-
Notifications
You must be signed in to change notification settings - Fork 8
/
dataset.py
379 lines (315 loc) · 14.8 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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
from typing import Dict, List, Optional, Tuple, TypedDict, Union
import numpy as np
import torch
from torch.tensor import Tensor
from tqdm import tqdm
from transformers import PreTrainedTokenizer
from neuro_comma.augmentation import AUGMENTATIONS
from neuro_comma.pretrained import TOKEN_IDX
class BaseDataset(torch.utils.data.Dataset):
def __init__(self, files: Union[str, List[str]],
tokenizer: PreTrainedTokenizer,
targets: Dict[str, int],
sequence_len: int,
token_style: str) -> None:
if isinstance(files, list):
self.data = []
for file in files:
self.data += self._parse_data(file, tokenizer, targets, sequence_len, token_style)
else:
self.data = self._parse_data(files, tokenizer, targets, sequence_len, token_style)
@classmethod
def _parse_data(cls,
file_path: str,
tokenizer: PreTrainedTokenizer,
targets: Dict[str, int],
seq_len: int,
token_style: str):
"""Parse file to train data
Args:
file_path (`str`): text file path that contains tokens and punctuations separated by tab in lines
tokenizer (`PreTrainedTokenizer`): tokenizer that will be used to further tokenize word for BERT like models
targets (`dict[str, int]`): dictionary of target: label
seq_len (`int`): maximum length of each sequence
token_style (`str`): For getting index of special tokens in pretrained.TOKEN_IDX
Returns:
list[Batch]: each having sequence_len punctuation_mask is used to ignore special indices like padding and intermediate sub-word token during evaluation
"""
with open(file_path, 'r', encoding='utf-8') as file:
x, y = [], []
for i, line in enumerate(file):
if (line := line.strip()):
token = line.rsplit('\t', 1)
if len(token) == 2:
x.append(token[0])
target = targets[token[1]]
y.append(target)
else:
continue
data = cls.parse_tokens(x, tokenizer, seq_len, token_style, y)
# cls._add_targets_to_data(y, data)
return data
@classmethod
def parse_tokens(cls,
tokens: Union[List[str], Tuple[str]],
tokenizer: PreTrainedTokenizer,
seq_len: int,
token_style: str,
targets: Optional[List[int]] = None,
debug: bool = True):
"""
Convert tokenized data for model prediction
Args:
tokens (`Union[list[str], tuple[str]]`): splited tokens
tokenizer (`PreTrainedTokenizer`): tokenizer which split tokens to subtokens
seq_len (`int`): sequence length
token_style (`str`): token_style from pretrained.TOKEN_IDX
Returns:
(`list[BatchWithoutTarget]`): list of bathces
```txt
tokens : [token token ##token PAD ]
x : [321 1233 23121 101 ]
y : [tar 0 tar 0 ]
y_mask : [1 0 1 0 ]
attn_mask : [1 1 1 0 ]
```
"""
data_items = []
# loop until end of the entire text
idx = 0
if debug:
pbar = tqdm(total=len(tokens))
while idx < len(tokens):
x = [TOKEN_IDX[token_style]['START_SEQ']]
w_id = [-1] # word indexes
y = [0]
y_mask = [1] if targets else [0]
# loop until we have required sequence length
# -1 because we will have a special end of sequence token at the end
while len(x) < seq_len - 1 and idx < len(tokens):
word_pieces = tokenizer.tokenize(tokens[idx])
# if taking these tokens exceeds sequence length we finish
# current sequence with padding
# then start next sequence from this token
if len(word_pieces) + len(x) >= seq_len:
break
for i in range(len(word_pieces) - 1):
x.append(tokenizer.convert_tokens_to_ids(word_pieces[i]))
w_id.append(idx)
y.append(0)
y_mask.append(0)
if len(word_pieces) > 0:
x.append(tokenizer.convert_tokens_to_ids(word_pieces[-1]))
else:
x.append(TOKEN_IDX[token_style]['UNK'])
w_id.append(idx)
if targets:
y.append(targets[idx])
else:
y.append(0)
y_mask.append(1)
