Upload tokenizer

fairseq_dictionary.py

@@ -0,0 +1,504 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is ported from fairseq:
+# https://github.com/facebookresearch/fairseq
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of the fairseq repo
+
+
+import os
+import re
+from collections import Counter
+from multiprocessing import Pool
+from typing import Iterable, List
+
+import torch
+
+
+def item(tensor):
+    # tpu-comment: making this a no-op for xla devices.
+    if torch.is_tensor(tensor) and tensor.device.type == "xla":
+        return tensor.detach()
+    if hasattr(tensor, "item"):
+        return tensor.item()
+    if hasattr(tensor, "__getitem__"):
+        return tensor[0]
+    return tensor
+
+
+def post_process(sentence: str, symbol: str):
+    if symbol == "sentencepiece":
+        sentence = sentence.replace(" ", "").replace("\u2581", " ").strip()
+    elif symbol == "wordpiece":
+        sentence = sentence.replace(" ", "").replace("_", " ").strip()
+    elif symbol == "letter":
+        sentence = sentence.replace(" ", "").replace("|", " ").strip()
+    elif symbol == "silence":
+        import re
+
+        sentence = sentence.replace("<SIL>", "")
+        sentence = re.sub(" +", " ", sentence).strip()
+    elif symbol == "_EOW":
+        sentence = sentence.replace(" ", "").replace("_EOW", " ").strip()
+    elif symbol in {"subword_nmt", "@@ ", "@@"}:
+        if symbol == "subword_nmt":
+            symbol = "@@ "
+        sentence = (sentence + " ").replace(symbol, "").rstrip()
+    elif symbol == "none":
+        pass
+    elif symbol is not None:
+        raise NotImplementedError(f"Unknown post_process option: {symbol}")
+    return sentence
+
+
+SPACE_NORMALIZER = re.compile(r"\s+")
+
+
+def tokenize_line(line):
+    line = SPACE_NORMALIZER.sub(" ", line)
+    line = line.strip()
+    return line.split()
+
+
+def _safe_readline(fd) -> str:
+    pos = fd.tell()
+    while True:
+        try:
+            return fd.readline()
+        except UnicodeDecodeError:
+            pos -= 1
+            fd.seek(pos)  # search where this character begins
+
+
+def find_offsets(filename: str, num_chunks: int) -> List[int]:
+    """
+    given a file and a number of chuncks, find the offsets in the file
+    to be able to chunk around full lines.
+    """
+    with open(filename, "r", encoding="utf-8") as f:
+        size = os.fstat(f.fileno()).st_size
+        chunk_size = size // num_chunks
+        offsets = [0 for _ in range(num_chunks + 1)]
+        for i in range(1, num_chunks):
+            f.seek(chunk_size * i)
+            _safe_readline(f)
+            offsets[i] = f.tell()
+        offsets[-1] = size
+        return offsets
+
+
+class ChunkLineIterator:
+    """
+    Iterator to properly iterate over lines of a file chunck.
+    """
+
+    def __init__(self, fd, start_offset: int, end_offset: int):
+        self._fd = fd
+        self._start_offset = start_offset
+        self._end_offset = end_offset
+
+    def __iter__(self) -> Iterable[str]:
+        self._fd.seek(self._start_offset)
+        # next(f) breaks f.tell(), hence readline() must be used
+        line = _safe_readline(self._fd)
+        while line:
+            pos = self._fd.tell()
+            # f.tell() does not always give the byte position in the file
+            # sometimes it skips to a very large number
+            # it is unlikely that through a normal read we go from
+            # end bytes to end + 2**32 bytes (4 GB) and this makes it unlikely
+            # that the procedure breaks by the undeterministic behavior of
+            # f.tell()
+            if (
+                self._end_offset > 0
+                and pos > self._end_offset
+                and pos < self._end_offset + 2 ** 32
+            ):
+                break
+            yield line
+            line = self._fd.readline()
+
+
+class Chunker:
+    """
+    contextmanager to read a chunck of a file line by line.
