dongxiaoqun commited on
Commit
2c60db7
1 Parent(s): 2d9766c

Upload tokenizers_pegasus.py

Browse files
Files changed (1) hide show
  1. tokenizers_pegasus.py +599 -0
tokenizers_pegasus.py ADDED
@@ -0,0 +1,599 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ from fengshen.examples.pegasus.data_utils import (
3
+ _is_control,
4
+ _is_punctuation,
5
+ _is_whitespace,
6
+ _is_chinese_char)
7
+ from transformers import PreTrainedTokenizer
8
+ from transformers import logging
9
+ from typing import List, Optional, Tuple, Union
10
+ import collections
11
+ import os
12
+ import unicodedata
13
+ import re
14
+ import jieba
15
+ import sys
16
+
17
+ sys.path.append("../../../../")
18
+
19
+ jieba.dt.tmp_dir = os.path.expanduser(
20
+ "/cognitive_comp/dongxiaoqun/software/jieba/tmp/")
21
+ # jieba.enable_parallel(8)
22
+ jieba.initialize()
23
+
24
+ logger = logging.get_logger(__name__)
25
+
26
+ VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
27
+
28
+
29
+ def load_vocab(vocab_file):
30
+ """Loads a vocabulary file into a dictionary."""
31
+ vocab = collections.OrderedDict()
32
+ with open(vocab_file, "r", encoding="utf-8") as reader:
33
+ tokens = reader.readlines()
34
+ for index, token in enumerate(tokens):
35
+ token = token.rstrip("\n")
36
+ vocab[token] = index
37
+ return vocab
38
+
39
+
40
+ def whitespace_tokenize(text):
41
+ """Runs basic whitespace cleaning and splitting on a piece of text."""
42
+ text = text.strip()
43
+ if not text:
44
+ return []
45
+ tokens = text.split()
46
+ return tokens
47
+
48
+
49
+ class PegasusTokenizer(PreTrainedTokenizer):
50
+ # copy from BertTokenizer
51
+ r"""
52
+ Construct a Pegasus tokenizer. Based on WordPiece.
53
+ This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
54
+ this superclass for more information regarding those methods.
55
+ Args:
56
+ vocab_file (`str`):
57
+ File containing the vocabulary.
58
+ do_lower_case (`bool`, *optional*, defaults to `True`):
59
+ Whether or not to lowercase the input when tokenizing.
60
+ do_basic_tokenize (`bool`, *optional*, defaults to `True`):
61
+ Whether or not to do basic tokenization before WordPiece.
62
+ never_split (`Iterable`, *optional*):
63
+ Collection of tokens which will never be split during tokenization. Only has an effect when
64
+ `do_basic_tokenize=True`
65
+ unk_token (`str`, *optional*, defaults to `"[UNK]"`):
66
+ The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
67
+ token instead.
68
+ sep_token (`str`, *optional*, defaults to `"[SEP]"`):
69
+ The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
70
+ sequence classification or for a text and a question for question answering. It is also used as the last
71
+ token of a sequence built with special tokens.
72
+ pad_token (`str`, *optional*, defaults to `"[PAD]"`):
73
+ The token used for padding, for example when batching sequences of different lengths.
74
+ cls_token (`str`, *optional*, defaults to `"[CLS]"`):
75
+ The classifier token which is used when doing sequence classification (classification of the whole sequence
76
+ instead of per-token classification). It is the first token of the sequence when built with special tokens.
77
+ mask_token (`str`, *optional*, defaults to `"[MASK]"`):
78
+ The token used for masking values. This is the token used when training this model with masked language
79
+ modeling. This is the token which the model will try to predict.
80
+ tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):
81
+ Whether or not to tokenize Chinese characters.
82
+ This should likely be deactivated for Japanese (see this
83
+ [issue](https://github.com/huggingface/transformers/issues/328)).
84
+ strip_accents (`bool`, *optional*):
85
+ Whether or not to strip all accents. If this option is not specified, then it will be determined by the
86
+ value for `lowercase` (as in the original BERT).
