Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2020 The Facebook AI Research 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 16import json17import os18from functools import lru_cache19from typing import Optional20 21import regex as re22 23from ...tokenization_utils import AddedToken, PreTrainedTokenizer24from ...utils import logging25 26 27logger = logging.get_logger(__name__)28 29 30VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt"}31 32# See all BART models at https://huggingface.co/models?filter=bart33 34 35@lru_cache36def bytes_to_unicode():37    """38    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control39    characters the bpe code barfs on.40 41    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab42    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for43    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup44    tables between utf-8 bytes and unicode strings.45    """46    bs = (47        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))48    )49    cs = bs[:]50    n = 051    for b in range(2**8):52        if b not in bs:53            bs.append(b)54            cs.append(2**8 + n)55            n += 156    cs = [chr(n) for n in cs]57    return dict(zip(bs, cs))58 59 60def get_pairs(word):61    """62    Return set of symbol pairs in a word.63 64    Word is represented as tuple of symbols (symbols being variable-length strings).65    """66    pairs = set()67    prev_char = word[0]68    for char in word[1:]:69        pairs.add((prev_char, char))70        prev_char = char71    return pairs72 73 74class BartTokenizer(PreTrainedTokenizer):75    """76    Constructs a BART tokenizer, which is smilar to the ROBERTa tokenizer, using byte-level Byte-Pair-Encoding.77 78    This tokenizer has been trained to treat spaces like parts of the tokens (a bit like sentencepiece) so a word will79    be encoded differently whether it is at the beginning of the sentence (without space) or not:80 81    ```python82    >>> from transformers import BartTokenizer83 84    >>> tokenizer = BartTokenizer.from_pretrained("facebook/bart-base")85    >>> tokenizer("Hello world")["input_ids"]86    [0, 31414, 232, 2]87 88    >>> tokenizer(" Hello world")["input_ids"]89    [0, 20920, 232, 2]90    ```91 92    You can get around that behavior by passing `add_prefix_space=True` when instantiating this tokenizer or when you93    call it on some text, but since the model was not pretrained this way, it might yield a decrease in performance.94 95    <Tip>96 97    When used with `is_split_into_words=True`, this tokenizer will add a space before each word (even the first one).98 99    </Tip>100 101    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to102    this superclass for more information regarding those methods.103 104    Args:105        vocab_file (`str`):106            Path to the vocabulary file.107        merges_file (`str`):108            Path to the merges file.109        errors (`str`, *optional*, defaults to `"replace"`):110            Paradigm to follow when decoding bytes to UTF-8. See111            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.112        bos_token (`str`, *optional*, defaults to `"<s>"`):113            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.114 115            <Tip>116 117            When building a sequence using special tokens, this is not the token that is used for the beginning of118            sequence. The token used is the `cls_token`.119 120            </Tip>121 122        eos_token (`str`, *optional*, defaults to `"</s>"`):123            The end of sequence token.124 125            <Tip>126 127            When building a sequence using special tokens, this is not the token that is used for the end of sequence.128            The token used is the `sep_token`.129 130            </Tip>131 132        sep_token (`str`, *optional*, defaults to `"</s>"`):133            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for134            sequence classification or for a text and a question for question answering. It is also used as the last135            token of a sequence built with special tokens.136        cls_token (`str`, *optional*, defaults to `"<s>"`):137            The classifier token which is used when doing sequence classification (classification of the whole sequence138            instead of per-token classification). It is the first token of the sequence when built with special tokens.139        unk_token (`str`, *optional*, defaults to `"<unk>"`):140            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this141            token instead.142        pad_token (`str`, *optional*, defaults to `"<pad>"`):143            The token used for padding, for example when batching sequences of different lengths.144        mask_token (`str`, *optional*, defaults to `"<mask>"`):145            The token used for masking values. This is the token used when training this model with masked language146            modeling. This is the token which the model will try to predict.147        add_prefix_space (`bool`, *optional*, defaults to `False`):148            Whether or not to add an initial space to the input. This allows to treat the leading word just as any149            other word. (BART tokenizer detect beginning of words by the preceding space).150    """151 152    vocab_files_names = VOCAB_FILES_NAMES153    model_input_names = ["input_ids", "attention_mask"]154 155    def __init__(156        self,157        vocab_file,158        merges_file,159        errors="replace",160        bos_token="<s>",161        eos_token="</s>",162        sep_token="</s>",163        cls_token="<s>",164        unk_token="<unk>",165        pad_token="<pad>",166        mask_token="<mask>",167        add_prefix_space=False,168        **kwargs,169    ):170        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token171        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token172        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token173        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token174        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token175        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token176 177        # Mask token behave like a normal word, i.e. include the space before it178        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token179 180        with open(vocab_file, encoding="utf-8") as vocab_handle:181            self.encoder = json.load(vocab_handle)182        self.decoder = {v: k for k, v in self.encoder.items()}183        self.errors = errors  # how to handle errors in decoding184        self.byte_encoder = bytes_to_unicode()185        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}186        with open(merges_file, encoding="utf-8") as merges_handle:187            bpe_merges = merges_handle.read().split("\n")[1:-1]188        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]189        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))190        self.cache = {}191        self.add_prefix_space = add_prefix_space192 193        # Should have added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions194        self.pat = re.compile(r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")195 196        super().