def main():
args_parser = argparse.ArgumentParser(
description='Tuning with graph-based parsing')
args_parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'],
help='architecture of rnn',
required=True)
args_parser.add_argument('--num_epochs',
type=int,
default=200,
help='Number of training epochs')
args_parser.add_argument('--batch_size',
type=int,
default=64,
help='Number of sentences in each batch')
args_parser.add_argument('--hidden_size',
type=int,
default=256,
help='Number of hidden units in RNN')
args_parser.add_argument('--arc_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--type_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--num_layers',
type=int,
default=1,
help='Number of layers of RNN')
args_parser.add_argument('--num_filters',
type=int,
default=50,
help='Number of filters in CNN')
args_parser.add_argument('--pos',
action='store_true',
help='use part-of-speech embedding.')
args_parser.add_argument('--char',
action='store_true',
help='use character embedding and CNN.')
args_parser.add_argument('--pos_dim',
type=int,
default=50,
help='Dimension of POS embeddings')
args_parser.add_argument('--char_dim',
type=int,
default=50,
help='Dimension of Character embeddings')
args_parser.add_argument('--opt',
choices=['adam', 'sgd', 'adamax'],
help='optimization algorithm')
args_parser.add_argument('--objective',
choices=['cross_entropy', 'crf'],
default='cross_entropy',
help='objective function of training procedure.')
args_parser.add_argument('--decode',
choices=['mst', 'greedy'],
help='decoding algorithm',
required=True)
args_parser.add_argument('--learning_rate',
type=float,
default=0.01,
help='Learning rate')
args_parser.add_argument('--decay_rate',
type=float,
default=0.05,
help='Decay rate of learning rate')
args_parser.add_argument('--clip',
type=float,
default=5.0,
help='gradient clipping')
args_parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for regularization')
args_parser.add_argument('--epsilon',
type=float,
default=1e-8,
help='epsilon for adam or adamax')
args_parser.add_argument('--p_rnn',
nargs=2,
type=float,
required=True,
help='dropout rate for RNN')
args_parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input embeddings')
args_parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
args_parser.add_argument('--schedule',
type=int,
help='schedule for learning rate decay')
args_parser.add_argument(
'--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
args_parser.add_argument('--punctuation',
nargs='+',
type=str,
help='List of punctuations')
args_parser.add_argument('--word_embedding',
choices=['glove', 'senna', 'sskip', 'polyglot'],
help='Embedding for words',
required=True)
args_parser.add_argument('--word_path',
help='path for word embedding dict')
args_parser.add_argument(
'--freeze',
action='store_true',
help='frozen the word embedding (disable fine-tuning).')
args_parser.add_argument('--char_embedding',
choices=['random', 'polyglot'],
help='Embedding for characters',
required=True)
args_parser.add_argument('--char_path',
help='path for character embedding dict')
args_parser.add_argument(
'--train') # "data/POS-penn/wsj/split1/wsj1.train.original"
args_parser.add_argument(
'--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original"
args_parser.add_argument(
'--test') # "data/POS-penn/wsj/split1/wsj1.test.original"
args_parser.add_argument('--model_path',
help='path for saving model file.',
required=True)
args_parser.add_argument('--model_name',
help='name for saving model file.',
required=True)
args = args_parser.parse_args()
logger = get_logger("GraphParser")
mode = args.mode
obj = args.objective
decoding = args.decode
train_path = args.train
dev_path = args.dev
test_path = args.test
model_path = args.model_path
model_name = args.model_name
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
arc_space = args.arc_space
type_space = args.type_space
num_layers = args.num_layers
num_filters = args.num_filters
learning_rate = args.learning_rate
opt = args.opt
momentum = 0.9
betas = (0.9, 0.9)
eps = args.epsilon
decay_rate = args.decay_rate
clip = args.clip
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
punctuation = args.punctuation
freeze = args.freeze
word_embedding = args.word_embedding
word_path = args.word_path
use_char = args.char
char_embedding = args.char_embedding
char_path = args.char_path
use_pos = args.pos
pos_dim = args.pos_dim
word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path)
char_dict = None
char_dim = args.char_dim
if char_embedding != 'random':
char_dict, char_dim = utils.load_embedding_dict(
char_embedding, char_path)
logger.info("Creating Alphabets")
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets(
alphabet_path,
train_path,
data_paths=[dev_path, test_path],
max_vocabulary_size=50000,
embedd_dict=word_dict)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
logger.info("Reading Data")
use_gpu = torch.cuda.is_available()
data_train = conllx_data.read_data_to_variable(train_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
use_gpu=use_gpu,
symbolic_root=True)
# data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
# num_data = sum([len(bucket) for bucket in data_train])
num_data = sum(data_train[1])
data_dev = conllx_data.read_data_to_variable(dev_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
use_gpu=use_gpu,
volatile=True,
symbolic_root=True)
data_test = conllx_data.read_data_to_variable(test_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
use_gpu=use_gpu,
volatile=True,
symbolic_root=True)
punct_set = None
if punctuation is not None:
punct_set = set(punctuation)
logger.info("punctuations(%d): %s" %
(len(punct_set), ' '.join(punct_set)))
def construct_word_embedding_table():
scale = np.sqrt(3.0 / word_dim)
table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in word_dict:
embedding = word_dict[word]
elif word.lower() in word_dict:
embedding = word_dict[word.lower()]
else:
embedding = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('word OOV: %d' % oov)
return torch.from_numpy(table)
def construct_char_embedding_table():
if char_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_chars, char_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, char_dim]).astype(np.float32)
oov = 0
for char, index, in char_alphabet.items():
if char in char_dict:
embedding = char_dict[char]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('character OOV: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
char_table = construct_char_embedding_table()
window = 3
if obj == 'cross_entropy':
network = BiRecurrentConvBiAffine(word_dim,
num_words,
char_dim,
num_chars,
pos_dim,
num_pos,
num_filters,
window,
mode,
hidden_size,
num_layers,
num_types,
arc_space,
type_space,
embedd_word=word_table,
embedd_char=char_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
biaffine=True,
pos=use_pos,
char=use_char)
elif obj == 'crf':
raise NotImplementedError
else:
raise RuntimeError('Unknown objective: %s' % obj)
if freeze:
network.word_embedd.freeze()
if use_gpu:
network.cuda()
pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
def generate_optimizer(opt, lr, params):
params = filter(lambda param: param.requires_grad, params)
if opt == 'adam':
return Adam(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
elif opt == 'sgd':
return SGD(params,
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
elif opt == 'adamax':
return Adamax(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
else:
raise ValueError('Unknown optimization algorithm: %s' % opt)
lr = learning_rate
optim = generate_optimizer(opt, lr, network.parameters())
opt_info = 'opt: %s, ' % opt
if opt == 'adam':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
elif opt == 'sgd':
opt_info += 'momentum=%.2f' % momentum
elif opt == 'adamax':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
word_status = 'frozen' if freeze else 'fine tune'
char_status = 'enabled' if use_char else 'disabled'
pos_status = 'enabled' if use_pos else 'disabled'
logger.info(
"Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" %
(word_dim, word_status, char_dim, char_status, pos_dim, pos_status))
logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window))
logger.info(
"RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" %
(mode, num_layers, hidden_size, arc_space, type_space))
logger.info(
"train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)"
% (obj, gamma, num_data, batch_size, clip, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
logger.info("decoding algorithm: %s" % decoding)
logger.info(opt_info)
num_batches = num_data / batch_size + 1
dev_ucorrect = 0.0
dev_lcorrect = 0.0
dev_ucomlpete_match = 0.0
dev_lcomplete_match = 0.0
dev_ucorrect_nopunc = 0.0
dev_lcorrect_nopunc = 0.0
dev_ucomlpete_match_nopunc = 0.0
dev_lcomplete_match_nopunc = 0.0
dev_root_correct = 0.0
best_epoch = 0
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_root_correct = 0.0
test_total = 0
test_total_nopunc = 0
test_total_inst = 0
test_total_root = 0
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
patient = 0
decay = 0
max_decay = 9
double_schedule_decay = 5
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): '
%
(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay))
train_err = 0.
train_err_arc = 0.
train_err_type = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
for batch in range(1, num_batches + 1):
word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_variable(
data_train, batch_size, unk_replace=unk_replace)
optim.zero_grad()
loss_arc, loss_type = network.loss(word,
char,
pos,
heads,
types,
mask=masks,
length=lengths)
loss = loss_arc + loss_type
loss.backward()
clip_grad_norm(network.parameters(), clip)
optim.step()
num_inst = word.size(
0) if obj == 'crf' else masks.data.sum() - word.size(0)
train_err += loss.data[0] * num_inst
train_err_arc += loss_arc.data[0] * num_inst
train_err_type += loss_type.data[0] * num_inst
train_total += num_inst
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
# update log
if batch % 10 == 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, arc: %.4f, type: %.4f, time left: %.2fs' % (
batch, num_batches, train_err / train_total, train_err_arc
/ train_total, train_err_type / 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, arc: %.4f, type: %.4f, time: %.2fs' %
(num_batches, train_err / train_total, train_err_arc / train_total,
train_err_type / train_total, time.time() - start_time))
# evaluate performance on dev data
network.eval()
pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch)
pred_writer.start(pred_filename)
gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch)
gold_writer.start(gold_filename)
dev_ucorr = 0.0
dev_lcorr = 0.0
dev_total = 0
dev_ucomlpete = 0.0
dev_lcomplete = 0.0
dev_ucorr_nopunc = 0.0
dev_lcorr_nopunc = 0.0
dev_total_nopunc = 0
dev_ucomlpete_nopunc = 0.0
dev_lcomplete_nopunc = 0.0
dev_root_corr = 0.0
dev_total_root = 0.0
dev_total_inst = 0.0
for batch in conllx_data.iterate_batch_variable(data_dev, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
dev_ucorr += ucorr
dev_lcorr += lcorr
dev_total += total
dev_ucomlpete += ucm
dev_lcomplete += lcm
dev_ucorr_nopunc += ucorr_nopunc
dev_lcorr_nopunc += lcorr_nopunc
dev_total_nopunc += total_nopunc
dev_ucomlpete_nopunc += ucm_nopunc
dev_lcomplete_nopunc += lcm_nopunc
dev_root_corr += corr_root
dev_total_root += total_root
dev_total_inst += num_inst
pred_writer.close()
gold_writer.close()
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total,
dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 /
dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc *
100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 /
dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' %
(dev_root_corr, dev_total_root,
dev_root_corr * 100 / dev_total_root))
if dev_lcorrect_nopunc < dev_lcorr_nopunc or (
dev_lcorrect_nopunc == dev_lcorr_nopunc
and dev_ucorrect_nopunc < dev_ucorr_nopunc):
dev_ucorrect_nopunc = dev_ucorr_nopunc
dev_lcorrect_nopunc = dev_lcorr_nopunc
dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc
dev_lcomplete_match_nopunc = dev_lcomplete_nopunc
dev_ucorrect = dev_ucorr
dev_lcorrect = dev_lcorr
dev_ucomlpete_match = dev_ucomlpete
dev_lcomplete_match = dev_lcomplete
dev_root_correct = dev_root_corr
best_epoch = epoch
patient = 0
# torch.save(network, model_name)
torch.save(network.state_dict(), model_name)
pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch)
pred_writer.start(pred_filename)
gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch)
gold_writer.start(gold_filename)
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
for batch in conllx_data.iterate_batch_variable(
data_test, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
pred_writer.close()
gold_writer.close()
else:
if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule:
# network = torch.load(model_name)
network.load_state_dict(torch.load(model_name))
lr = lr * decay_rate
optim = generate_optimizer(opt, lr, network.parameters())
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' %
decoding)
patient = 0
decay += 1
if decay % double_schedule_decay == 0:
schedule *= 2
else:
patient += 1
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect, dev_lcorrect, dev_total,
dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch))
print(
'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch))
print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' %
(dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch))
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 /
test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst,
test_lcomplete_match * 100 / test_total_inst, best_epoch))
print(
'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
%
(test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc,
test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst,
test_lcomplete_match_nopunc * 100 / test_total_inst, best_epoch))
print('best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' %
(test_root_correct, test_total_root,
test_root_correct * 100 / test_total_root, best_epoch))
print(
'============================================================================================================================'
)
if decay == max_decay:
break
p_in=kwargs['p_in'],
p_out=kwargs['p_out'],
p_rnn=kwargs['p_rnn'],
biaffine=kwargs['biaffine'],
pos=kwargs['pos'],
char=kwargs['char'])
char_alphabet.load(pretrained_dependency_parser_vocab_fn)
pos_alphabet.load(pretrained_dependency_parser_vocab_fn)
type_alphabet.load(pretrained_dependency_parser_vocab_fn)
word_alphabet.load(pretrained_dependency_parser_vocab_fn)
# char_alphabet.keep_growing = True
# word_alphabet.keep_growing = True
network.load_state_dict(parser)
#------------------------create data specific alphabets
print("Creating graph classifier Alphabets")
alphabet_path = os.path.join(custom_args.graph_alphabet_folder,
custom_args.dataset_name)
train_path, dev_path, test_path, _, _, _ = datasets.DATASET_FILES[
custom_args.dataset_name]
if custom_args.bio_embeddings != 'none':
if 'bio' in custom_args.bio_embeddings:
embedding_vocab_dict = datasets.load_bio_word_embedding_vocab(
custom_args.bio_embeddings)
elif 'glove' in custom_args.bio_embeddings:
embedding_vocab_dict = datasets.load_glove_word_embedding_vocab(
custom_args.bio_embeddings)
def main():
args_parser = argparse.ArgumentParser(
description='Tuning with graph-based parsing')
args_parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'],
help='architecture of rnn',
required=True)
args_parser.add_argument('--cuda', action='store_true', help='using GPU')
args_parser.add_argument('--num_epochs',
type=int,
default=200,
help='Number of training epochs')
args_parser.add_argument('--batch_size',
type=int,
default=64,
help='Number of sentences in each batch')
args_parser.add_argument('--hidden_size',
type=int,
default=256,
help='Number of hidden units in RNN')
args_parser.add_argument('--arc_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--type_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--num_layers',
type=int,
default=1,
help='Number of layers of RNN')
args_parser.add_argument('--num_filters',
type=int,
default=50,
help='Number of filters in CNN')
args_parser.add_argument('--pos',
action='store_true',
help='use part-of-speech embedding.')
args_parser.add_argument('--char',
action='store_true',
help='use character embedding and CNN.')
args_parser.add_argument('--pos_dim',
type=int,
default=50,
help='Dimension of POS embeddings')
args_parser.add_argument('--char_dim',
type=int,
default=50,
help='Dimension of Character embeddings')
args_parser.add_argument('--opt',
choices=['adam', 'sgd', 'adamax'],
help='optimization algorithm')
args_parser.add_argument('--objective',
choices=['cross_entropy', 'crf'],
default='cross_entropy',
help='objective function of training procedure.')
args_parser.add_argument('--decode',
choices=['mst', 'greedy'],
help='decoding algorithm',
required=True)
args_parser.add_argument('--learning_rate',
type=float,
default=0.01,
help='Learning rate')
args_parser.add_argument('--decay_rate',
type=float,
default=0.05,
help='Decay rate of learning rate')
args_parser.add_argument('--clip',
type=float,
default=5.0,
help='gradient clipping')
args_parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for regularization')
args_parser.add_argument('--epsilon',
type=float,
default=1e-8,
help='epsilon for adam or adamax')
args_parser.add_argument('--p_rnn',
nargs=2,
type=float,
required=True,
help='dropout rate for RNN')
args_parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input embeddings')
args_parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
args_parser.add_argument('--schedule',
type=int,
help='schedule for learning rate decay')
args_parser.add_argument(
'--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
args_parser.add_argument('--punctuation',
nargs='+',
type=str,
help='List of punctuations')
args_parser.add_argument('--word_embedding',
choices=['glove', 'senna', 'sskip', 'polyglot'],
help='Embedding for words',
required=True)
args_parser.add_argument('--word_path',
help='path for word embedding dict')
args_parser.add_argument(
'--freeze',
action='store_true',
help='frozen the word embedding (disable fine-tuning).')
