def build(cls, path, min_freq=2, fix_len=20, **kwargs):
r"""
Build a brand-new Parser, including initialization of all data fields and model parameters.
Args:
path (str):
The path of the model to be saved.
min_freq (str):
The minimum frequency needed to include a token in the vocabulary. Default: 2.
fix_len (int):
The max length of all subword pieces. The excess part of each piece will be truncated.
Required if using CharLSTM/BERT.
Default: 20.
kwargs (dict):
A dict holding the unconsumed arguments.
"""
args = Config(**locals())
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
os.makedirs(os.path.dirname(path) or './', exist_ok=True)
if os.path.exists(path) and not args.build:
parser = cls.load(**args)
parser.model = cls.MODEL(**parser.args)
parser.model.load_pretrained(parser.WORD.embed).to(args.device)
return parser
logger.info("Building the fields")
WORD = Field('words', pad=PAD, unk=UNK, bos=BOS, eos=EOS, lower=True)
TAG, CHAR, ELMO, BERT = None, None, None, None
if args.encoder == 'bert':
from transformers import (AutoTokenizer, GPT2Tokenizer,
GPT2TokenizerFast)
t = AutoTokenizer.from_pretrained(args.bert)
WORD = SubwordField(
'words',
pad=t.pad_token,
unk=t.unk_token,
bos=t.cls_token or t.cls_token,
eos=t.sep_token or t.sep_token,
fix_len=args.fix_len,
tokenize=t.tokenize,
fn=None if not isinstance(t,
(GPT2Tokenizer, GPT2TokenizerFast))
else lambda x: ' ' + x)
WORD.vocab = t.get_vocab()
else:
WORD = Field('words',
pad=PAD,
unk=UNK,
bos=BOS,
eos=EOS,
lower=True)
if 'tag' in args.feat:
TAG = Field('tags', bos=BOS, eos=EOS)
if 'char' in args.feat:
CHAR = SubwordField('chars',
pad=PAD,
unk=UNK,
bos=BOS,
eos=EOS,
fix_len=args.fix_len)
if 'elmo' in args.feat:
from allennlp.modules.elmo import batch_to_ids
ELMO = RawField('elmo')
ELMO.compose = lambda x: batch_to_ids(x).to(WORD.device)
if 'bert' in args.feat:
from transformers import (AutoTokenizer, GPT2Tokenizer,
GPT2TokenizerFast)
t = AutoTokenizer.from_pretrained(args.bert)
BERT = SubwordField(
'bert',
pad=t.pad_token,
unk=t.unk_token,
bos=t.cls_token or t.cls_token,
eos=t.sep_token or t.sep_token,
fix_len=args.fix_len,
tokenize=t.tokenize,
fn=None
if not isinstance(t, (GPT2Tokenizer, GPT2TokenizerFast))
else lambda x: ' ' + x)
BERT.vocab = t.get_vocab()
TREE = RawField('trees')
CHART = ChartField('charts')
transform = Tree(WORD=(WORD, CHAR, ELMO, BERT),
POS=TAG,
TREE=TREE,
CHART=CHART)
train = Dataset(transform, args.train)
if args.encoder != 'bert':
WORD.build(
train, args.min_freq,
(Embedding.load(args.embed, args.unk) if args.embed else None))
if TAG is not None:
TAG.build(train)
if CHAR is not None:
CHAR.build(train)
CHART.build(train)
args.update({
'n_words':
len(WORD.vocab) if args.encoder == 'bert' else WORD.vocab.n_init,
'n_labels':
len(CHART.vocab),
'n_tags':
len(TAG.vocab) if TAG is not None else None,
'n_chars':
len(CHAR.vocab) if CHAR is not None else None,
'char_pad_index':
CHAR.pad_index if CHAR is not None else None,
'bert_pad_index':
BERT.pad_index if BERT is not None else None,
'pad_index':
WORD.pad_index,
'unk_index':
WORD.unk_index,
'bos_index':
WORD.bos_index,
'eos_index':
WORD.eos_index
})
logger.info(f"{transform}")
logger.info("Building the model")
model = cls.MODEL(**args).load_pretrained(
WORD.embed if hasattr(WORD, 'embed') else None).to(args.device)
logger.info(f"{model}\n")
return cls(args, model, transform)
def build(cls, path,
optimizer_args={'lr': 2e-3, 'betas': (.9, .9), 'eps': 1e-12},
scheduler_args={'gamma': .75**(1/5000)},
min_freq=2,
fix_len=20,
**kwargs):
r"""
Build a brand-new Parser, including initialization of all data fields and model parameters.
