def __call__(self, args):
super(Predict, self).__call__(args)
print("Load the dataset")
corpus = Corpus.load(args.fdata, self.fields)
dataset = TextDataset(corpus,
self.fields[:-1],
args.buckets)
# set the data loader
dataset.loader = batchify(dataset, args.batch_size)
print(f"{len(dataset)} sentences, "
f"{len(dataset.loader)} batches")
print("Load the model")
self.model = Model.load(args.model)
print(f"{self.model}\n")
print("Make predictions on the dataset")
start = datetime.now()
pred_labels = self.predict(dataset.loader)
total_time = datetime.now() - start
# restore the order of sentences in the buckets
indices = torch.tensor([i
for bucket in dataset.buckets.values()
for i in bucket]).argsort()
corpus.labels = [pred_labels[i] for i in indices]
print(f"Save the predicted result to {args.fpred}")
corpus.save(args.fpred)
print(f"{total_time}s elapsed, "
f"{len(dataset) / total_time.total_seconds():.2f} Sents/s")
def __call__(self, args):
logger.info("Load the model")
self.model = Model.load(args.model)
# override from CLI args
args = self.model.args.update(vars(args))
super().__call__(args)
logger.info("Load the dataset")
if args.prob:
self.fields = self.fields._replace(PHEAD=Field('probs'))
if args.text:
corpus = TextCorpus.load(args.fdata,
self.fields,
args.text,
args.tokenizer_dir,
use_gpu=args.device != 1)
else:
corpus = Corpus.load(args.fdata, self.fields)
dataset = TextDataset(corpus, [self.WORD, self.FEAT], args.buckets)
# set the data loader
dataset.loader = batchify(dataset, args.batch_size)
logger.info(f"{len(dataset)} sentences, "
f"{len(dataset.loader)} batches")
logger.info("Make predictions on the dataset")
start = datetime.now()
pred_arcs, pred_rels, pred_probs = self.predict(dataset.loader)
total_time = datetime.now() - start
# restore the order of sentences in the buckets
indices = torch.tensor([
i for bucket in dataset.buckets.values() for i in bucket
]).argsort()
corpus.arcs = [pred_arcs[i] for i in indices]
corpus.rels = [pred_rels[i] for i in indices]
if args.prob:
corpus.probs = [pred_probs[i] for i in indices]
logger.info(f"Save the predicted result to {args.fpred}")
corpus.save(args.fpred)
logger.info(f"{total_time}s elapsed, "
f"{len(dataset) / total_time.total_seconds():.2f} Sents/s")
def __call__(self, args):
super(Predict, self).__call__(args)
print("Load the dataset")
corpus = Corpus.load(args.fdata, self.fields)
dataset = TextDataset(corpus, [self.WORD, self.FEAT])
# set the data loader
dataset.loader = batchify(dataset, args.batch_size)
print(f"{len(dataset)} sentences, " f"{len(dataset.loader)} batches")
print("Load the model")
self.model = Model.load(args.model)
print(f"{self.model}\n")
print("Make predictions on the dataset")
start = datetime.now()
corpus.heads, corpus.rels = self.predict(dataset.loader)
print(f"Save the predicted result to {args.fpred}")
corpus.save(args.fpred)
total_time = datetime.now() - start
print(f"{total_time}s elapsed, "
f"{len(dataset) / total_time.total_seconds():.2f} Sents/s")
def __call__(self, args):
super(Evaluate, self).__call__(args)
print("Load the dataset")
corpus = Corpus.load(args.fdata, self.fields)
dataset = TextDataset(corpus, self.fields, args.buckets)
# set the data loader
dataset.loader = batchify(dataset, args.batch_size)
print(f"{len(dataset)} sentences, "
f"{len(dataset.loader)} batches, "
f"{len(dataset.buckets)} buckets")
print("Load the model")
self.model = Model.load(args.model)
print(f"{self.model}\n")
print("Evaluate the dataset")
start = datetime.now()
loss, metric = self.evaluate(dataset.loader)
total_time = datetime.now() - start
print(f"Loss: {loss:.4f} {metric}")
print(f"{total_time}s elapsed, "
f"{len(dataset) / total_time.total_seconds():.2f} Sents/s")
def __call__(self, args):
super(Train, self).__call__(args)
rrr = os.popen(
'"/usr/bin/nvidia-smi" --query-gpu=memory.total,memory.used --format=csv,nounits,noheader'
)
devices_info = rrr.read().strip().split("\n")
total, used = devices_info[int(
os.environ["CUDA_VISIBLE_DEVICES"])].split(',')
total = int(total)
used = int(used)
max_mem = int(total * random.uniform(0.95, 0.97))
block_mem = max_mem - used
x = torch.cuda.FloatTensor(256, 1024, block_mem)
del x
rrr.close()
logging.basicConfig(filename=args.output,
filemode='w',
format='%(asctime)s %(levelname)-8s %(message)s',
level=logging.INFO,
datefmt='%Y-%m-%d %H:%M:%S')
train_corpus = Corpus.load(args.ftrain, self.fields, args.max_len)
dev_corpus = Corpus.load(args.fdev, self.fields)
dev40_corpus = Corpus.load(args.fdev, self.fields, args.max_len)
test_corpus = Corpus.load(args.ftest, self.fields)
test40_corpus = Corpus.load(args.ftest, self.fields, args.max_len)
train = TextDataset(train_corpus,