args_parser.add_argument('--char_embedding',
choices=['random', 'polyglot'],
help='Embedding for characters',
required=True)
args_parser.add_argument('--char_path',
help='path for character embedding dict')
args_parser.add_argument(
'--train') # "data/POS-penn/wsj/split1/wsj1.train.original"
args_parser.add_argument(
'--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original"
args_parser.add_argument(
'--test') # "data/POS-penn/wsj/split1/wsj1.test.original"
args_parser.add_argument('--model_path',
help='path for saving model file.',
required=True)
args_parser.add_argument('--model_name',
help='name for saving model file.',
required=True)
args_parser.add_argument('--seq2seq_save_path',
default='models/seq2seq/seq2seq_save_model',
type=str,
help='seq2seq_save_path')
args_parser.add_argument('--network_save_path',
default='models/seq2seq/network_save_model',
type=str,
help='network_save_path')
args_parser.add_argument('--seq2seq_load_path',
default='models/seq2seq/seq2seq_save_model',
type=str,
help='seq2seq_load_path')
args_parser.add_argument('--network_load_path',
default='models/seq2seq/network_save_model',
type=str,
help='network_load_path')
args_parser.add_argument('--rl_finetune_seq2seq_save_path',
default='models/rl_finetune/seq2seq_save_model',
type=str,
help='rl_finetune_seq2seq_save_path')
args_parser.add_argument('--rl_finetune_network_save_path',
default='models/rl_finetune/network_save_model',
type=str,
help='rl_finetune_network_save_path')
args_parser.add_argument('--rl_finetune_seq2seq_load_path',
default='models/rl_finetune/seq2seq_save_model',
type=str,
help='rl_finetune_seq2seq_load_path')
args_parser.add_argument('--rl_finetune_network_load_path',
default='models/rl_finetune/network_save_model',
type=str,
help='rl_finetune_network_load_path')
args = args_parser.parse_args()
# args.train = "data/ptb/dev.conllu"
# args.dev = "data/ptb/dev.conllu"
# args.test = "data/ptb/dev.conllu"
logger = get_logger("GraphParser")
TEST_PARSER_FLAG = False # TODO
# SEED = 0
# torch.manual_seed(SEED)
# torch.cuda.manual_seed(SEED)
mode = args.mode
obj = args.objective
decoding = args.decode
train_path = args.train
dev_path = args.dev
test_path = args.test
model_path = args.model_path
model_name = args.model_name
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
arc_space = args.arc_space
type_space = args.type_space
num_layers = args.num_layers
num_filters = args.num_filters
learning_rate = args.learning_rate
opt = args.opt
momentum = 0.9
betas = (0.9, 0.9)
eps = args.epsilon
decay_rate = args.decay_rate
clip = args.clip
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
punctuation = args.punctuation
freeze = args.freeze
word_embedding = args.word_embedding
word_path = args.word_path
use_char = args.char
char_embedding = args.char_embedding
char_path = args.char_path
use_pos = args.pos
pos_dim = args.pos_dim
word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path)
char_dict = None
char_dim = args.char_dim
if char_embedding != 'random':
char_dict, char_dim = utils.load_embedding_dict(
char_embedding, char_path)
logger.info("Creating Alphabets")
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_data.create_alphabets(
alphabet_path,
train_path,
data_paths=[dev_path, test_path],
max_vocabulary_size=100000,
embedd_dict=word_dict)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
logger.info("Reading Data")
device = torch.device(
'cuda:0'
) #torch.device('cuda:0') if args.cuda else torch.device('cpu') #TODO:8.8
data_train = conllx_data.read_data_to_tensor(train_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
# data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
# num_data = sum([len(bucket) for bucket in data_train])
num_data = sum(data_train[1])
data_dev = conllx_data.read_data_to_tensor(dev_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
data_test = conllx_data.read_data_to_tensor(test_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
punct_set = None
if punctuation is not None:
punct_set = set(punctuation)
logger.info("punctuations(%d): %s" %
(len(punct_set), ' '.join(punct_set)))
def construct_word_embedding_table():
scale = np.sqrt(3.0 / word_dim)
table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in word_dict:
embedding = word_dict[word]
elif word.lower() in word_dict:
embedding = word_dict[word.lower()]
else:
embedding = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('word OOV: %d' % oov)
return torch.from_numpy(table)
def construct_char_embedding_table():
if char_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_chars, char_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, char_dim]).astype(np.float32)
oov = 0
for char, index, in char_alphabet.items():
if char in char_dict:
embedding = char_dict[char]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('character OOV: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
char_table = construct_char_embedding_table()
# Pretrain structure prediction model (biaff model). model name: network
window = 3
if obj == 'cross_entropy':
network = BiRecurrentConvBiAffine(word_dim,
num_words,
char_dim,
num_chars,
pos_dim,
num_pos,
num_filters,
window,
mode,
hidden_size,
num_layers,
num_types,
arc_space,
type_space,
embedd_word=word_table,
embedd_char=char_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
biaffine=True,
pos=use_pos,
char=use_char)
elif obj == 'crf':
raise NotImplementedError
else:
raise RuntimeError('Unknown objective: %s' % obj)
def save_args():
arg_path = model_name + '.arg.json'
arguments = [
word_dim, num_words, char_dim, num_chars, pos_dim, num_pos,
num_filters, window, mode, hidden_size, num_layers, num_types,
arc_space, type_space
]
kwargs = {
'p_in': p_in,
'p_out': p_out,
'p_rnn': p_rnn,
'biaffine': True,
'pos': use_pos,
'char': use_char
}
json.dump({
'args': arguments,
'kwargs': kwargs
},
open(arg_path, 'w'),
indent=4)
if freeze:
freeze_embedding(network.word_embedd)
network = network.to(device)
save_args()
pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
def generate_optimizer(opt, lr, params):
params = filter(lambda param: param.requires_grad, params)
if opt == 'adam':
return Adam(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
elif opt == 'sgd':
return SGD(params,
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
elif opt == 'adamax':
return Adamax(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
else:
raise ValueError('Unknown optimization algorithm: %s' % opt)
lr = learning_rate
optim = generate_optimizer(opt, lr, network.parameters())
opt_info = 'opt: %s, ' % opt
if opt == 'adam':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
elif opt == 'sgd':
opt_info += 'momentum=%.2f' % momentum
elif opt == 'adamax':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
word_status = 'frozen' if freeze else 'fine tune'
char_status = 'enabled' if use_char else 'disabled'
pos_status = 'enabled' if use_pos else 'disabled'
logger.info(
"Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" %
(word_dim, word_status, char_dim, char_status, pos_dim, pos_status))
logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window))
logger.info(
"RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" %
(mode, num_layers, hidden_size, arc_space, type_space))
logger.info(
"train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)"
% (obj, gamma, num_data, batch_size, clip, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
logger.info("decoding algorithm: %s" % decoding)
logger.info(opt_info)
num_batches = num_data / batch_size + 1
dev_ucorrect = 0.0
dev_lcorrect = 0.0
dev_ucomlpete_match = 0.0
dev_lcomplete_match = 0.0
dev_ucorrect_nopunc = 0.0
dev_lcorrect_nopunc = 0.0
dev_ucomlpete_match_nopunc = 0.0
dev_lcomplete_match_nopunc = 0.0
dev_root_correct = 0.0
best_epoch = 0
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_root_correct = 0.0
test_total = 0
test_total_nopunc = 0
test_total_inst = 0
test_total_root = 0
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
print('Pretrain biaffine model.')
patient = 0
decay = 0
max_decay = 9
double_schedule_decay = 5
num_epochs = 1 # debug hanwj
network.load_state_dict(
torch.load('models/parsing/biaffine/network.pt')) # TODO: 7.13
network.to(device)
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): '
%
(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay))
train_err = 0.
train_err_arc = 0.
train_err_type = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
FLAG = True
for para in network.parameters():
if FLAG:
para.requires_grad = True
FLAG = False
else:
para.requires_grad = False
# network.state_dict().items()[0][1].requires_grad = True
# ori_word_embedding = network.parameters().next().detach().data.cpu().numpy().T # torch.Size([35374, 100])
# c_ = -2 * np.dot(ori_word_embedding.T, ori_word_embedding)
# a = np.sum(np.square(ori_word_embedding), axis=0).reshape((1, -1))
# b = a.T
# dist = a + b + c_
# np.save('base_attach/dist_counter_35374.npy', dist)
outf = 'base_attach/adv_sentences.txt'
wf = codecs.open(outf, 'w', encoding='utf8')
ori_outf = 'base_attach/ori_sentences.txt'
ori_wf = codecs.open(ori_outf, 'w', encoding='utf8')
ori_word_embedding = network.parameters().next().detach().data.cpu(
).numpy().T # torch.Size([35374, 100])
for batch in conllx_data.iterate_batch_tensor(data_dev, batch_size):
word, char, pos, heads, types, masks, lengths = batch
optim.zero_grad()
loss_arc, loss_type = network.loss(word,
char,
pos,
heads,
types,
mask=masks,
length=lengths)
loss = loss_arc #+ loss_type
loss = -loss
loss.backward()
clip_grad_norm_(network.parameters(), clip)
optim.step()
for batch_i in range(len(word)):
ori_wf.write(' '.join([
str(word_alphabet.get_instance(wordi))
for wordi in word[batch_i][1:] if not wordi == 1
]))
ori_wf.write('\n')
# adv_word_embedding = network.parameters().next().detach().data.cpu().numpy().T #torch.Size([35374, 100])
for batch_i in range(len(word)):
for batch_j in range(1, len(word[0])):
one_word = word[batch_i][batch_j].item()
if one_word == 1:
continue
src_word_idx = one_word
adv_vector = network.parameters().next().detach().data.cpu(
).numpy()[src_word_idx, :]
ori_word_embedding[:, src_word_idx] = adv_vector
c_ = -2 * np.dot(ori_word_embedding.T, adv_vector)
a = np.sum(np.square(ori_word_embedding),
axis=0) #.reshape((1, -1))
b = np.sum(np.square(adv_vector))
dist = a + c_ + b
neighbours, _ = glove_utils.pick_most_similar_words_from_vector(
src_word_idx,
adv_vector,
dist,
ret_count=1,
threshold=5)
word[batch_i][batch_j] = neighbours[0]
for batch_i in range(len(word)):
wf.write(' '.join([
str(word_alphabet.get_instance(wordi))
for wordi in word[batch_i][1:] if not wordi == 1
]))
wf.write('\n')
wf.close()
ori_wf.close()
# with torch.no_grad():
# num_inst = word.size(0) if obj == 'crf' else masks.sum() - word.size(0)
# train_err += loss * num_inst
# train_err_arc += loss_arc * num_inst
# train_err_type += loss_type * num_inst
# train_total += num_inst
#
# time_ave = (time.time() - start_time) / batch
# time_left = (num_batches - batch) * time_ave
#
# # update log
# if batch % 10 == 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, arc: %.4f, type: %.4f, time left: %.2fs' % (batch, num_batches, train_err / train_total,
# train_err_arc / train_total, train_err_type / 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, arc: %.4f, type: %.4f, time: %.2fs' % (num_batches, train_err / train_total,
# train_err_arc / train_total, train_err_type / train_total, time.time() - start_time))
# evaluate performance on dev data
if not TEST_PARSER_FLAG:
continue
with torch.no_grad():
network.eval()
pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch)
pred_writer.start(pred_filename)
gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch)
gold_writer.start(gold_filename)
dev_ucorr = 0.0
dev_lcorr = 0.0
dev_total = 0
dev_ucomlpete = 0.0
dev_lcomplete = 0.0
dev_ucorr_nopunc = 0.0
dev_lcorr_nopunc = 0.0
dev_total_nopunc = 0
dev_ucomlpete_nopunc = 0.0
dev_lcomplete_nopunc = 0.0
dev_root_corr = 0.0
dev_total_root = 0.0
dev_total_inst = 0.0
for batch in conllx_data.iterate_batch_tensor(
data_dev, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.cpu().numpy()
pos = pos.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.cpu().numpy()
types = types.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
dev_ucorr += ucorr
dev_lcorr += lcorr
dev_total += total
dev_ucomlpete += ucm
dev_lcomplete += lcm
dev_ucorr_nopunc += ucorr_nopunc
dev_lcorr_nopunc += lcorr_nopunc
dev_total_nopunc += total_nopunc
dev_ucomlpete_nopunc += ucm_nopunc
dev_lcomplete_nopunc += lcm_nopunc
dev_root_corr += corr_root
dev_total_root += total_root
dev_total_inst += num_inst
pred_writer.close()
gold_writer.close()
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 /
dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete *
100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
%
(dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc *
100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 /
dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' %
(dev_root_corr, dev_total_root,
dev_root_corr * 100 / dev_total_root))
if dev_lcorrect_nopunc < dev_lcorr_nopunc or (
dev_lcorrect_nopunc == dev_lcorr_nopunc
and dev_ucorrect_nopunc < dev_ucorr_nopunc):
dev_ucorrect_nopunc = dev_ucorr_nopunc
dev_lcorrect_nopunc = dev_lcorr_nopunc
dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc
dev_lcomplete_match_nopunc = dev_lcomplete_nopunc
dev_ucorrect = dev_ucorr
dev_lcorrect = dev_lcorr
dev_ucomlpete_match = dev_ucomlpete
dev_lcomplete_match = dev_lcomplete
dev_root_correct = dev_root_corr
best_epoch = epoch
patient = 0
# torch.save(network, model_name)
torch.save(network.state_dict(), model_name)
pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch)
pred_writer.start(pred_filename)
gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch)
gold_writer.start(gold_filename)
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
for batch in conllx_data.iterate_batch_tensor(
data_test, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.cpu().numpy()
pos = pos.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.cpu().numpy()
types = types.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
pred_writer.close()
gold_writer.close()
else:
if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule:
# network = torch.load(model_name)
network.load_state_dict(torch.load(model_name))
lr = lr * decay_rate
optim = generate_optimizer(opt, lr, network.parameters())
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' %
decoding)
patient = 0
decay += 1
if decay % double_schedule_decay == 0:
schedule *= 2
else:
patient += 1
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 /
dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch))
print(
'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst,
best_epoch))
print(
'best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)'
% (dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch))
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (test_ucorrect, test_lcorrect, test_total, test_ucorrect *
100 / test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst,
test_lcomplete_match * 100 / test_total_inst, best_epoch))
print(
'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
%
(test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc,
test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst,
test_lcomplete_match_nopunc * 100 / test_total_inst,
best_epoch))
print(
'best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)'
% (test_root_correct, test_total_root,
test_root_correct * 100 / test_total_root, best_epoch))
print(
'============================================================================================================================'
)
if decay == max_decay:
break
print('Baseline Attack.')
def run_biaffine(model_path, model_name, test_path, punct_set, use_gpu, logger,
args):
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
word_alphabet, char_alphabet, pos_alphabet, \
type_alphabet = conllx_data.create_alphabets(
alphabet_path,
None,
data_paths=[None, None],
max_vocabulary_size=50000,
embedd_dict=None
)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
decoding = args.decode
logger.info('use gpu: %s, decoding: %s' % (use_gpu, decoding))
device = torch.device('cuda') if use_gpu else torch.device('cpu')
data_test = aida_data.read_data_to_tensor(test_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
pred_writer = AIDAWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
logger.info('model: %s' % model_name)
def load_model_arguments_from_json():
arguments = json.load(open(arg_path, 'r'))
return arguments['args'], arguments['kwargs']
arg_path = model_name + '.arg.json'
model_args, model_kwargs = load_model_arguments_from_json()
network = BiRecurrentConvBiAffine(*model_args, **model_kwargs)
if torch.cuda.is_available():
map_location = lambda storage, loc: storage.cuda()
else:
map_location = 'cpu'
network.load_state_dict(torch.load(model_name, map_location=map_location))
if use_gpu:
network.cuda()
else:
network.cpu()
network.eval()
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
pred_writer.start(args.output_path)
sent = 0
start_time = time.time()
with torch.no_grad():
for batch in aida_data.iterate_batch_tensor(data_test, 1):
sys.stdout.write('Processing sentence: %d\n' % sent)
sys.stdout.flush()
sent += 1
word, char, pos, _, _, masks, lengths, segment_ids_words = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
segment_ids, segment_words = zip(*segment_ids_words)
pred_writer.write(segment_ids,
segment_words,
word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
pred_writer.close()
print('\ntime: %.2fs' % (time.time() - start_time))
def main():
args_parser = argparse.ArgumentParser(
description='Tuning with graph-based parsing')
args_parser.register('type', 'bool', str2bool)
args_parser.add_argument('--seed',
type=int,
default=1234,
help='random seed for reproducibility')
args_parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'],
help='architecture of rnn',
required=True)
args_parser.add_argument('--num_epochs',
type=int,
default=1000,
help='Number of training epochs')
args_parser.add_argument('--batch_size',
type=int,
default=64,
help='Number of sentences in each batch')
args_parser.add_argument('--hidden_size',
type=int,
default=256,
help='Number of hidden units in RNN')
args_parser.add_argument('--arc_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--type_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--num_layers',
type=int,
default=1,
help='Number of layers of encoder.')
args_parser.add_argument('--num_filters',
type=int,
default=50,
help='Number of filters in CNN')
args_parser.add_argument('--pos',
action='store_true',
help='use part-of-speech embedding.')
args_parser.add_argument('--char',
action='store_true',
help='use character embedding and CNN.')
args_parser.add_argument('--pos_dim',
type=int,
default=50,
help='Dimension of POS embeddings')
args_parser.add_argument('--char_dim',
type=int,
default=50,
help='Dimension of Character embeddings')
args_parser.add_argument('--opt',
choices=['adam', 'sgd', 'adamax'],
help='optimization algorithm')
args_parser.add_argument('--objective',
choices=['cross_entropy', 'crf'],
default='cross_entropy',
help='objective function of training procedure.')
args_parser.add_argument('--decode',
choices=['mst', 'greedy'],
default='mst',
help='decoding algorithm')
args_parser.add_argument('--learning_rate',
type=float,
default=0.01,
help='Learning rate')
# args_parser.add_argument('--decay_rate', type=float, default=0.05, help='Decay rate of learning rate')
args_parser.add_argument('--clip',
type=float,
default=5.0,
help='gradient clipping')
args_parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for regularization')
args_parser.add_argument('--epsilon',
type=float,
default=1e-8,
help='epsilon for adam or adamax')
args_parser.add_argument('--p_rnn',
nargs='+',
type=float,
required=True,
help='dropout rate for RNN')
args_parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input embeddings')
args_parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
# args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay')
args_parser.add_argument(
'--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
args_parser.add_argument('--punctuation',
nargs='+',
type=str,
help='List of punctuations')
args_parser.add_argument(
'--word_embedding',
choices=['word2vec', 'glove', 'senna', 'sskip', 'polyglot'],
help='Embedding for words',
required=True)
args_parser.add_argument('--word_path',
help='path for word embedding dict')
args_parser.add_argument(
'--freeze',
action='store_true',
help='frozen the word embedding (disable fine-tuning).')
args_parser.add_argument('--char_embedding',
choices=['random', 'polyglot'],
help='Embedding for characters',
required=True)
args_parser.add_argument('--char_path',
help='path for character embedding dict')
args_parser.add_argument('--data_dir', help='Data directory path')
args_parser.add_argument(
'--src_lang',
required=True,
help='Src language to train dependency parsing model')
args_parser.add_argument('--aux_lang',
nargs='+',
help='Language names for adversarial training')
args_parser.add_argument('--vocab_path',
help='path for prebuilt alphabets.',
default=None)
args_parser.add_argument('--model_path',
help='path for saving model file.',
required=True)
args_parser.add_argument('--model_name',
help='name for saving model file.',
required=True)
#
args_parser.add_argument('--attn_on_rnn',
action='store_true',
help='use self-attention on top of context RNN.')
args_parser.add_argument('--no_word',
type='bool',
default=False,
help='do not use word embedding.')
args_parser.add_argument('--use_bert',
type='bool',
default=False,
help='use multilingual BERT.')