87
+ """
88
+
89
+ vocab_files_names = VOCAB_FILES_NAMES
90
+ model_input_names = ["input_ids", "attention_mask"]
91
+
92
+ # pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
93
+ # pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
94
+ # max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
95
+
96
+ def __init__(self,
97
+ vocab_file,
98
+ do_lower_case=True,
99
+ do_basic_tokenize=True,
100
+ never_split=None,
101
+ pad_token="<pad>",
102
+ eos_token="</s>",
103
+ unk_token="<unk>",
104
+ mask_token="<mask_2>",
105
+ mask_token_sent="<mask_1>",
106
+ additional_special_tokens=None,
107
+ sep_token="[SEP]",
108
+ cls_token="[CLS]",
109
+ tokenize_chinese_chars=True,
110
+ strip_accents=None,
111
+ offset=100,
112
+ pre_tokenizer=lambda x: jieba.cut(x, HMM=False),
113
+ **kwargs):
114
+ self.offset = offset
115
+
116
+ if additional_special_tokens is not None:
117
+ if not isinstance(additional_special_tokens, list):
118
+ raise TypeError(
119
+ f"additional_special_tokens should be of type {type(list)}, \
120
+ but is {type(additional_special_tokens)}"
121
+ )
122
+
123
+ additional_special_tokens_extended = (
124
+ ([mask_token_sent] + additional_special_tokens)
125
+ if mask_token_sent not in additional_special_tokens
126
+ and mask_token_sent is not None else additional_special_tokens)
127
+
128
+ # fill additional tokens with ..., <unk_token_102> in case not all additional tokens are already taken
129
+ additional_special_tokens_extended += [
130
+ f"<unk_{i}>" for i in range(
131
+ len(additional_special_tokens_extended), self.offset - 1)
132
+ ]
133
+
134
+ if len(set(additional_special_tokens_extended)) != len(
135
+ additional_special_tokens_extended):
136
+ raise ValueError(
137
+ f"Please make sure that the provided additional_special_tokens \
138
+ do not contain an incorrectly shifted list of <unk_x> tokens. \
139
+ Found {additional_special_tokens_extended}."
140
+ )
141
+ additional_special_tokens = additional_special_tokens_extended
142
+ else:
143
+ additional_special_tokens = [
144
+ mask_token_sent
145
+ ] if mask_token_sent is not None else []
146
+ # additional_special_tokens += [f"<unk_{i}>" for i in range(3, self.offset)]
147
+
148
+ # print("additional_special_tokens: ", additional_special_tokens)
149
+
150
+ if not os.path.isfile(vocab_file):
151
+ raise ValueError(
152
+ f"Can't find a vocabulary file at path '{vocab_file}'. \
153
+ To load the vocabulary from a Google pretrained "
154
+ "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"
155
+ )
156
+
157
+ super().__init__(
158
+ do_lower_case=do_lower_case,
159
+ do_basic_tokenize=do_basic_tokenize,
160
+ never_split=never_split,
161
+ unk_token=unk_token,
162
+ sep_token=sep_token,
163
+ pad_token=pad_token,
164
+ cls_token=cls_token,
165
+ mask_token=mask_token,
166
+ eos_token=eos_token,
167
+ tokenize_chinese_chars=tokenize_chinese_chars,
168
+ additional_special_tokens=additional_special_tokens,
169
+ strip_accents=strip_accents,
170
+ **kwargs,
171
+ )
172
+
173
+ self.pre_tokenizer = pre_tokenizer
174
+ self.mask_token_sent = mask_token_sent
175
+ self.vocab = load_vocab(vocab_file)
176
+
177
+ self.vocab[self.eos_token] = self.vocab.pop("[unused1]")
178
+ # self.vocab[self.eos_token] = self.vocab.pop("[unused2]")
179
+ self.vocab[self.pad_token] = self.vocab.pop("[PAD]")
180
+ self.vocab[self.unk_token] = self.vocab.pop("[UNK]")
181
+
182
+ if self.mask_token_sent is not None:
183
+ self.vocab[self.mask_token] = self.vocab.pop("[unused3]")
184
+ self.vocab[self.mask_token_sent] = self.vocab.pop("[unused2]")
185
+
186
+ self.ids_to_tokens = collections.OrderedDict([