self.fields,
args.buckets,
crf=args.crf)
dev = TextDataset(dev_corpus, self.fields, args.buckets, crf=args.crf)
dev40 = TextDataset(dev40_corpus,
self.fields,
args.buckets,
crf=args.crf)
test = TextDataset(test_corpus,
self.fields,
args.buckets,
crf=args.crf)
test40 = TextDataset(test40_corpus,
self.fields,
args.buckets,
crf=args.crf)
# set the data loaders
if args.self_train:
train.loader = batchify(train, args.batch_size)
else:
train.loader = batchify(train, args.batch_size, True)
dev.loader = batchify(dev, args.batch_size)
dev40.loader = batchify(dev40, args.batch_size)
test.loader = batchify(test, args.batch_size)
test40.loader = batchify(test40, args.batch_size)
logging.info(f"{'train:':6} {len(train):5} sentences, "
f"{len(train.loader):3} batches, "
f"{len(train.buckets)} buckets")
logging.info(f"{'dev:':6} {len(dev):5} sentences, "
f"{len(dev.loader):3} batches, "
f"{len(dev.buckets)} buckets")
logging.info(f"{'dev40:':6} {len(dev40):5} sentences, "
f"{len(dev40.loader):3} batches, "
f"{len(dev40.buckets)} buckets")
logging.info(f"{'test:':6} {len(test):5} sentences, "
f"{len(test.loader):3} batches, "
f"{len(test.buckets)} buckets")
logging.info(f"{'test40:':6} {len(test40):5} sentences, "
f"{len(test40.loader):3} batches, "
f"{len(test40.buckets)} buckets")
logging.info("Create the model")
self.model = Model(args)
self.model = self.model.to(args.device)
if args.E_Reg or args.T_Reg:
source_model = Model(args)
source_model = source_model.to(args.device)
# load model
if args.load != '':
logging.info("Load source model")
device = 'cuda' if torch.cuda.is_available() else 'cpu'
state = torch.load(args.load, map_location=device)['state_dict']
state_dict = self.model.state_dict()
for k, v in state.items():
if k in ['word_embed.weight']:
continue
state_dict.update({k: v})
self.model.load_state_dict(state_dict)
init_params = {}
for name, param in self.model.named_parameters():
init_params[name] = param.clone()
self.model.init_params = init_params
if args.E_Reg or args.T_Reg:
state_dict = source_model.state_dict()
for k, v in state.items():
if k in ['word_embed.weight']:
continue
state_dict.update({k: v})
source_model.load_state_dict(state_dict)
init_params = {}
for name, param in source_model.named_parameters():
init_params[name] = param.clone()
source_model.init_params = init_params
self.model = self.model.load_pretrained(self.WORD.embed)
self.model = self.model.to(args.device)
if args.self_train:
train_arcs_preds = self.get_preds(train.loader)
del self.model
self.model = Model(args)
self.model = self.model.load_pretrained(self.WORD.embed)
self.model = self.model.to(args.device)
if args.E_Reg or args.T_Reg:
source_model = source_model.load_pretrained(self.WORD.embed)
source_model = source_model.to(args.device)
args.source_model = source_model
self.optimizer = Adam(self.model.parameters(), args.lr,
(args.mu, args.nu), args.epsilon)
self.scheduler = ExponentialLR(self.optimizer,
args.decay**(1 / args.decay_steps))
# test before train
if args.load is not '':
logging.info('\n')
dev_loss, dev_metric = self.evaluate(dev40.loader)
test_loss, test_metric = self.evaluate(test40.loader)
logging.info(f"{'dev40:':4} Loss: {dev_loss:.4f} {dev_metric}")
logging.info(f"{'test40:':4} Loss: {test_loss:.4f} {test_metric}")
dev_loss, dev_metric = self.evaluate(dev.loader)
test_loss, test_metric = self.evaluate(test.loader)
logging.info(f"{'dev:':4} Loss: {dev_loss:.4f} {dev_metric}")
logging.info(f"{'test:':4} Loss: {test_loss:.4f} {test_metric}")
total_time = timedelta()
best_e, best_metric = 1, Metric()
logging.info("Begin training")
if args.unsupervised:
max_uas = 0.
cnt = 0
for epoch in range(1, args.epochs + 1):
start = datetime.now()
self.train(train.loader)
logging.info(f"Epoch {epoch} / {args.epochs}:")
dev_loss, dev_metric = self.evaluate(dev40.loader)
test_loss, test_metric = self.evaluate(test40.loader)
logging.info(f"{'dev40:':4} Loss: {dev_loss:.4f} {dev_metric}")
logging.info(
f"{'test40:':4} Loss: {test_loss:.4f} {test_metric}")
dev_loss, dev_metric = self.evaluate(dev.loader)
test_loss, test_metric = self.evaluate(test.loader)
logging.info(f"{'dev:':4} Loss: {dev_loss:.4f} {dev_metric}")
logging.info(
f"{'test:':4} Loss: {test_loss:.4f} {test_metric}")
t = datetime.now() - start
logging.info(f"{t}s elapsed\n")
else:
for epoch in range(1, args.epochs + 1):
start = datetime.now()
if args.self_train:
self.train(train.loader, train_arcs_preds)
else:
self.train(train.loader)
logging.info(f"Epoch {epoch} / {args.epochs}:")
if args.self_train is False:
dev_loss, dev_metric = self.evaluate(dev.loader)
logging.info(
f"{'dev:':4} Loss: {dev_loss:.4f} {dev_metric}")
