File size: 8,129 Bytes
8e1984a | 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 | import os
import shutil
from typing import List, Optional, Tuple
import sentencepiece as spm
from transformers import PreTrainedTokenizer
class UnModelTokenizer(PreTrainedTokenizer):
"""SentencePiece tokenizer for the SAGE POND UnModel encoder.
The special-token arguments are explicit constructor parameters. This is
important for Transformers releases whose PythonBackend receives tokenizer
configuration values through ``from_pretrained``: values such as
``bos_token`` must not also remain in ``kwargs`` when passed to ``super``.
"""
vocab_files_names = {"vocab_file": "new_bpe.model"}
model_input_names = ["input_ids", "attention_mask"]
def __init__(
self,
vocab_file,
add_bos_token=False,
add_eos_token=False,
model_max_length=600,
pad_token=None,
bos_token=None,
eos_token=None,
unk_token=None,
clean_up_tokenization_spaces=False,
**kwargs,
):
if not vocab_file or not os.path.isfile(vocab_file):
raise FileNotFoundError(
"The tokenizer requires the SentencePiece model file 'new_bpe.model'."
)
self.vocab_file = os.path.abspath(vocab_file)
self.sp_model = spm.SentencePieceProcessor(model_file=self.vocab_file)
self.add_bos_token = bool(add_bos_token)
self.add_eos_token = bool(add_eos_token)
vocab_size = self.sp_model.get_piece_size()
if vocab_size != 40_000:
raise ValueError(
f"Expected a 40,000-piece SentencePiece model, got {vocab_size:,}."
)
def special_id(method_name, fallback):
value = getattr(self.sp_model, method_name)()
return value if value >= 0 else fallback
# These fallbacks match the IDs used by the original training script.
self._sp_pad_id = special_id("pad_id", 0)
self._sp_bos_id = special_id("bos_id", 2)
self._sp_eos_id = special_id("eos_id", 3)
self._sp_unk_id = special_id("unk_id", 1)
self._sp_special_tokens = {
"pad_token": self.sp_model.id_to_piece(self._sp_pad_id),
"bos_token": self.sp_model.id_to_piece(self._sp_bos_id),
"eos_token": self.sp_model.id_to_piece(self._sp_eos_id),
"unk_token": self.sp_model.id_to_piece(self._sp_unk_id),
}
for token_name, token_id in (
("pad_token", self._sp_pad_id),
("bos_token", self._sp_bos_id),
("eos_token", self._sp_eos_id),
("unk_token", self._sp_unk_id),
):
if token_id < 0 or token_id >= vocab_size:
raise ValueError(
f"{token_name} ID {token_id} is outside vocabulary size {vocab_size}."
)
# Validate values restored from tokenizer_config.json against the
# SentencePiece model. Keeping the same strings guarantees that token
# IDs remain aligned with the model's embedding table.
supplied_tokens = {
"pad_token": pad_token,
"bos_token": bos_token,
"eos_token": eos_token,
"unk_token": unk_token,
}
resolved_tokens = {}
for name, supplied in supplied_tokens.items():
expected = self._sp_special_tokens[name]
if supplied is None:
resolved_tokens[name] = expected
continue
supplied_content = getattr(supplied, "content", str(supplied))
if supplied_content != expected:
raise ValueError(
f"Configured {name}={supplied_content!r} does not match the "
f"SentencePiece token {expected!r}. The tokenizer config and "
"new_bpe.model must come from the same training run."
)
resolved_tokens[name] = supplied
super().__init__(
pad_token=resolved_tokens["pad_token"],
bos_token=resolved_tokens["bos_token"],
eos_token=resolved_tokens["eos_token"],
unk_token=resolved_tokens["unk_token"],
model_max_length=model_max_length,
clean_up_tokenization_spaces=clean_up_tokenization_spaces,
**kwargs,
)
# Ensure save_pretrained writes the dynamic AutoTokenizer mapping and
# the options needed when the tokenizer is reconstructed from the Hub.
self.init_kwargs["auto_map"] = {
"AutoTokenizer": ["tokenization_un_model.UnModelTokenizer", None]
}
self.init_kwargs["tokenizer_class"] = self.__class__.__name__
self.init_kwargs["add_bos_token"] = self.add_bos_token
self.init_kwargs["add_eos_token"] = self.add_eos_token
@property
def vocab_size(self):
return self.sp_model.get_piece_size()
def get_vocab(self):
vocab = {
self.sp_model.id_to_piece(token_id): token_id
for token_id in range(self.vocab_size)
}
vocab.update(self.added_tokens_encoder)
return vocab
def _tokenize(self, text, **kwargs):
return self.sp_model.encode(text, out_type=str)
def _convert_token_to_id(self, token):
token_id = self.sp_model.piece_to_id(token)
if token_id < 0 or token_id >= self.vocab_size:
return self._sp_unk_id
return token_id
def _convert_id_to_token(self, index):
if index < 0 or index >= self.vocab_size:
return self.unk_token
return self.sp_model.id_to_piece(index)
def convert_tokens_to_string(self, tokens: List[str]) -> str:
return self.sp_model.decode_pieces(tokens)
def build_inputs_with_special_tokens(
self,
token_ids_0: List[int],
token_ids_1: Optional[List[int]] = None,
) -> List[int]:
output = list(token_ids_0)
if self.add_bos_token:
output.insert(0, self.bos_token_id)
if self.add_eos_token:
output.append(self.eos_token_id)
if token_ids_1 is not None:
second = list(token_ids_1)
if self.add_bos_token:
second.insert(0, self.bos_token_id)
if self.add_eos_token:
second.append(self.eos_token_id)
output.extend(second)
return output
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]:
if already_has_special_tokens:
special_ids = set(self.all_special_ids)
return [1 if token_id in special_ids else 0 for token_id in token_ids_0]
mask = []
if self.add_bos_token:
mask.append(1)
mask.extend([0] * len(token_ids_0))
if self.add_eos_token:
mask.append(1)
if token_ids_1 is not None:
if self.add_bos_token:
mask.append(1)
mask.extend([0] * len(token_ids_1))
if self.add_eos_token:
mask.append(1)
return mask
def create_token_type_ids_from_sequences(
self,
token_ids_0: List[int],
token_ids_1: Optional[List[int]] = None,
) -> List[int]:
combined = self.build_inputs_with_special_tokens(token_ids_0, token_ids_1)
return [0] * len(combined)
def save_vocabulary(
self,
save_directory,
filename_prefix: Optional[str] = None,
) -> Tuple[str, ...]:
if not os.path.isdir(save_directory):
raise ValueError(f"Vocabulary save directory does not exist: {save_directory}")
filename = (
f"{filename_prefix}-new_bpe.model"
if filename_prefix
else "new_bpe.model"
)
destination = os.path.join(save_directory, filename)
if os.path.abspath(self.vocab_file) != os.path.abspath(destination):
shutil.copyfile(self.vocab_file, destination)
return (destination,)
UnModelTokenizer.register_for_auto_class("AutoTokenizer")
|