-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathmydatasets.py
187 lines (166 loc) · 7.32 KB
/
mydatasets.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
import random
from dataclasses import dataclass
import math
import datasets
from typing import Union, List, Tuple, Dict
from torch.nn.utils.rnn import pad_sequence
import torch
from torch.utils.data import Dataset
# from .arguments import DataArguments, RerankerTrainingArguments
from transformers import PreTrainedTokenizer, BatchEncoding
from transformers import DataCollatorWithPadding
import numpy as np
from tqdm import tqdm
import os
import random
from dataclasses import dataclass
import math
import datasets
from typing import Union, List, Tuple, Dict
from torch.nn.utils.rnn import pad_sequence
import torch
from torch.utils.data import Dataset
# from .arguments import DataArguments, RerankerTrainingArguments
from transformers import PreTrainedTokenizer, BatchEncoding
from transformers import DataCollatorWithPadding
import numpy as np
from tqdm import tqdm
import os
class HBERTPretrainedPointWiseDataset(Dataset):
def __init__(
self,
args,
tokenizer: PreTrainedTokenizer,
dataset_cache_dir,
dataset_script_dir,
max_seq_len,
):
self.max_seq_len = max_seq_len
train_file =os.path.abspath(args.train_file)
if os.path.isdir(train_file):
filenames = os.listdir(train_file)
train_files = [os.path.join(train_file, fn) for fn in filenames]
else:
train_files = train_file
block_size_10MB = 10<<20
print("start loading datasets, train_files: ", train_files)
# print(dataset_script_dir)
self.nlp_dataset = datasets.load_dataset(
f'{dataset_script_dir}/json.py',
data_files=train_files,
ignore_verifications=False,
cache_dir=dataset_cache_dir,
features=datasets.Features({
"text_tokens_idx":[datasets.Value("int32")],
"node_tokens_idx":[datasets.Value("int32")],
"inputs_type_idx":[datasets.Value("int32")],
"text_labels":[datasets.Value("int32")],
"node_labels":[datasets.Value("int32")],
"text_layer_index":[datasets.Value("int32")],
"node_layer_index":[datasets.Value("int32")],
"text_num":[datasets.Value("int32")],
"node_num":[datasets.Value("int32")],
"waiting_mask":[datasets.Value("int32")],
"position":[datasets.Value("int32")],
}),
block_size = block_size_10MB
)['train']
self.tok = tokenizer
self.SEP = [self.tok.sep_token_id]
self.args = args
self.total_len = len(self.nlp_dataset)
print("loading dataset ok! len of dataset,", self.total_len)
def __len__(self):
return self.total_len
def __getitem__(self, item):
data = self.nlp_dataset[item]
max_seq_len = self.max_seq_len
text_num = data['text_num']
node_num = data['node_num']
position_len = max([text_num[i]+node_num[i] for i in range(len(text_num))])
data = {
"token_input_ids":torch.LongTensor(data['text_tokens_idx']),
"node_input_ids":torch.LongTensor(data['node_tokens_idx']),
"inputs_type_idx":torch.LongTensor(data['inputs_type_idx']),
"token_labels":torch.LongTensor(data['text_labels']),
"node_labels":torch.LongTensor(data['node_labels']),
"token_layer_index":torch.LongTensor(data['text_layer_index']),
"node_layer_index":torch.LongTensor(data['node_layer_index']),
"seq_num":list(data['text_num']),
"node_num":list(data['node_num']),
"waiting_mask":torch.LongTensor(data['waiting_mask']),
"position":list(data['position']),
"position_len":position_len,
}
return BatchEncoding(data)
@dataclass
class HBERTPointCollator(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.
"""
def __call__(
self, features
):
max_seq_len = 256
max_node_len = 10
# print(features)
batch_size = len(features)
mlm_labels = []
inputs_type_idx = []
layer_index = []
waiting_mask = []
position = []
layer_num = len(features[0]['seq_num'])
token_input_ids = []
node_input_ids = []
inputs_type_idx = []
token_labels= []
node_labels= []
token_layer_index= []
node_layer_index= []
seq_num= []
node_num= []
waiting_mask= []
for i in range(batch_size):
one_position = features[i]['position']
position_len = features[i]['position_len']
one_position = [torch.LongTensor(one_position[j:j+position_len]) for j in range(0, len(one_position),position_len )]
assert len(one_position) == layer_num
position.extend(one_position)
token_input_ids.append(features[i]['token_input_ids'])
node_input_ids.append(features[i]['node_input_ids'])
inputs_type_idx.append(features[i]['inputs_type_idx'])
token_labels.append(features[i]['token_labels'])
token_layer_index.append(features[i]['token_layer_index'])
node_layer_index.append(features[i]['node_layer_index'])
seq_num.append(features[i]['seq_num'])
node_num.append(features[i]['node_num'])
waiting_mask.append(features[i]['waiting_mask'])
node_labels.append(features[i]['node_labels'])
del features[i]['token_input_ids']
del features[i]['node_input_ids']
del features[i]['inputs_type_idx']
del features[i]['token_labels']
del features[i]['node_labels']
del features[i]['token_layer_index']
del features[i]['node_layer_index']
del features[i]['seq_num']
del features[i]['node_num']
del features[i]['waiting_mask']
del features[i]['position']
del features[i]['position_len']
features = {}
features["token_input_ids"] = pad_sequence(token_input_ids, batch_first=True).view(batch_size,-1,max_seq_len)
features["node_input_ids"] = pad_sequence(node_input_ids, batch_first=True).view(batch_size,-1,max_node_len)
features["inputs_type_idx"] = pad_sequence(inputs_type_idx, batch_first=True).view(batch_size,-1,max_seq_len)
features["token_labels"] = pad_sequence(token_labels, batch_first=True).view(batch_size,-1,max_seq_len)
features["node_labels"] = pad_sequence(node_labels, batch_first=True).view(batch_size,-1,max_node_len)
features["token_layer_index"] = pad_sequence(token_layer_index, batch_first=True)
features["node_layer_index"] = pad_sequence(node_layer_index, batch_first=True)
features["seq_num"] = torch.LongTensor(seq_num)
features["node_num"] = torch.LongTensor(node_num)
features["waiting_mask"] = pad_sequence(waiting_mask, batch_first=True).view(batch_size,-1,max_node_len)
features["position"] = pad_sequence(position, batch_first=True).view(batch_size,layer_num,-1)
return features