-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
220 lines (175 loc) · 8.76 KB
/
dataset.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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
# adapted from https://github.com/PeterGriffinJin/Patton/blob/main/src/OpenLP/dataset/train_dataset.py
from dataclasses import dataclass
from transformers import DataCollatorWithPadding, DefaultDataCollator, PreTrainedTokenizer, DataCollatorForLanguageModeling
from datasets import load_dataset
from torch.utils.data import Dataset
from arguments import DataArguments
from transformers import Trainer
import os
import glob
import random
import pandas as pd
from typing import List, Tuple, Dict
import torch
class TrainDataset(Dataset):
def __init__(self, tokenizer: PreTrainedTokenizer, data_args, shuffle_seed: int = None, cache_dir: str = None, k=16) -> None:
super(TrainDataset, self).__init__()
self.data_files = [data_args.train_path] if data_args.train_dir is None else glob.glob(os.path.join(data_args.train_dir, "*.jsonl"))
self.dataset = load_dataset("json", data_files=self.data_files, streaming=False, cache_dir=cache_dir)["train"]
self.dataset = self.dataset.shuffle(seed=shuffle_seed) if shuffle_seed is not None else self.dataset
self.tokens = pd.read_csv(data_args.corpus_path, sep='\t', names=['asin', 'title'])['title'].values if data_args.corpus_path is not None else None
self.tokenizer = tokenizer
self.data_args = data_args
self.max_len = data_args.max_len
self.k = k #negative sampling
def create_one_example(self, text_encoding: List[int]):
item = self.tokenizer.encode_plus(
text_encoding,
truncation='only_first',
max_length=self.data_args.max_len,
padding=False,
return_attention_mask=False,
return_token_type_ids=False,
)
return item
def __len__(self): # __len__ is required by huggingface trainer
count = len(self.dataset)
return count
def process_multi(self, example):
encoded_query = self.create_one_example(example['q_text'])
encoded_key = self.create_one_example(example['k_text'])
encoded_query_n = [self.create_one_example(q_n) if q_n != [] else self.create_one_example([0]) for q_n in example['q_n_text']]
encoded_key_n = [self.create_one_example(k_n) if k_n != [] else self.create_one_example([0]) for k_n in example['k_n_text']]
query_n_mask = [1 if q_n != [] else 0 for q_n in example['q_n_text']]
key_n_mask = [1 if k_n != [] else 0 for k_n in example['k_n_text']]
return {"query": encoded_query, "key": encoded_key, 'query_n':encoded_query_n, 'key_n':encoded_key_n, 'query_n_mask':query_n_mask, 'key_n_mask':key_n_mask}
def process_fn(self, example):
encoded_query = self.create_one_example(example['q_text'])
encoded_key = self.create_one_example(example['k_text'])
return {"query": encoded_query, "key": encoded_key}
def __getitem__(self, index):
return self.process_fn(self.dataset[index])
class DPTrainDataset(TrainDataset):
def process_fn(self, example):
encoded_query = self.create_one_example(example['q_text'])
encoded_key = self.create_one_example(example['k_text'])
neg_sample = [self.create_one_example(neg) for neg in random.choices(self.tokens, k=self.k)]
return {"query": encoded_query, "key": encoded_key, "key_neg": neg_sample}
class EvalDataset(TrainDataset):
def __init__(self, tokenizer: PreTrainedTokenizer, data_args, shuffle_seed: int = None, cache_dir: str = None) -> None:
super(EvalDataset, self).__init__(tokenizer, data_args, None, cache_dir=cache_dir)
self.data_files = [data_args.eval_path]
self.dataset = load_dataset("json", data_files=self.data_files, streaming=False, cache_dir=cache_dir)["train"]
self.dataset = self.dataset.shuffle(seed=shuffle_seed) if shuffle_seed is not None else self.dataset
self.tokenizer = tokenizer
self.data_args = data_args
self.max_len = data_args.max_len
def process_fn(self, example):
encoded_query = self.create_one_example(example['q_text'])