187
+ (ids, tok) for tok, ids in self.vocab.items()
188
+ ])
189
+ self.do_basic_tokenize = do_basic_tokenize
190
+ if do_basic_tokenize:
191
+ self.basic_tokenizer = BasicTokenizer(
192
+ do_lower_case=do_lower_case,
193
+ never_split=never_split,
194
+ tokenize_chinese_chars=tokenize_chinese_chars,
195
+ strip_accents=strip_accents,
196
+ )
197
+ self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab,
198
+ unk_token=self.unk_token)
199
+
200
+ @property
201
+ def do_lower_case(self):
202
+ return self.basic_tokenizer.do_lower_case
203
+
204
+ @property
205
+ def vocab_size(self):
206
+ return len(self.vocab)
207
+
208
+ def get_vocab(self):
209
+ return dict(self.vocab, **self.added_tokens_encoder)
210
+
211
+ def _tokenize(self, text):
212
+ split_tokens = []
213
+ # print("pegasus_tokenizer: ", text)
214
+ for text in self.pre_tokenizer(text):
215
+ if text in self.vocab:
216
+ split_tokens.append(text)
217
+ else:
218
+ if self.do_basic_tokenize:
219
+ for token in self.basic_tokenizer.tokenize(
220
+ text, never_split=self.all_special_tokens):
221
+
222
+ # If the token is part of the never_split set
223
+ if token in self.basic_tokenizer.never_split:
224
+ split_tokens.append(token)
225
+ else:
226
+ split_tokens += self.wordpiece_tokenizer.tokenize(
227
+ token)
228
+ else:
229
+ split_tokens = self.wordpiece_tokenizer.tokenize(text)
230
+ return split_tokens
231
+
232
+ def _convert_token_to_id(self, token):
233
+ """Converts a token (str) in an id using the vocab."""
234
+ return self.vocab.get(token, self.vocab.get(self.unk_token))
235
+
236
+ def _convert_id_to_token(self, index):
237
+ """Converts an index (integer) in a token (str) using the vocab."""
238
+ return self.ids_to_tokens.get(index, self.unk_token)
239
+
240
+ @staticmethod
241
+ def _cjk_punctuation():
242
+ return u'\uff02\uff03\uff04\uff05\uff06\uff07\uff08\uff09\uff0a\uff0b\uff0c\uff0d\uff0f\uff1a\uff1b\uff1c\uff1d\
243
+ \uff1e\uff20\uff3b\uff3c\uff3d\uff3e\uff3f\uff40\uff5b\uff5c\uff5d\uff5e\uff5f\uff60\uff62\
244
+ \uff63\uff64\u3000\u3001\u3003\u3008\u3009\u300a\u300b\u300c\u300d\u300e\u300f\u3010\u3011\u3014\
245
+ \u3015\u3016\u3017\u3018\u3019\u301a\u301b\u301c\u301d\u301e\u301f\u3030\u303e\u303f\u2013\u2014\
246
+ \u2018\u2019\u201b\u201c\u201d\u201e\u201f\u2026\u2027\ufe4f\ufe51\ufe54\u00b7\uff01\uff1f\uff61\u3002'
247
+
248
+ def convert_ids_to_tokens(
249
+ self,
250
+ ids: Union[int, List[int]],
251
+ skip_special_tokens: bool = False) -> Union[str, List[str]]:
252
+ """
253
+ Converts a single index or a sequence of indices in a token or a sequence of tokens, using the vocabulary and
254
+ added tokens.
255
+ Args:
256
+ ids (`int` or `List[int]`):
257
+ The token id (or token ids) to convert to tokens.
258
+ skip_special_tokens (`bool`, *optional*, defaults to `False`):
259
+ Whether or not to remove special tokens in the decoding.
260
+ Returns:
261
+ `str` or `List[str]`: The decoded token(s).
262
+ """
263
+ if isinstance(ids, int):
264
+ if ids in self.added_tokens_decoder:
265
+ return self.added_tokens_decoder[ids]
266
+ else:
267
+ return self._convert_id_to_token(ids)
268
+ tokens = []
269
+ for index in ids:
270
+ index = int(index)
271
+ if skip_special_tokens and index in self.all_special_ids and index != 2:
272
+ continue
273
+ if index in self.added_tokens_decoder:
274
+ tokens.append(self.added_tokens_decoder[index])
275
+ else:
276
+ tokens.append(self._convert_id_to_token(index))
277
+ return tokens
278
+
279
+ def convert_tokens_to_string(self, tokens):
280
+ """Converts a sequence of tokens (string) in a single string."""