idx += 1
if debug:
pbar.update(1)
x.append(TOKEN_IDX[token_style]['END_SEQ'])
w_id.append(-1)
y.append(0)
if targets:
y_mask.append(1)
else:
y_mask.append(0)
# Fill with pad tokens
if len(x) < seq_len:
x = x + [TOKEN_IDX[token_style]['PAD'] for _ in range(seq_len - len(x))]
w_id = w_id + [-100 for _ in range(seq_len - len(w_id))]
y = y + [0 for _ in range(seq_len - len(y))]
y_mask = y_mask + [0 for _ in range(seq_len - len(y_mask))]
attn_mask = [1 if token != TOKEN_IDX[token_style]['PAD'] else 0 for token in x]
data_items.append([x, w_id, attn_mask, y, y_mask])
if debug:
pbar.close()
return data_items
@classmethod
def _add_targets_to_data(cls,
targets: List[int],
data: List[List[List[int]]]) -> None:
targets = targets.copy()
for idx, batch in tqdm(enumerate(data), total=len(data)):
y = []
y_mask = []
word_ids = batch[1]
for i, w_id in enumerate(word_ids):
if w_id == -100:
y.append(0)
y_mask.append(0)
elif w_id == -1:
y.append(0)
y_mask.append(1)
elif w_id != word_ids[i + 1]:
y.append(targets.pop(0))
y_mask.append(1)
else:
y.append(0)
y_mask.append(0)
data[idx].extend([y, y_mask])
def __len__(self) -> int:
return len(self.data)
def __getitem__(self, index: int) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
x = self.data[index][0]
attn_mask = self.data[index][2]
y = self.data[index][3]
y_mask = self.data[index][4]
x = torch.tensor(x)
attn_mask = torch.tensor(attn_mask)
y = torch.tensor(y)
y_mask = torch.tensor(y_mask)
return x, y, attn_mask, y_mask
class RepunctDataset(BaseDataset):
def __init__(self,
files: Union[str, List[str]],
tokenizer: PreTrainedTokenizer,
targets: Dict[str, int],
sequence_len: int,
token_style: str,
is_train=False,
augment_rate=0.,
augment_type='substitute') -> None:
"""Preprocess data for restore punctuation
Args:
files (`Union[str, list[str]]`): single file or list of text files containing tokens and punctuations separated by tab in lines
tokenizer (`PreTrainedTokenizer`): tokenizer that will be used to further tokenize word for BERT like models
targets (`dict[str, int]`): dict with targets
sequence_len (`int`): length of each sequence
token_style (`str`): For getting index of special tokens in pretrained.TOKEN_IDX
is_train (`bool, optional`): if false do not apply augmentation. Defaults to False.
augment_rate (`float, optional`): percent of data which should be augmented. Defaults to 0.0.
augment_type (`str, optional`): augmentation type. Defaults to 'substitute'.
"""
super().__init__(files, tokenizer, targets, sequence_len, token_style)
self.sequence_len = sequence_len
self.token_style = token_style
self.is_train = is_train
self.augment_type = augment_type
self.augment_rate = augment_rate
# TODO: remove method
@classmethod
def _parse_data_old(cls,
file_path: str,
tokenizer: PreTrainedTokenizer,
targets: Dict[str, int],
seq_len: int,
token_style: str):
"""Parse file to train data
Args:
file_path (`str`): text file path that contains tokens and punctuations separated by tab in lines
tokenizer (`PreTrainedTokenizer`): tokenizer that will be used to further tokenize word for BERT like models
targets (`dict[str, int]`): dictionary of target: label
seq_len (`int`): maximum length of each sequence
token_style (`str`): For getting index of special tokens in pretrained.TOKEN_IDX
Returns:
list[Batch]: each having sequence_len punctuation_mask is used to ignore special indices like padding and intermediate sub-word token during evaluation
"""
data_items = []
with open(file_path, 'r', encoding='utf-8') as f:
lines = [line for line in f.read().split('\n') if line.strip()]
idx = 0
# loop until end of the entire text
while idx < len(lines):
x = [TOKEN_IDX[token_style]['START_SEQ']]
y = [0]