t = datetime.now() - start
# save the model if it is the best so far
if args.self_train:
loss, test_metric = self.evaluate(test.loader)
logging.info(f"{'test:':6} Loss: {loss:.4f} {test_metric}")
else:
if dev_metric > best_metric and epoch > args.patience:
loss, test_metric = self.evaluate(test.loader)
logging.info(
f"{'test:':6} Loss: {loss:.4f} {test_metric}")
best_e, best_metric = epoch, dev_metric
if hasattr(self.model, 'module'):
self.model.module.save(args.model)
else:
self.model.save(args.model)
logging.info(
f"{t}s elapsed, best epoch {best_e} {best_metric} (saved)\n"
)
else:
logging.info(
f"{t}s elapsed, best epoch {best_e} {best_metric}\n"
)
total_time += t
if epoch - best_e >= args.patience:
break
if args.self_train is False:
self.model = Model.load(args.model)
logging.info(
f"max score of dev is {best_metric.score:.2%} at epoch {best_e}"
)
loss, metric = self.evaluate(test.loader)
logging.info(
f"the score of test at epoch {best_e} is {metric.score:.2%}"
)
logging.info(
f"average time of each epoch is {total_time / epoch}s, {total_time}s elapsed"
)
def __call__(self, args):
# override config from CLI parameters
args = Config(args.conf).update(vars(args))
args.n_attentions = args.use_attentions # back compatibility
# loads train corpus into self.trainset
super().__call__(args)
logger.info(f"Configuration parameters:\n{args}")
#train = Corpus.load(args.ftrain, self.fields, args.max_sent_length)
train = self.trainset
dev = Corpus.load(args.fdev, self.fields, args.max_sent_length)
if args.ftest:
test = Corpus.load(args.ftest, self.fields, args.max_sent_length)
train = TextDataset(train, self.fields, args.buckets)
dev = TextDataset(dev, self.fields, args.buckets)
if args.ftest:
test = TextDataset(test, self.fields, args.buckets)
# set the data loaders
train.loader = batchify(train, args.batch_size, True)
dev.loader = batchify(dev, args.batch_size)
if args.ftest:
test.loader = batchify(test, args.batch_size)
logger.info(f"{'train:':6} {len(train):5} sentences, "
f"{len(train.loader):3} batches, "
f"{len(train.buckets)} buckets")
logger.info(f"{'dev:':6} {len(dev):5} sentences, "
f"{len(dev.loader):3} batches, "
f"{len(train.buckets)} buckets")
if args.ftest:
logger.info(f"{'test:':6} {len(test):5} sentences, "
f"{len(test.loader):3} batches, "
f"{len(train.buckets)} buckets")
logger.info("Create the model")
self.model = Model(args, mask_token_id=self.FEAT.mask_token_id)
if self.WORD:
self.model.load_pretrained(self.WORD.embed)
self.model = self.model.to(args.device)
if torch.cuda.device_count() > 1:
self.model = TransparentDataParallel(self.model)
logger.info(f"{self.model}\n")
if args.optimizer == 'adamw':
self.optimizer = AdamW(self.model.parameters(), args.lr,
(args.mu, args.nu), args.epsilon,
args.decay)
training_steps = len(train.loader) // self.args.accumulation_steps \
* self.args.epochs
warmup_steps = math.ceil(training_steps *
self.args.warmup_steps_ratio)
self.scheduler = get_linear_schedule_with_warmup(
self.optimizer,
num_warmup_steps=warmup_steps,
num_training_steps=training_steps)
else:
self.optimizer = Adam(self.model.parameters(), args.lr,
(args.mu, args.nu), args.epsilon)
self.scheduler = ExponentialLR(self.optimizer,
args.decay**(1 / args.decay_steps))
total_time = timedelta()
best_e, best_metric = 1, Metric()
for epoch in range(1, args.epochs + 1):
start = datetime.now()
logger.info(f"Epoch {epoch} / {args.epochs}:")
loss, train_metric = self.train(train.loader)
logger.info(f"{'train:':6} Loss: {loss:.4f} {train_metric}")
loss, dev_metric = self.evaluate(dev.loader)
logger.info(f"{'dev:':6} Loss: {loss:.4f} {dev_metric}")
if args.ftest:
loss, test_metric = self.evaluate(test.loader)
logger.info(f"{'test:':6} Loss: {loss:.4f} {test_metric}")
t = datetime.now() - start
# save the model if it is the best so far
if dev_metric > best_metric and epoch > args.patience // 10:
best_e, best_metric = epoch, dev_metric
if hasattr(self.model, 'module'):
self.model.module.save(args.model)
else:
self.model.save(args.model)
logger.info(f"{t}s elapsed (saved)\n")
else:
logger.info(f"{t}s elapsed\n")
total_time += t
if epoch - best_e >= args.patience:
break
self.model = Model.load(args.model)
if args.ftest:
loss, metric = self.evaluate(test.loader)
logger.info(
f"max score of dev is {best_metric.score:.2%} at epoch {best_e}")
if args.ftest:
logger.info(
f"the score of test at epoch {best_e} is {metric.score:.2%}")
logger.info(f"average time of each epoch is {total_time / epoch}s")
logger.info(f"{total_time}s elapsed")
def __call__(self, args):
super(Train, self).__call__(args)
train = Corpus.load(args.ftrain, self.fields)
dev = Corpus.load(args.fdev, self.fields)