__init__(197            errors=errors,198            bos_token=bos_token,199            eos_token=eos_token,200            unk_token=unk_token,201            sep_token=sep_token,202            cls_token=cls_token,203            pad_token=pad_token,204            mask_token=mask_token,205            add_prefix_space=add_prefix_space,206            **kwargs,207        )208 209    @property210    def vocab_size(self):211        return len(self.encoder)212 213    def get_vocab(self):214        return dict(self.encoder, **self.added_tokens_encoder)215 216    def bpe(self, token):217        if token in self.cache:218            return self.cache[token]219        word = tuple(token)220        pairs = get_pairs(word)221 222        if not pairs:223            return token224 225        while True:226            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))227            if bigram not in self.bpe_ranks:228                break229            first, second = bigram230            new_word = []231            i = 0232            while i < len(word):233                try:234                    j = word.index(first, i)235                except ValueError:236                    new_word.extend(word[i:])237                    break238                else:239                    new_word.extend(word[i:j])240                    i = j241 242                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:243                    new_word.append(first + second)244                    i += 2245                else:246                    new_word.append(word[i])247                    i += 1248            new_word = tuple(new_word)249            word = new_word250            if len(word) == 1:251                break252            else:253                pairs = get_pairs(word)254        word = " ".join(word)255        self.cache[token] = word256        return word257 258    def _tokenize(self, text):259        """Tokenize a string."""260        bpe_tokens = []261        for token in re.findall(self.pat, text):262            token = "".join(263                self.byte_encoder[b] for b in token.encode("utf-8")264            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)265            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))266        return bpe_tokens267 268    def _convert_token_to_id(self, token):269        """Converts a token (str) in an id using the vocab."""270        return self.encoder.get(token, self.encoder.get(self.unk_token))271 272    def _convert_id_to_token(self, index):273        """Converts an index (integer) in a token (str) using the vocab."""274        return self.decoder.get(index)275 276    def convert_tokens_to_string(self, tokens):277        """Converts a sequence of tokens (string) in a single string."""278        text = "".join(tokens)279        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)280        return text281 282    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple[str]:283        if not os.path.isdir(save_directory):284            logger.error(f"Vocabulary path ({save_directory}) should be a directory")285            return286        vocab_file = os.path.join(287            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]288        )289        merge_file = os.path.join(290            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]291        )292 293        with open(vocab_file, "w", encoding="utf-8") as f:294            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")295 296        index = 0297        with open(merge_file, "w", encoding="utf-8") as writer:298            writer.write("#version: 0.2\n")299            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):300                if index != token_index:301                    logger.warning(302                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."303                        " Please check that the tokenizer is not corrupted!"304                    )305                    index = token_index306                writer.write(" ".join(bpe_tokens) + "\n")307                index += 1308 309        return vocab_file, merge_file310 311    def build_inputs_with_special_tokens(312        self, token_ids_0: list[int], token_ids_1: Optional[list[int]] = None313    ) -> list[int]:314        """315        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and316        adding special tokens. A BART sequence has the following format:317 318        - single sequence: `<s> X </s>`319        - pair of sequences: `<s> A </s></s> B </s>`320 321        Args:322            token_ids_0 (`list[int]`):323                List of IDs to which the special tokens will be added.324            token_ids_1 (`list[int]`, *optional*):325                Optional second list of IDs for sequence pairs.326 327        Returns:328            `list[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.329        """330        if token_ids_1 is None:331            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]332        cls = [self.cls_token_id]333        sep = [self.sep_token_id]334        return cls + token_ids_0 + sep + sep + token_ids_1 + sep335 336    def get_special_tokens_mask(337        self, token_ids_0: list[int], token_ids_1: Optional[list[int]] = None, already_has_special_tokens: bool = False338    ) -> list[int]:339        """340        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding341        special tokens using the tokenizer `prepare_for_model` method.342 343        Args:344            token_ids_0 (`list[int]`):345                List of IDs.346            token_ids_1 (`list[int]`, *optional*):347                Optional second list of IDs for sequence pairs.348            already_has_special_tokens (`bool`, *optional*, defaults to `False`):349                Whether or not the token list is already formatted with special tokens for the model.350 351        Returns:352            `list[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.353        """354        if already_has_special_tokens:355            return super().get_special_tokens_mask(356                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True357            )358 359        if token_ids_1 is None:360            return [1] + ([0] * len(token_ids_0)) + [1]361        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]362 363    def create_token_type_ids_from_sequences(364        self, token_ids_0: list[int], token_ids_1: Optional[list[int]] = None365    ) -> list[int]:366        """367        Create a mask from the two sequences passed to be used in a sequence-pair classification task. BART does not368        make use of token type ids, therefore a list of zeros is returned.369 370        Args:371            token_ids_0 (`list[int]`):372                List of IDs.373            token_ids_1 (`list[int]`, *optional*):374                Optional second list of IDs for sequence pairs.375 376        Returns:377            `list[int]`: List of zeros.378        """379        sep = [self.sep_token_id]380        cls = [self.cls_token_id]381 382        if token_ids_1 is None:383            return len(cls + token_ids_0 + sep) * [0]384        return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]385 386    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):387        add_prefix_space = kwargs.pop("add_prefix_space", self.add_prefix_space)388        if (is_split_into_words or add_prefix_space) and (len(text) > 0 and not text[0].isspace()):389            text = " " + text390        return (text, kwargs)391 392 393__all__ = ["BartTokenizer"]394