y_mask = [1] # which positions we need to consider while evaluating i.e., ignore pad or sub tokens
# loop until we have required sequence length
# -1 because we will have a special end of sequence token at the end
while len(x) < seq_len - 1 and idx < len(lines):
word, target = lines[idx].rsplit('\t', 1)
tokens = tokenizer.tokenize(word)
# if taking these tokens exceeds sequence length we finish current sequence with padding
# then start next sequence from this token
if len(tokens) + len(x) >= seq_len:
break
else:
for i in range(len(tokens) - 1):
x.append(tokenizer.convert_tokens_to_ids(tokens[i]))
y.append(0)
y_mask.append(0)
if len(tokens) > 0:
x.append(tokenizer.convert_tokens_to_ids(tokens[-1]))
else:
x.append(TOKEN_IDX[token_style]['UNK'])
y.append(targets[target])
y_mask.append(1)
idx += 1
x.append(TOKEN_IDX[token_style]['END_SEQ'])
y.append(0)
y_mask.append(1)
# Fill with pad tokens
if len(x) < seq_len:
x = x + [TOKEN_IDX[token_style]['PAD'] for _ in range(seq_len - len(x))]
y = y + [0 for _ in range(seq_len - len(y))]
y_mask = y_mask + [0 for _ in range(seq_len - len(y_mask))]
attn_mask = [1 if token != TOKEN_IDX[token_style]['PAD'] else 0 for token in x]
data_items.append([x, y, attn_mask, y_mask])
return data_items
def _augment(self, x, y, y_mask):
x_aug = []
y_aug = []
y_mask_aug = []
for i in range(len(x)):
r = np.random.rand()
if r < self.augment_rate:
AUGMENTATIONS[self.augment_type](x, y, y_mask, x_aug, y_aug, y_mask_aug, i, self.token_style)
else:
x_aug.append(x[i])
y_aug.append(y[i])
y_mask_aug.append(y_mask[i])
if len(x_aug) > self.sequence_len:
# len increased due to insert
x_aug = x_aug[:self.sequence_len]
y_aug = y_aug[:self.sequence_len]
y_mask_aug = y_mask_aug[:self.sequence_len]
elif len(x_aug) < self.sequence_len:
# len decreased due to delete
x_aug = x_aug + [TOKEN_IDX[self.token_style]['PAD'] for _ in range(self.sequence_len - len(x_aug))]
y_aug = y_aug + [0 for _ in range(self.sequence_len - len(y_aug))]
y_mask_aug = y_mask_aug + [0 for _ in range(self.sequence_len - len(y_mask_aug))]
attn_mask = [1 if token != TOKEN_IDX[self.token_style]['PAD'] else 0 for token in x]
return x_aug, y_aug, attn_mask, y_mask_aug
def __getitem__(self, index: int) -> Tuple[Tensor, Tensor, Tensor, Tensor]:
x = self.data[index][0]
attn_mask = self.data[index][2]
y = self.data[index][3]
y_mask = self.data[index][4]
if self.is_train and self.augment_rate > 0:
x, y, attn_mask, y_mask = self._augment(x, y, y_mask)
x = torch.tensor(x)
attn_mask = torch.tensor(attn_mask)
y = torch.tensor(y)
y_mask = torch.tensor(y_mask)
return x, y, attn_mask, y_mask
if __name__ == "__main__":
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('DeepPavlov/rubert-base-cased-sentence')
file_path = '/media/sviperm/9740514d-d8c8-4f3e-afee-16ce6923340c2/sviperm/Documents/voicetextassistant.ai/contextual-mistakes/data/debug-data/valid'
targets = {'O': 0, 'M': 1}
seq_len = 8
token_style = 'bert'
data = RepunctDataset.parse_tokens(['казнить', 'нельзя', 'помиловать'],
tokenizer,
seq_len,
token_style,
[0, 1, 2],
debug=False)
print(tokenizer.convert_ids_to_tokens(data[0][0]))
print(data)
# old_data = RepunctDataset._parse_data_old(file_path, tokenizer, targets, seq_len, token_style)
# new_data = BaseDataset._parse_data(file_path, tokenizer, targets, seq_len, token_style)
# assert len(old_data) == len(new_data)
# assert old_data[0][0] == new_data[0][0]
# assert old_data[0][1] == new_data[0][3]
# assert old_data[0][2] == new_data[0][2]
# assert old_data[0][3] == new_data[0][4]
# assert old_data[1][0] == new_data[1][0]
# assert old_data[1][1] == new_data[1][3]
# assert old_data[1][2] == new_data[1][2]
# assert old_data[1][3] == new_data[1][4]
# print('All clear')