-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataloader_utils.py
104 lines (83 loc) · 4.19 KB
/
dataloader_utils.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
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import random
import pandas as pd
import tqdm
# import nltk
random.seed(2024)
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
ROBERTA_MAX_TEXT_LENGTH=512 # maximum text length: 512
GPT2_MAX_TEXT_LENGTH=1024
def load_base_model_tokenizer(model_name,cache_dir):
"""load the language models and their corresponding tokenizers"""
base_model_kwargs = {}
if 'gpt-j' in model_name or 'neox' in model_name:
base_model_kwargs.update(dict(torch_dtype=torch.float16))
if 'gpt-j' in model_name:
base_model_kwargs.update(dict(revision='float16'))
base_model = AutoModelForCausalLM.from_pretrained(
model_name,
**base_model_kwargs,
cache_dir=cache_dir
).to(DEVICE)
optional_tok_kwargs = {}
if "facebook/opt-" in model_name:
print("Using non-fast tokenizer for OPT")
optional_tok_kwargs['fast'] = False
base_tokenizer = AutoTokenizer.from_pretrained(model_name, **optional_tok_kwargs, cache_dir=cache_dir)
base_tokenizer.pad_token_id = base_tokenizer.eos_token_id
return base_model, base_tokenizer
def merge_sentences(original_sentences, generated_sentences, segment_num=2):
if len(generated_sentences)==0:
print("no generated sentences")
return original_sentences, [0 for i in range(len(original_sentences))], 0
if len(generated_sentences)<segment_num:
# print("generated_sentences length:", len(generated_sentences))
segment_num=len(generated_sentences) # in this case, segment_size=1, fix the bug for separate characters
segment_size = 1
else:
segment_size = len(generated_sentences)//segment_num
divided_generated_sentences = []
for i in range(segment_num):
if i==segment_num-1:
divided_generated_sentences.append(generated_sentences[i*segment_size:])
else:
divided_generated_sentences.append(generated_sentences[i*segment_size:(i+1)*segment_size])
total_len = len(original_sentences)+len(generated_sentences)
try:
random_positions = random.sample(range(1,len(original_sentences)), segment_num)
except:
print("length of original sentences:",len(original_sentences))
print("number of segment:", segment_num)
if len(original_sentences)==1:
segment_num=1
random_positions=[1]
else:
segment_num = len(original_sentences)-1
random_positions=random.sample(range(1,len(original_sentences)), segment_num)
return_sentences = []
return_labels = []
segment_id = 0
for sentence_id in range(len(original_sentences)):
if sentence_id not in random_positions:
return_sentences.append(original_sentences[sentence_id])
return_labels.append(0) # 0 represents human-written text
else:
return_sentences+=divided_generated_sentences[segment_id]
return_labels+=[1]*len(divided_generated_sentences[segment_id])
segment_id+=1
return return_sentences, return_labels, segment_num
def get_roberta_feature(sample, roberta_detector, roberta_tokenizer):
sample_article_id = sample["article_id"]
sample_original_article = "\n\n".join(sample["original_paragraphs"])
sample_manipulated_article = "\n\n".join(sample["merge_paragraphs"])
sample_config_dict = sample["config_dict"]
sample_manipulated_article_token = roberta_tokenizer(sample_manipulated_article,
padding='max_length', # longest, max_length, False
truncation=True,
max_length=512,
return_tensors="pt").to(DEVICE) # (1, text_length), text_length should be smaller than 512
sample_manipulated_article_embeddings = roberta_detector(**sample_manipulated_article_token,
output_hidden_states=True, return_dict=True)
last_hidden_state = sample_manipulated_article_embeddings['hidden_states'][-1] # (1,512,1024)
return last_hidden_state