281
+ # for token in
282
+ # tokens = tokens or self.ids_to_tokens(ids)
283
+ # tokens = [token for token in tokens if not self._is_special(token)]
284
+
285
+ text = ''
286
+ for i, token in enumerate(tokens):
287
+ if token[:2] == '##':
288
+ text += token[2:]
289
+ elif len(token) == 1 and _is_chinese_char(ord(token)):
290
+ text += token
291
+ elif len(token) == 1 and _is_punctuation(token):
292
+ text += token
293
+ text += ' '
294
+ elif i > 0 and _is_chinese_char(ord(text[-1])):
295
+ text += token
296
+ elif tokens == "</s>":
297
+ continue
298
+ else:
299
+ text += ' '
300
+ text += token
301
+
302
+ text = re.sub(' +', ' ', text)
303
+ text = re.sub('\' (re|m|s|t|ve|d|ll) ', '\'\\1 ', text)
304
+ punctuation = re.sub(' +', '', self._cjk_punctuation()).strip() + '+-/={(<['
305
+ punctuation_regex = '|'.join([re.escape(p) for p in punctuation])
306
+ punctuation_regex = '(%s) ' % punctuation_regex
307
+ text = re.sub(punctuation_regex, '\\1', text)
308
+ text = re.sub(r'(\d\.) (\d)', '\\1\\2', text)
309
+
310
+ return text.strip()
311
+ # out_string = " ".join(tokens).replace(" ##", "").strip()
312
+
313
+ def build_inputs_with_special_tokens(
314
+ self,
315
+ token_ids_0: List[int],
316
+ token_ids_1: Optional[List[int]] = None) -> List[int]:
317
+ """
318
+ Build model inputs from a sequence or a pair of sequences for sequence classification tasks by concatenating
319
+ and adding special tokens. A PEGASUS sequence has the following format, where `X` represents the sequence:
320
+ - single sequence: `X </s>`
321
+ - pair of sequences: `A B </s>` (not intended use)
322
+ BOS is never used. Pairs of sequences are not the expected use case, but they will be handled without a
323
+ separator.
324
+ Args:
325
+ token_ids_0 (`List[int]`):
326
+ List of IDs to which the special tokens will be added.
327
+ token_ids_1 (`List[int]`, *optional*):
328
+ Optional second list of IDs for sequence pairs.
329
+ Returns:
330
+ `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
331
+ """
332
+ if token_ids_1 is None:
333
+ return token_ids_0 + [self.eos_token_id]
334
+ return token_ids_0 + token_ids_1 + [self.eos_token_id]
335
+
336
+ def _special_token_mask(self, seq):
337
+ all_special_ids = set(
338
+ self.all_special_ids) # call it once instead of inside list comp
339
+ # all_special_ids.remove(self.unk_token_id) # <unk> is only sometimes special
340
+
341
+ return [1 if x in all_special_ids else 0 for x in seq]
342
+
343
+ def get_special_tokens_mask(
344
+ self,
345
+ token_ids_0: List[int],
346
+ token_ids_1: Optional[List[int]] = None,
347
+ already_has_special_tokens: bool = False) -> List[int]:
348
+ """
349
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
350
+ special tokens using the tokenizer `prepare_for_model` method.
351
+ Args:
352
+ token_ids_0 (`List[int]`):
353
+ List of IDs.
354
+ token_ids_1 (`List[int]`, *optional*):
355
+ Optional second list of IDs for sequence pairs.
356
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
357
+ Whether or not the token list is already formatted with special tokens for the model.
358
+ Returns:
359
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
360
+ """
361
+
362
+ if already_has_special_tokens:
363
+ return self._special_token_mask(token_ids_0)
364
+ elif token_ids_1 is None:
365
+ return self._special_token_mask(token_ids_0) + [self.eos_token_id]
366
+ else:
367
+ return self._special_token_mask(token_ids_0 +
368
+ token_ids_1) + [self.eos_token_id]
369
+
370
+ def num_special_tokens_to_add(self, pair=False):
371
+ """Just EOS"""
372
+ return 1
373
+
374
+ def save_vocabulary(self,
375
+ save_directory: str,
376
+ filename_prefix: Optional[str] = None) -> Tuple[str]:
377
+ index = 0
378
+ if os.path.isdir(save_directory):
379
+ vocab_file = os.path.join(
380
+ save_directory,
381
+ (filename_prefix + "-" if filename_prefix else "") +
382
+ VOCAB_FILES_NAMES["vocab_file"])
383
+ else:
384
+ vocab_file = (filename_prefix +
385
+ "-" if filename_prefix else "") + save_directory
386
+ with open(vocab_file, "w", encoding="utf-8") as writer:
387
+ for token, token_index in sorted(self.vocab.items(),
388
+ key=lambda kv: kv[1]):
389
+ if index != token_index:
390
+ logger.warning(
391
+ f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."
