-
Notifications
You must be signed in to change notification settings - Fork 7
/
bert_dataset.py
370 lines (316 loc) · 14.8 KB
/
bert_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
import random
from tqdm import tqdm
import torch
from torch.utils.data import Dataset
class BERTDataset(Dataset):
def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8", corpus_lines=None, on_memory=False):
self.vocab = tokenizer.vocab
self.tokenizer = tokenizer
self.seq_len = seq_len
self.on_memory = on_memory
self.corpus_lines = corpus_lines # number of non-empty lines in input corpus
self.corpus_path = corpus_path
self.encoding = encoding
self.current_doc = 0 # to avoid random sentence from same doc
# for loading samples directly from file
self.sample_counter = 0 # used to keep track of full epochs on file
self.line_buffer = None # keep second sentence of a pair in memory and use as first sentence in next pair
# for loading samples in memory
self.current_random_doc = 0
self.num_docs = 0
self.sample_to_doc = [] # map sample index to doc and line
# load samples into memory
if on_memory:
self.all_docs = []
doc = []
self.corpus_lines = 0
with open(corpus_path, "r", encoding=encoding) as f:
for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
line = line.strip()
if line == "":
self.all_docs.append(doc)
doc = []
# remove last added sample because there won't be a subsequent line anymore in the doc
self.sample_to_doc.pop()
else:
# store as one sample
sample = {"doc_id": len(self.all_docs),
"line": len(doc)}
self.sample_to_doc.append(sample)
doc.append(line)
self.corpus_lines = self.corpus_lines + 1
# if last row in file is not empty
if self.all_docs[-1] != doc:
self.all_docs.append(doc)
self.sample_to_doc.pop()
self.num_docs = len(self.all_docs)
# load samples later lazily from disk
else:
if self.corpus_lines is None:
with open(corpus_path, "r", encoding=encoding) as f:
self.corpus_lines = 0
for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
if line.strip() == "":
self.num_docs += 1
else:
self.corpus_lines += 1
# if doc does not end with empty line
if line.strip() != "":
self.num_docs += 1
self.file = open(corpus_path, "r", encoding=encoding)
self.random_file = open(corpus_path, "r", encoding=encoding)
def __len__(self):
# last line of doc won't be used, because there's no "nextSentence". Additionally, we start counting at 0.
return self.corpus_lines - self.num_docs - 1
def __getitem__(self, item):
cur_id = self.sample_counter
self.sample_counter += 1
if not self.on_memory:
# after one epoch we start again from beginning of file
if cur_id != 0 and (cur_id % len(self) == 0):
self.file.close()
self.file = open(self.corpus_path, "r", encoding=self.encoding)
t1, t2, is_next_label = self.random_sent(item)
# tokenize
tokens_a = self.tokenizer.tokenize(t1)
tokens_b = self.tokenizer.tokenize(t2)
# combine to one sample
cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, tokens_b=tokens_b, is_next=is_next_label)
# transform sample to features
cur_features = convert_example_to_features(cur_example, self.seq_len, self.tokenizer)
cur_tensors = dict(
input_ids=torch.tensor(cur_features.input_ids),
attention_mask=torch.tensor(cur_features.input_mask),
token_type_ids=torch.tensor(cur_features.segment_ids),
labels=torch.tensor(cur_features.lm_label_ids),
next_sentence_label=torch.tensor(cur_features.is_next))
return cur_tensors
def random_sent(self, index):
"""
Get one sample from corpus consisting of two sentences. With prob. 50% these are two subsequent sentences
from one doc. With 50% the second sentence will be a random one from another doc.
:param index: int, index of sample.
:return: (str, str, int), sentence 1, sentence 2, isNextSentence Label
"""
t1, t2 = self.get_corpus_line(index)
if random.random() > 0.5:
label = 0
else:
while True:
t2 = self.get_random_line()
if len(t2) > 0:
break
label = 1
assert len(t1) > 0, "empty t1 in random_send"
assert len(t2) > 0, "empty t2 in random_send"
return t1, t2, label
def get_corpus_line(self, item):
"""
Get one sample from corpus consisting of a pair of two subsequent lines from the same doc.