Args:
path (str):
The path of the model to be saved.
optimizer_args (dict):
Arguments for creating an optimizer.
scheduler_args (dict):
Arguments for creating a scheduler.
min_freq (str):
The minimum frequency needed to include a token in the vocabulary. Default: 2.
fix_len (int):
The max length of all subword pieces. The excess part of each piece will be truncated.
Required if using CharLSTM/BERT.
Default: 20.
kwargs (dict):
A dict holding the unconsumed arguments.
"""
args = Config(**locals())
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
os.makedirs(os.path.dirname(path), exist_ok=True)
if os.path.exists(path) and not args.build:
parser = cls.load(**args)
parser.model = cls.MODEL(**parser.args)
parser.model.load_pretrained(parser.WORD.embed).to(args.device)
return parser
logger.info("Building the fields")
WORD = Field('words', pad=pad, unk=unk, bos=bos, eos=eos, lower=True)
if args.feat == 'char':
FEAT = SubwordField('chars', pad=pad, unk=unk, bos=bos, eos=eos, fix_len=args.fix_len)
elif args.feat == 'bert':
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(args.bert)
FEAT = SubwordField('bert',
pad=tokenizer.pad_token,
unk=tokenizer.unk_token,
bos=tokenizer.cls_token or tokenizer.cls_token,
eos=tokenizer.sep_token or tokenizer.sep_token,
fix_len=args.fix_len,
tokenize=tokenizer.tokenize)
FEAT.vocab = tokenizer.get_vocab()
else:
FEAT = Field('tags', bos=bos, eos=eos)
TREE = RawField('trees')
CHART = ChartField('charts')
if args.feat in ('char', 'bert'):
transform = Tree(WORD=(WORD, FEAT), TREE=TREE, CHART=CHART)
else:
transform = Tree(WORD=WORD, POS=FEAT, TREE=TREE, CHART=CHART)
train = Dataset(transform, args.train)
WORD.build(train, args.min_freq, (Embedding.load(args.embed, args.unk) if args.embed else None))
FEAT.build(train)
CHART.build(train)
args.update({
'n_words': WORD.vocab.n_init,
'n_feats': len(FEAT.vocab),
'n_labels': len(CHART.vocab),
'pad_index': WORD.pad_index,
'unk_index': WORD.unk_index,
'bos_index': WORD.bos_index,
'eos_index': WORD.eos_index,
'feat_pad_index': FEAT.pad_index
})
logger.info(f"{transform}")
logger.info("Building the model")
model = cls.MODEL(**args).load_pretrained(WORD.embed).to(args.device)
logger.info(f"{model}\n")
optimizer = Adam(model.parameters(), **optimizer_args)
scheduler = ExponentialLR(optimizer, **scheduler_args)
return cls(args, model, transform, optimizer, scheduler)
def build(cls, path, min_freq=2, fix_len=20, **kwargs):
r"""
Build a brand-new Parser, including initialization of all data fields and model parameters.
Args:
path (str):
The path of the model to be saved.
min_freq (str):
The minimum frequency needed to include a token in the vocabulary. Default: 2.
fix_len (int):
The max length of all subword pieces. The excess part of each piece will be truncated.
Required if using CharLSTM/BERT.
Default: 20.
kwargs (dict):
A dict holding the unconsumed arguments.