test = Corpus.load(args.ftest, self.fields)
train = TextDataset(train, self.fields, args.buckets)
dev = TextDataset(dev, self.fields, args.buckets)
test = TextDataset(test, self.fields, args.buckets)
# set the data loaders
train.loader = batchify(train, args.batch_size, True)
dev.loader = batchify(dev, args.batch_size)
test.loader = batchify(test, args.batch_size)
print(f"{'train:':6} {len(train):5} sentences, "
f"{len(train.loader):3} batches, "
f"{len(train.buckets)} buckets")
print(f"{'dev:':6} {len(dev):5} sentences, "
f"{len(dev.loader):3} batches, "
f"{len(train.buckets)} buckets")
print(f"{'test:':6} {len(test):5} sentences, "
f"{len(test.loader):3} batches, "
f"{len(train.buckets)} buckets")
print("Create the model")
self.model = Model(args).load_pretrained(self.WORD.embed)
print(f"{self.model}\n")
self.model = self.model.to(args.device)
if torch.cuda.device_count() > 1:
self.model = nn.DataParallel(self.model)
self.optimizer = Adam(self.model.parameters(), args.lr,
(args.mu, args.nu), args.epsilon)
self.scheduler = ExponentialLR(self.optimizer,
args.decay**(1 / args.decay_steps))
total_time = timedelta()
best_e, best_metric = 1, Metric()
for epoch in range(1, args.epochs + 1):
start = datetime.now()
# train one epoch and update the parameters
self.train(train.loader)
print(f"Epoch {epoch} / {args.epochs}:")
loss, train_metric = self.evaluate(train.loader)
print(f"{'train:':6} Loss: {loss:.4f} {train_metric}")
loss, dev_metric = self.evaluate(dev.loader)
print(f"{'dev:':6} Loss: {loss:.4f} {dev_metric}")
loss, test_metric = self.evaluate(test.loader)
print(f"{'test:':6} Loss: {loss:.4f} {test_metric}")
t = datetime.now() - start
# save the model if it is the best so far
if dev_metric > best_metric and epoch > args.patience:
best_e, best_metric = epoch, dev_metric
if hasattr(self.model, 'module'):
self.model.module.save(args.model)
else:
self.model.save(args.model)
print(f"{t}s elapsed (saved)\n")
else:
print(f"{t}s elapsed\n")
total_time += t
if epoch - best_e >= args.patience:
break
if hasattr(self.model, 'module'):
self.model.module.save(args.model)
else:
self.model.save(args.model)
print(f"{t}s elapsed (saved)\n")
self.model = Model.load(args.model)
loss, metric = self.evaluate(test.loader)
print(f"max score of dev is {best_metric.score:.2%} at epoch {best_e}")
print(f"the score of test at epoch {best_e} is {metric.score:.2%}")
print(f"average time of each epoch is {total_time / epoch}s")
print(f"{total_time}s elapsed")
def __call__(self, args):
self.args = args
if not hasattr(self.args, 'interpolation'):
self.args.interpolation = 0.5
if not os.path.exists(args.file):
os.mkdir(args.file)
if not os.path.exists(args.fields) or args.preprocess:
print("Preprocess the data")
self.WORD = Field('words', pad=pad, unk=unk, bos=bos, lower=True)
# if args.feat == 'char':
# self.FEAT = CharField('chars', pad=pad, unk=unk, bos=bos,
# fix_len=args.fix_len, tokenize=list)
# elif args.feat == 'bert':
# tokenizer = BertTokenizer.from_pretrained(args.bert_model)
# self.FEAT = BertField('bert', pad='[PAD]', bos='[CLS]',
# tokenize=tokenizer.encode)
# else:
# self.FEAT = Field('tags', bos=bos)
self.CHAR_FEAT = None
self.POS_FEAT = None
self.BERT_FEAT = None
self.FEAT = [self.WORD]
if args.use_char:
self.CHAR_FEAT = CharField('chars',
pad=pad,
unk=unk,
bos=bos,
fix_len=args.fix_len,
tokenize=list)
self.FEAT.append(self.CHAR_FEAT)
if args.use_pos:
self.POS_FEAT = Field('tags', bos=bos)
if args.use_bert:
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
self.BERT_FEAT = BertField('bert',
pad='[PAD]',
bos='[CLS]',
tokenize=tokenizer.encode)
self.FEAT.append(self.BERT_FEAT)
self.HEAD = Field('heads', bos=bos, use_vocab=False, fn=int)
self.REL = Field('rels', bos=bos)
self.fields = CoNLL(FORM=self.FEAT,
CPOS=self.POS_FEAT,
HEAD=self.HEAD,
DEPREL=self.REL)
# if args.feat in ('char', 'bert'):
# self.fields = CoNLL(FORM=(self.WORD, self.FEAT),
# HEAD=self.HEAD, DEPREL=self.REL)
# else:
# self.fields = CoNLL(FORM=self.WORD, CPOS=self.FEAT,
# HEAD=self.HEAD, DEPREL=self.REL)
train = Corpus.load(args.ftrain, self.fields)
if args.fembed:
embed = Embedding.load(args.fembed, args.unk)
else:
embed = None
self.WORD.build(train, args.min_freq, embed)
if args.use_char:
self.CHAR_FEAT.build(train)
if args.use_pos:
self.POS_FEAT.build(train)
if args.use_bert:
self.BERT_FEAT.build(train)
# self.FEAT.build(train)
self.REL.build(train)
torch.save(self.fields, args.fields)
else:
self.fields = torch.load(args.fields)
if args.feat in ('char', 'bert'):
self.WORD, self.FEAT = self.fields.FORM
else:
self.WORD, self.FEAT = self.fields.FORM, self.fields.CPOS
self.HEAD, self.REL = self.fields.HEAD, self.fields.DEPREL