392
+ " Please check that the vocabulary is not corrupted!")
393
+ index = token_index
394
+ writer.write(token + "\n")
395
+ index += 1
396
+ return (vocab_file, )
397
+
398
+
399
+ class BasicTokenizer(object):
400
+ """
401
+ Constructs a BasicTokenizer that will run basic tokenization (punctuation splitting, lower casing, etc.).
402
+ Args:
403
+ do_lower_case (`bool`, *optional*, defaults to `True`):
404
+ Whether or not to lowercase the input when tokenizing.
405
+ never_split (`Iterable`, *optional*):
406
+ Collection of tokens which will never be split during tokenization. Only has an effect when
407
+ `do_basic_tokenize=True`
408
+ tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):
409
+ Whether or not to tokenize Chinese characters.
410
+ This should likely be deactivated for Japanese (see this
411
+ [issue](https://github.com/huggingface/transformers/issues/328)).
412
+ strip_accents: (`bool`, *optional*):
413
+ Whether or not to strip all accents. If this option is not specified, then it will be determined by the
414
+ value for `lowercase` (as in the original BERT).
415
+ """
416
+
417
+ def __init__(self,
418
+ do_lower_case=True,
419
+ never_split=None,
420
+ tokenize_chinese_chars=True,
421
+ strip_accents=None):
422
+ if never_split is None:
423
+ never_split = []
424
+ self.do_lower_case = do_lower_case
425
+ self.never_split = set(never_split)
426
+ self.tokenize_chinese_chars = tokenize_chinese_chars
427
+ self.strip_accents = strip_accents
428
+
429
+ def tokenize(self, text, never_split=None):
430
+ """
431
+ Basic Tokenization of a piece of text. Split on "white spaces" only, for sub-word tokenization, see
432
+ WordPieceTokenizer.
433
+ Args:
434
+ never_split (`List[str]`, *optional*)
435
+ Kept for backward compatibility purposes. Now implemented directly at the base class level (see
436
+ [`PreTrainedTokenizer.tokenize`]) List of token not to split.
437
+ """
438
+ # union() returns a new set by concatenating the two sets.
439
+ never_split = self.never_split.union(
440
+ set(never_split)) if never_split else self.never_split
441
+ text = self._clean_text(text)
442
+
443
+ # This was added on November 1st, 2018 for the multilingual and Chinese
444
+ # models. This is also applied to the English models now, but it doesn't
445
+ # matter since the English models were not trained on any Chinese data
446
+ # and generally don't have any Chinese data in them (there are Chinese
447
+ # characters in the vocabulary because Wikipedia does have some Chinese
448
+ # words in the English Wikipedia.).
449
+ if self.tokenize_chinese_chars:
450
+ text = self._tokenize_chinese_chars(text)
451
+ orig_tokens = whitespace_tokenize(text)
452
+ split_tokens = []
453
+ for token in orig_tokens:
454
+ if token not in never_split:
455
+ if self.do_lower_case:
456
+ token = token.lower()
457
+ if self.strip_accents is not False:
458
+ token = self._run_strip_accents(token)
459
+ elif self.strip_accents:
460
+ token = self._run_strip_accents(token)
461
+ split_tokens.extend(self._run_split_on_punc(token, never_split))
462
+
463
+ output_tokens = whitespace_tokenize(" ".join(split_tokens))
464
+ return output_tokens
465
+
466
+ def _run_strip_accents(self, text):
467
+ """Strips accents from a piece of text."""
468
+ text = unicodedata.normalize("NFD", text)
469
+ output = []
470
+ for char in text:
471
+ cat = unicodedata.category(char)
472
+ if cat == "Mn":
473
+ continue
474
+ output.append(char)
475
+ return "".join(output)
476
+
477
+ def _run_split_on_punc(self, text, never_split=None):
478
+ """Splits punctuation on a piece of text."""
