def __init__(self,
trn=False,
tst=False,
input_data=False,
def_data=False,
def_y=False,
classes_per_it=60,
num_support=5,
iterations=100,
target_frames=False):
with open(
dir_path +
'/../koreanframenet/resource/info/fn1.7_frame2idx.json',
'r') as f:
self.frame2idx = json.load(f)
with open(
dir_path +
'/../koreanframenet/resource/info/fn1.7_frame_definitions.json',
'r') as f:
self.frame2definition = json.load(f)
if target_frames:
self.target_frames = target_frames
else:
with open(dir_path + '/../data/target_frames.json', 'r') as f:
self.target_frames = json.load(f)
self.idx2frame = dict(
zip(self.frame2idx.values(), self.frame2idx.keys()))
# self.trn = trn
# self.tst = tst
# self.trn_y = self.get_y(self.trn)
# self.tst_y = self.get_y(self.tst)
# self.trn_data = trn_data
# self.tst_data = tst_data
self.bert_io = utils.for_BERT(mode='train', language='multi')
if def_data:
self.def_data = def_data
self.def_y = def_y
else:
self.def_data, self.def_y = self.bert_io.convert_to_bert_input_label_definition(
self.frame2definition, self.frame2idx)
self.classes_per_it = classes_per_it
self.num_support = num_support
self.iterations = iterations
(t_hour, t_min) = divmod(t_min, 60)
result = '{}hour:{}min:{}sec'.format(t_hour, t_min, t_sec)
return result
# In[3]:
try:
dir_path = os.path.dirname(os.path.abspath(__file__))
except:
dir_path = '.'
# In[4]:
bert_io = utils.for_BERT(mode='train', language='multi')
# In[5]:
frameBERT_dir = '/disk/data/models/frameBERT/frameBERT_en'
frameBERT = FrameBERT.from_pretrained(frameBERT_dir,
num_senses=len(bert_io.sense2idx),
num_args=len(bert_io.bio_arg2idx),
lufrmap=bert_io.lufrmap,
frargmap=bert_io.bio_frargmap)
frameBERT.to(device)
frameBERT.eval()
# In[6]:
epochs = 50
print('\nFineTuning Multilingual')
print('### TRAINING')
print('MODEL:', srl)
print('LANGUAGE:', language)
print('PRETRAINED BERT:', PRETRAINED_MODEL)
print('training data:')
print('\t(ko):', len(trn))
print('BATCH_SIZE:', batch_size)
print('MAX_LEN:', MAX_LEN)
print('')
bert_io = utils.for_BERT(mode='train',
srl=srl,
language=language,
masking=masking,
fnversion=fnversion,
pretrained=PRETRAINED_MODEL)
train(
retrain=True,
pretrained_dir='/disk/data/models/dict_framenet/enModel-with-exemplar/9/')
# In[ ]:
# by 25% (4460)
# model_dir = '/disk/data/models/framenet/mulModel-25/'
# epochs = 50
# trn, dev, tst = dataio.load_data(srl=srl, language='ko')
def __init__(self,
fnversion=1.1,
language='ko',
masking=True,
srl='framenet',
model_path=False,
gold_pred=False,
viterbi=False,
tgt=True,
pretrained='bert-base-multilingual-cased'):
self.fnversion = fnversion
self.language = language
self.masking = masking
self.srl = srl
self.gold_pred = gold_pred
self.viterbi = viterbi
self.pretrained = pretrained
self.tgt = tgt #using <tgt> and </tgt> as a special token
if self.masking == True:
self.targetid = target_identifier.targetIdentifier()
else:
self.targetid = target_identifier.targetIdentifier(only_lu=False)
if self.srl == 'propbank-dp':
self.viterbi = False
self.masking = False
print('srl model:', self.srl)
print('language:', self.language)
print('version:', self.fnversion)
print('using viterbi:', self.viterbi)
print('using masking:', self.masking)
print('pretrained BERT:', self.pretrained)
print('using TGT special token:', self.tgt)
self.bert_io = utils.for_BERT(mode='predict',
srl=self.srl,
language=self.language,
masking=self.masking,
fnversion=self.fnversion,
pretrained=self.pretrained)
#load model
if model_path:
self.model_path = model_path
else:
print('model_path={your_model_dir}')
# self.model = torch.load(model_path, map_location=device)
self.model = BertForJointShallowSemanticParsing.from_pretrained(
self.model_path,
num_senses=len(self.bert_io.sense2idx),
num_args=len(self.bert_io.bio_arg2idx),
lufrmap=self.bert_io.lufrmap,
masking=self.masking,
frargmap=self.bert_io.bio_frargmap)
self.model.to(device)
print('...loaded model path:', self.model_path)
# self.model = BertForJointShallowSemanticParsing
self.model.eval()
print(self.model_path)
print('...model is loaded')
# trainsition parameter for vitervi decoding
if self.srl != 'propbank-dp':
self.transition_param = inference.get_transition_params(
self.bert_io.idx2bio_arg.values())