encoded_key = self.create_one_example(example['k_text'])
return {"query": encoded_query, "key": encoded_key}
def __getitem__(self, index):
return self.process_fn(self.dataset[index])
class TrainNCCDataset(Dataset):
def __init__(self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, trainer: Trainer = None, shuffle_seed: int = None, cache_dir: str = None) -> None:
super(TrainNCCDataset, self).__init__()
self.data_files = [data_args.train_path] if data_args.train_dir is None else glob.glob(os.path.join(data_args.train_dir, "*.jsonl"))
self.dataset = load_dataset("json", data_files=self.data_files, streaming=False, cache_dir=cache_dir)["train"]
self.dataset = self.dataset.shuffle(seed=shuffle_seed) if shuffle_seed is not None else self.dataset
self.tokenizer = tokenizer
self.data_args = data_args
self.max_len = data_args.max_len
self.trainer = trainer
def create_one_example(self, text_encoding: List[int]):
item = self.tokenizer.encode_plus(
text_encoding,
truncation='only_first',
max_length=self.data_args.max_len,
padding=False,
return_attention_mask=False,
return_token_type_ids=False,
)
return item
def __len__(self): # __len__ is required by huggingface trainer
count = len(self.dataset)
return count
def process_fn(self, example):
encoded_query = self.create_one_example(example['q_text'])
return {"query": encoded_query, 'label':example['label']}
def __getitem__(self, index):
return self.process_fn(self.dataset[index])
class EvalNCCDataset(TrainNCCDataset):
def __init__(self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, shuffle_seed: int = None, cache_dir: str = None, eval=True) -> None:
super(EvalNCCDataset, self).__init__(tokenizer, data_args, None, cache_dir=cache_dir)
self.data_files = [data_args.eval_path] if eval else [data_args.test_path]
self.dataset = load_dataset("json", data_files=self.data_files, streaming=False, cache_dir=cache_dir)["train"]
self.dataset = self.dataset.shuffle(seed=shuffle_seed) if shuffle_seed is not None else self.dataset
self.tokenizer = tokenizer
self.data_args = data_args
self.max_len = data_args.max_len
def __getitem__(self, index):
return self.process_fn(self.dataset[index])
@dataclass
class TrainCollator(DataCollatorWithPadding):
"""
Wrapper that does conversion from List[Tuple[encode_qry, encode_psg]] to List[qry], List[psg]
and pass batch separately to the actual collator.
Abstract out data detail for the model.
"""
max_len: int = 32
def __call__(self, features):
qq = [f["query"] for f in features]
kk = [f["key"] for f in features]
if isinstance(qq[0], list):
qq = sum(qq, [])
if isinstance(kk[0], list):
kk = sum(kk, [])
q_collated = self.tokenizer.pad(
qq,
padding='max_length',
max_length=self.max_len,
return_tensors="pt",
)
k_collated = self.tokenizer.pad(
kk,
padding='max_length',
max_length=self.max_len,
return_tensors="pt",
)
if "key_neg" in features[0]:
vv = [f["key_neg"] for f in features]
if isinstance(vv[0], list):
# vv = sum(vv, [])
# [[1,2,3], [a,b,c], [x,y,z]] -> [1,a,x,2,b,y,3,c,z]
# easy to compute the gradient norms for each tuple
vv = [item for sublist in vv for item in sublist]
v_collated = self.tokenizer.pad(
vv,
padding='max_length',
max_length=self.max_len,
return_tensors="pt",
)
return q_collated, k_collated, v_collated
else:
return q_collated, k_collated
@dataclass
class TrainNCCCollator(DataCollatorWithPadding):
"""
Wrapper that does conversion from List[Tuple[encode_qry, encode_psg]] to List[qry], List[psg]
and pass batch separately to the actual collator.
Abstract out data detail for the model.
"""
max_len: int = 32
def __call__(self, features):
qq = [f["query"] for f in features]
labels = [f["label"] for f in features]
if isinstance(qq[0], list):
qq = sum(qq, [])
q_collated = self.tokenizer.pad(
qq,
padding='max_length',
max_length=self.max_len,
return_tensors="pt",
)
labels = torch.LongTensor(labels)
return {'center_input': q_collated}, labels