"""
args = Config(**locals())
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
os.makedirs(os.path.dirname(path) or './', exist_ok=True)
if os.path.exists(path) and not args.build:
parser = cls.load(**args)
parser.model = cls.MODEL(**parser.args)
parser.model.load_pretrained(parser.WORD.embed).to(args.device)
return parser
logger.info("Building the fields")
TAG, CHAR, BERT = None, None, None
if args.encoder != 'lstm':
from transformers import (AutoTokenizer, GPT2Tokenizer,
GPT2TokenizerFast)
t = AutoTokenizer.from_pretrained(args.bert)
WORD = SubwordField('words',
pad=t.pad_token,
unk=t.unk_token,
bos=t.bos_token or t.cls_token,
fix_len=args.fix_len,
tokenize=t.tokenize,
fn=None if not isinstance(t, (GPT2Tokenizer, GPT2TokenizerFast)) else lambda x: ' '+x)
WORD.vocab = t.get_vocab()
else:
WORD = Field('words', pad=pad, unk=unk, bos=bos, lower=True)
if 'tag' in args.feat:
TAG = Field('tags', bos=bos)
if 'char' in args.feat:
CHAR = SubwordField('chars', pad=pad, unk=unk, bos=bos, fix_len=args.fix_len)
if 'bert' in args.feat:
from transformers import (AutoTokenizer, GPT2Tokenizer,
GPT2TokenizerFast)
t = AutoTokenizer.from_pretrained(args.bert)
BERT = SubwordField('bert',
pad=t.pad_token,
unk=t.unk_token,
bos=t.bos_token or t.cls_token,
fix_len=args.fix_len,
tokenize=t.tokenize,
fn=None if not isinstance(t, (GPT2Tokenizer, GPT2TokenizerFast)) else lambda x: ' '+x)
BERT.vocab = t.get_vocab()
TEXT = RawField('texts')
ARC = Field('arcs', bos=bos, use_vocab=False, fn=CoNLL.get_arcs)
SIB = ChartField('sibs', bos=bos, use_vocab=False, fn=CoNLL.get_sibs)
REL = Field('rels', bos=bos)
transform = CoNLL(FORM=(WORD, TEXT, CHAR, BERT), CPOS=TAG, HEAD=(ARC, SIB), DEPREL=REL)
train = Dataset(transform, args.train)
if args.encoder == 'lstm':
WORD.build(train, args.min_freq, (Embedding.load(args.embed, args.unk) if args.embed else None))
if TAG is not None:
TAG.build(train)
if CHAR is not None:
CHAR.build(train)
REL.build(train)
args.update({
'n_words': len(WORD.vocab) if args.encoder != 'lstm' else WORD.vocab.n_init,
'n_rels': len(REL.vocab),
'n_tags': len(TAG.vocab) if TAG is not None else None,
'n_chars': len(CHAR.vocab) if CHAR is not None else None,
'char_pad_index': CHAR.pad_index if CHAR is not None else None,
'bert_pad_index': BERT.pad_index if BERT is not None else None,
'pad_index': WORD.pad_index,
'unk_index': WORD.unk_index,
'bos_index': WORD.bos_index
})
logger.info(f"{transform}")
logger.info("Building the model")
model = cls.MODEL(**args).load_pretrained(WORD.embed if hasattr(WORD, 'embed') else None).to(args.device)
logger.info(f"{model}\n")
return cls(args, model, transform)
def build(cls, path, min_freq=2, fix_len=20, **kwargs):
r"""
Build a brand-new Parser, including initialization of all data fields and model parameters.
Args:
path (str):
The path of the model to be saved.
min_freq (str):
The minimum frequency needed to include a token in the vocabulary. Default: 2.
fix_len (int):
The max length of all subword pieces. The excess part of each piece will be truncated.
Required if using CharLSTM/BERT.
Default: 20.
kwargs (dict):
A dict holding the unconsumed arguments.
"""
args = Config(**locals())
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
os.makedirs(os.path.dirname(path), exist_ok=True)
if os.path.exists(path) and not args.build:
# 加载已有模型
parser = cls.load(**args)
parser.model = cls.MODEL(**parser.args)
parser.model.load_pretrained(parser.WORD.embed).to(args.device)
return parser
logger.info("Building the fields")
WORD = Field('words', pad=pad, unk=unk, lower=True)
if args.feat == 'char':
FEAT = SubwordField('chars',
pad=pad,
unk=unk,
fix_len=args.fix_len)
# 怎么用bert，学习一下
elif args.feat == 'bert':
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(args.bert)
FEAT = SubwordField('bert',
pad=tokenizer.pad_token,
unk=tokenizer.unk_token,
fix_len=args.fix_len,
tokenize=tokenizer.tokenize)
FEAT.vocab = tokenizer.get_vocab()
else:
FEAT = Field('tags')
EDGE = ChartField('edges', use_vocab=False, fn=CoNLL.get_edges)
LABEL = ChartField('labels', fn=CoNLL.get_labels)
# 对于图的边和标签的抽取，定义了新的field
if args.feat in ('char', 'bert'):
transform = CoNLL(FORM=(WORD, FEAT), PHEAD=(EDGE, LABEL))
else:
transform = CoNLL(FORM=WORD, POS=FEAT, PHEAD=(EDGE, LABEL))
train = Dataset(transform, args.train)
WORD.build(
train, args.min_freq,
(Embedding.load(args.embed, args.unk) if args.embed else None))
FEAT.build(train)
LABEL.build(train)
args.update({
'n_words': WORD.vocab.n_init,
'n_feats': len(FEAT.vocab),
'n_labels': len(LABEL.vocab),
'pad_index': WORD.pad_index,
'unk_index': WORD.unk_index,
'feat_pad_index': FEAT.pad_index
})
model = cls.MODEL(**args)
model.load_pretrained(WORD.embed).to(args.device)
return cls(args, model, transform)
def build(cls,
path,
optimizer_args={
'lr': 1e-3,
'betas': (.0, .95),
'eps': 1e-12,
'weight_decay': 3e-9
},
scheduler_args={'gamma': .75**(1 / 5000)},
min_freq=7,
fix_len=20,
**kwargs):
r"""
Build a brand-new Parser, including initialization of all data fields and model parameters.
