % align token_ids += [nlc_data.PAD_ID] * pad ones += [0] * pad source = np.array(token_ids).reshape([-1, 1]) mask = np.array(ones).reshape([-1, 1]) return source, mask
def decode():
# Prepare NLC data.
global reverse_vocab, vocab, lm
if FLAGS.lmfile is not None:
print("Loading Language model from %s" % FLAGS.lmfile)
lm = kenlm.LanguageModel(FLAGS.lmfile)
print("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(),
FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, reverse_vocab = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
print("Vocabulary size: %d" % vocab_size)
with tf.Session() as sess:
print("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
# import codecs
# outfile = open('predictions.txt','w')
# outfile2 = open('predictions_all.txt','w')
# outfile = open('lambda_train.txt','w')
# infile = open(FLAGS.data_dir+'/'+FLAGS.tokenizer.lower()+'/test.y.txt')
# lines = infile.readlines()
# index = 0
# with codecs.open(FLAGS.data_dir+'/'+FLAGS.tokenizer.lower()+'/test.x.txt', encoding='utf-8') as fr:
# with codecs.open('ytc_test.txt', encoding='utf-8') as fr:
# for sent in fr:
# print("Original: ", sent.strip().encode('utf-8'))
# output_sent,all_sents,all_prob,all_lmscore = fix_sent(model, sess, sent.encode('utf-8'))
# print("Revised: ", output_sent)
# print('*'*30)
# outfile.write(output_sent+'\n')
# outfile2.write('\t'.join(all_sents)+'\n')
# correct_sent = lines[index].strip('\n').strip('\r')
# for i in range(len(all_sents)):
# if all_sents[i] == correct_sent:
# outfile.write('10 qid:'+str(index)+' 1:'+str(all_prob[i])+' 2:'+str(all_lmscore[i])+' #'+all_sents[i]+'\n')
# else:
# outfile.write('0 qid:'+str(index)+' 1:'+str(all_prob[i])+' 2:'+str(all_lmscore[i])+' #'+all_sents[i]+'\n')
# index+=1
while True:
sent = raw_input("Enter a sentence: ")
output_sent, _, _, _ = fix_sent(model, sess, sent)
print("Candidate: ", output_sent)
def setup_batch_decode(sess):
# decode for dev-sets, in batches
global reverse_vocab, vocab, lm
if FLAGS.lmfile is not None:
print("Loading Language model from %s" % FLAGS.lmfile)
lm = kenlm.LanguageModel(FLAGS.lmfile)
print("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(), FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS), other_dev_path="/deep/group/nlp_data/nlc_data/ourdev/bpe")
vocab, reverse_vocab = nlc_data.initialize_vocabulary(vocab_path, bpe=(FLAGS.tokenizer.lower()=="bpe"))
vocab_size = len(vocab)
print("Vocabulary size: %d" % vocab_size)
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
return model, x_dev, y_dev
def setup_batch_decode(sess):
# decode for dev-sets, in batches
global reverse_vocab, vocab, lm
if FLAGS.lmfile is not None:
print("Loading Language model from %s" % FLAGS.lmfile)
lm = kenlm.LanguageModel(FLAGS.lmfile)
print("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(), FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS), other_dev_path="/deep/group/nlp_data/nlc_data/ourdev/bpe")
vocab, reverse_vocab = nlc_data.initialize_vocabulary(vocab_path, bpe=(FLAGS.tokenizer.lower()=="bpe"))
vocab_size = len(vocab)
print("Vocabulary size: %d" % vocab_size)
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
return model, x_dev, y_dev
def decode():