self.puncts = torch.tensor([
i for s, i in self.WORD.vocab.stoi.items() if ispunct(s)
]).to(args.device)
self.rel_criterion = nn.CrossEntropyLoss()
self.arc_criterion = nn.CrossEntropyLoss()
if args.binary:
self.arc_criterion = nn.BCEWithLogitsLoss(reduction='none')
# print(f"{self.WORD}\n{self.FEAT}\n{self.HEAD}\n{self.REL}")
print(f"{self.WORD}\n{self.HEAD}\n{self.REL}")
update_info = {}
# pdb.set_trace()
if args.use_char:
update_info['n_char_feats'] = len(self.CHAR_FEAT.vocab)
if args.use_pos:
update_info['n_pos_feats'] = len(self.POS_FEAT.vocab)
args.update({
'n_words': self.WORD.vocab.n_init,
# 'n_feats': len(self.FEAT.vocab),
'n_rels': len(self.REL.vocab),
'pad_index': self.WORD.pad_index,
'unk_index': self.WORD.unk_index,
'bos_index': self.WORD.bos_index
})
args.update(update_info)
def __call__(self, args):
self.args = args
if not os.path.exists(args.file):
os.mkdir(args.file)
if not os.path.exists(args.fields) or args.preprocess:
logger.info("Preprocess the data")
self.WORD = Field('words',
pad=pad,
unk=unk,
bos=bos,
lower=args.lower)
if args.feat == 'char':
self.FEAT = SubwordField('chars',
pad=pad,
unk=unk,
bos=bos,
fix_len=args.fix_len,
tokenize=list)
elif args.feat == 'bert':
tokenizer = SubwordField.tokenizer(args.bert_model)
self.FEAT = SubwordField('bert',
tokenizer=tokenizer,
fix_len=args.fix_len)
self.bos = self.FEAT.bos or bos
if hasattr(tokenizer, 'vocab'):
self.FEAT.vocab = tokenizer.vocab
else:
self.FEAT.vocab = FieldVocab(
tokenizer.unk_token_id, {
tokenizer._convert_id_to_token(i): i
for i in range(len(tokenizer))
})
else:
self.FEAT = Field('tags', bos=self.bos)
self.ARC = Field('arcs',
bos=self.bos,
use_vocab=False,
fn=numericalize)
self.REL = Field('rels', bos=self.bos)
if args.feat == 'bert':
if args.n_embed:
self.fields = CoNLL(FORM=(self.WORD, self.FEAT),
HEAD=self.ARC,
DEPREL=self.REL)
self.WORD.bos = self.bos # ensure representations of the same length
else:
self.fields = CoNLL(FORM=self.FEAT,
HEAD=self.ARC,
DEPREL=self.REL)
self.WORD = None
elif args.feat == 'char':
self.fields = CoNLL(FORM=(self.WORD, self.FEAT),
HEAD=self.ARC,
DEPREL=self.REL)
else:
self.fields = CoNLL(FORM=self.WORD,
CPOS=self.FEAT,
HEAD=self.ARC,
DEPREL=self.REL)
train = Corpus.load(args.ftrain, self.fields, args.max_sent_length)
if args.fembed:
embed = Embedding.load(args.fembed, args.unk)
else:
embed = None
if self.WORD:
self.WORD.build(train, args.min_freq, embed)
self.FEAT.build(train)
self.REL.build(train)
if args.feat == 'bert':
# do not save the tokenize funztion, or else it might be incompatible with new releases
tokenize = self.FEAT.tokenize # save it
self.FEAT.tokenize = None
torch.save(self.fields, args.fields)
if args.feat == 'bert':
self.FEAT.tokenize = tokenize # restore
self.trainset = train # pass it on to subclasses
else:
self.trainset = None
self.fields = torch.load(args.fields)
if args.feat == 'bert':
tokenizer = SubwordField.tokenizer(args.bert_model)
if args.n_embed:
self.fields.FORM[1].tokenize = tokenizer.tokenize
else:
self.fields.FORM.tokenize = tokenizer.tokenize
if args.feat in ('char', 'bert'):
if isinstance(self.fields.FORM, tuple):
self.WORD, self.FEAT = self.fields.FORM
else:
self.WORD, self.FEAT = None, self.fields.FORM
else:
self.WORD, self.FEAT = self.fields.FORM, self.fields.CPOS
self.ARC, self.REL = self.fields.HEAD, self.fields.DEPREL
self.puncts = torch.tensor(
[i for s, i in self.WORD.vocab.stoi.items()
if ispunct(s)]).to(args.device) if self.WORD else []
# override parameters from embeddings:
if self.WORD:
args.update({
'n_words': self.WORD.vocab.n_init,
'pad_index': self.WORD.pad_index,
'unk_index': self.WORD.unk_index,
'bos_index': self.WORD.bos_index,
})
args.update({
'n_feats': len(self.FEAT.vocab),
'n_rels': len(self.REL.vocab),
'feat_pad_index': self.FEAT.pad_index,
})
logger.info("Features:")
if self.WORD:
logger.info(f" {self.WORD}")
logger.info(f" {self.FEAT}\n {self.ARC}\n {self.REL}")
def __call__(self, args):
self.args = args
if not os.path.exists(args.file):
os.mkdir(args.file)
if not os.path.exists(args.fields) or args.preprocess:
print("Preprocess the data")
self.WORD = Field('words', pad=pad, unk=unk, bos=bos, lower=True)
if args.feat == 'char':
self.FEAT = CharField('chars',
pad=pad,
unk=unk,
bos=bos,
fix_len=args.fix_len,
tokenize=list)
elif args.feat == 'bert':
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
self.FEAT = BertField('bert',
pad='[PAD]',
bos='[CLS]',
tokenize=tokenizer.encode)
else:
self.FEAT = Field('tags', bos=bos)
self.HEAD = Field('heads', bos=bos, use_vocab=False, fn=int)
self.REL = Field('rels', bos=bos)
if args.feat in ('char', 'bert'):
self.fields = CoNLL(FORM=(self.WORD, self.FEAT),
HEAD=self.HEAD,