+    """
+
+    def __init__(self, path: str, start_offset: int, end_offset: int):
+        self.path = path
+        self.start_offset = start_offset
+        self.end_offset = end_offset
+
+    def __enter__(self) -> ChunkLineIterator:
+        self.fd = open(self.path, "r", encoding="utf-8")
+        return ChunkLineIterator(self.fd, self.start_offset, self.end_offset)
+
+    def __exit__(self, exc_type, exc_val, exc_tb) -> None:
+        self.fd.close()
+
+
+class Dictionary:
+    """A mapping from symbols to consecutive integers"""
+
+    def __init__(
+        self,
+        *,  # begin keyword-only arguments
+        bos="<s>",
+        pad="<pad>",
+        eos="</s>",
+        unk="<unk>",
+        extra_special_symbols=None,
+    ):
+        self.bos_word, self.unk_word, self.pad_word, self.eos_word = bos, unk, pad, eos
+        self.symbols = []
+        self.count = []
+        self.indices = {}
+        self.bos_index = self.add_symbol(bos)
+        self.pad_index = self.add_symbol(pad)
+        self.eos_index = self.add_symbol(eos)
+        self.unk_index = self.add_symbol(unk)
+        if extra_special_symbols:
+            for s in extra_special_symbols:
+                self.add_symbol(s)
+        self.nspecial = len(self.symbols)
+
+    def __eq__(self, other):
+        return self.indices == other.indices
+
+    def __getitem__(self, idx):
+        if idx < len(self.symbols):
+            return self.symbols[idx]
+        return self.unk_word
+
+    def get_count(self, idx):
+        return self.count[idx]
+
+    def __len__(self):
+        """Returns the number of symbols in the dictionary"""
+        return len(self.symbols)
+
+    def __contains__(self, sym):
+        return sym in self.indices
+
+    def index(self, sym):
+        """Returns the index of the specified symbol"""
+        assert isinstance(sym, str)
+        if sym in self.indices:
+            return self.indices[sym]
+        return self.unk_index
+
+    def string(
+        self,
+        tensor,
+        bpe_symbol=None,
+        escape_unk=False,
+        extra_symbols_to_ignore=None,
+        unk_string=None,
+        include_eos=False,
+        separator=" ",
+    ):
+        """Helper for converting a tensor of token indices to a string.
+
+        Can optionally remove BPE symbols or escape <unk> words.
+        """
+        if torch.is_tensor(tensor) and tensor.dim() == 2:
+            return "\n".join(
+                self.string(
+                    t,
+                    bpe_symbol,
+                    escape_unk,
+                    extra_symbols_to_ignore,
+                    include_eos=include_eos,
+                )
+                for t in tensor
+            )
+
+        extra_symbols_to_ignore = set(extra_symbols_to_ignore or [])
+        if not include_eos:
+            extra_symbols_to_ignore.add(self.eos())
+
+        def token_string(i):
+            if i == self.unk():
+                if unk_string is not None:
+                    return unk_string
+                else:
+                    return self.unk_string(escape_unk)
+            else:
+                return self[i]
+
+        if hasattr(self, "bos_index"):
+            extra_symbols_to_ignore.add(self.bos())
+
+        sent = separator.join(
+            token_string(i)
+            for i in tensor
+            if item(i) not in extra_symbols_to_ignore
+        )
+
+        return post_process(sent, bpe_symbol)
+
+    def unk_string(self, escape=False):
+        """Return unknown string, optionally escaped as: <<unk>>"""
+        if escape:
+            return "<{}>".format(self.unk_word)
+        else:
+            return self.unk_word
+
+    def add_symbol(self, word, n=1, overwrite=False):
+        """Adds a word to the dictionary"""
+        if word in self.indices and not overwrite:
+            idx = self.indices[word]
+            self.count[idx] = self.count[idx] + n
+            return idx
+        else:
+            idx = len(self.symbols)
+            self.indices[word] = idx
+            self.symbols.append(word)
+            self.count.append(n)
+            return idx
+
+    def update(self, new_dict):
+        """Updates counts from new dictionary."""