# Prepare NLC data.
global reverse_vocab, vocab, lm
if FLAGS.lmfile is not None:
print("Loading Language model from %s" % FLAGS.lmfile)
lm = kenlm.LanguageModel(FLAGS.lmfile)
print("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(), FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, reverse_vocab = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
print("Vocabulary size: %d" % vocab_size)
with tf.Session() as sess:
print("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
if FLAGS.interactive:
while True:
sent = raw_input("Enter a sentence: ")
if sent == 'exit':
exit(0)
output_sent = fix_sent(model, sess, sent.decode('utf-8'))
print("Candidate: ", output_sent)
else:
test_x_data = os.path.join(FLAGS.data_dir, FLAGS.tokenizer.lower()+'/test.x.txt')
if not os.path.exists(test_x_data):
print("Please provide {} to test.".format(test_x_data))
exit(-1)
with codecs.open(test_x_data, encoding='utf-8') as fr:
for sent in fr:
print("Original: ", sent.strip().encode('utf-8'))
output_sent = fix_sent(model, sess, sent1)
print("Revised: ", output_sent.encode('utf-8'))
print('*'*30)
def minimize_language(args):
tokenizer = util.get_tokenizer(args.tokenizer_name)
minimize_partition('dev', 'v4_gold_conll', args, tokenizer)
minimize_partition('test', 'v4_gold_conll', args, tokenizer)
minimize_partition('train', 'v4_gold_conll', args, tokenizer)
nargs='?',
default=256)
parser.add_argument('-ep', '--epochs', type=int, nargs='?', default=7)
args = parser.parse_args()
batch_size = args.batch_size
epochs = args.epochs
# 1. data process
logger.info("start train...")
x_train, y_train, x_dev, y_dev, x_test, y_test = get_x_and_y_for_train_dev_test(
)
tokenizer = util.get_tokenizer(num_words=constant.MAX_NUM_WORDS,
texts=list(x_train) + list(x_dev))
train_sequences, dev_sequences, test_sequences = util.get_train_dev_test_sequences(
tokenizer, x_train, x_dev, x_test)
padded_train_sequences, padded_dev_sequences, padded_test_sequences = util.get_train_dev_test_padded_sequences(
maxlen=constant.MAX_LEN,
train_sequences=train_sequences,
dev_sequences=dev_sequences,
test_sequences=test_sequences)
# 2. embedding_matrix
from config import glove_embedding_data_path
num_words = constant.MAX_NUM_WORDS
def train():
"""Train a translation model using NLC data."""
# Prepare NLC data.
logging.info("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(),
FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, _ = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
logging.info("Vocabulary size: %d" % vocab_size)
if not os.path.exists(FLAGS.train_dir):
os.makedirs(FLAGS.train_dir)
file_handler = logging.FileHandler("{0}/log.txt".format(FLAGS.train_dir))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.train_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
logging.info("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
logging.info('Initial validation cost: %f' %
validate(model, sess, x_dev, y_dev))
if False:
tic = time.time()
params = tf.trainable_variables()
num_params = sum(
map(lambda t: np.prod(tf.shape(t.value()).eval()), params))
toc = time.time()
print("Number of params: %d (retreival took %f secs)" %
(num_params, toc - tic))
epoch = 0
best_epoch = 0
previous_losses = []
exp_cost = None
exp_length = None
exp_norm = None
total_iters = 0
start_time = time.time()
while (FLAGS.epochs == 0 or epoch < FLAGS.epochs):
epoch += 1
current_step = 0
## Train
epoch_tic = time.time()
for source_tokens, source_mask, target_tokens, target_mask in pair_iter(
x_train, y_train, FLAGS.batch_size, FLAGS.num_layers):