#
# lrate schedule with warmup in the first iter.
args_parser.add_argument('--use_warmup_schedule',
type='bool',
default=False,
help="Use warmup lrate schedule.")
args_parser.add_argument('--decay_rate',
type=float,
default=0.75,
help='Decay rate of learning rate')
args_parser.add_argument('--max_decay',
type=int,
default=9,
help='Number of decays before stop')
args_parser.add_argument('--schedule',
type=int,
help='schedule for learning rate decay')
args_parser.add_argument('--double_schedule_decay',
type=int,
default=5,
help='Number of decays to double schedule')
args_parser.add_argument(
'--check_dev',
type=int,
default=5,
help='Check development performance in every n\'th iteration')
# encoder selection
args_parser.add_argument('--encoder_type',
choices=['Transformer', 'RNN', 'SelfAttn'],
default='RNN',
help='do not use context RNN.')
args_parser.add_argument(
'--pool_type',
default='mean',
choices=['max', 'mean', 'weight'],
help='pool type to form fixed length vector from word embeddings')
# Tansformer encoder
args_parser.add_argument(
'--trans_hid_size',
type=int,
default=1024,
help='#hidden units in point-wise feed-forward in transformer')
args_parser.add_argument(
'--d_k',
type=int,
default=64,
help='d_k for multi-head-attention in transformer encoder')
args_parser.add_argument(
'--d_v',
type=int,
default=64,
help='d_v for multi-head-attention in transformer encoder')
args_parser.add_argument('--num_head',
type=int,
default=8,
help='Value of h in multi-head attention')
args_parser.add_argument(
'--use_all_encoder_layers',
type='bool',
default=False,
help='Use a weighted representations of all encoder layers')
# - positional
args_parser.add_argument(
'--enc_use_neg_dist',
action='store_true',
help="Use negative distance for enc's relational-distance embedding.")
args_parser.add_argument(
'--enc_clip_dist',
type=int,
default=0,
help="The clipping distance for relative position features.")
args_parser.add_argument('--position_dim',
type=int,
default=50,
help='Dimension of Position embeddings.')
args_parser.add_argument(
'--position_embed_num',
type=int,
default=200,
help=
'Minimum value of position embedding num, which usually is max-sent-length.'
)
args_parser.add_argument('--train_position',
action='store_true',
help='train positional encoding for transformer.')
args_parser.add_argument('--input_concat_embeds',
action='store_true',
help="Concat input embeddings, otherwise add.")
args_parser.add_argument('--input_concat_position',
action='store_true',
help="Concat position embeddings, otherwise add.")
args_parser.add_argument(
'--partitioned',
type='bool',
default=False,
help=
"Partition the content and positional attention for multi-head attention."
)
args_parser.add_argument(
'--partition_type',
choices=['content-position', 'lexical-delexical'],
default='content-position',
help="How to apply partition in the self-attention.")
#
args_parser.add_argument(
'--train_len_thresh',
type=int,
default=100,
help='In training, discard sentences longer than this.')
#
# regarding adversarial training
args_parser.add_argument('--pre_model_path',
type=str,
default=None,
help='Path of the pretrained model.')
args_parser.add_argument('--pre_model_name',
type=str,
default=None,
help='Name of the pretrained model.')
args_parser.add_argument('--adv_training',
type='bool',
default=False,
help='Use adversarial training.')
args_parser.add_argument(
'--lambdaG',
type=float,
default=0.001,
help='Scaling parameter to control generator loss.')
args_parser.add_argument('--discriminator',
choices=['weak', 'not-so-weak', 'strong'],
default='weak',
help='architecture of the discriminator')
args_parser.add_argument(
'--delay',
type=int,
default=0,
help='Number of epochs to be run first for the source task')
args_parser.add_argument(
'--n_critic',
type=int,
default=5,
help='Number of training steps for discriminator per iter')
args_parser.add_argument(
'--clip_disc',
type=float,
default=5.0,
help='Lower and upper clip value for disc. weights')
args_parser.add_argument('--debug',
type='bool',
default=False,
help='Use debug portion of the training data')
args_parser.add_argument('--train_level',
type=str,
default='word',
choices=['word', 'sent'],
help='Use X-level adversarial training')
args_parser.add_argument('--train_type',
type=str,
default='GAN',
choices=['GR', 'GAN', 'WGAN'],
help='Type of adversarial training')
#
# regarding motivational training
args_parser.add_argument(
'--motivate',
type='bool',
default=False,
help='This is opposite of the adversarial training')
#
args = args_parser.parse_args()
# fix data-prepare seed
random.seed(1234)
np.random.seed(1234)
# model's seed
torch.manual_seed(args.seed)
# if output directory doesn't exist, create it
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
logger = get_logger("GraphParser")
logger.info('\ncommand-line params : {0}\n'.format(sys.argv[1:]))
logger.info('{0}\n'.format(args))
logger.info("Visible GPUs: %s", str(os.environ["CUDA_VISIBLE_DEVICES"]))
args.parallel = False
if torch.cuda.device_count() > 1:
args.parallel = True
mode = args.mode
obj = args.objective
decoding = args.decode
train_path = args.data_dir + args.src_lang + "_train.debug.1_10.conllu" \
if args.debug else args.data_dir + args.src_lang + '_train.conllu'
dev_path = args.data_dir + args.src_lang + "_dev.conllu"
test_path = args.data_dir + args.src_lang + "_test.conllu"
#
vocab_path = args.vocab_path if args.vocab_path is not None else args.model_path
model_path = args.model_path
model_name = args.model_name
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
arc_space = args.arc_space
type_space = args.type_space
num_layers = args.num_layers
num_filters = args.num_filters
learning_rate = args.learning_rate
opt = args.opt
momentum = 0.9
betas = (0.9, 0.9)
eps = args.epsilon
decay_rate = args.decay_rate
clip = args.clip
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
punctuation = args.punctuation
freeze = args.freeze
use_word_emb = not args.no_word
word_embedding = args.word_embedding
word_path = args.word_path
use_char = args.char
char_embedding = args.char_embedding
char_path = args.char_path
attn_on_rnn = args.attn_on_rnn
encoder_type = args.encoder_type
if attn_on_rnn:
assert encoder_type == 'RNN'
t_types = (args.adv_training, args.motivate)
t_count = sum(1 for tt in t_types if tt)
if t_count > 1:
assert False, "Only one of: adv_training or motivate can be true"
# ------------------- Loading/initializing embeddings -------------------- #
use_pos = args.pos
pos_dim = args.pos_dim
word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path)
char_dict = None
char_dim = args.char_dim
if char_embedding != 'random':
char_dict, char_dim = utils.load_embedding_dict(
char_embedding, char_path)
logger.info("Creating Alphabets")
alphabet_path = os.path.join(vocab_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
# TODO (WARNING): must build vocabs previously
assert os.path.isdir(alphabet_path), "should have build vocabs previously"
word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets(
alphabet_path,
train_path,
data_paths=[dev_path, test_path],
max_vocabulary_size=50000,
embedd_dict=word_dict)
max_sent_length = max(max_sent_length, args.position_embed_num)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
# ------------------------------------------------------------------------- #
# --------------------- Loading/building the model ------------------------ #
logger.info("Reading Data")
use_gpu = torch.cuda.is_available()
def construct_word_embedding_table():
scale = np.sqrt(3.0 / word_dim)
table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in word_dict:
embedding = word_dict[word]
elif word.lower() in word_dict:
embedding = word_dict[word.lower()]
else:
embedding = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('word OOV: %d' % oov)
return torch.from_numpy(table)
def construct_char_embedding_table():
if char_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_chars, char_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, char_dim]).astype(np.float32)
oov = 0
for char, index, in char_alphabet.items():
if char in char_dict:
embedding = char_dict[char]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('character OOV: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table() if use_word_emb else None
char_table = construct_char_embedding_table() if use_char else None
def load_model_arguments_from_json():
arguments = json.load(open(pre_model_path, 'r'))
return arguments['args'], arguments['kwargs']
window = 3
if obj == 'cross_entropy':
if args.pre_model_path and args.pre_model_name:
pre_model_name = os.path.join(args.pre_model_path,
args.pre_model_name)
pre_model_path = pre_model_name + '.arg.json'
model_args, kwargs = load_model_arguments_from_json()
network = BiRecurrentConvBiAffine(use_gpu=use_gpu,
*model_args,
**kwargs)
network.load_state_dict(torch.load(pre_model_name))
logger.info("Model reloaded from %s" % pre_model_path)
# Adjust the word embedding layer
if network.embedder.word_embedd is not None:
network.embedder.word_embedd = nn.Embedding(num_words,
word_dim,
_weight=word_table)
else:
network = BiRecurrentConvBiAffine(
word_dim,
num_words,
char_dim,
num_chars,
pos_dim,
num_pos,
num_filters,
window,
mode,
hidden_size,
num_layers,
num_types,
arc_space,
type_space,
embedd_word=word_table,
embedd_char=char_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
biaffine=True,
pos=use_pos,
char=use_char,
train_position=args.train_position,
encoder_type=encoder_type,
trans_hid_size=args.trans_hid_size,
d_k=args.d_k,
d_v=args.d_v,
num_head=args.num_head,
enc_use_neg_dist=args.enc_use_neg_dist,
enc_clip_dist=args.enc_clip_dist,
position_dim=args.position_dim,
max_sent_length=max_sent_length,
use_gpu=use_gpu,
use_word_emb=use_word_emb,
input_concat_embeds=args.input_concat_embeds,
input_concat_position=args.input_concat_position,
attn_on_rnn=attn_on_rnn,
partitioned=args.partitioned,
partition_type=args.partition_type,
use_all_encoder_layers=args.use_all_encoder_layers,
use_bert=args.use_bert)
elif obj == 'crf':
raise NotImplementedError
else:
raise RuntimeError('Unknown objective: %s' % obj)
# ------------------------------------------------------------------------- #
# --------------------- Loading data -------------------------------------- #
train_data = dict()
dev_data = dict()
test_data = dict()
num_data = dict()
lang_ids = dict()
reverse_lang_ids = dict()
# ===== the reading =============================================
def _read_one(path, is_train):
lang_id = guess_language_id(path)
logger.info("Reading: guess that the language of file %s is %s." %
(path, lang_id))
one_data = conllx_data.read_data_to_variable(
path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
use_gpu=False,
volatile=(not is_train),
symbolic_root=True,
lang_id=lang_id,
use_bert=args.use_bert,
len_thresh=(args.train_len_thresh if is_train else 100000))
return one_data
data_train = _read_one(train_path, True)
train_data[args.src_lang] = data_train
num_data[args.src_lang] = sum(data_train[1])
lang_ids[args.src_lang] = len(lang_ids)
reverse_lang_ids[lang_ids[args.src_lang]] = args.src_lang
data_dev = _read_one(dev_path, False)
data_test = _read_one(test_path, False)
dev_data[args.src_lang] = data_dev
test_data[args.src_lang] = data_test
# ===============================================================
# ===== reading data for adversarial training ===================
if t_count > 0:
for language in args.aux_lang:
aux_train_path = args.data_dir + language + "_train.debug.1_10.conllu" \
if args.debug else args.data_dir + language + '_train.conllu'
aux_train_data = _read_one(aux_train_path, True)
num_data[language] = sum(aux_train_data[1])
train_data[language] = aux_train_data
lang_ids[language] = len(lang_ids)
reverse_lang_ids[lang_ids[language]] = language
# ===============================================================
punct_set = None
if punctuation is not None:
punct_set = set(punctuation)
logger.info("punctuations(%d): %s" %
(len(punct_set), ' '.join(punct_set)))
def save_args():
arg_path = model_name + '.arg.json'
arguments = [
word_dim, num_words, char_dim, num_chars, pos_dim, num_pos,
num_filters, window, mode, hidden_size, num_layers, num_types,
arc_space, type_space
]
kwargs = {
'p_in': p_in,
'p_out': p_out,
'p_rnn': p_rnn,
'biaffine': True,
'pos': use_pos,
'char': use_char,
'train_position': args.train_position,
'encoder_type': args.encoder_type,
'trans_hid_size': args.trans_hid_size,
'd_k': args.d_k,
'd_v': args.d_v,
'num_head': args.num_head,
'enc_use_neg_dist': args.enc_use_neg_dist,
'enc_clip_dist': args.enc_clip_dist,
'position_dim': args.position_dim,
'max_sent_length': max_sent_length,
'use_word_emb': use_word_emb,
'input_concat_embeds': args.input_concat_embeds,
'input_concat_position': args.input_concat_position,
'attn_on_rnn': attn_on_rnn,
'partitioned': args.partitioned,
'partition_type': args.partition_type,
'use_all_encoder_layers': args.use_all_encoder_layers,
'use_bert': args.use_bert
}
json.dump({
'args': arguments,
'kwargs': kwargs
},
open(arg_path, 'w'),
indent=4)
if use_word_emb and freeze:
freeze_embedding(network.embedder.word_embedd)
if args.parallel:
network = torch.nn.DataParallel(network)
if use_gpu:
network = network.cuda()
save_args()
param_dict = {}
encoder = network.module.encoder if args.parallel else network.encoder
for name, param in encoder.named_parameters():
if param.requires_grad:
param_dict[name] = np.prod(param.size())
total_params = np.sum(list(param_dict.values()))
logger.info('Total Encoder Parameters = %d' % total_params)
# ------------------------------------------------------------------------- #
# =============================================
if args.adv_training:
disc_feat_size = network.module.encoder.output_dim if args.parallel else network.encoder.output_dim
reverse_grad = args.train_type == 'GR'
nclass = len(lang_ids) if args.train_type == 'GR' else 1
kwargs = {
'input_size': disc_feat_size,
'disc_type': args.discriminator,
'train_level': args.train_level,
'train_type': args.train_type,
'reverse_grad': reverse_grad,
'soft_label': True,
'nclass': nclass,
'scale': args.lambdaG,
'use_gpu': use_gpu,
'opt': 'adam',
'lr': 0.001,
'betas': (0.9, 0.999),
'gamma': 0,
'eps': 1e-8,
'momentum': 0,
'clip_disc': args.clip_disc
}
AdvAgent = Adversarial(**kwargs)
if use_gpu:
AdvAgent.cuda()
elif args.motivate:
disc_feat_size = network.module.encoder.output_dim if args.parallel else network.encoder.output_dim
nclass = len(lang_ids)
kwargs = {
'input_size': disc_feat_size,
'disc_type': args.discriminator,
'train_level': args.train_level,
'nclass': nclass,
'scale': args.lambdaG,
'use_gpu': use_gpu,
'opt': 'adam',
'lr': 0.001,
'betas': (0.9, 0.999),
'gamma': 0,
'eps': 1e-8,
'momentum': 0,
'clip_disc': args.clip_disc
}
MtvAgent = Motivator(**kwargs)
if use_gpu:
MtvAgent.cuda()
# =============================================
# --------------------- Initializing the optimizer ------------------------ #
lr = learning_rate
optim = generate_optimizer(opt, lr, network.parameters(), betas, gamma,
eps, momentum)
opt_info = 'opt: %s, ' % opt
if opt == 'adam':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
elif opt == 'sgd':
opt_info += 'momentum=%.2f' % momentum
elif opt == 'adamax':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
# =============================================
total_data = min(num_data.values())
word_status = 'frozen' if freeze else 'fine tune'
char_status = 'enabled' if use_char else 'disabled'
pos_status = 'enabled' if use_pos else 'disabled'
logger.info(
"Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" %
(word_dim, word_status, char_dim, char_status, pos_dim, pos_status))
logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window))
logger.info(
"RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" %
(mode, num_layers, hidden_size, arc_space, type_space))
logger.info(
"train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)"
% (obj, gamma, total_data, batch_size, clip, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
logger.info("decoding algorithm: %s" % decoding)
logger.info(opt_info)
# ------------------------------------------------------------------------- #
# --------------------- Form the mini-batches ----------------------------- #
num_batches = total_data // batch_size + 1
aux_lang = []
if t_count > 0:
for language in args.aux_lang:
aux_lang.extend([language] * num_data[language])
assert num_data[args.src_lang] <= len(aux_lang)
# ------------------------------------------------------------------------- #
dev_ucorrect = 0.0
dev_lcorrect = 0.0
dev_ucomlpete_match = 0.0
dev_lcomplete_match = 0.0
dev_ucorrect_nopunc = 0.0
dev_lcorrect_nopunc = 0.0
dev_ucomlpete_match_nopunc = 0.0
dev_lcomplete_match_nopunc = 0.0
dev_root_correct = 0.0
best_epoch = 0
if decoding == 'greedy':
decode = network.module.decode if args.parallel else network.decode
elif decoding == 'mst':
decode = network.module.decode_mst if args.parallel else network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
patient = 0
decay = 0
max_decay = args.max_decay
double_schedule_decay = args.double_schedule_decay
# lrate schedule
step_num = 0
use_warmup_schedule = args.use_warmup_schedule
if use_warmup_schedule:
logger.info("Use warmup lrate for the first epoch, from 0 up to %s." %
(lr, ))
skip_adv_tuning = 0
loss_fn = network.module.loss if args.parallel else network.loss
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): '
%
(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay))
train_err = 0.
train_err_arc = 0.
train_err_type = 0.
train_total = 0.
