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:
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
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
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}")