+        for word in new_dict.symbols:
+            idx2 = new_dict.indices[word]
+            if word in self.indices:
+                idx = self.indices[word]
+                self.count[idx] = self.count[idx] + new_dict.count[idx2]
+            else:
+                idx = len(self.symbols)
+                self.indices[word] = idx
+                self.symbols.append(word)
+                self.count.append(new_dict.count[idx2])
+
+    def finalize(self, threshold=-1, nwords=-1, padding_factor=8):
+        """Sort symbols by frequency in descending order, ignoring special ones.
+
+        Args:
+            - threshold defines the minimum word count
+            - nwords defines the total number of words in the final dictionary,
+                including special symbols
+            - padding_factor can be used to pad the dictionary size to be a
+                multiple of 8, which is important on some hardware (e.g., Nvidia
+                Tensor Cores).
+        """
+        if nwords <= 0:
+            nwords = len(self)
+
+        new_indices = dict(zip(self.symbols[: self.nspecial], range(self.nspecial)))
+        new_symbols = self.symbols[: self.nspecial]
+        new_count = self.count[: self.nspecial]
+
+        c = Counter(
+            dict(
+                sorted(zip(self.symbols[self.nspecial:], self.count[self.nspecial:]))
+            )
+        )
+        for symbol, count in c.most_common(nwords - self.nspecial):
+            if count >= threshold:
+                new_indices[symbol] = len(new_symbols)
+                new_symbols.append(symbol)
+                new_count.append(count)
+            else:
+                break
+
+        assert len(new_symbols) == len(new_indices)
+
+        self.count = list(new_count)
+        self.symbols = list(new_symbols)
+        self.indices = new_indices
+
+        self.pad_to_multiple_(padding_factor)
+
+    def pad_to_multiple_(self, padding_factor):
+        """Pad Dictionary size to be a multiple of *padding_factor*."""
+        if padding_factor > 1:
+            i = 0
+            while len(self) % padding_factor != 0:
+                symbol = "madeupword{:04d}".format(i)
+                self.add_symbol(symbol, n=0)
+                i += 1
+
+    def bos(self):
+        """Helper to get index of beginning-of-sentence symbol"""
+        return self.bos_index
+
+    def pad(self):
+        """Helper to get index of pad symbol"""
+        return self.pad_index
+
+    def eos(self):
+        """Helper to get index of end-of-sentence symbol"""
+        return self.eos_index
+
+    def unk(self):
+        """Helper to get index of unk symbol"""
+        return self.unk_index
+
+    @classmethod
+    def load(cls, f):
+        """Loads the dictionary from a text file with the format:
+
+        ```
+        <symbol0> <count0>
+        <symbol1> <count1>
+        ...
+        ```
+        """
+        d = cls()
+        d.add_from_file(f)
+        return d
+
+    def add_from_file(self, f):
+        """
+        Loads a pre-existing dictionary from a text file and adds its symbols
+        to this instance.
+        """
+        if isinstance(f, str):
+            try:
+                with open(f, "r", encoding="utf-8") as fd:
+                    self.add_from_file(fd)
+            except FileNotFoundError as fnfe:
+                raise fnfe
+            except UnicodeError:
+                raise Exception(
+                    "Incorrect encoding detected in {}, please "
+                    "rebuild the dataset".format(f)
+                )
+            return
+
+        lines = f.readlines()
+        indices_start_line = self._load_meta(lines)
+
+        for line in lines[indices_start_line:]:
+            try:
+                line, field = line.rstrip().rsplit(" ", 1)
+                if field == "#fairseq:overwrite":
+                    overwrite = True
+                    line, field = line.rsplit(" ", 1)
+                else:
+                    overwrite = False
+                count = int(field)
+                word = line
+                if word in self and not overwrite:
+                    raise RuntimeError(
+                        "Duplicate word found when loading Dictionary: '{}'. "
+                        "Duplicate words can overwrite earlier ones by adding the "
+                        "#fairseq:overwrite flag at the end of the corresponding row "