start_time = time.time()
num_back = 0
skip_adv_tuning += 1
loss_d_real, loss_d_fake = [], []
acc_d_real, acc_d_fake, = [], []
gen_loss, parsing_loss = [], []
disent_loss = []
if t_count > 0 and skip_adv_tuning > args.delay:
batch_size = args.batch_size // 2
num_batches = total_data // batch_size + 1
# ---------------------- Sample the mini-batches -------------------------- #
if t_count > 0:
sampled_aux_lang = random.sample(aux_lang, num_batches)
lang_in_batch = [(args.src_lang, sampled_aux_lang[k])
for k in range(num_batches)]
else:
lang_in_batch = [(args.src_lang, None) for _ in range(num_batches)]
assert len(lang_in_batch) == num_batches
# ------------------------------------------------------------------------- #
network.train()
warmup_factor = (lr + 0.) / num_batches
for batch in range(1, num_batches + 1):
update_generator = True
update_discriminator = False
# lrate schedule (before each step)
step_num += 1
if use_warmup_schedule and epoch <= 1:
cur_lrate = warmup_factor * step_num
# set lr
for param_group in optim.param_groups:
param_group['lr'] = cur_lrate
# considering source language as real and auxiliary languages as fake
real_lang, fake_lang = lang_in_batch[batch - 1]
real_idx, fake_idx = lang_ids.get(real_lang), lang_ids.get(
fake_lang, -1)
#
word, char, pos, heads, types, masks, lengths, bert_inputs = conllx_data.get_batch_variable(
train_data[real_lang], batch_size, unk_replace=unk_replace)
if use_gpu:
word = word.cuda()
char = char.cuda()
pos = pos.cuda()
heads = heads.cuda()
types = types.cuda()
masks = masks.cuda()
lengths = lengths.cuda()
if bert_inputs[0] is not None:
bert_inputs[0] = bert_inputs[0].cuda()
bert_inputs[1] = bert_inputs[1].cuda()
bert_inputs[2] = bert_inputs[2].cuda()
real_enc = network(word,
char,
pos,
input_bert=bert_inputs,
mask=masks,
length=lengths,
hx=None)
# ========== Update the discriminator ==========
if t_count > 0 and skip_adv_tuning > args.delay:
# fake examples = 0
word_f, char_f, pos_f, heads_f, types_f, masks_f, lengths_f, bert_inputs = conllx_data.get_batch_variable(
train_data[fake_lang], batch_size, unk_replace=unk_replace)
if use_gpu:
word_f = word_f.cuda()
char_f = char_f.cuda()
pos_f = pos_f.cuda()
heads_f = heads_f.cuda()
types_f = types_f.cuda()
masks_f = masks_f.cuda()
lengths_f = lengths_f.cuda()
if bert_inputs[0] is not None:
bert_inputs[0] = bert_inputs[0].cuda()
bert_inputs[1] = bert_inputs[1].cuda()
bert_inputs[2] = bert_inputs[2].cuda()
fake_enc = network(word_f,
char_f,
pos_f,
input_bert=bert_inputs,
mask=masks_f,
length=lengths_f,
hx=None)
# TODO: temporary crack
if t_count > 0 and skip_adv_tuning > args.delay:
# skip discriminator training for '|n_critic|' iterations if 'n_critic' < 0
if args.n_critic > 0 or (batch - 1) % (-1 *
args.n_critic) == 0:
update_discriminator = True
if update_discriminator:
if args.adv_training:
real_loss, fake_loss, real_acc, fake_acc = AdvAgent.update(
real_enc['output'].detach(),
fake_enc['output'].detach(), real_idx, fake_idx)
loss_d_real.append(real_loss)
loss_d_fake.append(fake_loss)
acc_d_real.append(real_acc)
acc_d_fake.append(fake_acc)
elif args.motivate:
real_loss, fake_loss, real_acc, fake_acc = MtvAgent.update(
real_enc['output'].detach(),
fake_enc['output'].detach(), real_idx, fake_idx)
loss_d_real.append(real_loss)
loss_d_fake.append(fake_loss)
acc_d_real.append(real_acc)
acc_d_fake.append(fake_acc)
else:
raise NotImplementedError()
if args.n_critic > 0 and (batch - 1) % args.n_critic != 0:
update_generator = False
# ==============================================
# =========== Update the generator =============
if update_generator:
others_loss = None
if args.adv_training and skip_adv_tuning > args.delay:
# for GAN: L_G= L_parsing - (lambda_G * L_D)
# for GR : L_G= L_parsing + L_D
others_loss = AdvAgent.gen_loss(real_enc['output'],
fake_enc['output'],
real_idx, fake_idx)
gen_loss.append(others_loss.item())
elif args.motivate and skip_adv_tuning > args.delay:
others_loss = MtvAgent.gen_loss(real_enc['output'],
fake_enc['output'],
real_idx, fake_idx)
gen_loss.append(others_loss.item())
optim.zero_grad()
loss_arc, loss_type = loss_fn(real_enc['output'],
heads,
types,
mask=masks,
length=lengths)
loss = loss_arc + loss_type
num_inst = word.size(
0) if obj == 'crf' else masks.sum() - word.size(0)
train_err += loss.item() * num_inst
train_err_arc += loss_arc.item() * num_inst
train_err_type += loss_type.item() * num_inst
train_total += num_inst
parsing_loss.append(loss.item())
if others_loss is not None:
loss = loss + others_loss
loss.backward()
clip_grad_norm_(network.parameters(), clip)
optim.step()
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
if (args.adv_training
or args.motivate) and skip_adv_tuning > args.delay:
logger.info(
'epoch: %d train: %d loss: %.4f, arc: %.4f, type: %.4f, dis_loss: (%.2f, %.2f), dis_acc: (%.2f, %.2f), '
'gen_loss: %.2f, time: %.2fs' %
(epoch, num_batches, train_err / train_total,
train_err_arc / train_total, train_err_type / train_total,
sum(loss_d_real) / len(loss_d_real), sum(loss_d_fake) /
len(loss_d_fake), sum(acc_d_real) / len(acc_d_real),
sum(acc_d_fake) / len(acc_d_fake),
sum(gen_loss) / len(gen_loss), time.time() - start_time))
else:
logger.info(
'epoch: %d train: %d loss: %.4f, arc: %.4f, type: %.4f, time: %.2fs'
% (epoch, num_batches, train_err / train_total,
train_err_arc / train_total, train_err_type / train_total,
time.time() - start_time))
################# Validation on Dependency Parsing Only #################################
if epoch % args.check_dev != 0:
continue
with torch.no_grad():
# evaluate performance on dev data
network.eval()
dev_ucorr = 0.0
dev_lcorr = 0.0
dev_total = 0
dev_ucomlpete = 0.0
dev_lcomplete = 0.0
dev_ucorr_nopunc = 0.0
dev_lcorr_nopunc = 0.0
dev_total_nopunc = 0
dev_ucomlpete_nopunc = 0.0
dev_lcomplete_nopunc = 0.0
dev_root_corr = 0.0
dev_total_root = 0.0
dev_total_inst = 0.0
for lang, data_dev in dev_data.items():
for batch in conllx_data.iterate_batch_variable(
data_dev, batch_size):
word, char, pos, heads, types, masks, lengths, bert_inputs = batch
if use_gpu:
word = word.cuda()
char = char.cuda()
pos = pos.cuda()
heads = heads.cuda()
types = types.cuda()
masks = masks.cuda()
lengths = lengths.cuda()
if bert_inputs[0] is not None:
bert_inputs[0] = bert_inputs[0].cuda()
bert_inputs[1] = bert_inputs[1].cuda()
bert_inputs[2] = bert_inputs[2].cuda()
heads_pred, types_pred = decode(
word,
char,
pos,
input_bert=bert_inputs,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.cpu().numpy()
pos = pos.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.cpu().numpy()
types = types.cpu().numpy()
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
dev_ucorr += ucorr
dev_lcorr += lcorr
dev_total += total
dev_ucomlpete += ucm
dev_lcomplete += lcm
dev_ucorr_nopunc += ucorr_nopunc
dev_lcorr_nopunc += lcorr_nopunc
dev_total_nopunc += total_nopunc
dev_ucomlpete_nopunc += ucm_nopunc
dev_lcomplete_nopunc += lcm_nopunc
dev_root_corr += corr_root
dev_total_root += total_root
dev_total_inst += num_inst
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 /
dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete *
100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
%
(dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc *
100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 /
dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' %
(dev_root_corr, dev_total_root,
dev_root_corr * 100 / dev_total_root))
if dev_lcorrect_nopunc < dev_lcorr_nopunc or (
dev_lcorrect_nopunc == dev_lcorr_nopunc
and dev_ucorrect_nopunc < dev_ucorr_nopunc):
dev_ucorrect_nopunc = dev_ucorr_nopunc
dev_lcorrect_nopunc = dev_lcorr_nopunc
dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc
dev_lcomplete_match_nopunc = dev_lcomplete_nopunc
dev_ucorrect = dev_ucorr
dev_lcorrect = dev_lcorr
dev_ucomlpete_match = dev_ucomlpete
dev_lcomplete_match = dev_lcomplete
dev_root_correct = dev_root_corr
best_epoch = epoch
patient = 0
state_dict = network.module.state_dict(
) if args.parallel else network.state_dict()
torch.save(state_dict, model_name)
else:
if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule:
state_dict = torch.load(model_name)
if args.parallel:
network.module.load_state_dict(state_dict)
else:
network.load_state_dict(state_dict)
lr = lr * decay_rate
optim = generate_optimizer(opt, lr, network.parameters(),
betas, gamma, eps, momentum)
if decoding == 'greedy':
decode = network.module.decode if args.parallel else network.decode
elif decoding == 'mst':
decode = network.module.decode_mst if args.parallel else network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' %
decoding)
patient = 0
decay += 1
if decay % double_schedule_decay == 0:
schedule *= 2
else:
patient += 1
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 /
dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch))
print(
'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst,
best_epoch))
print(
'best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)'
% (dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch))
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
if decay == max_decay:
break
torch.cuda.empty_cache() # release memory that can be released
def biaffine(model_path, model_name, test_path, punct_set, use_gpu, logger, args):
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
word_alphabet, char_alphabet, pos_alphabet, \
type_alphabet = conllx_data.create_alphabets(alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
decoding = args.decode
logger.info('use gpu: %s, decoding: %s' % (use_gpu, decoding))
data_test = conllx_data.read_data_to_variable(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet,
use_gpu=use_gpu, volatile=True, symbolic_root=True)
pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
logger.info('model: %s' % model_name)
def load_model_arguments_from_json():
arguments = json.load(open(arg_path, 'r'))
return arguments['args'], arguments['kwargs']
arg_path = model_name + '.arg.json'
args, kwargs = load_model_arguments_from_json()
network = BiRecurrentConvBiAffine(*args, **kwargs)
network.load_state_dict(torch.load(model_name))
if use_gpu:
network.cuda()
else:
network.cpu()
network.eval()
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
pred_writer.start('tmp/analyze_pred_%s' % str(uid))
gold_writer.start('tmp/analyze_gold_%s' % str(uid))
sent = 0
start_time = time.time()
for batch in conllx_data.iterate_batch_variable(data_test, 1):
sys.stdout.write('%d, ' % sent)
sys.stdout.flush()
sent += 1
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True)
gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types, word_alphabet, pos_alphabet, lengths, punct_set=punct_set, symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
pred_writer.close()
gold_writer.close()
print('\ntime: %.2fs' % (time.time() - start_time))
print('test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst))
print('test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst))
print('test Root: corr: %d, total: %d, acc: %.2f%%' % (
test_root_correct, test_total_root, test_root_correct * 100 / test_total_root))
def train_main(args):
logger = get_logger("GraphParser")
mode = args.model_type
obj = args.objective
decoding = args.decode
train_path = args.train_data
dev_path = args.dev_data
test_path = args.test_data
model_path = args.model_path
model_name = args.model_name
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
arc_space = args.arc_space
type_space = args.type_space
num_layers = args.num_layers
num_filters = args.num_filters
learning_rate = args.learning_rate
opt = args.opt
momentum = 0.9
betas = (0.9, 0.9)
eps = args.epsilon
decay_rate = args.decay_rate
clip = args.clip
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
punctuation = args.punctuation
freeze = args.freeze
word_embedding = args.word_embedding
word_path = args.word_path
use_char = args.char
char_embedding = args.char_embedding
char_path = args.char_path
use_pos = args.pos
pos_dim = args.pos_dim
word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path)
char_dict = None
char_dim = args.char_dim
if char_embedding != 'random':
char_dict, char_dim = utils.load_embedding_dict(
char_embedding, char_path)
logger.info("Creating Alphabets")
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet) = conllx_data.create_alphabets(
alphabet_path,
train_path,
data_paths=[dev_path, test_path],
max_vocabulary_size=100000,
embedd_dict=word_dict)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
logger.info("Reading Data")
device = torch.device('cuda') if args.gpu else torch.device('cpu')
data_train = conllx_data.read_data_to_tensor(train_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
# data_train = conllx_data.read_data(train_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
# num_data = sum([len(bucket) for bucket in data_train])
num_data = sum(data_train[1])
data_dev = conllx_data.read_data_to_tensor(dev_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
data_test = conllx_data.read_data_to_tensor(test_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
symbolic_root=True,
device=device)
punct_set = None
if punctuation is not None:
punct_set = set(punctuation)
logger.info("punctuations(%d): %s" %
(len(punct_set), ' '.join(punct_set)))
def construct_word_embedding_table():
scale = np.sqrt(3.0 / word_dim)
table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32)
if freeze:
table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim
]).astype(np.float32)
else:
table[conllx_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in word_dict:
embedding = word_dict[word]
elif word.lower() in word_dict:
embedding = word_dict[word.lower()]
else:
if freeze:
embedding = np.zeros([1, word_dim]).astype(np.float32)
else:
embedding = np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('word OOV: %d' % oov)
return torch.from_numpy(table)
def construct_char_embedding_table():
if char_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_chars, char_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, char_dim]).astype(np.float32)
oov = 0
for char, index, in char_alphabet.items():
if char in char_dict:
embedding = char_dict[char]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('character OOV: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
char_table = construct_char_embedding_table()
window = 3
if obj == 'cross_entropy':
network = BiRecurrentConvBiAffine(word_dim,
num_words,
char_dim,
num_chars,
pos_dim,
num_pos,
num_filters,
window,
mode,
hidden_size,
num_layers,
num_types,
arc_space,
type_space,
embedd_word=word_table,
embedd_char=char_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
biaffine=True,
pos=use_pos,
char=use_char)
elif obj == 'crf':
raise NotImplementedError
else:
raise RuntimeError('Unknown objective: %s' % obj)
def save_args():
arg_path = model_name + '.arg.json'
arguments = [
word_dim, num_words, char_dim, num_chars, pos_dim, num_pos,
num_filters, window, mode, hidden_size, num_layers, num_types,
arc_space, type_space
]
kwargs = {
'p_in': p_in,
'p_out': p_out,
'p_rnn': p_rnn,
'biaffine': True,
'pos': use_pos,
'char': use_char
}
json.dump({
'args': arguments,
'kwargs': kwargs
},
open(arg_path, 'w'),
indent=4)
if freeze:
freeze_embedding(network.word_embedd)
network = network.to(device)
save_args()
# pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
# gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
def generate_optimizer(opt, lr, params):
params = filter(lambda param: param.requires_grad, params)
if opt == 'adam':
return Adam(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
elif opt == 'sgd':
return SGD(params,
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
elif opt == 'adamax':
return Adamax(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
else:
raise ValueError('Unknown optimization algorithm: %s' % opt)
lr = learning_rate
optim = generate_optimizer(opt, lr, network.parameters())
opt_info = 'opt: %s, ' % opt
if opt == 'adam':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
elif opt == 'sgd':
opt_info += 'momentum=%.2f' % momentum
elif opt == 'adamax':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
word_status = 'frozen' if freeze else 'fine tune'
char_status = 'enabled' if use_char else 'disabled'
pos_status = 'enabled' if use_pos else 'disabled'
logger.info(
"Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" %
(word_dim, word_status, char_dim, char_status, pos_dim, pos_status))
logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window))
logger.info(
"RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" %
(mode, num_layers, hidden_size, arc_space, type_space))
logger.info(
"train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)"
% (obj, gamma, num_data, batch_size, clip, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
logger.info("decoding algorithm: %s" % decoding)
logger.info(opt_info)
num_batches = num_data / batch_size + 1
dev_ucorrect = 0.0
dev_lcorrect = 0.0
dev_ucomlpete_match = 0.0
dev_lcomplete_match = 0.0
dev_ucorrect_nopunc = 0.0
dev_lcorrect_nopunc = 0.0
dev_ucomlpete_match_nopunc = 0.0
dev_lcomplete_match_nopunc = 0.0
dev_root_correct = 0.0
best_epoch = 0
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_root_correct = 0.0
test_total = 0
test_total_nopunc = 0
test_total_inst = 0
test_total_root = 0
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
patient = 0
decay = 0
max_decay = 9
double_schedule_decay = 5
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s, optim: %s, lr=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): '
%
(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay))
train_err = 0.
train_err_arc = 0.
