def __init__(self,
data,
opt,
mode='train',
image_features=None,
fixed_answers_entry=None):
self.opt = opt
self.data = []
self.mode = mode
assert mode in ['train', 'dev', 'test']
error_samples = []
for datum in data:
if len(datum['annotated_question']['word']) == 0:
error_samples.append(datum['question_id'])
continue
if mode != 'test' and len(datum['orign_answers']) == 0:
error_samples.append(datum['question_id'])
continue
self.data.append(datum)
# global self.opt
log.info('Remove {} samples for empty question or answers: {}'.format(
len(error_samples), error_samples))
self.opt = opt
self.set_dataset()
if 'DEBUG' in self.opt:
self.debug = True
else:
self.debug = False
self.debug_dataset()
self.img_features_cache = {}
self.image_features = image_features
self.fixed_answers_entry = fixed_answers_entry
if 'ES_ocr' in self.opt:
self.ocr_name_list = [self.opt['ES_ocr']] + self.ocr_name_list
self.es_ocr_len = int(self.opt['ES_ocr_len'])
self.es_sort_way = self.opt['ES_sort_way']
log.info('Using OCR from: {}'.format(self.ocr_name_list))
log.info('Using OD from: {}'.format(self.od_name_list))
if 'BERT' in self.opt:
if 'BERT_LARGE' in self.opt:
log.debug('Using BERT Large model')
tokenizer_file = os.path.join(
self.opt['datadir'], self.opt['BERT_large_tokenizer_file'])
log.debug('Loading tokenizer from {}'.format(tokenizer_file))
self.bert_tokenizer = BertTokenizer.from_pretrained(
tokenizer_file)
else:
log.debug('Using BERT base model')
tokenizer_file = os.path.join(self.opt['datadir'],
self.opt['BERT_tokenizer_file'])
log.debug('Loading tokenizer from {}'.format(tokenizer_file))
self.bert_tokenizer = BertTokenizer.from_pretrained(
tokenizer_file)
def __init__(self, opt, data, use_cuda, vocab, char_vocab, evaluation=False):
# file_name = os.path.join(self.spacyDir, 'coqa-' + dataset_label + '-preprocessed.json')
self.data = data
self.use_cuda = use_cuda
self.vocab = vocab
self.char_vocab = char_vocab
self.evaluation = evaluation
self.opt = opt
if 'PREV_ANS' in self.opt:
self.prev_ans = self.opt['PREV_ANS']
else:
self.prev_ans = 2
if 'PREV_QUES' in self.opt:
self.prev_ques = self.opt['PREV_QUES']
else:
self.prev_ques = 0
self.use_char_cnn = 'CHAR_CNN' in self.opt
self.bert_tokenizer = None
if 'BERT' in self.opt:
if 'BERT_LARGE' in opt:
print('Using BERT Large model')
tokenizer_file = os.path.join(opt['datadir'], opt['BERT_large_tokenizer_file'])
print('Loading tokenizer from', tokenizer_file)
self.bert_tokenizer = BertTokenizer.from_pretrained(tokenizer_file)
else:
print('Using BERT base model')
tokenizer_file = os.path.join(opt['datadir'], opt['BERT_tokenizer_file'])
print('Loading tokenizer from', tokenizer_file)
self.bert_tokenizer = BertTokenizer.from_pretrained(tokenizer_file)
self.answer_span_in_context = 'ANSWER_SPAN_IN_CONTEXT_FEATURE' in self.opt
self.ques_max_len = (30 + 1) * self.prev_ans + (25 + 1) * (self.prev_ques + 1)
self.char_max_len = 30
print('*****************')
print('prev_ques :', self.prev_ques)
print('prev_ans :', self.prev_ans)
print('ques_max_len:', self.ques_max_len)
print('*****************')
c2id = {c: i for i, c in enumerate(char_vocab)}
# random shuffle for training
if not evaluation:
indices = list(range(len(self.data)))
random.shuffle(indices)
self.data = [self.data[i] for i in indices]
def __init__(self, opt, data, use_cuda, vocab, char_vocab, train_img_id2idx, train_img_features, train_img_spatials, val_img_id2idx, val_img_features, val_img_spatials, mod='train'):
# file_name = os.path.join(self.spacyDir, 'coqa-' + dataset_label + '-preprocessed.json')
self.data = data
self.use_cuda = use_cuda
self.vocab = vocab
self.char_vocab = char_vocab
self.train_img_features = train_img_features
self.train_img_id2idx = train_img_id2idx
self.train_img_spatials = train_img_spatials
self.val_img_features = val_img_features
self.val_img_spatials = val_img_spatials
self.val_img_id2idx = val_img_id2idx
self.img_feature_dim = opt['img_fea_dim']
self.img_spatial_dim = opt['img_spa_dim']
self.img_fea_num = opt['img_fea_num']
self.use_img_feature = 'img_feature' in opt
if 'ModelParallel' in opt:
self.bert_cuda = 'cuda:{}'.format(opt['ModelParallel'][-1])
self.main_cuda = 'cuda:{}'.format(opt['ModelParallel'][0])
self.ocr_name_list = opt['ocr_name_list'].split(',')
if 'ES_ocr' in opt:
self.ocr_name_list = [opt['ES_ocr']] + self.ocr_name_list
self.es_ocr_len = int(opt['ES_ocr_len'])
self.es_sort_way = opt['ES_sort_way']
else:
self.es_ocr_len = None
error_ocr_name = []
for ocr_name in self.ocr_name_list:
if ocr_name not in self.data[0]:
error_ocr_name.append(ocr_name)
if len(error_ocr_name) != 0:
log.error('OCR name ERROR: ' + str(error_ocr_name))
assert False
else:
log.info('Using OCR from: ' + str(self.ocr_name_list))
self.mod = mod
# if mod == 'train':
# self.evaluation = False
# else:
# self.evaluation = True
self.opt = opt
if 'PREV_ANS' in self.opt:
self.prev_ans = self.opt['PREV_ANS']
else:
self.prev_ans = 2
if 'PREV_QUES' in self.opt:
self.prev_ques = self.opt['PREV_QUES']
else:
self.prev_ques = 0
self.use_char_cnn = 'CHAR_CNN' in self.opt
self.bert_tokenizer = None
if 'BERT' in self.opt:
if 'BERT_LARGE' in opt:
print('Using BERT Large model')
tokenizer_file = os.path.join(opt['datadir'], opt['BERT_large_tokenizer_file'])
print('Loading tokenizer from', tokenizer_file)
self.bert_tokenizer = BertTokenizer.from_pretrained(tokenizer_file)
else:
# print('Using BERT base model')
tokenizer_file = os.path.join(opt['datadir'], opt['BERT_tokenizer_file'])
# print('Loading tokenizer from', tokenizer_file)
self.bert_tokenizer = BertTokenizer.from_pretrained(tokenizer_file)
self.answer_span_in_context = 'ANSWER_SPAN_IN_CONTEXT_FEATURE' in self.opt
self.ques_max_len = (30 + 1) * self.prev_ans + (25 + 1) * (self.prev_ques + 1)
self.char_max_len = 30
# print('*****************')
# print('prev_ques :', self.prev_ques)
# print('prev_ans :', self.prev_ans)
# print('ques_max_len:', self.ques_max_len)
# print('*****************')
c2id = {c: i for i, c in enumerate(char_vocab)}
self.od_name_list = opt['od_name_list'].split(',')
error_od_name = []
for od_name in self.od_name_list:
if od_name not in self.data[0]:
error_od_name.append(od_name)
if len(error_od_name) != 0:
log.error('OD name ERROR: ' + error_ocr_name)
assert False
else:
log.info('Using OD from: ' + str(self.od_name_list))