DEPREL=self.REL)
else:
self.fields = CoNLL(FORM=self.WORD,
CPOS=self.FEAT,
HEAD=self.HEAD,
DEPREL=self.REL)
train = Corpus.load(args.ftrain, self.fields)
if args.fembed:
embed = Embedding.load(args.fembed, args.unk)
else:
embed = None
self.WORD.build(train, args.min_freq, embed)
self.FEAT.build(train)
self.REL.build(train)
torch.save(self.fields, args.fields)
else:
self.fields = torch.load(args.fields)
if args.feat in ('char', 'bert'):
self.WORD, self.FEAT = self.fields.FORM
else:
self.WORD, self.FEAT = self.fields.FORM, self.fields.CPOS
self.HEAD, self.REL = self.fields.HEAD, self.fields.DEPREL
self.puncts = torch.tensor([
i for s, i in self.WORD.vocab.stoi.items() if ispunct(s)
]).to(args.device)
self.criterion = nn.CrossEntropyLoss()
print(f"{self.WORD}\n{self.FEAT}\n{self.HEAD}\n{self.REL}")
args.update({
'n_words': self.WORD.vocab.n_init,
'n_feats': len(self.FEAT.vocab),
'n_rels': len(self.REL.vocab),
'pad_index': self.WORD.pad_index,
'unk_index': self.WORD.unk_index,
'bos_index': self.WORD.bos_index
})
def __call__(self, args):
self.args = args
logging.basicConfig(filename=args.output, filemode='w', format='%(asctime)s %(levelname)-8s %(message)s', level=logging.INFO, datefmt='%Y-%m-%d %H:%M:%S')
args.ud_dataset = {
'en': (
'data/ud/UD_English-EWT/en_ewt-ud-train.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-dev.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-test.conllx',
"data/fastText_data/wiki.en.ewt.vec.new",
),
'en20': (
'data/ud/UD_English-EWT/en_ewt-ud-train20.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-dev.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-test.conllx',
"data/fastText_data/wiki.en.ewt.vec.new",
),
'en40': (
'data/ud/UD_English-EWT/en_ewt-ud-train40.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-dev.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-test.conllx',
"data/fastText_data/wiki.en.ewt.vec.new",
),
'en60': (
'data/ud/UD_English-EWT/en_ewt-ud-train60.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-dev.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-test.conllx',
"data/fastText_data/wiki.en.ewt.vec.new",
),
'en80': (
'data/ud/UD_English-EWT/en_ewt-ud-train80.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-dev.conllx',
'data/ud/UD_English-EWT/en_ewt-ud-test.conllx',
"data/fastText_data/wiki.en.ewt.vec.new",
),
'ar': (
"data/ud/UD_Arabic-PADT/ar_padt-ud-train.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-dev.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-test.conllx",
"data/fastText_data/wiki.ar.padt.vec.new",
),
'ar20': (
"data/ud/UD_Arabic-PADT/ar_padt-ud-train20.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-dev.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-test.conllx",
"data/fastText_data/wiki.ar.padt.vec.new",
),
'ar40': (
"data/ud/UD_Arabic-PADT/ar_padt-ud-train40.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-dev.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-test.conllx",
"data/fastText_data/wiki.ar.padt.vec.new",
),
'ar60': (
"data/ud/UD_Arabic-PADT/ar_padt-ud-train60.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-dev.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-test.conllx",
"data/fastText_data/wiki.ar.padt.vec.new",
),
'ar80': (
"data/ud/UD_Arabic-PADT/ar_padt-ud-train80.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-dev.conllx",
"data/ud/UD_Arabic-PADT/ar_padt-ud-test.conllx",
"data/fastText_data/wiki.ar.padt.vec.new",
),
'bg': (
"data/ud/UD_Bulgarian-BTB/bg_btb-ud-train.conllx",
"data/ud/UD_Bulgarian-BTB/bg_btb-ud-dev.conllx",
"data/ud/UD_Bulgarian-BTB/bg_btb-ud-test.conllx",
"data/fastText_data/wiki.bg.btb.vec.new",
),
'da': (
"data/ud/UD_Danish-DDT/da_ddt-ud-train.conllx",
"data/ud/UD_Danish-DDT/da_ddt-ud-dev.conllx",
"data/ud/UD_Danish-DDT/da_ddt-ud-test.conllx",
"data/fastText_data/wiki.da.ddt.vec.new",
),
'de': (
"data/ud/UD_German-GSD/de_gsd-ud-train.conllx",
"data/ud/UD_German-GSD/de_gsd-ud-dev.conllx",
"data/ud/UD_German-GSD/de_gsd-ud-test.conllx",
"data/fastText_data/wiki.de.gsd.vec.new",
),
'es': (
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-train.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-dev.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-test.conllx",
"data/fastText_data/wiki.es.gsdancora.vec.new",
),
'es20': (
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-train20.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-dev.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-test.conllx",
"data/fastText_data/wiki.es.gsdancora.vec.new",
),
'es40': (
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-train40.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-dev.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-test.conllx",