train_err_type = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
for batch in range(1, num_batches + 1):
word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_tensor(
data_train, batch_size, unk_replace=unk_replace)
optim.zero_grad()
loss_arc, loss_type = network.loss(word,
char,
pos,
heads,
types,
mask=masks,
length=lengths)
loss = loss_arc + loss_type
loss.backward()
clip_grad_norm_(network.parameters(), clip)
optim.step()
with torch.no_grad():
num_inst = word.size(
0) if obj == 'crf' else masks.sum() - word.size(0)
train_err += loss * num_inst
train_err_arc += loss_arc * num_inst
train_err_type += loss_type * num_inst
train_total += num_inst
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
# update log
if batch % 10 == 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, arc: %.4f, type: %.4f, time left: %.2fs' % (
batch, num_batches, train_err / train_total, train_err_arc
/ train_total, train_err_type / 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, arc: %.4f, type: %.4f, time: %.2fs' %
(num_batches, train_err / train_total, train_err_arc / train_total,
train_err_type / train_total, time.time() - start_time))
# evaluate performance on dev data
with torch.no_grad():
network.eval()
# pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch)
# pred_writer.start(pred_filename)
# gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch)
# gold_writer.start(gold_filename)
dev_ucorr = 0.0
dev_lcorr = 0.0
dev_total = 0
dev_ucomlpete = 0.0
dev_lcomplete = 0.0
dev_ucorr_nopunc = 0.0
dev_lcorr_nopunc = 0.0
dev_total_nopunc = 0
dev_ucomlpete_nopunc = 0.0
dev_lcomplete_nopunc = 0.0
dev_root_corr = 0.0
dev_total_root = 0.0
dev_total_inst = 0.0
for batch in conllx_data.iterate_batch_tensor(
data_dev, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.cpu().numpy()
pos = pos.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.cpu().numpy()
types = types.cpu().numpy()
# pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True)
# gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
dev_ucorr += ucorr
dev_lcorr += lcorr
dev_total += total
dev_ucomlpete += ucm
dev_lcomplete += lcm
dev_ucorr_nopunc += ucorr_nopunc
dev_lcorr_nopunc += lcorr_nopunc
dev_total_nopunc += total_nopunc
dev_ucomlpete_nopunc += ucm_nopunc
dev_lcomplete_nopunc += lcm_nopunc
dev_root_corr += corr_root
dev_total_root += total_root
dev_total_inst += num_inst
# pred_writer.close()
# gold_writer.close()
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 /
dev_total, dev_lcorr * 100 / dev_total, dev_ucomlpete *
100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
%
(dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc *
100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 /
dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' %
(dev_root_corr, dev_total_root,
dev_root_corr * 100 / dev_total_root))
if (dev_lcorrect_nopunc < dev_lcorr_nopunc
or (dev_lcorrect_nopunc == dev_lcorr_nopunc
and dev_ucorrect_nopunc < dev_ucorr_nopunc)):
dev_ucorrect_nopunc = dev_ucorr_nopunc
dev_lcorrect_nopunc = dev_lcorr_nopunc
dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc
dev_lcomplete_match_nopunc = dev_lcomplete_nopunc
dev_ucorrect = dev_ucorr
dev_lcorrect = dev_lcorr
dev_ucomlpete_match = dev_ucomlpete
dev_lcomplete_match = dev_lcomplete
dev_root_correct = dev_root_corr
best_epoch = epoch
patient = 0
# torch.save(network, model_name)
torch.save(network.state_dict(), model_name)
# pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch)
# pred_writer.start(pred_filename)
# gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch)
# gold_writer.start(gold_filename)
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
for batch in conllx_data.iterate_batch_tensor(
data_test, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.cpu().numpy()
pos = pos.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.cpu().numpy()
types = types.cpu().numpy()
# pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True)
# gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True)
(stats, stats_nopunc, stats_root,
num_inst) = parser.eval(word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
# pred_writer.close()
# gold_writer.close()
else:
if (dev_ucorr_nopunc * 100 / dev_total_nopunc <
dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5
or patient >= schedule):
# network = torch.load(model_name)
network.load_state_dict(torch.load(model_name))
lr = lr * decay_rate
optim = generate_optimizer(opt, lr, network.parameters())
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' %
decoding)
patient = 0
decay += 1
if decay % double_schedule_decay == 0:
schedule *= 2
else:
patient += 1
print('-' * 124)
print(
'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect, dev_lcorrect, dev_total, dev_ucorrect * 100 /
dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch))
print(
'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst,
best_epoch))
print(
'best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)'
% (dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch))
print('-' * 124)
print(
'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (test_ucorrect, test_lcorrect, test_total, test_ucorrect *
100 / test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst,
test_lcomplete_match * 100 / test_total_inst, best_epoch))
print(
'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
%
(test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc,
test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst,
test_lcomplete_match_nopunc * 100 / test_total_inst,
best_epoch))
print(
'best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)'
% (test_root_correct, test_total_root,
test_root_correct * 100 / test_total_root, best_epoch))
print('=' * 124)
if decay == max_decay:
break
def biaffine(model_path, model_name, test_path, punct_set, use_gpu, logger, args):
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets(alphabet_path,
None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None)
# word_alphabet, char_alphabet, pos_alphabet, type_alphabet = create_alphabets(alphabet_path,
# None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
decoding = args.decode
out_filename = args.out_filename
constraints_method = args.constraints_method
constraintFile = args.constraint_file
ratioFile = args.ratio_file
tolerance = args.tolerance
gamma = args.gamma
the_language = args.mt_log[9:11]
mt_log = open(args.mt_log, 'a')
summary_log = open(args.summary_log, 'a')
logger.info('use gpu: %s, decoding: %s' % (use_gpu, decoding))
#
extra_embeds_arr = augment_with_extra_embedding(word_alphabet, args.extra_embed, args.extra_embed_src, test_path, logger)
# ===== the reading
def _read_one(path, is_train):
lang_id = guess_language_id(path)
logger.info("Reading: guess that the language of file %s is %s." % (path, lang_id))
one_data = conllx_data.read_data_to_variable(path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet, use_gpu=use_gpu, volatile=(not is_train), symbolic_root=True, lang_id=lang_id)
return one_data
data_test = _read_one(test_path, False)
# data_test = conllx_data.read_data_to_variable(test_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet,
# use_gpu=use_gpu, volatile=True, symbolic_root=True)
pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
logger.info('model: %s' % model_name)
def load_model_arguments_from_json():
arguments = json.load(open(arg_path, 'r'))
return arguments['args'], arguments['kwargs']
arg_path = model_name + '.arg.json'
args, kwargs = load_model_arguments_from_json()
network = BiRecurrentConvBiAffine(use_gpu=use_gpu, *args, **kwargs)
network.load_state_dict(torch.load(model_name))
#
augment_network_embed(word_alphabet.size(), network, extra_embeds_arr)
if use_gpu:
network.cuda()
else:
network.cpu()
network.eval()
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
elif decoding == 'proj':
decode = network.decode_proj
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
# pred_writer.start('tmp/analyze_pred_%s' % str(uid))
# gold_writer.start('tmp/analyze_gold_%s' % str(uid))
# pred_writer.start(model_path + out_filename + '_pred')
# gold_writer.start(model_path + out_filename + '_gold')
pred_writer.start(out_filename + '_pred')
gold_writer.start(out_filename + '_gold')
sent = 0
start_time = time.time()
constraints = []
mt_log.write("=====================%s, Ablation 2================\n"%(constraints_method))
summary_log.write("==========================%s, Ablation 2=============\n"%(constraints_method))
if ratioFile == 'WALS':
import pickle as pk
cFile = open(constraintFile, 'rb')
WALS_data = pk.load(cFile)
for idx in ['85A', '87A', '89A']:
constraint = Constraint(0,0,0)
extra_const = constraint.load_WALS(idx, WALS_data[the_language][idx], pos_alphabet, method=constraints_method)
constraints.append(constraint)
if extra_const:
constraints.append(extra_const)
constraint = Constraint(0,0,0)
extra_const = constraint.load_WALS_unary(WALS_data[the_language], pos_alphabet, method=constraints_method)
if extra_const:
constraints.append(extra_const)
constraints.append(constraint)
elif ratioFile == 'None':
summary_log.write("=================No it is baseline================\n")
mt_log.write("==================No it is baseline==============\n")
else:
cFile = open(constraintFile, 'r')
for line in cFile:
if len(line.strip()) < 2:
break
pos1, pos2 = line.strip().split('\t')
constraint = Constraint(0,0,0)
constraint.load(pos1, pos2, ratioFile, pos_alphabet)
constraints.append(constraint)
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
arc_list = []
type_list = []
length_list = []
pos_list = []
for batch in conllx_data.iterate_batch_variable(data_test, 1):
word, char, pos, heads, types, masks, lengths = batch
out_arc, out_type, length = network.pretrain_constraint(word, char, pos, mask=masks, length=lengths, leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
arc_list += list(out_arc)
type_list += list(out_type)
length_list += list(length)
pos_list += list(pos)
if constraints_method == 'binary':
train_constraints = network.binary_constraints
if constraints_method == 'Lagrange':
train_constraints = network.Lagrange_constraints
if constraints_method == 'PR':
train_constraints = network.PR_constraints
train_constraints(arc_list, type_list, length_list, pos_list, constraints, tolerance, mt_log, gamma=gamma)
for batch in conllx_data.iterate_batch_variable(data_test, 1):
#sys.stdout.write('%d, ' % sent)
#sys.stdout.flush()
sent += 1
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(word, char, pos, mask=masks, length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS, constraints=constraints, method=constraints_method, gamma=gamma)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word, pos, heads_pred, types_pred, lengths, symbolic_root=True)
gold_writer.write(word, pos, heads, types, lengths, symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(word, pos, heads_pred, types_pred, heads, types,
word_alphabet, pos_alphabet, lengths,
punct_set=punct_set, symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
print('\ntime: %.2fs' % (time.time() - start_time))
print('test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst))
print('test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst, test_lcomplete_match_nopunc * 100 / test_total_inst))
print('test Root: corr: %d, total: %d, acc: %.2f%%' % (
test_root_correct, test_total_root, test_root_correct * 100 / test_total_root))
mt_log.write('uas: %.2f, las: %.2f\n'%(test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc))
summary_log.write('%s: %.2f %.2f\n'%(the_language, test_ucorrect_nopunc * 100 / test_total_nopunc, test_lcorrect_nopunc * 100 / test_total_nopunc))
pred_writer.close()
gold_writer.close()
class BiaffineModel(object):
def __init__(self, model_path, model_name):
print("................................................")
print("LOADING Biaffine Model")
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
self.word_alpha, self.char_alpha, self.tag_alpha, self.type_alpha = conllx_data.create_alphabets(alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None)
self.id2word = {v: k for k, v in self.word_alpha.instance2index.iteritems()}
num_words = self.word_alpha.size()
num_chars = self.char_alpha.size()
num_pos = self.tag_alpha.size()
num_types = self.type_alpha.size()
print("Word Alphabet Size: %d" % num_words)
print("Character Alphabet Size: %d" % num_chars)
print("POS Alphabet Size: %d" % num_pos)
print("Type Alphabet Size: %d" % num_types)
def load_model_arguments_from_json():
arguments = json.load(open(arg_path, 'r'))
return arguments['args'], arguments['kwargs']
arg_path = model_name + '.arg.json'
args, kwargs = load_model_arguments_from_json()
self.network = BiRecurrentConvBiAffine(*args, **kwargs)
self.network.load_state_dict(torch.load(model_name))
self.network.id2word = self.id2word
self.network.cuda()
self.network.eval()
def prepare_data(self, sentences, use_gpu=True):
ret_value = []
for sentence in sentences:
inst_size = sentence.length()
data = None
max_len = 0
bucket = 0
for bucket_size in _buckets:
if inst_size < bucket_size:
bucket = bucket_size
data = [sentence.word_ids, sentence.seq_char_ids, sentence.tag_ids]
max_len = max([len(seq_char) for seq_char in sentence.seq_chars])
break
if data is None: # meaning the sentence is too long, we cut it into 300 length
bucket = _buckets[-1]
data = [sentence.word_ids[:bucket], sentence.seq_char_ids[:bucket], sentence.tag_ids[:bucket]]
max_len = max([len(seq_char) for seq_char in sentence.seq_chars])
char_length = min(utils.MAX_CHAR_LENGTH, max_len + utils.NUM_CHAR_PAD)
wid_inputs = np.empty([1, bucket], dtype=np.int64)
cid_inputs = np.empty([1, bucket, char_length], dtype=np.int64)
pid_inputs = np.empty([1, bucket], dtype=np.int64)
masks = np.zeros([1, bucket], dtype=np.float32)
single = np.zeros([1, bucket], dtype=np.int64)
lengths = np.empty(bucket, dtype=np.int64)
wids = data[0]
cid_seqs = data[1]
pids = data[2]
inst_size = len(wids)
lengths[0] = inst_size
# word ids
wid_inputs[0, :inst_size] = wids
wid_inputs[0, inst_size:] = PAD_ID_WORD
for c, cids in enumerate(cid_seqs):
limit = len(cids)
if limit > char_length: limit = char_length
try:
cid_inputs[0, c, :limit] = cids[:limit]
cid_inputs[0, c, limit:] = PAD_ID_CHAR
except:
import ipdb; ipdb.set_trace()
cid_inputs[0, inst_size:, :] = PAD_ID_CHAR
# pos ids
pid_inputs[0, :inst_size] = pids
pid_inputs[0, inst_size:] = PAD_ID_TAG
# masks
masks[0, :inst_size] = 1.0
for j, wid in enumerate(wids):
if self.word_alpha.is_singleton(wid):
single[0, j] = 1
words = Variable(torch.from_numpy(wid_inputs), volatile=False)
chars = Variable(torch.from_numpy(cid_inputs), volatile=False)
pos = Variable(torch.from_numpy(pid_inputs), volatile=False)
masks = Variable(torch.from_numpy(masks), volatile=False)
single = Variable(torch.from_numpy(single), volatile=False)
lengths = torch.from_numpy(lengths)
if use_gpu:
words = words.cuda()
chars = chars.cuda()
pos = pos.cuda()
masks = masks.cuda()
single = single.cuda()
lengths = lengths.cuda()
index = slice(0,1)
ret_value.append((words[index], chars[index], pos[index], masks[index], lengths[index], sentence.words, sentence.edu_ids))
return ret_value
def get_syntax_feature(self, data_test, sentences):
sent = 0
syntax_features = []
for data in data_test:
cur_length = len(sentences[sent].words)
word, char, pos, masks, lengths, original_words, edu_ids = data
sent += 1
syntax_feature = self.network.get_syntax_feature(original_words, word, char, pos, mask=masks, length=lengths)
_ , sent_len, dim = syntax_feature.shape
if sent_len != cur_length:
assert sent_len < cur_length
diff = cur_length - sent_len
zeros = Variable(torch.zeros(1, diff, dim)).type(torch.FloatTensor).cuda()
syntax_feature = torch.cat([syntax_feature, zeros], dim=1)
syntax_features.append(syntax_feature)
return syntax_features
def main():
args_parser = argparse.ArgumentParser(
description='Tuning with stack pointer parser')
args_parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'],
help='architecture of rnn',
required=True)
args_parser.add_argument('--num_epochs',
type=int,
default=200,
help='Number of training epochs')
args_parser.add_argument('--batch_size',
type=int,
default=64,
help='Number of sentences in each batch')
#args_parser.add_argument('--decoder_input_size', type=int, default=256, help='Number of input units in decoder RNN.')
args_parser.add_argument('--hidden_size',
type=int,
default=256,
help='Number of hidden units in RNN')
args_parser.add_argument('--arc_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--type_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--encoder_layers',
type=int,
default=1,
help='Number of layers of encoder RNN')
#args_parser.add_argument('--decoder_layers', type=int, default=1, help='Number of layers of decoder RNN')
args_parser.add_argument('--num_filters',
type=int,
default=50,
help='Number of filters in CNN')
# NOTE: action='store_true' is just to set ON
args_parser.add_argument('--pos',
action='store_true',
help='use part-of-speech embedding.')
args_parser.add_argument('--char',
action='store_true',
help='use character embedding and CNN.')
args_parser.add_argument('--pos_dim',
type=int,
default=50,
help='Dimension of POS embeddings')
args_parser.add_argument('--char_dim',
type=int,
default=50,
help='Dimension of Character embeddings')
# NOTE: arg MUST be one of choices(when specified)
args_parser.add_argument('--opt',
choices=['adam', 'sgd', 'adamax'],
help='optimization algorithm')
args_parser.add_argument('--learning_rate',
type=float,
default=0.001,
help='Learning rate')
args_parser.add_argument('--decay_rate',
type=float,
default=0.75,
help='Decay rate of learning rate')
args_parser.add_argument('--max_decay',
type=int,
default=9,
help='Number of decays before stop')
args_parser.add_argument('--double_schedule_decay',
type=int,
default=5,
help='Number of decays to double schedule')
args_parser.add_argument('--clip',
type=float,
default=1.0,
help='gradient clipping')
args_parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for regularization')
args_parser.add_argument('--epsilon',
type=float,
default=1e-8,
help='epsilon for adam or adamax')
args_parser.add_argument('--coverage',
type=float,
default=0.0,
help='weight for coverage loss')
args_parser.add_argument('--p_rnn',
nargs=2,
type=float,
required=True,
help='dropout rate for RNN')
args_parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input embeddings')
args_parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
args_parser.add_argument('--label_smooth',
type=float,
default=1.0,
help='weight of label smoothing method')
args_parser.add_argument('--skipConnect',
action='store_true',
help='use skip connection for decoder RNN.')
args_parser.add_argument('--grandPar',
action='store_true',
help='use grand parent.')
args_parser.add_argument('--sibling',
action='store_true',
help='use sibling.')
args_parser.add_argument(
'--prior_order',
choices=['inside_out', 'left2right', 'deep_first', 'shallow_first'],
help='prior order of children.',
required=True)
args_parser.add_argument('--schedule',
type=int,
help='schedule for learning rate decay')
args_parser.add_argument(
'--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
args_parser.add_argument('--punctuation',
nargs='+',
type=str,
help='List of punctuations')
args_parser.add_argument('--beam',
type=int,
default=1,
help='Beam size for decoding')
args_parser.add_argument(
'--word_embedding',
choices=['glove', 'senna', 'sskip', 'polyglot', 'NNLM'],
help='Embedding for words',
required=True)
args_parser.add_argument('--word_path',
help='path for word embedding dict')
args_parser.add_argument(
'--freeze',
action='store_true',
help='frozen the word embedding (disable fine-tuning).')
args_parser.add_argument('--char_embedding',
choices=['random', 'polyglot'],
help='Embedding for characters',
required=True)
args_parser.add_argument('--char_path',
help='path for character embedding dict')
args_parser.add_argument(
'--train') # "data/POS-penn/wsj/split1/wsj1.train.original"
args_parser.add_argument(
'--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original"
args_parser.add_argument(
'--test') # "data/POS-penn/wsj/split1/wsj1.test.original"
args_parser.add_argument('--model_path',
help='path for saving model file.',
required=True)
args_parser.add_argument('--model_name',
help='name for saving model file.',
required=True)
# TODO: to include in logging process
args_parser.add_argument('--pos_embedding',
choices=[1, 2, 4],
type=int,
help='Embedding method for korean POS tag',
default=2)
args_parser.add_argument('--pos_path', help='path for pos embedding dict')
args_parser.add_argument('--elmo',
action='store_true',
help='use elmo embedding.')
args_parser.add_argument('--elmo_path',
help='path for elmo embedding model.')
args_parser.add_argument('--elmo_dim',
type=int,
help='dimension for elmo embedding model')
#args_parser.add_argument('--fine_tune_path', help='fine tune starting from this state_dict')
args_parser.add_argument('--model_version',
help='previous model version to load')
#hoon : bert
args_parser.add_argument(
'--bert', action='store_true',
help='use elmo embedding.') # true if use bert(hoon)
args_parser.add_argument(
'--etri_train',
help='path for etri data of bert') # etri train path(hoon)
args_parser.add_argument(
'--etri_dev', help='path for etri data of bert') # etri dev path(hoon)
args_parser.add_argument('--bert_path',
help='path for bert embedding model.') # yjyj
args_parser.add_argument('--bert_dim',
type=int,
help='dimension for bert embedding model') # yjyj
args_parser.add_argument('--bert_learning_rate',
type=float,
default=5e-5,
help='Bert Learning rate')
args_parser.add_argument('--decode',
choices=['mst', 'greedy'],
help='decoding algorithm',
required=True) #yj
args_parser.add_argument('--objective',
choices=['cross_entropy', 'crf'],
default='cross_entropy',
help='objective function of training procedure.')