# Get a batch and make a step.
tic = time.time()
grad_norm, cost, param_norm = model.train(
sess, source_tokens, source_mask, target_tokens,
target_mask)
toc = time.time()
iter_time = toc - tic
total_iters += np.sum(target_mask)
tps = total_iters / (time.time() - start_time)
current_step += 1
lengths = np.sum(target_mask, axis=0)
mean_length = np.mean(lengths)
std_length = np.std(lengths)
if not exp_cost:
exp_cost = cost
exp_length = mean_length
exp_norm = grad_norm
else:
exp_cost = 0.99 * exp_cost + 0.01 * cost
exp_length = 0.99 * exp_length + 0.01 * mean_length
exp_norm = 0.99 * exp_norm + 0.01 * grad_norm
cost = cost / mean_length
if current_step % FLAGS.print_every == 0:
logging.info(
'epoch %d, iter %d, cost %f, exp_cost %f, grad norm %f, param norm %f, tps %f, length mean/std %f/%f'
%
(epoch, current_step, cost, exp_cost / exp_length,
grad_norm, param_norm, tps, mean_length, std_length))
epoch_toc = time.time()
## Checkpoint
checkpoint_path = os.path.join(FLAGS.train_dir, "best.ckpt")
## Validate
valid_cost = validate(model, sess, x_dev, y_dev)
logging.info("Epoch %d Validation cost: %f time: %f" %
(epoch, valid_cost, epoch_toc - epoch_tic))
if len(previous_losses) > 2 and valid_cost > previous_losses[-1]:
logging.info("Annealing learning rate by %f" %
FLAGS.learning_rate_decay_factor)
sess.run(model.learning_rate_decay_op)
model.saver.restore(sess,
checkpoint_path + ("-%d" % best_epoch))
else:
previous_losses.append(valid_cost)
best_epoch = epoch
model.saver.save(sess, checkpoint_path, global_step=epoch)
sys.stdout.flush()
def train():
"""Train a translation model using NLC data."""
# Prepare NLC data.
logging.info("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(), FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, _ = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
logging.info("Vocabulary size: %d" % vocab_size)
if not os.path.exists(FLAGS.train_dir):
os.makedirs(FLAGS.train_dir)
file_handler = logging.FileHandler("{0}/log.txt".format(FLAGS.train_dir))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.train_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
logging.info("Creating %d layers of %d units." % (FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
logging.info('Initial validation cost: %f' % validate(model, sess, x_dev, y_dev))
if False:
tic = time.time()
params = tf.trainable_variables()
num_params = sum(map(lambda t: np.prod(tf.shape(t.value()).eval()), params))
toc = time.time()
print ("Number of params: %d (retreival took %f secs)" % (num_params, toc - tic))
epoch = 0
best_epoch = 0
previous_losses = []
exp_cost = None
exp_length = None
exp_norm = None
while (FLAGS.epochs == 0 or epoch < FLAGS.epochs):
epoch += 1
current_step = 0
## Train
epoch_tic = time.time()
for source_tokens, source_mask, target_tokens, target_mask in pair_iter(x_train, y_train, FLAGS.batch_size, FLAGS.num_layers):
# Get a batch and make a step.
tic = time.time()
grad_norm, cost, param_norm = model.train(sess, source_tokens, source_mask, target_tokens, target_mask)
toc = time.time()
iter_time = toc - tic
current_step += 1
lengths = np.sum(target_mask, axis=0)
mean_length = np.mean(lengths)
std_length = np.std(lengths)
if not exp_cost:
exp_cost = cost
exp_length = mean_length
exp_norm = grad_norm
else:
exp_cost = 0.99*exp_cost + 0.01*cost
exp_length = 0.99*exp_length + 0.01*mean_length
exp_norm = 0.99*exp_norm + 0.01*grad_norm
cost = cost / mean_length
if current_step % FLAGS.print_every == 0:
logging.info('epoch %d, iter %d, cost %f, exp_cost %f, grad norm %f, param norm %f, batch time %f, length mean/std %f/%f' %
(epoch, current_step, cost, exp_cost / exp_length, grad_norm, param_norm, iter_time, mean_length, std_length))
epoch_toc = time.time()
## Checkpoint
checkpoint_path = os.path.join(FLAGS.train_dir, "best.ckpt")
## Validate
valid_cost = validate(model, sess, x_dev, y_dev)
logging.info("Epoch %d Validation cost: %f time: %f" % (epoch, valid_cost, epoch_toc - epoch_tic))
if len(previous_losses) > 2 and valid_cost > previous_losses[-1]:
logging.info("Annealing learning rate by %f" % FLAGS.learning_rate_decay_factor)
sess.run(model.learning_rate_decay_op)
model.saver.restore(sess, checkpoint_path + ("-%d" % best_epoch))
else:
previous_losses.append(valid_cost)
best_epoch = epoch
model.saver.save(sess, checkpoint_path, global_step=epoch)
sys.stdout.flush()
def train():
"""Train a translation model using NLC data."""