Args:
path (str):
The path of the model to be saved.
optimizer_args (dict):
Arguments for creating an optimizer.
scheduler_args (dict):
Arguments for creating a scheduler.
min_freq (str):
The minimum frequency needed to include a token in the vocabulary. Default:7.
fix_len (int):
The max length of all subword pieces. The excess part of each piece will be truncated.
Required if using CharLSTM/BERT.
Default: 20.
kwargs (dict):
A dict holding the unconsumed arguments.
"""
args = Config(**locals())
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
os.makedirs(os.path.dirname(path), exist_ok=True)
if os.path.exists(path) and not args.build:
parser = cls.load(**args)
parser.model = cls.MODEL(**parser.args)
parser.model.load_pretrained(parser.WORD.embed).to(args.device)
return parser
logger.info("Building the fields")
WORD = Field('words', pad=pad, unk=unk, bos=bos, lower=True)
TAG, CHAR, LEMMA, BERT = None, None, None, None
if 'tag' in args.feat:
TAG = Field('tags', bos=bos)
if 'char' in args.feat:
CHAR = SubwordField('chars',
pad=pad,
unk=unk,
bos=bos,
fix_len=args.fix_len)
if 'lemma' in args.feat:
LEMMA = Field('lemmas', pad=pad, unk=unk, bos=bos, lower=True)
if 'bert' in args.feat:
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(args.bert)
BERT = SubwordField('bert',
pad=tokenizer.pad_token,
unk=tokenizer.unk_token,
bos=tokenizer.bos_token or tokenizer.cls_token,
fix_len=args.fix_len,
tokenize=tokenizer.tokenize)
BERT.vocab = tokenizer.get_vocab()
EDGE = ChartField('edges', use_vocab=False, fn=CoNLL.get_edges)
LABEL = ChartField('labels', fn=CoNLL.get_labels)
transform = CoNLL(FORM=(WORD, CHAR, BERT),
LEMMA=LEMMA,
POS=TAG,
PHEAD=(EDGE, LABEL))
train = Dataset(transform, args.train)
WORD.build(
train, args.min_freq,
(Embedding.load(args.embed, args.unk) if args.embed else None))
if TAG is not None:
TAG.build(train)
if CHAR is not None:
CHAR.build(train)
if LEMMA is not None:
LEMMA.build(train)
LABEL.build(train)
args.update({
'n_words': WORD.vocab.n_init,
'n_labels': len(LABEL.vocab),
'n_tags': len(TAG.vocab) if TAG is not None else None,
'n_chars': len(CHAR.vocab) if CHAR is not None else None,
'char_pad_index': CHAR.pad_index if CHAR is not None else None,
'n_lemmas': len(LEMMA.vocab) if LEMMA is not None else None,
'bert_pad_index': BERT.pad_index if BERT is not None else None,
'pad_index': WORD.pad_index,
'unk_index': WORD.unk_index
})
logger.info(f"{transform}")
logger.info("Building the model")
model = cls.MODEL(**args).load_pretrained(WORD.embed).to(args.device)
logger.info(f"{model}\n")
optimizer = Adam(model.parameters(), **optimizer_args)
scheduler = ExponentialLR(optimizer, **scheduler_args)
return cls(args, model, transform, optimizer, scheduler)