args = args_parser.parse_args()
logger = get_logger("PtrParser")
mode = args.mode
train_path = args.train
dev_path = args.dev
test_path = args.test
model_path = args.model_path + uid + '/' # for numerous experiments
model_name = args.model_name
num_epochs = args.num_epochs
batch_size = args.batch_size
#input_size_decoder = args.decoder_input_size
hidden_size = args.hidden_size
arc_space = args.arc_space
type_space = args.type_space
encoder_layers = args.encoder_layers
#decoder_layers = args.decoder_layers
num_filters = args.num_filters
learning_rate = args.learning_rate
opt = args.opt
momentum = 0.9
betas = (0.9, 0.9)
eps = args.epsilon
decay_rate = args.decay_rate
clip = args.clip
gamma = args.gamma
cov = args.coverage
schedule = args.schedule
p_rnn = tuple(args.p_rnn)
p_in = args.p_in
p_out = args.p_out
label_smooth = args.label_smooth
unk_replace = args.unk_replace
prior_order = args.prior_order
skipConnect = args.skipConnect
grandPar = args.grandPar
sibling = args.sibling
beam = args.beam
punctuation = args.punctuation
freeze = args.freeze
word_embedding = args.word_embedding
word_path = args.word_path
use_char = args.char
char_embedding = args.char_embedding
# QUESTION: pretrained vector for char?
char_path = args.char_path
use_pos = False
pos_embedding = args.pos_embedding
pos_path = args.pos_path
pos_dict = None
pos_dim = args.pos_dim # NOTE pretrain 있을 경우 pos_dim은 그거 따라감
if pos_path is not None:
pos_dict, pos_dim = utils.load_embedding_dict(
word_embedding,
pos_path) # NOTE 임시적으로 word_embedding(NNLM)이랑 같은 형식
word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path)
char_dict = None
char_dim = args.char_dim
if char_embedding != 'random':
char_dict, char_dim = utils.load_embedding_dict(
char_embedding, char_path)
use_elmo = args.elmo
elmo_path = args.elmo_path
elmo_dim = args.elmo_dim
#fine_tune_path = args.fine_tune_path
#bert(hoon)
use_bert = args.bert
#bert yj
bert_path = args.bert_path
bert_dim = args.bert_dim
bert_lr = args.bert_learning_rate
etri_train_path = args.etri_train
etri_dev_path = args.etri_dev
obj = args.objective
decoding = args.decode
logger.info("Creating Alphabets")
alphabet_path = os.path.join(model_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
# min_occurence=1
data_paths = [dev_path, test_path] if test_path else [dev_path]
word_alphabet, char_alphabet, pos_alphabet, type_alphabet = conllx_stacked_data.create_alphabets(
alphabet_path,
train_path,
data_paths=data_paths,
max_vocabulary_size=50000,
pos_embedding=pos_embedding,
embedd_dict=word_dict)
num_words = word_alphabet.size() # 30268
num_chars = char_alphabet.size() # 3545
num_pos = pos_alphabet.size() # 46
num_types = type_alphabet.size() # 39
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
logger.info("Reading Data")
use_gpu = torch.cuda.is_available()
# data is a list of tuple containing tensors, etc ...
data_train = conllx_stacked_data.read_stacked_data_to_variable(
train_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
pos_embedding,
use_gpu=1,
prior_order=prior_order,
elmo=use_elmo,
bert=use_bert,
etri_path=etri_train_path)
num_data = sum(data_train[2])
data_dev = conllx_stacked_data.read_stacked_data_to_variable(
dev_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
pos_embedding,
use_gpu=use_gpu,
volatile=True,
prior_order=prior_order,
elmo=use_elmo,
bert=use_bert,
etri_path=etri_dev_path)
if test_path:
data_test = conllx_stacked_data.read_stacked_data_to_variable(
test_path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
pos_embedding,
use_gpu=use_gpu,
volatile=True,
prior_order=prior_order,
elmo=use_elmo)
punct_set = None
if punctuation is not None:
punct_set = set(punctuation)
logger.info("punctuations(%d): %s" %
(len(punct_set), ' '.join(punct_set)))
def construct_word_embedding_table():
scale = np.sqrt(3.0 / word_dim)
table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32)
# NOTE: UNK 관리!
table[conllx_stacked_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov = 0
for word, index in list(word_alphabet.items()):
if word in word_dict:
embedding = word_dict[word]
elif word.lower() in word_dict:
embedding = word_dict[word.lower()]
else:
# NOTE: words not in pretrained are set to random
embedding = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('word OOV: %d' % oov)
return torch.from_numpy(table)
def construct_char_embedding_table():
if char_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_chars, char_dim], dtype=np.float32)
table[conllx_stacked_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, char_dim]).astype(np.float32)
oov = 0
#for char, index, in char_alphabet.items():
for char, index in list(char_alphabet.items()):
if char in char_dict:
embedding = char_dict[char]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('character OOV: %d' % oov)
return torch.from_numpy(table)
def construct_pos_embedding_table():
if pos_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_pos, pos_dim], dtype=np.float32)
for pos, index in list(pos_alphabet.items()):
if pos in pos_dict:
embedding = pos_dict[pos]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
table[index, :] = embedding
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
char_table = construct_char_embedding_table()
pos_table = construct_pos_embedding_table()
window = 3
# yj 수정
# network = StackPtrNet(word_dim, num_words, char_dim, num_chars, pos_dim, num_pos, num_filters, window,
# mode, input_size_decoder, hidden_size, encoder_layers, decoder_layers,
# num_types, arc_space, type_space, pos_embedding,
# embedd_word=word_table, embedd_char=char_table, embedd_pos=pos_table, p_in=p_in, p_out=p_out,
# p_rnn=p_rnn, biaffine=True, pos=use_pos, char=use_char, elmo=use_elmo, prior_order=prior_order,
# skipConnect=skipConnect, grandPar=grandPar, sibling=sibling, elmo_path=elmo_path, elmo_dim=elmo_dim,
# bert = use_bert, bert_path=bert_path, bert_dim=bert_dim)
network = BiRecurrentConvBiAffine(word_dim,
num_words,
char_dim,
num_chars,
pos_dim,
num_pos,
num_filters,
window,
mode,
hidden_size,
encoder_layers,
num_types,
arc_space,
type_space,
embedd_word=word_table,
embedd_char=char_table,
embedd_pos=pos_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
biaffine=True,
pos=use_pos,
char=use_char,
elmo=use_elmo,
elmo_path=elmo_path,
elmo_dim=elmo_dim,
bert=use_bert,
bert_path=bert_path,
bert_dim=bert_dim)
# if fine_tune_path is not None:
# pretrained_dict = torch.load(fine_tune_path)
# model_dict = network.state_dict()
# # select
# #model_dict['pos_embedd.weight'] = pretrained_dict['pos_embedd.weight']
# model_dict['word_embedd.weight'] = pretrained_dict['word_embedd.weight']
# #model_dict['char_embedd.weight'] = pretrained_dict['char_embedd.weight']
# network.load_state_dict(model_dict)
model_ver = args.model_version
if model_ver is not None:
savePath = args.model_path + model_ver + 'network.pt'
network.load_state_dict(torch.load(savePath))
logger.info('Load model: %s' % (model_ver))
def save_args():
arg_path = model_name + '.arg.json'
arguments = [
word_dim, num_words, char_dim, num_chars, pos_dim, num_pos,
num_filters, window, mode, hidden_size, encoder_layers, num_types,
arc_space, type_space, pos_embedding
]
kwargs = {
'p_in': p_in,
'p_out': p_out,
'p_rnn': p_rnn,
'biaffine': True,
'pos': use_pos,
'char': use_char,
'elmo': use_elmo,
'bert': use_bert
}
json.dump({
'args': arguments,
'kwargs': kwargs
},
open(arg_path, 'w', encoding="utf-8"),
indent=4)
with open(arg_path + '.raw_args', 'w', encoding="utf-8") as f:
f.write(str(args))
if freeze:
network.word_embedd.freeze()
if use_gpu:
network.cuda()
save_args()
pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet, pos_embedding)
gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet, pos_embedding)
def generate_optimizer(opt, lr, params):
# params = [param for name, param in params if param.requires_grad]
params = [param for name, param in params]
if True:
return AdamW(params, lr=lr, betas=betas, weight_decay=gamma)
if opt == 'adam':
return Adam(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
elif opt == 'sgd':
return SGD(params,
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
elif opt == 'adamax':
return Adamax(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
else:
raise ValueError('Unknown optimization algorithm: %s' % opt)
# 우선 huggingface 기본 bert option으로 수정
def generate_bert_optimizer(t_total, bert_lr, model):
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
'weight_decay':
gamma
}, {
'params': [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
optimizer = AdamW(optimizer_grouped_parameters, lr=bert_lr, eps=1e-8)
scheduler = WarmupLinearSchedule(optimizer,
warmup_steps=0,
t_total=t_total)
return scheduler, optimizer
lr = learning_rate
if use_bert:
scheduler, optim = generate_bert_optimizer(
len(data_train) * num_epochs, lr, network)
#optim = generate_optimizer(opt, lr, network.named_parameters())
opt_info = 'opt: %s, ' % opt
if opt == 'adam':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
elif opt == 'sgd':
opt_info += 'momentum=%.2f' % momentum
elif opt == 'adamax':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
word_status = 'frozen' if freeze else 'fine tune'
char_status = 'enabled' if use_char else 'disabled'
pos_status = 'enabled' if use_pos else 'disabled'
logger.info(
"Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" %
(word_dim, word_status, char_dim, char_status, pos_dim, pos_status))
logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window))
#logger.info("RNN: %s, num_layer=(%d, %d), input_dec=%d, hidden=%d, arc_space=%d, type_space=%d" % (mode, encoder_layers, decoder_layers, input_size_decoder, hidden_size, arc_space, type_space))
logger.info(
"train: cov: %.1f, (#data: %d, batch: %d, clip: %.2f, label_smooth: %.2f, unk_repl: %.2f)"
% (cov, num_data, batch_size, clip, label_smooth, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
logger.info('prior order: %s, grand parent: %s, sibling: %s, ' %
(prior_order, grandPar, sibling))
logger.info('skip connect: %s, beam: %d' % (skipConnect, beam))
logger.info(opt_info)
num_batches = int(num_data / batch_size + 1) # kwon
dev_ucorrect = 0.0
dev_lcorrect = 0.0
dev_ucomlpete_match = 0.0
dev_lcomplete_match = 0.0
dev_ucorrect_nopunc = 0.0
dev_lcorrect_nopunc = 0.0
dev_ucomlpete_match_nopunc = 0.0
dev_lcomplete_match_nopunc = 0.0
dev_root_correct = 0.0
best_epoch = 0
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_root_correct = 0.0
test_total = 0
test_total_nopunc = 0
test_total_inst = 0
test_total_root = 0
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
patient = 0
decay = 0
max_decay = args.max_decay
double_schedule_decay = args.double_schedule_decay
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): '
%
(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay))
train_err = 0.
train_err_arc = 0.
train_err_type = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
for batch in range(1, num_batches + 1):
# load data
input_encoder, _ = conllx_stacked_data.get_batch_stacked_variable(
data_train,
batch_size,
pos_embedding,
unk_replace=unk_replace,
elmo=use_elmo,
bert=use_bert)
word_elmo = None
if use_elmo:
word, char, pos, heads, types, masks, lengths, word_elmo, word_bert = input_encoder
else:
word, char, pos, heads, types, masks, lengths, word_bert = input_encoder
#stacked_heads, children, sibling, stacked_types, skip_connect, masks_d, lengths_d = input_decoder
optim.zero_grad()
# yjyj
loss_arc, loss_type, bert_word_feature_ids, bert_morp_feature_ids = network.loss(
word,
char,
pos,
heads,
types,
mask=masks,
length=lengths,
input_word_bert=word_bert)
# loss_arc_leaf, loss_arc_non_leaf, \
# loss_type_leaf, loss_type_non_leaf, \
# loss_cov, num_leaf, num_non_leaf = network.loss(word, char, pos, heads, stacked_heads, children, sibling, stacked_types, label_smooth, skip_connect=skip_connect, mask_e=masks_e, \
# length_e=lengths_e, mask_d=masks_d, length_d=lengths_d, input_word_elmo = word_elmo, input_word_bert = word_bert)
# loss_arc = loss_arc_leaf + loss_arc_non_leaf
# loss_type = loss_type_leaf + loss_type_non_leaf
# loss = loss_arc + loss_type + cov * loss_cov # cov is set to 0 by default
loss = loss_arc + loss_type
loss.backward()
clip_grad_norm_(network.parameters(), clip)
optim.step()
if use_bert:
pass
#bert_optim.step()
#scheduler.step()
num_inst = word.size(
0) if obj == 'crf' else masks.data.sum() - word.size(0)
train_err += loss.item() * num_inst
train_err_arc += loss_arc.item() * num_inst
train_err_type += loss_type.item() * num_inst
train_total += num_inst
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
# yjyj
# num_leaf = num_leaf.item()
# num_non_leaf = num_non_leaf.item()
# train_err_arc_leaf += loss_arc_leaf.item() * num_leaf
# train_err_arc_non_leaf += loss_arc_non_leaf.item() * num_non_leaf
#
# train_err_type_leaf += loss_type_leaf.item() * num_leaf
# train_err_type_non_leaf += loss_type_non_leaf.item() * num_non_leaf
#
# train_err_cov += loss_cov.item() * (num_leaf + num_non_leaf)
# train_total_leaf += num_leaf
# train_total_non_leaf += num_non_leaf
#
# time_ave = (time.time() - start_time) / batch
# time_left = (num_batches - batch) * time_ave
# update log
# update log
if batch % 10 == 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, arc: %.4f, type: %.4f, time left: %.2fs' % (
batch, num_batches, train_err / train_total, train_err_arc
/ train_total, train_err_type / 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, arc: %.4f, type: %.4f, time: %.2fs' %
(num_batches, train_err / train_total, train_err_arc / train_total,
train_err_type / train_total, time.time() - start_time))
# yjyj
# if batch % 10 == 0:
# sys.stdout.write("\b" * num_back)
# sys.stdout.write(" " * num_back)
# sys.stdout.write("\b" * num_back)
# err_arc_leaf = train_err_arc_leaf / train_total_leaf
# err_arc_non_leaf = train_err_arc_non_leaf / train_total_non_leaf
# err_arc = err_arc_leaf + err_arc_non_leaf
#
# err_type_leaf = train_err_type_leaf / train_total_leaf
# err_type_non_leaf = train_err_type_non_leaf / train_total_non_leaf
# err_type = err_type_leaf + err_type_non_leaf
#
# err_cov = train_err_cov / (train_total_leaf + train_total_non_leaf)
#
# err = err_arc + err_type + cov * err_cov
# log_info = 'train: %d/%d loss (leaf, non_leaf): %.4f, arc: %.4f (%.4f, %.4f), type: %.4f (%.4f, %.4f), coverage: %.4f, time left (estimated): %.2fs' % (
# batch, num_batches, err, err_arc, err_arc_leaf, err_arc_non_leaf, err_type, err_type_leaf, err_type_non_leaf, err_cov, 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)
# err_arc_leaf = train_err_arc_leaf / train_total_leaf
# err_arc_non_leaf = train_err_arc_non_leaf / train_total_non_leaf
# err_arc = err_arc_leaf + err_arc_non_leaf
#
# err_type_leaf = train_err_type_leaf / train_total_leaf
# err_type_non_leaf = train_err_type_non_leaf / train_total_non_leaf
# err_type = err_type_leaf + err_type_non_leaf
#
# err_cov = train_err_cov / (train_total_leaf + train_total_non_leaf)
#
# err = err_arc + err_type + cov * err_cov
# print('train: %d loss (leaf, non_leaf): %.4f, arc: %.4f (%.4f, %.4f), type: %.4f (%.4f, %.4f), coverage: %.4f, time: %.2fs' % (
# num_batches, err, err_arc, err_arc_leaf, err_arc_non_leaf, err_type, err_type_leaf, err_type_non_leaf, err_cov, time.time() - start_time))
# evaluate performance on dev data
network.eval()
pred_filename = model_path + 'tmp/pred_dev%d' % (epoch)
pred_writer.start(pred_filename)
gold_filename = model_path + 'tmp/gold_dev%d' % (epoch)
gold_writer.start(gold_filename)
dev_ucorr = 0.0
dev_lcorr = 0.0
dev_total = 0
dev_ucomlpete = 0.0
dev_lcomplete = 0.0
dev_ucorr_nopunc = 0.0
dev_lcorr_nopunc = 0.0
dev_total_nopunc = 0
dev_ucomlpete_nopunc = 0.0
dev_lcomplete_nopunc = 0.0
dev_root_corr = 0.0
dev_total_root = 0.0
dev_total_inst = 0.0
for batch in conllx_stacked_data.iterate_batch_stacked_variable(
data_dev, batch_size, pos_embedding, type='dev',
elmo=use_elmo):
input_encoder, _ = batch
#@TODO 여기 input word elmo랑 input word bert 처리
if use_elmo:
word, char, pos, heads, types, masks, lengths, word_elmo, word_bert = input_encoder
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_stacked_data.NUM_SYMBOLIC_TAGS)
# heads_pred, types_pred, _, _ = network.decode(word, char, pos, input_word_elmo=word_elmo, mask=masks,
# length=lengths, beam=beam,
# leading_symbolic=conllx_stacked_data.NUM_SYMBOLIC_TAGS, input_word_bert=word_bert)
else:
word, char, pos, heads, types, masks, lengths, word_bert = input_encoder
heads_pred, types_pred, bert_word_feature_ids, bert_morp_feature_ids = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_stacked_data.NUM_SYMBOLIC_TAGS,
input_word_bert=word_bert)
# heads_pred, types_pred, _, _ = network.decode(word, char, pos, mask=masks, length=lengths, beam=beam,
# leading_symbolic=conllx_stacked_data.NUM_SYMBOLIC_TAGS, input_word_bert=word_bert)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser_bpe.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True,
bert_word_feature_ids=bert_word_feature_ids,
bert_morp_feature_ids=bert_morp_feature_ids)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
dev_ucorr += ucorr
dev_lcorr += lcorr
dev_total += total
dev_ucomlpete += ucm
dev_lcomplete += lcm
dev_ucorr_nopunc += ucorr_nopunc
dev_lcorr_nopunc += lcorr_nopunc
dev_total_nopunc += total_nopunc
dev_ucomlpete_nopunc += ucm_nopunc
dev_lcomplete_nopunc += lcm_nopunc
dev_root_corr += corr_root
dev_total_root += total_root
dev_total_inst += num_inst
pred_writer.close()
gold_writer.close()
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total,
dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 /
dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc *
100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 /
dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' %
(dev_root_corr, dev_total_root,
dev_root_corr * 100 / dev_total_root))
if dev_ucorrect_nopunc * 1.5 + dev_lcorrect_nopunc < dev_ucorr_nopunc * 1.5 + dev_lcorr_nopunc:
dev_ucorrect_nopunc = dev_ucorr_nopunc
dev_lcorrect_nopunc = dev_lcorr_nopunc
dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc
dev_lcomplete_match_nopunc = dev_lcomplete_nopunc
dev_ucorrect = dev_ucorr
dev_lcorrect = dev_lcorr
dev_ucomlpete_match = dev_ucomlpete
dev_lcomplete_match = dev_lcomplete
dev_root_correct = dev_root_corr
best_epoch = epoch
patient = 0
# torch.save(network, model_name)
torch.save(network.state_dict(), model_name)