# Prepare NLC data.
logging.info("Preparing NLC data in %s" % FLAGS.data_dir)
x_train, y_train, x_dev, y_dev, vocab_path = nlc_data.prepare_nlc_data(
FLAGS.data_dir + '/' + FLAGS.tokenizer.lower(),
FLAGS.max_vocab_size,
tokenizer=get_tokenizer(FLAGS))
vocab, _ = nlc_data.initialize_vocabulary(vocab_path)
vocab_size = len(vocab)
logging.info("Vocabulary size: %d" % vocab_size)
if not os.path.exists(FLAGS.train_dir):
os.makedirs(FLAGS.train_dir)
file_handler = logging.FileHandler("{0}/log.txt".format(FLAGS.train_dir))
logging.getLogger().addHandler(file_handler)
print(vars(FLAGS))
with open(os.path.join(FLAGS.train_dir, "flags.json"), 'w') as fout:
json.dump(FLAGS.__flags, fout)
with tf.Session() as sess:
logging.info("Creating %d layers of %d units." %
(FLAGS.num_layers, FLAGS.size))
model = create_model(sess, vocab_size, False)
tic = time.time()
params = tf.trainable_variables()
num_params = sum(
map(lambda t: np.prod(tf.shape(t.value()).eval()), params))
toc = time.time()
print("Number of params: %d (retreival took %f secs)" %
(num_params, toc - tic))
epoch = 0
best_epoch = 0
train_costs = []
valid_costs = []
previous_valid_losses = []
while (FLAGS.epochs == 0 or epoch < FLAGS.epochs):
epoch += 1
current_step = 0
epoch_cost = 0
epoch_tic = time.time()
for source_tokens, source_mask, target_tokens, target_mask in pair_iter(
x_train, y_train, FLAGS.batch_size, FLAGS.num_layers):
# Get a batch and make a step.
grad_norm, cost, param_norm = model.train(
sess, source_tokens, source_mask, target_tokens,
target_mask)
lengths = np.sum(target_mask, axis=0)
mean_length = np.mean(lengths)
std_length = np.std(lengths)
cost = cost / mean_length
epoch_cost += cost
current_step += 1
if current_step % FLAGS.print_every == 0:
logging.info(
'epoch %d, iter %d, cost %f, length mean/std %f/%f' %
(epoch, current_step, cost, mean_length, std_length))
# One epoch average train cost
train_costs.append(epoch_cost / current_step)
# After one epoch average validate cost
epoch_toc = time.time()
epoch_time = epoch_toc - epoch_tic
valid_cost = validate(model, sess, x_dev, y_dev)
valid_costs.append(valid_cost)
logging.info("Epoch %d Validation cost: %f time: %2fs" %
(epoch, valid_cost, epoch_time))
# Checkpoint
checkpoint_path = os.path.join(FLAGS.train_dir, "best.ckpt")
if len(previous_valid_losses
) > 2 and valid_cost > previous_valid_losses[-1]:
logging.info("Annealing learning rate by %f" %
FLAGS.learning_rate_decay_factor)
sess.run(model.learning_rate_decay_op)
model.saver.restore(sess,
checkpoint_path + ("-%d" % best_epoch))
else:
previous_valid_losses.append(valid_cost)
best_epoch = epoch
model.saver.save(sess, checkpoint_path, global_step=epoch)
import cPickle as pickle
pickle.dump([train_costs, valid_costs], open('costs_data.pkl', 'wb'))