479
+ if never_split is not None and text in never_split:
480
+ return [text]
481
+ chars = list(text)
482
+ i = 0
483
+ start_new_word = True
484
+ output = []
485
+ while i < len(chars):
486
+ char = chars[i]
487
+ if _is_punctuation(char):
488
+ output.append([char])
489
+ start_new_word = True
490
+ else:
491
+ if start_new_word:
492
+ output.append([])
493
+ start_new_word = False
494
+ output[-1].append(char)
495
+ i += 1
496
+
497
+ return ["".join(x) for x in output]
498
+
499
+ def _tokenize_chinese_chars(self, text):
500
+ """Adds whitespace around any CJK character."""
501
+ output = []
502
+ for char in text:
503
+ cp = ord(char)
504
+ if self._is_chinese_char(cp):
505
+ output.append(" ")
506
+ output.append(char)
507
+ output.append(" ")
508
+ else:
509
+ output.append(char)
510
+ return "".join(output)
511
+
512
+ def _is_chinese_char(self, cp):
513
+ """Checks whether CP is the codepoint of a CJK character."""
514
+ # This defines a "chinese character" as anything in the CJK Unicode block:
515
+ # https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
516
+ #
517
+ # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
518
+ # despite its name. The modern Korean Hangul alphabet is a different block,
519
+ # as is Japanese Hiragana and Katakana. Those alphabets are used to write
520
+ # space-separated words, so they are not treated specially and handled
521
+ # like the all of the other languages.
522
+ if ((cp >= 0x4E00 and cp <= 0x9FFF)
523
+ or (cp >= 0x3400 and cp <= 0x4DBF) #
524
+ or (cp >= 0x20000 and cp <= 0x2A6DF) #
525
+ or (cp >= 0x2A700 and cp <= 0x2B73F) #
526
+ or (cp >= 0x2B740 and cp <= 0x2B81F) #
527
+ or (cp >= 0x2B820 and cp <= 0x2CEAF) #
528
+ or (cp >= 0xF900 and cp <= 0xFAFF)
529
+ or (cp >= 0x2F800 and cp <= 0x2FA1F)): #
530
+ return True
531
+
532
+ return False
533
+
534
+ def _clean_text(self, text):
535
+ """Performs invalid character removal and whitespace cleanup on text."""
536
+ output = []
537
+ for char in text:
538
+ cp = ord(char)
539
+ if cp == 0 or cp == 0xFFFD or _is_control(char):
540
+ continue
541
+ if _is_whitespace(char):
542
+ output.append(" ")
543
+ else:
544
+ output.append(char)
545
+ return "".join(output)
546
+
547
+
548
+ class WordpieceTokenizer(object):
549
+ """Runs WordPiece tokenization."""
550
+
551
+ def __init__(self, vocab, unk_token, max_input_chars_per_word=100):
552
+ self.vocab = vocab
553
+ self.unk_token = unk_token
554
+ self.max_input_chars_per_word = max_input_chars_per_word
555
+
556
+ def tokenize(self, text):
557
+ """
558
+ Tokenizes a piece of text into its word pieces. This uses a greedy longest-match-first algorithm to perform
559
+ tokenization using the given vocabulary.
560
+ For example, `input = "unaffable"` wil return as output `["un", "##aff", "##able"]`.
561
+ Args:
562
+ text: A single token or whitespace separated tokens. This should have
563
+ already been passed through *BasicTokenizer*.
564
+ Returns:
565
+ A list of wordpiece tokens.
566
+ """
567
+
568
+ output_tokens = []
569
+ for token in whitespace_tokenize(text):
570
+ chars = list(token)
571
+ if len(chars) > self.max_input_chars_per_word:
572
+ output_tokens.append(self.unk_token)
573
+ continue
574
+
575
+ is_bad = False
576
+ start = 0
577
+ sub_tokens = []
578
+ while start < len(chars):
579
+ end = len(chars)
580
+ cur_substr = None
581
+ while start < end:
582
+ substr = "".join(chars[start:end])
583
+ if start > 0:
584
+ substr = "##" + substr
585
+ if substr in self.vocab:
586
+ cur_substr = substr
587
+ break
588
+ end -= 1
589
+ if cur_substr is None:
590
+ is_bad = True
591
+ break
592
+ sub_tokens.append(cur_substr)
593
+ start = end
594
+
595
+ if is_bad:
596
+ output_tokens.append(self.unk_token)
597
+ else:
598
+ output_tokens.extend(sub_tokens)
599
+ return output_tokens