# save embedding to txt
# FIXME format!
#with open(model_path + 'embedding.txt', 'w') as f:
# for word, idx in word_alphabet.items():
# embedding = network.word_embedd.weight[idx, :]
# f.write('{}\t{}\n'.format(word, embedding))
if test_path:
pred_filename = model_path + 'tmp/%spred_test%d' % (str(uid),
epoch)
pred_writer.start(pred_filename)
gold_filename = model_path + 'tmp/%sgold_test%d' % (str(uid),
epoch)
gold_writer.start(gold_filename)
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
for batch in conllx_stacked_data.iterate_batch_stacked_variable(
data_test, batch_size, pos_embedding, type='dev'):
input_encoder, _ = batch
word, char, pos, heads, types, masks, lengths = input_encoder
# yjyj
# heads_pred, types_pred, _, _ = network.decode(word, char, pos, mask=masks, length=lengths, beam=beam, leading_symbolic=conllx_stacked_data.NUM_SYMBOLIC_TAGS)
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_stacked_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser_bpe.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
pred_writer.close()
gold_writer.close()
else:
if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule:
# network = torch.load(model_name)
network.load_state_dict(torch.load(model_name))
lr = lr * decay_rate
# = generate_optimizer(opt, lr, network.named_parameters())
optim = generate_bert_optimizer(opt, lr, network)
patient = 0
decay += 1
if decay % double_schedule_decay == 0:
schedule *= 2
else:
patient += 1
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect, dev_lcorrect, dev_total,
dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch))
print(
'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch))
print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' %
(dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch))
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
if test_path:
print(
'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (test_ucorrect, test_lcorrect, test_total, test_ucorrect *
100 / test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst,
test_lcomplete_match * 100 / test_total_inst, best_epoch))
print(
'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
%
(test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc,
test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst,
test_lcomplete_match_nopunc * 100 / test_total_inst,
best_epoch))
print(
'best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)'
% (test_root_correct, test_total_root,
test_root_correct * 100 / test_total_root, best_epoch))
print(
'============================================================================================================================'
)
if decay == max_decay:
break
def save_result():
result_path = model_name + '.result.txt'
best_dev_Punc = 'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (
dev_ucorrect, dev_lcorrect, dev_total,
dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch)
best_dev_noPunc = 'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)' % (
dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch)
best_dev_Root = 'best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' % (
dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch)
f = open(result_path, 'w')
f.write(str(best_dev_Punc.encode('utf-8')) + '\n')
f.write(str(best_dev_noPunc.encode('utf-8')) + '\n')
f.write(str(best_dev_Root.encode('utf-8')))
f.close()
save_result()
def biaffine(model_path, model_name, pre_model_path, pre_model_name, use_gpu, logger, args):
alphabet_path = os.path.join(pre_model_path, 'alphabets/')
logger.info("Alphabet Path: %s" % alphabet_path)
pre_model_name = os.path.join(pre_model_path, pre_model_name)
model_name = os.path.join(model_path, model_name)
# Load pre-created alphabets
word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets(
alphabet_path, None, data_paths=[None, None], max_vocabulary_size=50000, embedd_dict=None)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
logger.info('use gpu: %s' % (use_gpu))
if args.test_lang:
extra_embed = args.embed_dir + ("wiki.multi.%s.vec" % args.test_lang)
extra_word_dict, _ = load_embedding_dict('word2vec', extra_embed)
test_path = args.data_dir + args.test_lang + '_test.conllu'
extra_embeds_arr = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger)
else:
extra_embeds_arr = []
for language in args.langs:
extra_embed = args.embed_dir + ("wiki.multi.%s.vec" % language)
extra_word_dict, _ = load_embedding_dict('word2vec', extra_embed)
test_path = args.data_dir + language + '_train.conllu'
embeds_arr1 = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger)
test_path = args.data_dir + language + '_dev.conllu'
embeds_arr2 = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger)
test_path = args.data_dir + language + '_test.conllu'
embeds_arr3 = augment_with_extra_embedding(word_alphabet, extra_word_dict, test_path, logger)
extra_embeds_arr.extend(embeds_arr1 + embeds_arr2 + embeds_arr3)
# ------------------------------------------------------------------------- #
# --------------------- Loading model ------------------------------------- #
def load_model_arguments_from_json():
arguments = json.load(open(arg_path, 'r'))
return arguments['args'], arguments['kwargs']
arg_path = pre_model_name + '.arg.json'
margs, kwargs = load_model_arguments_from_json()
network = BiRecurrentConvBiAffine(use_gpu=use_gpu, *margs, **kwargs)
network.load_state_dict(torch.load(pre_model_name))
args.use_bert = kwargs.get('use_bert', False)
#
augment_network_embed(word_alphabet.size(), network, extra_embeds_arr)
network.eval()
logger.info('model: %s' % pre_model_name)
# Freeze the network
for p in network.parameters():
p.requires_grad = False
nclass = args.nclass
classifier = nn.Sequential(
nn.Linear(network.encoder.output_dim, 512),
nn.Linear(512, nclass)
)
if use_gpu:
network.cuda()
classifier.cuda()
else:
network.cpu()
classifier.cpu()
batch_size = args.batch_size
# ===== the reading
def _read_one(path, is_train=False, max_size=None):
lang_id = guess_language_id(path)
logger.info("Reading: guess that the language of file %s is %s." % (path, lang_id))
one_data = conllx_data.read_data_to_variable(path, word_alphabet, char_alphabet, pos_alphabet,
type_alphabet, use_gpu=use_gpu, volatile=(not is_train),
use_bert=args.use_bert, symbolic_root=True, lang_id=lang_id,
max_size=max_size)
return one_data
def compute_accuracy(data, lang_idx):
total_corr, total = 0, 0
classifier.eval()
with torch.no_grad():
for batch in conllx_data.iterate_batch_variable(data, batch_size):
word, char, pos, _, _, masks, lengths, bert_inputs = batch
if use_gpu:
word = word.cuda()
char = char.cuda()
pos = pos.cuda()
masks = masks.cuda()
lengths = lengths.cuda()
if bert_inputs[0] is not None:
bert_inputs[0] = bert_inputs[0].cuda()
bert_inputs[1] = bert_inputs[1].cuda()
bert_inputs[2] = bert_inputs[2].cuda()
output = network.forward(word, char, pos, input_bert=bert_inputs,
mask=masks, length=lengths, hx=None)
output = output['output'].detach()
if args.train_level == 'word':
output = classifier(output)
output = output.contiguous().view(-1, output.size(2))
else:
output = torch.mean(output, dim=1)
output = classifier(output)
preds = output.max(1)[1].cpu()
labels = torch.LongTensor([lang_idx])
labels = labels.expand(*preds.size())
n_correct = preds.eq(labels).sum().item()
total_corr += n_correct
total += output.size(0)
return {'total_corr': total_corr, 'total': total}
if args.test_lang:
classifier.load_state_dict(torch.load(model_name))
path = args.data_dir + args.test_lang + '_train.conllu'
test_data = _read_one(path)
# TODO: fixed indexing is not GOOD
lang_idx = 0 if args.test_lang == args.src_lang else 1
result = compute_accuracy(test_data, lang_idx)
accuracy = (result['total_corr'] * 100.0) / result['total']
logger.info('[Classifier performance] Language: %s || accuracy: %.2f%%' % (args.test_lang, accuracy))
else:
# if output directory doesn't exist, create it
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# --------------------- Loading data -------------------------------------- #
train_data = dict()
dev_data = dict()
test_data = dict()
num_data = dict()
lang_ids = dict()
reverse_lang_ids = dict()
# loading language data
for language in args.langs:
lang_ids[language] = len(lang_ids)
reverse_lang_ids[lang_ids[language]] = language
train_path = args.data_dir + language + '_train.conllu'
# Utilize at most 10000 examples
tmp_data = _read_one(train_path, max_size=10000)
num_data[language] = sum(tmp_data[1])
train_data[language] = tmp_data
dev_path = args.data_dir + language + '_dev.conllu'
tmp_data = _read_one(dev_path)
dev_data[language] = tmp_data
test_path = args.data_dir + language + '_test.conllu'
tmp_data = _read_one(test_path)
test_data[language] = tmp_data
# ------------------------------------------------------------------------- #
optim = torch.optim.Adam(classifier.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()
def compute_loss(lang_name, land_idx):
word, char, pos, _, _, masks, lengths, bert_inputs = conllx_data.get_batch_variable(train_data[lang_name],
batch_size,
unk_replace=0.5)
if use_gpu:
word = word.cuda()
char = char.cuda()
pos = pos.cuda()
masks = masks.cuda()
lengths = lengths.cuda()
if bert_inputs[0] is not None:
bert_inputs[0] = bert_inputs[0].cuda()
bert_inputs[1] = bert_inputs[1].cuda()
bert_inputs[2] = bert_inputs[2].cuda()
output = network.forward(word, char, pos, input_bert=bert_inputs,
mask=masks, length=lengths, hx=None)
output = output['output'].detach()
if args.train_level == 'word':
output = classifier(output)
output = output.contiguous().view(-1, output.size(2))
else:
output = torch.mean(output, dim=1)
output = classifier(output)
labels = torch.empty(output.size(0)).fill_(land_idx).type_as(output).long()
loss = criterion(output, labels)
return loss
# ---------------------- Form the mini-batches -------------------------- #
num_batches = 0
batch_lang_labels = []
for lang in args.langs:
nbatches = num_data[lang] // batch_size + 1
batch_lang_labels.extend([lang] * nbatches)
num_batches += nbatches
assert len(batch_lang_labels) == num_batches
# ------------------------------------------------------------------------- #
best_dev_accuracy = 0
patience = 0
for epoch in range(1, args.num_epochs + 1):
# shuffling the data
lang_in_batch = copy.copy(batch_lang_labels)
random.shuffle(lang_in_batch)
classifier.train()
for batch in range(1, num_batches + 1):
lang_name = lang_in_batch[batch - 1]
lang_id = lang_ids.get(lang_name)
loss = compute_loss(lang_name, lang_id)
loss.backward()
optim.step()
# Validation
avg_acc = dict()
for dev_lang in dev_data.keys():
lang_idx = lang_ids.get(dev_lang)
result = compute_accuracy(dev_data[dev_lang], lang_idx)
accuracy = (result['total_corr'] * 100.0) / result['total']
avg_acc[dev_lang] = accuracy
acc = ', '.join('%s: %.2f' % (key, val) for (key, val) in avg_acc.items())
logger.info('Epoch: %d, Performance[%s]' % (epoch, acc))
avg_acc = sum(avg_acc.values()) / len(avg_acc)
if best_dev_accuracy < avg_acc:
best_dev_accuracy = avg_acc
patience = 0
state_dict = classifier.state_dict()
torch.save(state_dict, model_name)
else:
patience += 1
if patience >= 5:
break
# Testing
logger.info('Testing model %s' % pre_model_name)
total_corr, total = 0, 0
for test_lang in UD_languages:
if test_lang in test_data:
lang_idx = lang_ids.get(test_lang)
result = compute_accuracy(test_data[test_lang], lang_idx)
accuracy = (result['total_corr'] * 100.0) / result['total']
print('[LANG]: %s, [ACC]: %.2f' % (test_lang.upper(), accuracy))
total_corr += result['total_corr']
total += result['total']
print('[Avg. Performance]: %.2f' % ((total_corr * 100.0) / total))
def main():
args_parser = argparse.ArgumentParser(
description='Tuning with graph-based parsing')
args_parser.add_argument('--seed',
type=int,
default=1234,
help='random seed for reproducibility')
args_parser.add_argument('--mode',
choices=['RNN', 'LSTM', 'GRU', 'FastLSTM'],
help='architecture of rnn',
required=True)
args_parser.add_argument('--num_epochs',
type=int,
default=1000,
help='Number of training epochs')
args_parser.add_argument('--batch_size',
type=int,
default=64,
help='Number of sentences in each batch')
args_parser.add_argument('--hidden_size',
type=int,
default=256,
help='Number of hidden units in RNN')
args_parser.add_argument('--arc_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--type_space',
type=int,
default=128,
help='Dimension of tag space')
args_parser.add_argument('--num_layers',
type=int,
default=1,
help='Number of layers of encoder.')
args_parser.add_argument('--num_filters',
type=int,
default=50,
help='Number of filters in CNN')
args_parser.add_argument('--pos',
action='store_true',
help='use part-of-speech embedding.')
args_parser.add_argument('--char',
action='store_true',
help='use character embedding and CNN.')
args_parser.add_argument('--pos_dim',
type=int,
default=50,
help='Dimension of POS embeddings')
args_parser.add_argument('--char_dim',
type=int,
default=50,
help='Dimension of Character embeddings')
args_parser.add_argument('--opt',
choices=['adam', 'sgd', 'adamax'],
help='optimization algorithm')
args_parser.add_argument('--objective',
choices=['cross_entropy', 'crf'],
default='cross_entropy',
help='objective function of training procedure.')
args_parser.add_argument('--decode',
choices=['mst', 'greedy'],
default='mst',
help='decoding algorithm')
args_parser.add_argument('--learning_rate',
type=float,
default=0.01,
help='Learning rate')
# args_parser.add_argument('--decay_rate', type=float, default=0.05, help='Decay rate of learning rate')
args_parser.add_argument('--clip',
type=float,
default=5.0,
help='gradient clipping')
args_parser.add_argument('--gamma',
type=float,
default=0.0,
help='weight for regularization')
args_parser.add_argument('--epsilon',
type=float,
default=1e-8,
help='epsilon for adam or adamax')
args_parser.add_argument('--p_rnn',
nargs='+',
type=float,
required=True,
help='dropout rate for RNN')
args_parser.add_argument('--p_in',
type=float,
default=0.33,
help='dropout rate for input embeddings')
args_parser.add_argument('--p_out',
type=float,
default=0.33,
help='dropout rate for output layer')
# args_parser.add_argument('--schedule', type=int, help='schedule for learning rate decay')
args_parser.add_argument(
'--unk_replace',
type=float,
default=0.,
help='The rate to replace a singleton word with UNK')
args_parser.add_argument('--punctuation',
nargs='+',
type=str,
help='List of punctuations')
args_parser.add_argument(
'--word_embedding',
choices=['word2vec', 'glove', 'senna', 'sskip', 'polyglot'],
help='Embedding for words',
required=True)
args_parser.add_argument('--word_path',
help='path for word embedding dict')
args_parser.add_argument(
'--freeze',
action='store_true',
help='frozen the word embedding (disable fine-tuning).')
args_parser.add_argument('--char_embedding',
choices=['random', 'polyglot'],
help='Embedding for characters',
required=True)
args_parser.add_argument('--char_path',
help='path for character embedding dict')
args_parser.add_argument(
'--train') # "data/POS-penn/wsj/split1/wsj1.train.original"
args_parser.add_argument(
'--dev') # "data/POS-penn/wsj/split1/wsj1.dev.original"
args_parser.add_argument(
'--test') # "data/POS-penn/wsj/split1/wsj1.test.original"
args_parser.add_argument('--vocab_path',
help='path for prebuilt alphabets.',
default=None)
args_parser.add_argument('--model_path',
help='path for saving model file.',
required=True)
args_parser.add_argument('--model_name',
help='name for saving model file.',
required=True)
#
args_parser.add_argument('--no_word',
action='store_true',
help='do not use word embedding.')
#
# lrate schedule with warmup in the first iter.
args_parser.add_argument('--use_warmup_schedule',
action='store_true',
help="Use warmup lrate schedule.")
args_parser.add_argument('--decay_rate',
type=float,
default=0.75,
help='Decay rate of learning rate')
args_parser.add_argument('--max_decay',
type=int,
default=9,
help='Number of decays before stop')
args_parser.add_argument('--schedule',
type=int,
help='schedule for learning rate decay')
args_parser.add_argument('--double_schedule_decay',
type=int,
default=5,
help='Number of decays to double schedule')
args_parser.add_argument(
'--check_dev',
type=int,
default=5,
help='Check development performance in every n\'th iteration')
# Tansformer encoder
args_parser.add_argument('--no_CoRNN',
action='store_true',
help='do not use context RNN.')
args_parser.add_argument(
'--trans_hid_size',
type=int,
default=1024,
help='#hidden units in point-wise feed-forward in transformer')
args_parser.add_argument(
'--d_k',
type=int,
default=64,
help='d_k for multi-head-attention in transformer encoder')
args_parser.add_argument(
'--d_v',
type=int,
default=64,
help='d_v for multi-head-attention in transformer encoder')
args_parser.add_argument('--multi_head_attn',
action='store_true',
help='use multi-head-attention.')
args_parser.add_argument('--num_head',
type=int,
default=8,
help='Value of h in multi-head attention')
# - positional
args_parser.add_argument(
'--enc_use_neg_dist',
action='store_true',
help="Use negative distance for enc's relational-distance embedding.")
args_parser.add_argument(
'--enc_clip_dist',
type=int,
default=0,
help="The clipping distance for relative position features.")
args_parser.add_argument('--position_dim',
type=int,
default=50,
help='Dimension of Position embeddings.')
args_parser.add_argument(
'--position_embed_num',
type=int,
default=200,
help=
'Minimum value of position embedding num, which usually is max-sent-length.'
)
args_parser.add_argument('--train_position',
action='store_true',
help='train positional encoding for transformer.')
