Aluode/PerceptionLabPortable
0
1# coding=utf-82# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8# http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""Tokenization classes for Bert."""16 17import collections18import os19import unicodedata20from typing import Optional21 22from ...tokenization_utils import PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace23from ...utils import logging24 25 26logger = logging.get_logger(__name__)27 28VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}29 30 31def load_vocab(vocab_file):32 """Loads a vocabulary file into a dictionary."""33 vocab = collections.OrderedDict()34 with open(vocab_file, "r", encoding="utf-8") as reader:35 tokens = reader.readlines()36 for index, token in enumerate(tokens):37 token = token.rstrip("\n")38 vocab[token] = index39 return vocab40 41 42def whitespace_tokenize(text):43 """Runs basic whitespace cleaning and splitting on a piece of text."""44 text = text.strip()45 if not text:46 return []47 tokens = text.split()48 return tokens49 50 51class BertTokenizer(PreTrainedTokenizer):52 r"""53 Construct a BERT tokenizer. Based on WordPiece.54 55 This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to56 this superclass for more information regarding those methods.57 58 Args:59 vocab_file (`str`):60 File containing the vocabulary.61 do_lower_case (`bool`, *optional*, defaults to `True`):62 Whether or not to lowercase the input when tokenizing.63 do_basic_tokenize (`bool`, *optional*, defaults to `True`):64 Whether or not to do basic tokenization before WordPiece.65 never_split (`Iterable`, *optional*):66 Collection of tokens which will never be split during tokenization. Only has an effect when67 `do_basic_tokenize=True`68 unk_token (`str`, *optional*, defaults to `"[UNK]"`):69 The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this70 token instead.71 sep_token (`str`, *optional*, defaults to `"[SEP]"`):72 The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for73 sequence classification or for a text and a question for question answering. It is also used as the last74 token of a sequence built with special tokens.75 pad_token (`str`, *optional*, defaults to `"[PAD]"`):76 The token used for padding, for example when batching sequences of different lengths.77 cls_token (`str`, *optional*, defaults to `"[CLS]"`):78 The classifier token which is used when doing sequence classification (classification of the whole sequence79 instead of per-token classification). It is the first token of the sequence when built with special tokens.80 mask_token (`str`, *optional*, defaults to `"[MASK]"`):81 The token used for masking values. This is the token used when training this model with masked language82 modeling. This is the token which the model will try to predict.83 tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):84 Whether or not to tokenize Chinese characters.85 86 This should likely be deactivated for Japanese (see this87 [issue](https://github.com/huggingface/transformers/issues/328)).88 strip_accents (`bool`, *optional*):89 Whether or not to strip all accents. If this option is not specified, then it will be determined by the90 value for `lowercase` (as in the original BERT).91 clean_up_tokenization_spaces (`bool`, *optional*, defaults to `True`):92 Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like93 extra spaces.94 """95 96 vocab_files_names = VOCAB_FILES_NAMES97 98 def __init__(99 self,100 vocab_file,101 do_lower_case=True,102 do_basic_tokenize=True,103 never_split=None,104 unk_token="[UNK]",105 sep_token="[SEP]",106 pad_token="[PAD]",107 cls_token="[CLS]",108 mask_token="[MASK]",109 tokenize_chinese_chars=True,110 strip_accents=None,111 clean_up_tokenization_spaces=True,112 **kwargs,113 ):114 if not os.path.isfile(vocab_file):115 raise ValueError(116 f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google pretrained"117 " model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`"118 )119 self.vocab = load_vocab(vocab_file)120 self.ids_to_tokens = collections.OrderedDict([(ids, tok) for tok, ids in self.vocab.items()])121 self.do_basic_tokenize = do_basic_tokenize122 if do_basic_tokenize:123 self.basic_tokenizer = BasicTokenizer(124 do_lower_case=do_lower_case,125 never_split=never_split,126 tokenize_chinese_chars=tokenize_chinese_chars,127 strip_accents=strip_accents,128 )129 130 self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=str(unk_token))131 132 super().__init__(133 do_lower_case=do_lower_case,134 do_basic_tokenize=do_basic_tokenize,135 never_split=never_split,136 unk_token=unk_token,137 sep_token=sep_token,138 pad_token=pad_token,139 cls_token=cls_token,140 mask_token=mask_token,141 tokenize_chinese_chars=tokenize_chinese_chars,142 strip_accents=strip_accents,143 clean_up_tokenization_spaces=clean_up_tokenization_spaces,144 **kwargs,145 )146 147 @property148 def do_lower_case(self):149 return self.basic_tokenizer.do_lower_case150 151 @property152 def vocab_size(self):153 return len(self.vocab)154 155 def get_vocab(self):156 return dict(self.vocab, **self.added_tokens_encoder)157 158 def _tokenize(self, text, split_special_tokens=False):159 split_tokens = []160 if self.do_basic_tokenize:161 for token in self.basic_tokenizer.tokenize(162 text, never_split=self.all_special_tokens if not split_special_tokens else