+                        "in the dictionary file. If using the Camembert model, please "
+                        "download an updated copy of the model file.".format(word)
+                    )
+                self.add_symbol(word, n=count, overwrite=overwrite)
+            except ValueError:
+                raise ValueError(
+                    f"Incorrect dictionary format, expected '<token> <cnt> [flags]': \"{line}\""
+                )
+
+    def _save(self, f, kv_iterator):
+        if isinstance(f, str):
+            os.makedirs(os.path.dirname(f), exist_ok=True)
+            with open(f, "w", encoding="utf-8") as fd:
+                return self.save(fd)
+        for k, v in kv_iterator:
+            print("{} {}".format(k, v), file=f)
+
+    def _get_meta(self):
+        return [], []
+
+    def _load_meta(self, lines):
+        return 0
+
+    def save(self, f):
+        """Stores dictionary into a text file"""
+        ex_keys, ex_vals = self._get_meta()
+        self._save(
+            f,
+            zip(
+                ex_keys + self.symbols[self.nspecial:],
+                ex_vals + self.count[self.nspecial:],
+            ),
+        )
+
+    def dummy_sentence(self, length):
+        t = torch.Tensor(length).uniform_(self.nspecial + 1, len(self)).long()
+        t[-1] = self.eos()
+        return t
+
+    def encode_line(
+        self,
+        line,
+        line_tokenizer=tokenize_line,
+        add_if_not_exist=True,
+        consumer=None,
+        append_eos=True,
+        reverse_order=False,
+    ) -> torch.IntTensor:
+        words = line_tokenizer(line)
+        if reverse_order:
+            words = list(reversed(words))
+        nwords = len(words)
+        ids = torch.IntTensor(nwords + 1 if append_eos else nwords)
+
+        for i, word in enumerate(words):
+            if add_if_not_exist:
+                idx = self.add_symbol(word)
+            else:
+                idx = self.index(word)
+            if consumer is not None:
+                consumer(word, idx)
+            ids[i] = idx
+        if append_eos:
+            ids[nwords] = self.eos_index
+        return ids
+
+    @staticmethod
+    def _add_file_to_dictionary_single_worker(
+        filename,
+        tokenize,
+        eos_word,
+        start_offset,
+        end_offset,
+    ):
+        counter = Counter()
+        with Chunker(filename, start_offset, end_offset) as line_iterator:
+            for line in line_iterator:
+                for word in tokenize(line):
+                    counter.update([word])
+                counter.update([eos_word])
+        return counter
+
+    @staticmethod
+    def add_file_to_dictionary(filename, dict, tokenize, num_workers):
+        def merge_result(counter):
+            for w, c in sorted(counter.items()):
+                dict.add_symbol(w, c)
+
+        local_file = filename
+        offsets = find_offsets(local_file, num_workers)
+        if num_workers > 1:
+            chunks = zip(offsets, offsets[1:])
+            pool = Pool(processes=num_workers)
+            results = []
+            for (start_offset, end_offset) in chunks:
+                results.append(
+                    pool.apply_async(
+                        Dictionary._add_file_to_dictionary_single_worker,
+                        (
+                            local_file,
+                            tokenize,
+                            dict.eos_word,
+                            start_offset,
+                            end_offset,
+                        ),
+                    )
+                )
+            pool.close()
+            pool.join()
+            for r in results:
+                merge_result(r.get())
+        else:
+            merge_result(
+                Dictionary._add_file_to_dictionary_single_worker(
+                    local_file, tokenize, dict.eos_word, offsets[0], offsets[1]
+                )
+            )

special_tokens_map.json

@@ -0,0 +1,8 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

tokenization_ast_t5.py

@@ -0,0 +1,178 @@
+import os
+import re
+from shutil import copyfile
+from typing import List, Optional
+
+import tiktoken
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.utils import logging
+
+from .fairseq_dictionary import Dictionary
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "dict_path": "dict.txt"
+}
+
+GPT4ENC = tiktoken.encoding_for_model("gpt-4")
+
+
+class GPT4Dictionary:
+    def __init__(self):
+        self.vocab = {}
