-
Notifications
You must be signed in to change notification settings - Fork 0
/
make_tokenized_data.py
181 lines (144 loc) · 6.36 KB
/
make_tokenized_data.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
from transformers import BertTokenizer,DebertaV2Tokenizer
from tqdm import tqdm
from dataset import ProcessedIdDataset
import argparse
import torch
import os
def tokenized_and_process(data, args,mode='train'):
if 'deberta' in args.model_type:
_tokenizer=DebertaV2Tokenizer.from_pretrained(args.model_type)
else:
_tokenizer=BertTokenizer.from_pretrained(args.model_type)
text_list=[]
text_id_list=[]
aspect_question_list=[]
aspect_question_id_list=[]
opinion_answer_list=[]
opinion_question_list=[]
opinion_question_id_list=[]
aspect_answer_list=[]
ignore_indexes=[]
sentiment_list=[]
header_fmt='Tokenize data {:>5s}'
for sample in tqdm(data,desc=f'{header_fmt.format(mode.upper())}'):
##Temp data
###Text
temp_text=sample.text_tokens
text_list.append(temp_text)
##Aspect question
aspect_question=sample.aspect_queries
aspect_question_list.append(aspect_question)
##Opinion question
opinion_question=sample.opinion_queries
opinion_question_list.append(opinion_question)
ignore_index=[]
###Convert tokens to ids
if 'deberta' not in args.model_type:
##The text
text_ids=_tokenizer.convert_tokens_to_ids(
[word.lower() for word in temp_text]
)
text_id_list.append(text_ids)
##Aspect question
aspect_queries_ids=_tokenizer.convert_tokens_to_ids(
[word.lower() for word in aspect_question]
)
aspect_question_id_list.append(aspect_queries_ids)
##Opinion question
opinion_queries_ids=_tokenizer.convert_tokens_to_ids(
[word.lower() for word in opinion_question]
)
opinion_question_id_list.append(opinion_queries_ids)
assert len(text_ids)==len(sample.aspect_answers)==len(sample.opinion_answers)
#Apsect answer
aspect_answer_list.append(sample.aspect_answers)
#Opinion answer
opinion_answer_list.append(sample.opinion_answers)
#Sentiment
sentiment_list.append(sample.sentiments)
else:
aspect_answer=[]
opinion_answer=[]
sentiment=[]
ignore_index=[]
##The text
text_ids=_tokenizer.encode(' '.join(temp_text).lower(),add_special_tokens=False)
text_id_list.append(text_ids)
##Aspect question
aspect_queries_ids=_tokenizer.encode(' '.join(aspect_question).lower(),add_special_tokens=False)
aspect_question_id_list.append(aspect_queries_ids)
##Opinion question
opinion_queries_ids=_tokenizer.encode(' '.join(opinion_question).lower(),add_special_tokens=False)
opinion_question_id_list.append(opinion_queries_ids)
##Điều chỉnh lại nhãn cho phù hợp với encode của deberta
temp_text_ids=[]
for ind,tok in enumerate(temp_text):
ids=_tokenizer.encode(tok.lower(),add_special_tokens=False)
temp_text_ids+=ids
aspect_answer.append(sample.aspect_answers[ind])
opinion_answer.append(sample.opinion_answers[ind])
sentiment.append(sample.sentiments[ind])
ignore_index.append(0)
for _ in range(len(ids[1:])):
ignore_index.append(-1)
aspect_answer.append(-1)
opinion_answer.append(-1)
sentiment.append(-1)
assert temp_text_ids==text_ids ##Đảm bảo giữa phần encode từng từ và encode nguyên câu là giống nhau
assert len(ignore_index)==len(aspect_answer)==len(opinion_answer)==len(sentiment)==len(text_ids) ###Đảm bảo phần nhãn nằm đúng ở các vị trí
#Apsect answer
aspect_answer_list.append(aspect_answer)
#Opinion answer
opinion_answer_list.append(opinion_answer)
#Sentiment
sentiment_list.append(sentiment)
##Ignore_indexes
ignore_indexes.append(ignore_index)
result={
'texts':text_list,
'texts_ids':text_id_list,
'aspect_questions':aspect_question_list,
'aspect_questions_ids':aspect_question_id_list,
'opinion_answers':opinion_answer_list,
'opinion_questions':opinion_question_list,
'opinion_questions_ids':opinion_question_id_list,
'aspect_answers':aspect_answer_list,
'sentiments':sentiment_list,
'ignore_indexes':ignore_indexes
}
final_data=ProcessedIdDataset(result)
return final_data
if __name__=='__main__':
parser = argparse.ArgumentParser(description='Processing data')
##Define path where save unprocessed data and where to save processed data
parser.add_argument('--data_path', type=str, default="./data/14resV2/preprocess")
parser.add_argument('--output_path', type=str, default="./data/14resV2/preprocess")
parser.add_argument('--model_type',type=str,default='microsoft/deberta-v3-xsmall')
args=parser.parse_args()
train_data_path = f"{args.data_path}/train_PREPROCESSED.pt"
dev_data_path = f"{args.data_path}/dev_PREPROCESSED.pt"
test_data_path = f"{args.data_path}/test_PREPROCESSED.pt"
train_data=torch.load(train_data_path)
dev_data=torch.load(dev_data_path)
test_data=torch.load(test_data_path)
'''##Making tokenize data before preprocess to id
train_tokenized,train_max_len=tokenize_data(train_data,version=args_version,mode='train')
dev_tokenized,dev_max_len=tokenize_data(dev_data,version=args_version,mode='dev')
test_tokenized,test_max_len=tokenize_data(test_data,version=args_version,mode='test')'''
'''print(f"train_max_len : {train_max_len}")
print(f"dev_max_len : {dev_max_len}")
print(f"test_max_len : {test_max_len}")'''
##Processing tokenied data to ids
train_preprocess = tokenized_and_process(train_data,args,mode='train')
dev_preprocess = tokenized_and_process(dev_data, args,mode='dev')
test_preprocess = tokenized_and_process(test_data, args,mode='test')
##Saving preprocessing full data
if not os.path.exists(args.output_path):
os.makedirs(args.output_path)
output_path=f'{args.output_path}/data_deberta_v3_xsmall.pt'
print(f"Saved data : `{output_path}`.")
torch.save({
'train':train_preprocess,
'dev':dev_preprocess,
'test':test_preprocess
},output_path)