None163 ):164 # If the token is part of the never_split set165 if token in self.basic_tokenizer.never_split:166 split_tokens.append(token)167 else:168 split_tokens += self.wordpiece_tokenizer.tokenize(token)169 else:170 split_tokens = self.wordpiece_tokenizer.tokenize(text)171 return split_tokens172 173 def _convert_token_to_id(self, token):174 """Converts a token (str) in an id using the vocab."""175 return self.vocab.get(token, self.vocab.get(self.unk_token))176 177 def _convert_id_to_token(self, index):178 """Converts an index (integer) in a token (str) using the vocab."""179 return self.ids_to_tokens.get(index, self.unk_token)180 181 def convert_tokens_to_string(self, tokens):182 """Converts a sequence of tokens (string) in a single string."""183 out_string = " ".join(tokens).replace(" ##", "").strip()184 return out_string185 186 def build_inputs_with_special_tokens(187 self, token_ids_0: list[int], token_ids_1: Optional[list[int]] = None188 ) -> list[int]:189 """190 Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and191 adding special tokens. A BERT sequence has the following format:192 193 - single sequence: `[CLS] X [SEP]`194 - pair of sequences: `[CLS] A [SEP] B [SEP]`195 196 Args:197 token_ids_0 (`List[int]`):198 List of IDs to which the special tokens will be added.199 token_ids_1 (`List[int]`, *optional*):200 Optional second list of IDs for sequence pairs.201 202 Returns:203 `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.204 """205 if token_ids_1 is None:206 return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]207 cls = [self.cls_token_id]208 sep = [self.sep_token_id]209 return cls + token_ids_0 + sep + token_ids_1 + sep210 211 def get_special_tokens_mask(212 self, token_ids_0: list[int], token_ids_1: Optional[list[int]] = None, already_has_special_tokens: bool = False213 ) -> list[int]:214 """215 Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding216 special tokens using the tokenizer `prepare_for_model` method.217 218 Args:219 token_ids_0 (`List[int]`):220 List of IDs.221 token_ids_1 (`List[int]`, *optional*):222 Optional second list of IDs for sequence pairs.223 already_has_special_tokens (`bool`, *optional*, defaults to `False`):224 Whether or not the token list is already formatted with special tokens for the model.225 226 Returns:227 `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.228 """229 230 if already_has_special_tokens:231 return super().get_special_tokens_mask(232 token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True233 )234 235 if token_ids_1 is not None:236 return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]237 return [1] + ([0] * len(token_ids_0)) + [1]238 239 def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple[str]:240 index = 0241 if os.path.isdir(save_directory):242 vocab_file = os.path.join(243 save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]244 )245 else:246 vocab_file = (filename_prefix + "-" if filename_prefix else "") + save_directory247 with open(vocab_file, "w", encoding="utf-8") as writer:248 for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):249 if index != token_index:250 logger.warning(251 f"Saving vocabulary to {vocab_file}: vocabulary indices are not consecutive."252 " Please check that the vocabulary is not corrupted!"253 )254 index = token_index255 writer.write(token + "\n")256 index += 1257 return (vocab_file,)258 259 260class BasicTokenizer:261 """262 Constructs a BasicTokenizer that will run basic tokenization (punctuation splitting, lower casing, etc.).263 264 Args:265 do_lower_case (`bool`, *optional*, defaults to `True`):266 Whether or not to lowercase the input when tokenizing.267 never_split (`Iterable`, *optional*):268 Collection of tokens which will never be split during tokenization. Only has an effect when269 `do_basic_tokenize=True`270 tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):271 Whether or not to tokenize Chinese characters.272 273 This should likely be deactivated for Japanese (see this274 [issue](https://github.com/huggingface/transformers/issues/328)).275 strip_accents (`bool`, *optional*):276 Whether or not to strip all accents. If this option is not specified, then it will be determined by the277 value for `lowercase` (as in the original BERT).278 do_split_on_punc (`bool`, *optional*, defaults to `True`):279 In some instances we want to skip the basic punctuation splitting so that later tokenization can capture280 the full context of the words, such as contractions.281 """282 283 def __init__(284 self,285 do_lower_case=True,286 never_split=None,287 tokenize_chinese_chars=True,288 strip_accents=None,289 do_split_on_punc=True,290 ):291 if never_split is None:292 never_split = []293 self.do_lower_case = do_lower_case294 self.never_split = set(never_split)295 self.tokenize_chinese_chars = tokenize_chinese_chars296 self.strip_accents = strip_accents297 self.do_split_on_punc = do_split_on_punc298 299 def tokenize(self, text, never_split=None):300 """301 Basic Tokenization of a piece of text. For sub-word tokenization, see WordPieceTokenizer.302 303 Args:304 never_split (`List[str]`, *optional*)305 Kept for backward compatibility purposes. Now implemented directly at the base class level (see306 [`PreTrainedTokenizer.tokenize`]) List of token not to split.307 """308 # union() returns a