+        self.words = {}
+        self.vocab_cnt = 4
+        for i in range(GPT4ENC.n_vocab):
+            try:
+                w = GPT4ENC.decode_single_token_bytes(i)
+                self.vocab[w] = self.vocab_cnt
+                self.words[self.vocab_cnt] = w
+                self.vocab_cnt += 1
+            except KeyError:
+                pass
+        self.eos_index = 2
+        self.words[2] = b"</s>"
+        self.sentinel_start = self.vocab_cnt
+        for i in range(1000):
+            self.words[self.sentinel_start + i] = f"<sen{i:03d}>".encode("utf-8")
+
+    def index(self, w):
+        assert w in self.vocab
+        return self.vocab[w]
+
+    def __getitem__(self, i):
+        if i in self.words:
+            return self.words[i]
+        else:
+            return b""
+
+
+class ASTT5Tokenizer(PreTrainedTokenizer):
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        dict_path,
+        n_sentinel_tokens=0,
+        bos_token="<s>",
+        eos_token="</s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        **kwargs
+    ) -> None:
+
+        self.dict_path = dict_path
+        self.tik_dict = GPT4Dictionary()
+        self.fs_dict = Dictionary.load(dict_path)
+        self.fs_dict_sentinel_start = len(self.fs_dict)
+        for i in range(n_sentinel_tokens):
+            self.fs_dict.add_symbol(f'<sen{i:03d}>')
+
+        if "sep_token" in kwargs:
+            assert kwargs["sep_token"] == eos_token
+            kwargs.pop("sep_token")
+        if "cls_token" in kwargs:
+            assert kwargs["cls_token"] == bos_token
+            kwargs.pop("cls_token")
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            sep_token=eos_token,
+            cls_token=bos_token,
+            n_sentinel_tokens=n_sentinel_tokens,
+            **kwargs, )
+
+    @property
+    def vocab_size(self):
+        return len(self.fs_dict)
+
+    def get_vocab(self):
+        return self.fs_dict.indices
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        mask_0 = [(0 if w < self.fs_dict_sentinel_start else 1) for w in token_ids_0]
+        mask_1 = [(0 if w < self.fs_dict_sentinel_start else 1) for w in token_ids_1]
+        if token_ids_1 is None:
+            return mask_0 + [1]
+        return mask_0 + [1] + mask_1 + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        sep = [self.sep_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + sep) * [0]
+        return len(token_ids_0 + sep + token_ids_1 + sep) * [0]
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        if token_ids_1 is None:
+            return token_ids_0 + [self.sep_token_id]
+        sep = [self.sep_token_id]
+        return token_ids_0 + sep + token_ids_1 + sep
+
+    def _tokenize(self, text: str) -> List[str]:
+        parts = re.split(r"(<sen\d+>)", text)
+        tokenized = []
+        for part in parts:
+            if re.match(r"<sen\d+>", part):
+                tokenized.append(part)
+            else:
+                tokenized.extend(
+                    [
+                        self.fs_dict[self.tik_dict.index(w)]
+                        for w in GPT4ENC.decode_tokens_bytes(GPT4ENC.encode_ordinary(part))
+                    ]
+                )
+        return tokenized
+
+    def _convert_token_to_id(self, token):
+        return self.fs_dict.index(token)
+
+    def _convert_id_to_token(self, index):
+        return self.fs_dict[index]
+
+    def convert_tokens_to_string(self, tokens):
+        token_bytes = b"".join([self.tik_dict[self.fs_dict.index(token)] for token in tokens])
+        return token_bytes.decode("utf-8", errors="ignore")
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None):
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_dict_path = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["dict_path"]
+        )
+
+        if os.path.abspath(self.dict_path) != os.path.abspath(out_dict_path):
+            copyfile(self.dict_path, out_dict_path)
+            logger.info(f"Copy from {self.dict_path} to {out_dict_path}")
+
+        return (out_dict_path,)

tokenizer_config.json

@@ -0,0 +1,52 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_ast_t5.ASTT5Tokenizer",
+      null
+    ]
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "model_max_length": 1000000000000000019884624838656,
+  "n_sentinel_tokens": 1000,
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "tokenizer_class": "ASTT5Tokenizer",
+  "unk_token": "<unk>"
+}