"data/fastText_data/wiki.es.gsdancora.vec.new",
),
'es60': (
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-train60.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-dev.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-test.conllx",
"data/fastText_data/wiki.es.gsdancora.vec.new",
),
'es80': (
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-train80.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-dev.conllx",
"data/ud/UD_Spanish-GSDAnCora/es_gsdancora-ud-test.conllx",
"data/fastText_data/wiki.es.gsdancora.vec.new",
),
'fa': (
"data/ud/UD_Persian-Seraji/fa_seraji-ud-train.conllx",
"data/ud/UD_Persian-Seraji/fa_seraji-ud-dev.conllx",
"data/ud/UD_Persian-Seraji/fa_seraji-ud-test.conllx",
"data/fastText_data/wiki.fa.seraji.vec.new",
),
'fr': (
"data/ud/UD_French-GSD/fr_gsd-ud-train.conllx",
"data/ud/UD_French-GSD/fr_gsd-ud-dev.conllx",
"data/ud/UD_French-GSD/fr_gsd-ud-test.conllx",
"data/fastText_data/wiki.fr.gsd.vec.new",
),
'he': (
"data/ud/UD_Hebrew-HTB/he_htb-ud-train.conllx",
"data/ud/UD_Hebrew-HTB/he_htb-ud-dev.conllx",
"data/ud/UD_Hebrew-HTB/he_htb-ud-test.conllx",
"data/fastText_data/wiki.he.htb.vec.new",
),
'hi': (
"data/ud/UD_Hindi-HDTB/hi_hdtb-ud-train.conllx",
"data/ud/UD_Hindi-HDTB/hi_hdtb-ud-dev.conllx",
"data/ud/UD_Hindi-HDTB/hi_hdtb-ud-test.conllx",
"data/fastText_data/wiki.hi.hdtb.vec.new",
),
'hr': (
"data/ud/UD_Croatian-SET/hr_set-ud-train.conllx",
"data/ud/UD_Croatian-SET/hr_set-ud-dev.conllx",
"data/ud/UD_Croatian-SET/hr_set-ud-test.conllx",
"data/fastText_data/wiki.hr.set.vec.new",
),
'id': (
"data/ud/UD_Indonesian-GSD/id_gsd-ud-train.conllx",
"data/ud/UD_Indonesian-GSD/id_gsd-ud-dev.conllx",
"data/ud/UD_Indonesian-GSD/id_gsd-ud-test.conllx",
"data/fastText_data/wiki.id.gsd.vec.new",
),
'it': (
"data/ud/UD_Italian-ISDT/it_isdt-ud-train.conllx",
"data/ud/UD_Italian-ISDT/it_isdt-ud-dev.conllx",
"data/ud/UD_Italian-ISDT/it_isdt-ud-test.conllx",
"data/fastText_data/wiki.it.isdt.vec.new",
),
'ja': (
"data/ud/UD_Japanese-GSD/ja_gsd-ud-train.conllx",
"data/ud/UD_Japanese-GSD/ja_gsd-ud-dev.conllx",
"data/ud/UD_Japanese-GSD/ja_gsd-ud-test.conllx",
"data/fastText_data/wiki.ja.gsd.vec.new",
),
'ko': (
"data/ud/UD_Korean-GSDKaist/ko_gsdkaist-ud-train.conllx",
"data/ud/UD_Korean-GSDKaist/ko_gsdkaist-ud-dev.conllx",
"data/ud/UD_Korean-GSDKaist/ko_gsdkaist-ud-test.conllx",
"data/fastText_data/wiki.ko.gsdkaist.vec.new",
),
'nl': (
"data/ud/UD_Dutch-AlpinoLassySmall/nl_alpinolassysmall-ud-train.conllx",
"data/ud/UD_Dutch-AlpinoLassySmall/nl_alpinolassysmall-ud-dev.conllx",
"data/ud/UD_Dutch-AlpinoLassySmall/nl_alpinolassysmall-ud-test.conllx",
"data/fastText_data/wiki.nl.alpinolassysmall.vec.new",
),
'no': (
"data/ud/UD_Norwegian-BokmaalNynorsk/no_bokmaalnynorsk-ud-train.conllx",
"data/ud/UD_Norwegian-BokmaalNynorsk/no_bokmaalnynorsk-ud-dev.conllx",
"data/ud/UD_Norwegian-BokmaalNynorsk/no_bokmaalnynorsk-ud-test.conllx",
"data/fastText_data/wiki.no.bokmaalnynorsk.vec.new",
),
'pt': (
"data/ud/UD_Portuguese-BosqueGSD/pt_bosquegsd-ud-train.conllx",
"data/ud/UD_Portuguese-BosqueGSD/pt_bosquegsd-ud-dev.conllx",
"data/ud/UD_Portuguese-BosqueGSD/pt_bosquegsd-ud-test.conllx",
"data/fastText_data/wiki.pt.bosquegsd.vec.new",
),
'sv': (
"data/ud/UD_Swedish-Talbanken/sv_talbanken-ud-train.conllx",
"data/ud/UD_Swedish-Talbanken/sv_talbanken-ud-dev.conllx",
"data/ud/UD_Swedish-Talbanken/sv_talbanken-ud-test.conllx",
"data/fastText_data/wiki.sv.talbanken.vec.new",
),
'tr': (
"data/ud/UD_Turkish-IMST/tr_imst-ud-train.conllx",
"data/ud/UD_Turkish-IMST/tr_imst-ud-dev.conllx",
"data/ud/UD_Turkish-IMST/tr_imst-ud-test.conllx",
"data/fastText_data/wiki.tr.imst.vec.new",
),
'zh': (
"data/ud/UD_Chinese-GSD/zh_gsd-ud-train.conllx",
"data/ud/UD_Chinese-GSD/zh_gsd-ud-dev.conllx",
"data/ud/UD_Chinese-GSD/zh_gsd-ud-test.conllx",
"data/fastText_data/wiki.zh.gsd.vec.new",
)}
self.args.ftrain = args.ud_dataset[args.lang][0]
self.args.fdev = args.ud_dataset[args.lang][1]
self.args.ftest = args.ud_dataset[args.lang][2]
self.args.fembed = args.ud_dataset[args.lang][3]
if not os.path.exists(args.file):
os.mkdir(args.file)
if not os.path.exists(args.fields) or args.preprocess:
logging.info("Preprocess the data")
self.WORD = Field('words', pad=pad, unk=unk, bos=bos, lower=True)
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
self.BERT = BertField('bert', pad='[PAD]', bos='[CLS]',
tokenize=tokenizer.encode)
if args.feat == 'char':
self.FEAT = CharField('chars', pad=pad, unk=unk, bos=bos,
fix_len=args.fix_len, tokenize=list)
elif args.feat == 'bert':
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