new set by concatenating the two sets.309 never_split = self.never_split.union(set(never_split)) if never_split else self.never_split310 text = self._clean_text(text)311 312 # This was added on November 1st, 2018 for the multilingual and Chinese313 # models. This is also applied to the English models now, but it doesn't314 # matter since the English models were not trained on any Chinese data315 # and generally don't have any Chinese data in them (there are Chinese316 # characters in the vocabulary because Wikipedia does have some Chinese317 # words in the English Wikipedia.).318 if self.tokenize_chinese_chars:319 text = self._tokenize_chinese_chars(text)320 # prevents treating the same character with different unicode codepoints as different characters321 unicode_normalized_text = unicodedata.normalize("NFC", text)322 orig_tokens = whitespace_tokenize(unicode_normalized_text)323 split_tokens = []324 for token in orig_tokens:325 if token not in never_split:326 if self.do_lower_case:327 token = token.lower()328 if self.strip_accents is not False:329 token = self._run_strip_accents(token)330 elif self.strip_accents:331 token = self._run_strip_accents(token)332 split_tokens.extend(self._run_split_on_punc(token, never_split))333 334 output_tokens = whitespace_tokenize(" ".join(split_tokens))335 return output_tokens336 337 def _run_strip_accents(self, text):338 """Strips accents from a piece of text."""339 text = unicodedata.normalize("NFD", text)340 output = []341 for char in text:342 cat = unicodedata.category(char)343 if cat == "Mn":344 continue345 output.append(char)346 return "".join(output)347 348 def _run_split_on_punc(self, text, never_split=None):349 """Splits punctuation on a piece of text."""350 if not self.do_split_on_punc or (never_split is not None and text in never_split):351 return [text]352 chars = list(text)353 i = 0354 start_new_word = True355 output = []356 while i < len(chars):357 char = chars[i]358 if _is_punctuation(char):359 output.append([char])360 start_new_word = True361 else:362 if start_new_word:363 output.append([])364 start_new_word = False365 output[-1].append(char)366 i += 1367 368 return ["".join(x) for x in output]369 370 def _tokenize_chinese_chars(self, text):371 """Adds whitespace around any CJK character."""372 output = []373 for char in text:374 cp = ord(char)375 if self._is_chinese_char(cp):376 output.append(" ")377 output.append(char)378 output.append(" ")379 else:380 output.append(char)381 return "".join(output)382 383 def _is_chinese_char(self, cp):384 """Checks whether CP is the codepoint of a CJK character."""385 # This defines a "chinese character" as anything in the CJK Unicode block:386 # https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)387 #388 # Note that the CJK Unicode block is NOT all Japanese and Korean characters,389 # despite its name. The modern Korean Hangul alphabet is a different block,390 # as is Japanese Hiragana and Katakana. Those alphabets are used to write391 # space-separated words, so they are not treated specially and handled392 # like the all of the other languages.393 if (394 (cp >= 0x4E00 and cp <= 0x9FFF)395 or (cp >= 0x3400 and cp <= 0x4DBF)396 or (cp >= 0x20000 and cp <= 0x2A6DF)397 or (cp >= 0x2A700 and cp <= 0x2B73F)398 or (cp >= 0x2B740 and cp <= 0x2B81F)399 or (cp >= 0x2B820 and cp <= 0x2CEAF)400 or (cp >= 0xF900 and cp <= 0xFAFF)401 or (cp >= 0x2F800 and cp <= 0x2FA1F)402 ):403 return True404 405 return False406 407 def _clean_text(self, text):408 """Performs invalid character removal and whitespace cleanup on text."""409 output = []410 for char in text:411 cp = ord(char)412 if cp == 0 or cp == 0xFFFD or _is_control(char):413 continue414 if _is_whitespace(char):415 output.append(" ")416 else:417 output.append(char)418 return "".join(output)419 420 421class WordpieceTokenizer:422 """Runs WordPiece tokenization."""423 424 def __init__(self, vocab, unk_token, max_input_chars_per_word=100):425 self.vocab = vocab426 self.unk_token = unk_token427 self.max_input_chars_per_word = max_input_chars_per_word428 429 def tokenize(self, text):430 """431 Tokenizes a piece of text into its word pieces. This uses a greedy longest-match-first algorithm to perform432 tokenization using the given vocabulary.433 434 For example, `input = "unaffable"` will return as output `["un", "##aff", "##able"]`.435 436 Args:437 text: A single token or whitespace separated tokens. This should have438 already been passed through *BasicTokenizer*.439 440 Returns:441 A list of wordpiece tokens.442 """443 444 output_tokens = []445 for token in whitespace_tokenize(text):446 chars = list(token)447 if len(chars) > self.max_input_chars_per_word:448 output_tokens.append(self.unk_token)449 continue450 451 is_bad = False452 start = 0453 sub_tokens = []454 while start < len(chars):455 end = len(chars)456 cur_substr = None457 while start < end:458 substr = "".join(chars[start:end])459 if start > 0:460 substr = "##" + substr461 if substr in self.vocab:462 cur_substr = substr463 break464 end -= 1465 if cur_substr is None:466 is_bad = True467 break468 sub_tokens.append(cur_substr)469 start = end470 471 if is_bad:472 output_tokens.append(self.unk_token)473 else:474 output_tokens.extend(sub_tokens)475 return output_tokens476 477 478__all__ = ["BasicTokenizer", "BertTokenizer", "WordpieceTokenizer"]479