#
args_parser.add_argument(
'--train_len_thresh',
type=int,
default=100,
help='In training, discard sentences longer than this.')
#
args = args_parser.parse_args()
# fix data-prepare seed
random.seed(1234)
np.random.seed(1234)
# model's seed
torch.manual_seed(args.seed)
logger = get_logger("GraphParser")
mode = args.mode
obj = args.objective
decoding = args.decode
train_path = args.train
dev_path = args.dev
test_path = args.test
model_path = args.model_path
model_name = args.model_name
num_epochs = args.num_epochs
batch_size = args.batch_size
hidden_size = args.hidden_size
arc_space = args.arc_space
type_space = args.type_space
num_layers = args.num_layers
num_filters = args.num_filters
learning_rate = args.learning_rate
opt = args.opt
momentum = 0.9
betas = (0.9, 0.9)
eps = args.epsilon
decay_rate = args.decay_rate
clip = args.clip
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
punctuation = args.punctuation
freeze = args.freeze
word_embedding = args.word_embedding
word_path = args.word_path
use_char = args.char
char_embedding = args.char_embedding
char_path = args.char_path
use_pos = args.pos
pos_dim = args.pos_dim
word_dict, word_dim = utils.load_embedding_dict(word_embedding, word_path)
char_dict = None
char_dim = args.char_dim
if char_embedding != 'random':
char_dict, char_dim = utils.load_embedding_dict(
char_embedding, char_path)
#
vocab_path = args.vocab_path if args.vocab_path is not None else args.model_path
logger.info("Creating Alphabets")
alphabet_path = os.path.join(vocab_path, 'alphabets/')
model_name = os.path.join(model_path, model_name)
# todo(warn): exactly same for loading vocabs
word_alphabet, char_alphabet, pos_alphabet, type_alphabet, max_sent_length = conllx_data.create_alphabets(
alphabet_path,
train_path,
data_paths=[dev_path, test_path],
max_vocabulary_size=50000,
embedd_dict=word_dict)
max_sent_length = max(max_sent_length, args.position_embed_num)
num_words = word_alphabet.size()
num_chars = char_alphabet.size()
num_pos = pos_alphabet.size()
num_types = type_alphabet.size()
logger.info("Word Alphabet Size: %d" % num_words)
logger.info("Character Alphabet Size: %d" % num_chars)
logger.info("POS Alphabet Size: %d" % num_pos)
logger.info("Type Alphabet Size: %d" % num_types)
logger.info("Reading Data")
use_gpu = torch.cuda.is_available()
# ===== the reading
def _read_one(path, is_train):
lang_id = guess_language_id(path)
logger.info("Reading: guess that the language of file %s is %s." %
(path, lang_id))
one_data = conllx_data.read_data_to_variable(
path,
word_alphabet,
char_alphabet,
pos_alphabet,
type_alphabet,
use_gpu=use_gpu,
volatile=(not is_train),
symbolic_root=True,
lang_id=lang_id,
len_thresh=(args.train_len_thresh if is_train else 100000))
return one_data
data_train = _read_one(train_path, True)
num_data = sum(data_train[1])
data_dev = _read_one(dev_path, False)
data_test = _read_one(test_path, False)
# =====
punct_set = None
if punctuation is not None:
punct_set = set(punctuation)
logger.info("punctuations(%d): %s" %
(len(punct_set), ' '.join(punct_set)))
def construct_word_embedding_table():
scale = np.sqrt(3.0 / word_dim)
table = np.empty([word_alphabet.size(), word_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov = 0
for word, index in word_alphabet.items():
if word in word_dict:
embedding = word_dict[word]
elif word.lower() in word_dict:
embedding = word_dict[word.lower()]
else:
embedding = np.zeros([1, word_dim]).astype(
np.float32) if freeze else np.random.uniform(
-scale, scale, [1, word_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('word OOV: %d' % oov)
return torch.from_numpy(table)
def construct_char_embedding_table():
if char_dict is None:
return None
scale = np.sqrt(3.0 / char_dim)
table = np.empty([num_chars, char_dim], dtype=np.float32)
table[conllx_data.UNK_ID, :] = np.random.uniform(
-scale, scale, [1, char_dim]).astype(np.float32)
oov = 0
for char, index, in char_alphabet.items():
if char in char_dict:
embedding = char_dict[char]
else:
embedding = np.random.uniform(-scale, scale,
[1, char_dim]).astype(np.float32)
oov += 1
table[index, :] = embedding
print('character OOV: %d' % oov)
return torch.from_numpy(table)
word_table = construct_word_embedding_table()
char_table = construct_char_embedding_table()
window = 3
if obj == 'cross_entropy':
network = BiRecurrentConvBiAffine(
word_dim,
num_words,
char_dim,
num_chars,
pos_dim,
num_pos,
num_filters,
window,
mode,
hidden_size,
num_layers,
num_types,
arc_space,
type_space,
embedd_word=word_table,
embedd_char=char_table,
p_in=p_in,
p_out=p_out,
p_rnn=p_rnn,
biaffine=True,
pos=use_pos,
char=use_char,
train_position=args.train_position,
use_con_rnn=(not args.no_CoRNN),
trans_hid_size=args.trans_hid_size,
d_k=args.d_k,
d_v=args.d_v,
multi_head_attn=args.multi_head_attn,
num_head=args.num_head,
enc_use_neg_dist=args.enc_use_neg_dist,
enc_clip_dist=args.enc_clip_dist,
position_dim=args.position_dim,
max_sent_length=max_sent_length,
use_gpu=use_gpu,
no_word=args.no_word)
elif obj == 'crf':
raise NotImplementedError
else:
raise RuntimeError('Unknown objective: %s' % obj)
def save_args():
arg_path = model_name + '.arg.json'
arguments = [
word_dim, num_words, char_dim, num_chars, pos_dim, num_pos,
num_filters, window, mode, hidden_size, num_layers, num_types,
arc_space, type_space
]
kwargs = {
'p_in': p_in,
'p_out': p_out,
'p_rnn': p_rnn,
'biaffine': True,
'pos': use_pos,
'char': use_char,
'train_position': args.train_position,
'use_con_rnn': (not args.no_CoRNN),
'trans_hid_size': args.trans_hid_size,
'd_k': args.d_k,
'd_v': args.d_v,
'multi_head_attn': args.multi_head_attn,
'num_head': args.num_head,
'enc_use_neg_dist': args.enc_use_neg_dist,
'enc_clip_dist': args.enc_clip_dist,
'position_dim': args.position_dim,
'max_sent_length': max_sent_length,
'no_word': args.no_word
}
json.dump({
'args': arguments,
'kwargs': kwargs
},
open(arg_path, 'w'),
indent=4)
if freeze:
network.word_embedd.freeze()
if use_gpu:
network.cuda()
save_args()
pred_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
gold_writer = CoNLLXWriter(word_alphabet, char_alphabet, pos_alphabet,
type_alphabet)
def generate_optimizer(opt, lr, params):
params = filter(lambda param: param.requires_grad, params)
if opt == 'adam':
return Adam(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
elif opt == 'sgd':
return SGD(params,
lr=lr,
momentum=momentum,
weight_decay=gamma,
nesterov=True)
elif opt == 'adamax':
return Adamax(params,
lr=lr,
betas=betas,
weight_decay=gamma,
eps=eps)
else:
raise ValueError('Unknown optimization algorithm: %s' % opt)
lr = learning_rate
optim = generate_optimizer(opt, lr, network.parameters())
opt_info = 'opt: %s, ' % opt
if opt == 'adam':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
elif opt == 'sgd':
opt_info += 'momentum=%.2f' % momentum
elif opt == 'adamax':
opt_info += 'betas=%s, eps=%.1e' % (betas, eps)
word_status = 'frozen' if freeze else 'fine tune'
char_status = 'enabled' if use_char else 'disabled'
pos_status = 'enabled' if use_pos else 'disabled'
logger.info(
"Embedding dim: word=%d (%s), char=%d (%s), pos=%d (%s)" %
(word_dim, word_status, char_dim, char_status, pos_dim, pos_status))
logger.info("CNN: filter=%d, kernel=%d" % (num_filters, window))
logger.info(
"RNN: %s, num_layer=%d, hidden=%d, arc_space=%d, type_space=%d" %
(mode, num_layers, hidden_size, arc_space, type_space))
logger.info(
"train: obj: %s, l2: %f, (#data: %d, batch: %d, clip: %.2f, unk replace: %.2f)"
% (obj, gamma, num_data, batch_size, clip, unk_replace))
logger.info("dropout(in, out, rnn): (%.2f, %.2f, %s)" %
(p_in, p_out, p_rnn))
logger.info("decoding algorithm: %s" % decoding)
logger.info(opt_info)
num_batches = num_data / batch_size + 1
dev_ucorrect = 0.0
dev_lcorrect = 0.0
dev_ucomlpete_match = 0.0
dev_lcomplete_match = 0.0
dev_ucorrect_nopunc = 0.0
dev_lcorrect_nopunc = 0.0
dev_ucomlpete_match_nopunc = 0.0
dev_lcomplete_match_nopunc = 0.0
dev_root_correct = 0.0
best_epoch = 0
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_root_correct = 0.0
test_total = 0
test_total_nopunc = 0
test_total_inst = 0
test_total_root = 0
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' % decoding)
patient = 0
decay = 0
max_decay = args.max_decay
double_schedule_decay = args.double_schedule_decay
# lrate schedule
step_num = 0
use_warmup_schedule = args.use_warmup_schedule
warmup_factor = (lr + 0.) / num_batches
if use_warmup_schedule:
logger.info("Use warmup lrate for the first epoch, from 0 up to %s." %
(lr, ))
#
for epoch in range(1, num_epochs + 1):
print(
'Epoch %d (%s, optim: %s, learning rate=%.6f, eps=%.1e, decay rate=%.2f (schedule=%d, patient=%d, decay=%d)): '
%
(epoch, mode, opt, lr, eps, decay_rate, schedule, patient, decay))
train_err = 0.
train_err_arc = 0.
train_err_type = 0.
train_total = 0.
start_time = time.time()
num_back = 0
network.train()
for batch in range(1, num_batches + 1):
# lrate schedule (before each step)
step_num += 1
if use_warmup_schedule and epoch <= 1:
cur_lrate = warmup_factor * step_num
# set lr
for param_group in optim.param_groups:
param_group['lr'] = cur_lrate
#
word, char, pos, heads, types, masks, lengths = conllx_data.get_batch_variable(
data_train, batch_size, unk_replace=unk_replace)
optim.zero_grad()
loss_arc, loss_type = network.loss(word,
char,
pos,
heads,
types,
mask=masks,
length=lengths)
loss = loss_arc + loss_type
loss.backward()
clip_grad_norm(network.parameters(), clip)
optim.step()
num_inst = word.size(
0) if obj == 'crf' else masks.data.sum() - word.size(0)
train_err += loss.data[0] * num_inst
train_err_arc += loss_arc.data[0] * num_inst
train_err_type += loss_type.data[0] * num_inst
train_total += num_inst
time_ave = (time.time() - start_time) / batch
time_left = (num_batches - batch) * time_ave
# update log
if batch % 10 == 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, arc: %.4f, type: %.4f, time left: %.2fs' % (
batch, num_batches, train_err / train_total, train_err_arc
/ train_total, train_err_type / 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, arc: %.4f, type: %.4f, time: %.2fs' %
(num_batches, train_err / train_total, train_err_arc / train_total,
train_err_type / train_total, time.time() - start_time))
################################################################################################
if epoch % args.check_dev != 0:
continue
# evaluate performance on dev data
network.eval()
pred_filename = 'tmp/%spred_dev%d' % (str(uid), epoch)
pred_writer.start(pred_filename)
gold_filename = 'tmp/%sgold_dev%d' % (str(uid), epoch)
gold_writer.start(gold_filename)
dev_ucorr = 0.0
dev_lcorr = 0.0
dev_total = 0
dev_ucomlpete = 0.0
dev_lcomplete = 0.0
dev_ucorr_nopunc = 0.0
dev_lcorr_nopunc = 0.0
dev_total_nopunc = 0
dev_ucomlpete_nopunc = 0.0
dev_lcomplete_nopunc = 0.0
dev_root_corr = 0.0
dev_total_root = 0.0
dev_total_inst = 0.0
for batch in conllx_data.iterate_batch_variable(data_dev, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
dev_ucorr += ucorr
dev_lcorr += lcorr
dev_total += total
dev_ucomlpete += ucm
dev_lcomplete += lcm
dev_ucorr_nopunc += ucorr_nopunc
dev_lcorr_nopunc += lcorr_nopunc
dev_total_nopunc += total_nopunc
dev_ucomlpete_nopunc += ucm_nopunc
dev_lcomplete_nopunc += lcm_nopunc
dev_root_corr += corr_root
dev_total_root += total_root
dev_total_inst += num_inst
pred_writer.close()
gold_writer.close()
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total,
dev_lcorr * 100 / dev_total, dev_ucomlpete * 100 /
dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%'
% (dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc, dev_lcorr_nopunc *
100 / dev_total_nopunc, dev_ucomlpete_nopunc * 100 /
dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' %
(dev_root_corr, dev_total_root,
dev_root_corr * 100 / dev_total_root))
if dev_lcorrect_nopunc < dev_lcorr_nopunc or (
dev_lcorrect_nopunc == dev_lcorr_nopunc
and dev_ucorrect_nopunc < dev_ucorr_nopunc):
dev_ucorrect_nopunc = dev_ucorr_nopunc
dev_lcorrect_nopunc = dev_lcorr_nopunc
dev_ucomlpete_match_nopunc = dev_ucomlpete_nopunc
dev_lcomplete_match_nopunc = dev_lcomplete_nopunc
dev_ucorrect = dev_ucorr
dev_lcorrect = dev_lcorr
dev_ucomlpete_match = dev_ucomlpete
dev_lcomplete_match = dev_lcomplete
dev_root_correct = dev_root_corr
best_epoch = epoch
patient = 0
# torch.save(network, model_name)
torch.save(network.state_dict(), model_name)
pred_filename = 'tmp/%spred_test%d' % (str(uid), epoch)
pred_writer.start(pred_filename)
gold_filename = 'tmp/%sgold_test%d' % (str(uid), epoch)
gold_writer.start(gold_filename)
test_ucorrect = 0.0
test_lcorrect = 0.0
test_ucomlpete_match = 0.0
test_lcomplete_match = 0.0
test_total = 0
test_ucorrect_nopunc = 0.0
test_lcorrect_nopunc = 0.0
test_ucomlpete_match_nopunc = 0.0
test_lcomplete_match_nopunc = 0.0
test_total_nopunc = 0
test_total_inst = 0
test_root_correct = 0.0
test_total_root = 0
for batch in conllx_data.iterate_batch_variable(
data_test, batch_size):
word, char, pos, heads, types, masks, lengths = batch
heads_pred, types_pred = decode(
word,
char,
pos,
mask=masks,
length=lengths,
leading_symbolic=conllx_data.NUM_SYMBOLIC_TAGS)
word = word.data.cpu().numpy()
pos = pos.data.cpu().numpy()
lengths = lengths.cpu().numpy()
heads = heads.data.cpu().numpy()
types = types.data.cpu().numpy()
pred_writer.write(word,
pos,
heads_pred,
types_pred,
lengths,
symbolic_root=True)
gold_writer.write(word,
pos,
heads,
types,
lengths,
symbolic_root=True)
stats, stats_nopunc, stats_root, num_inst = parser.eval(
word,
pos,
heads_pred,
types_pred,
heads,
types,
word_alphabet,
pos_alphabet,
lengths,
punct_set=punct_set,
symbolic_root=True)
ucorr, lcorr, total, ucm, lcm = stats
ucorr_nopunc, lcorr_nopunc, total_nopunc, ucm_nopunc, lcm_nopunc = stats_nopunc
corr_root, total_root = stats_root
test_ucorrect += ucorr
test_lcorrect += lcorr
test_total += total
test_ucomlpete_match += ucm
test_lcomplete_match += lcm
test_ucorrect_nopunc += ucorr_nopunc
test_lcorrect_nopunc += lcorr_nopunc
test_total_nopunc += total_nopunc
test_ucomlpete_match_nopunc += ucm_nopunc
test_lcomplete_match_nopunc += lcm_nopunc
test_root_correct += corr_root
test_total_root += total_root
test_total_inst += num_inst
pred_writer.close()
gold_writer.close()
else:
if dev_ucorr_nopunc * 100 / dev_total_nopunc < dev_ucorrect_nopunc * 100 / dev_total_nopunc - 5 or patient >= schedule:
# network = torch.load(model_name)
network.load_state_dict(torch.load(model_name))
lr = lr * decay_rate
optim = generate_optimizer(opt, lr, network.parameters())
if decoding == 'greedy':
decode = network.decode
elif decoding == 'mst':
decode = network.decode_mst
else:
raise ValueError('Unknown decoding algorithm: %s' %
decoding)
patient = 0
decay += 1
if decay % double_schedule_decay == 0:
schedule *= 2
else:
patient += 1
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best dev W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect, dev_lcorrect, dev_total,
dev_ucorrect * 100 / dev_total, dev_lcorrect * 100 / dev_total,
dev_ucomlpete_match * 100 / dev_total_inst,
dev_lcomplete_match * 100 / dev_total_inst, best_epoch))
print(
'best dev Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (dev_ucorrect_nopunc, dev_lcorrect_nopunc, dev_total_nopunc,
dev_ucorrect_nopunc * 100 / dev_total_nopunc,
dev_lcorrect_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_match_nopunc * 100 / dev_total_inst,
dev_lcomplete_match_nopunc * 100 / dev_total_inst, best_epoch))
print('best dev Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' %
(dev_root_correct, dev_total_root,
dev_root_correct * 100 / dev_total_root, best_epoch))
print(
'----------------------------------------------------------------------------------------------------------------------------'
)
print(
'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
% (test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 /
test_total, test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst,
test_lcomplete_match * 100 / test_total_inst, best_epoch))
print(
'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% (epoch: %d)'
%
(test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc,
test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst,
test_lcomplete_match_nopunc * 100 / test_total_inst, best_epoch))
print('best test Root: corr: %d, total: %d, acc: %.2f%% (epoch: %d)' %
(test_root_correct, test_total_root,
test_root_correct * 100 / test_total_root, best_epoch))
print(
'============================================================================================================================'
)
if decay == max_decay:
break