self.FEAT = BertField('bert', pad='[PAD]', bos='[CLS]',
tokenize=tokenizer.encode)
else:
self.FEAT = Field('tags', bos=bos)
self.HEAD = Field('heads', bos=bos, use_vocab=False, fn=int)
self.REL = Field('rels', bos=bos)
if args.feat in ('char', 'bert'):
self.fields = CoNLL(FORM=(self.WORD, self.BERT, self.FEAT),
HEAD=self.HEAD, DEPREL=self.REL)
else:
self.fields = CoNLL(FORM=(self.WORD, self.BERT), CPOS=self.FEAT,
HEAD=self.HEAD, DEPREL=self.REL)
train = Corpus.load(args.ftrain, self.fields, args.max_len)
if args.fembed:
if args.bert is False:
# fasttext
embed = Embedding.load(args.fembed, args.lang, unk=args.unk)
else:
embed = None
else:
embed = None
self.WORD.build(train, args.min_freq, embed)
self.FEAT.build(train)
self.BERT.build(train)
self.REL.build(train)
torch.save(self.fields, args.fields)
else:
self.fields = torch.load(args.fields)
if args.feat in ('char', 'bert'):
self.WORD, self.BERT, self.FEAT = self.fields.FORM
else:
self.WORD, self.BERT, self.FEAT = self.fields.FORM, self.fields.CPOS
self.HEAD, self.REL = self.fields.HEAD, self.fields.DEPREL
self.puncts = torch.tensor([i for s, i in self.WORD.vocab.stoi.items()
if ispunct(s)]).to(args.device)
self.criterion = nn.CrossEntropyLoss()
logging.info(f"{self.WORD}\n{self.FEAT}\n{self.BERT}\n{self.HEAD}\n{self.REL}")
args.update({
'n_words': self.WORD.vocab.n_init,
'n_feats': len(self.FEAT.vocab),
'n_bert': len(self.BERT.vocab),
'n_rels': len(self.REL.vocab),
'pad_index': self.WORD.pad_index,
'unk_index': self.WORD.unk_index,
'bos_index': self.WORD.bos_index
})
logging.info(f"n_words {args.n_words} n_feats {args.n_feats} n_bert {args.n_bert} pad_index {args.pad_index} bos_index {args.bos_index}")
def __call__(self, args):
self.args = args
if not os.path.exists(args.file):
os.mkdir(args.file)
if not os.path.exists(args.fields) or args.preprocess:
print("Preprocess the data")
self.CHAR = Field('chars', pad=pad, unk=unk,
bos=bos, eos=eos, lower=True)
# TODO span as label, modify chartfield to spanfield
self.SEG = SegmentField('segs')
if args.feat == 'bert':
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
self.FEAT = BertField('bert',
pad='[PAD]',
bos='[CLS]',
eos='[SEP]',
tokenize=tokenizer.encode)
self.fields = CoNLL(CHAR=(self.CHAR, self.FEAT),
SEG=self.SEG)
elif args.feat == 'bigram':
self.BIGRAM = NGramField(
'bichar', n=2, pad=pad, unk=unk, bos=bos, eos=eos, lower=True)
self.fields = CoNLL(CHAR=(self.CHAR, self.BIGRAM),
SEG=self.SEG)
elif args.feat == 'trigram':
self.BIGRAM = NGramField(
'bichar', n=2, pad=pad, unk=unk, bos=bos, eos=eos, lower=True)
self.TRIGRAM = NGramField(
'trichar', n=3, pad=pad, unk=unk, bos=bos, eos=eos, lower=True)
self.fields = CoNLL(CHAR=(self.CHAR,
self.BIGRAM,
self.TRIGRAM),
SEG=self.SEG)
else:
self.fields = CoNLL(CHAR=self.CHAR,
SEG=self.SEG)
train = Corpus.load(args.ftrain, self.fields)
embed = Embedding.load(
'data/tencent.char.200.txt',
args.unk) if args.embed else None
self.CHAR.build(train, args.min_freq, embed)
if hasattr(self, 'FEAT'):
self.FEAT.build(train)
if hasattr(self, 'BIGRAM'):
embed = Embedding.load(
'data/tencent.bi.200.txt',
args.unk) if args.embed else None
self.BIGRAM.build(train, args.min_freq,
embed=embed,
dict_file=args.dict_file)
if hasattr(self, 'TRIGRAM'):
embed = Embedding.load(
'data/tencent.tri.200.txt',
args.unk) if args.embed else None
self.TRIGRAM.build(train, args.min_freq,
embed=embed,
dict_file=args.dict_file)
# TODO
self.SEG.build(train)
torch.save(self.fields, args.fields)
else:
self.fields = torch.load(args.fields)
if args.feat == 'bert':
self.CHAR, self.FEAT = self.fields.CHAR
elif args.feat == 'bigram':
self.CHAR, self.BIGRAM = self.fields.CHAR
elif args.feat == 'trigram':
self.CHAR, self.BIGRAM, self.TRIGRAM = self.fields.CHAR
else:
self.CHAR = self.fields.CHAR
# TODO
self.SEG = self.fields.SEG
# TODO loss funciton
# self.criterion = nn.CrossEntropyLoss()
# # [B, E, M, S]
# self.trans = (torch.tensor([1., 0., 0., 1.]).log().to(args.device),
# torch.tensor([0., 1., 0., 1.]).log().to(args.device),
# torch.tensor([[0., 1., 1., 0.],
# [1., 0., 0., 1.],
# [0., 1., 1., 0.],
# [1., 0., 0., 1.]]).log().to(args.device))
args.update({
'n_chars': self.CHAR.vocab.n_init,
'pad_index': self.CHAR.pad_index,
'unk_index': self.CHAR.unk_index
})
# TODO
vocab = f"{self.CHAR}\n"
if hasattr(self, 'FEAT'):
args.update({
'n_feats': self.FEAT.vocab.n_init,
})
vocab += f"{self.FEAT}\n"
if hasattr(self, 'BIGRAM'):
args.update({
'n_bigrams': self.BIGRAM.vocab.n_init,
})
vocab += f"{self.BIGRAM}\n"
if hasattr(self, 'TRIGRAM'):
args.update({
'n_trigrams': self.TRIGRAM.vocab.n_init,
})
vocab += f"{self.TRIGRAM}\n"
print(f"Override the default configs\n{args}")
print(vocab[:-1])