:param item: int, index of sample.
:return: (str, str), two subsequent sentences from corpus
"""
t1 = ""
t2 = ""
assert item < self.corpus_lines, "out of dataset bound"
if self.on_memory:
sample = self.sample_to_doc[item]
t1 = self.all_docs[sample["doc_id"]][sample["line"]]
t2 = self.all_docs[sample["doc_id"]][sample["line"]+1]
# used later to avoid random nextSentence from same doc
self.current_doc = sample["doc_id"]
return t1, t2
else:
if self.line_buffer is None:
# read first non-empty line of file
while t1 == "" or t2 == "":
t1 = next(self.file).strip()
t2 = next(self.file).strip()
else:
# use t2 from previous iteration as new t1
t1 = self.line_buffer
t2 = next(self.file).strip()
# skip empty rows that are used for separating documents and keep track of current doc id
while t2 == "" or t1 == "":
t1 = next(self.file).strip()
t2 = next(self.file).strip()
self.current_doc = self.current_doc+1
self.line_buffer = t2
assert t1 != "", "t1 empty"
assert t2 != "", "t2 empty"
return t1, t2
def get_random_line(self):
"""
Get random line from another document for nextSentence task.
:return: str, content of one line
"""
# Similar to original tf repo: This outer loop should rarely go for more than one iteration for large
# corpora. However, just to be careful, we try to make sure that
# the random document is not the same as the document we're processing.
for _ in range(10):
if self.on_memory:
rand_doc_idx = random.randint(0, len(self.all_docs)-1)
rand_doc = self.all_docs[rand_doc_idx]
line = rand_doc[random.randrange(len(rand_doc))]
else:
rand_index = random.randint(1, self.corpus_lines if self.corpus_lines < 1000 else 1000)
# pick random line
for _ in range(rand_index):
line = self.get_next_line()
# check if our picked random line is really from another doc like we want it to be
if self.current_random_doc != self.current_doc:
break
return line
def get_next_line(self):
""" Gets next line of random_file and starts over when reaching end of file"""
try:
line = next(self.random_file).strip()
# keep track of which document we are currently looking at to later avoid having the same doc as t1
if line == "":
self.current_random_doc = self.current_random_doc + 1
line = next(self.random_file).strip()
except StopIteration:
self.random_file.close()
self.random_file = open(self.corpus_path, "r", encoding=self.encoding)
line = next(self.random_file).strip()
return line
class InputExample(object):
"""A single training/test example for the language model."""
def __init__(self, guid, tokens_a, tokens_b=None, is_next=None, lm_labels=None):
"""Constructs a InputExample.
Args:
guid: Unique id for the example.
tokens_a: string. The untokenized text of the first sequence. For single
sequence tasks, only this sequence must be specified.
tokens_b: (Optional) string. The untokenized text of the second sequence.
Only must be specified for sequence pair tasks.
"""
self.guid = guid
self.tokens_a = tokens_a
self.tokens_b = tokens_b
self.is_next = is_next # nextSentence
self.lm_labels = lm_labels # masked words for language model
class InputFeatures(object):
"""A single set of features of data."""
def __init__(self, input_ids, input_mask, segment_ids, is_next, lm_label_ids):
self.input_ids = input_ids
self.input_mask = input_mask
self.segment_ids = segment_ids
self.is_next = is_next
self.lm_label_ids = lm_label_ids
def random_word(tokens, tokenizer):
"""
Masking some random tokens for Language Model task with probabilities as in the original BERT paper.
:param tokens: list of str, tokenized sentence.
:param tokenizer: Tokenizer, object used for tokenization (we need it's vocab here)
:return: (list of str, list of int), masked tokens and related labels for LM prediction
"""
output_label = []
for i, token in enumerate(tokens):
prob = random.random()
# mask token with 15% probability
if prob < 0.15:
prob /= 0.15
# 80% randomly change token to mask token
if prob < 0.8:
tokens[i] = "[MASK]"
# 10% randomly change token to random token
elif prob < 0.9:
tokens[i] = random.choice(list(tokenizer.vocab.items()))[0]
# -> rest 10% randomly keep current token
# append current token to output (we will predict these later)
try:
output_label.append(tokenizer.vocab[token])
except KeyError:
# For unknown words (should not occur with BPE vocab)
output_label.append(tokenizer.vocab["[UNK]"])
else:
# no masking token (will be ignored by loss function later)
output_label.append(-100)
return tokens, output_label
def convert_example_to_features(example, max_seq_length, tokenizer):
"""
Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample with
IDs, LM labels, input_mask, CLS and SEP tokens etc.
:param example: InputExample, containing sentence input as strings and is_next label
:param max_seq_length: int, maximum length of sequence.
:param tokenizer: Tokenizer
:return: InputFeatures, containing all inputs and labels of one sample as IDs (as used for model training)
"""
tokens_a = example.tokens_a
tokens_b = example.tokens_b
# Modifies `tokens_a` and `tokens_b` in place so that the total
# length is less than the specified length.
# Account for [CLS], [SEP], [SEP] with "- 3"
_truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
tokens_a, t1_label = random_word(tokens_a, tokenizer)
tokens_b, t2_label = random_word(tokens_b, tokenizer)
# concatenate lm labels and account for CLS, SEP, SEP
lm_label_ids = ([-100] + t1_label + [-100] + t2_label + [-100])
# The convention in BERT is:
# (a) For sequence pairs:
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
# (b) For single sequences:
# tokens: [CLS] the dog is hairy . [SEP]
# type_ids: 0 0 0 0 0 0 0
#
# Where "type_ids" are used to indicate whether this is the first
# sequence or the second sequence. The embedding vectors for `type=0` and
# `type=1` were learned during pre-training and are added to the wordpiece
# embedding vector (and position vector). This is not *strictly* necessary
# since the [SEP] token unambigiously separates the sequences, but it makes
# it easier for the model to learn the concept of sequences.
#
# For classification tasks, the first vector (corresponding to [CLS]) is
# used as as the "sentence vector". Note that this only makes sense because
# the entire model is fine-tuned.
tokens = []
segment_ids = []
tokens.append("[CLS]")
segment_ids.append(0)
for token in tokens_a:
tokens.append(token)
segment_ids.append(0)
tokens.append("[SEP]")
segment_ids.append(0)
assert len(tokens_b) > 0, "empty tokens_b"
for token in tokens_b:
tokens.append(token)
segment_ids.append(1)
tokens.append("[SEP]")
segment_ids.append(1)
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1] * len(input_ids)
# Zero-pad up to the sequence length.
while len(input_ids) < max_seq_length:
input_ids.append(0)
input_mask.append(0)
segment_ids.append(0)
lm_label_ids.append(-100)
assert len(input_ids) == max_seq_length, "input_ids not equal to max len"
assert len(input_mask) == max_seq_length, "input_mask not equal to max len"
assert len(segment_ids) == max_seq_length, "segment_ids not equal to max len"
assert len(lm_label_ids) == max_seq_length, "lm_label_ids not equal to max len"
features = InputFeatures(input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
lm_label_ids=lm_label_ids,
is_next=example.is_next)
return features
def _truncate_seq_pair(tokens_a, tokens_b, max_length):
"""Truncates a sequence pair in place to the maximum length."""
# This is a simple heuristic which will always truncate the longer sequence
# one token at a time. This makes more sense than truncating an equal percent
# of tokens from each, since if one sequence is very short then each token
# that's truncated likely contains more information than a longer sequence.
while True:
total_length = len(tokens_a) + len(tokens_b)
if total_length <= max_length:
break
if len(tokens_a) > len(tokens_b):
tokens_a.pop()
else:
tokens_b.pop()