LoneStriker commited on
Commit
9c08e4c
1 Parent(s): 238cf58

Upload tokenization_yi.py

Browse files
Files changed (1) hide show
  1. tokenization_yi.py +255 -0
tokenization_yi.py ADDED
@@ -0,0 +1,255 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from shutil import copyfile
3
+ from typing import Any, Dict, List, Optional, Tuple
4
+
5
+ import sentencepiece as spm
6
+ from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
7
+ from transformers.utils import logging
8
+
9
+ logger = logging.get_logger(__name__)
10
+
11
+ VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
12
+
13
+ PRETRAINED_VOCAB_FILES_MAP = {
14
+ "vocab_file": {},
15
+ "tokenizer_file": {},
16
+ }
17
+ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
18
+
19
+
20
+ class YiTokenizer(PreTrainedTokenizer):
21
+ """
22
+ Construct a Yi tokenizer. Based on byte-level Byte-Pair-Encoding.
23
+
24
+ Args:
25
+ vocab_file (`str`):
26
+ Path to the vocabulary file.
27
+ """
28
+
29
+ vocab_files_names = VOCAB_FILES_NAMES
30
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
31
+ max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
32
+ model_input_names = ["input_ids", "attention_mask"]
33
+
34
+ def __init__(
35
+ self,
36
+ vocab_file,
37
+ unk_token="<unk>",
38
+ bos_token="<|startoftext|>",
39
+ eos_token="<|endoftext|>",
40
+ pad_token="<unk>",
41
+ sp_model_kwargs: Optional[Dict[str, Any]] = None,
42
+ add_bos_token=True,
43
+ add_eos_token=False,
44
+ clean_up_tokenization_spaces=False,
45
+ **kwargs,
46
+ ):
47
+ self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
48
+ bos_token = (
49
+ AddedToken(bos_token, lstrip=False, rstrip=False)
50
+ if isinstance(bos_token, str)
51
+ else bos_token
52
+ )
53
+ eos_token = (
54
+ AddedToken(eos_token, lstrip=False, rstrip=False)
55
+ if isinstance(eos_token, str)
56
+ else eos_token
57
+ )
58
+ unk_token = (
59
+ AddedToken(unk_token, lstrip=False, rstrip=False)
60
+ if isinstance(unk_token, str)
61
+ else unk_token
62
+ )
63
+ pad_token = (
64
+ AddedToken(pad_token, lstrip=False, rstrip=False)
65
+ if isinstance(pad_token, str)
66
+ else pad_token
67
+ )
68
+ self.vocab_file = vocab_file
69
+ self.add_bos_token = add_bos_token
70
+ self.add_eos_token = add_eos_token
71
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
72
+ self.sp_model.Load(vocab_file)
73
+ super().__init__(
74
+ bos_token=bos_token,
75
+ eos_token=eos_token,
76
+ unk_token=unk_token,
77
+ pad_token=pad_token,
78
+ add_bos_token=add_bos_token,
79
+ add_eos_token=add_eos_token,
80
+ sp_model_kwargs=self.sp_model_kwargs,
81
+ clean_up_tokenization_spaces=clean_up_tokenization_spaces,
82
+ **kwargs,
83
+ )
84
+
85
+ def __getstate__(self):
86
+ state = self.__dict__.copy()
87
+ state["sp_model"] = None
88
+ return state
89
+
90
+ def __setstate__(self, d):
91
+ self.__dict__ = d
92
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
93
+ self.sp_model.Load(self.vocab_file)
94
+
95
+ @property
96
+ def vocab_size(self):
97
+ """Returns vocab size"""
98
+ return self.sp_model.get_piece_size()
99
+
100
+ def get_vocab(self):
101
+ """Returns vocab as a dict"""
102
+ vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
103
+ vocab.update(self.added_tokens_encoder)
104
+ return vocab
105
+
106
+ def _tokenize(self, text):
107
+ """Returns a tokenized string."""
108
+ return self.sp_model.encode(text, out_type=str)
109
+
110
+ def _convert_token_to_id(self, token):
111
+ """Converts a token (str) in an id using the vocab."""
112
+ return self.sp_model.piece_to_id(token)
113
+
114
+ def _convert_id_to_token(self, index):
115
+ """Converts an index (integer) in a token (str) using the vocab."""
116
+ token = self.sp_model.IdToPiece(index)
117
+ return token
118
+
119
+ def convert_tokens_to_string(self, tokens):
120
+ """Converts a sequence of tokens (string) in a single string."""
121
+ current_sub_tokens = []
122
+ out_string = ""
123
+ prev_is_special = False
124
+ for i, token in enumerate(tokens):
125
+ # make sure that special tokens are not decoded using sentencepiece model
126
+ if token in self.all_special_tokens:
127
+ if not prev_is_special and i != 0:
128
+ out_string += " "
129
+ out_string += self.sp_model.decode(current_sub_tokens) + token
130
+ prev_is_special = True
131
+ current_sub_tokens = []
132
+ else:
133
+ current_sub_tokens.append(token)
134
+ prev_is_special = False
135
+ out_string += self.sp_model.decode(current_sub_tokens)
136
+ return out_string
137
+
138
+ def save_vocabulary(
139
+ self, save_directory, filename_prefix: Optional[str] = None
140
+ ) -> Tuple[str]:
141
+ """
142
+ Save the vocabulary and special tokens file to a directory.
143
+
144
+ Args:
145
+ save_directory (`str`):
146
+ The directory in which to save the vocabulary.
147
+
148
+ Returns:
149
+ `Tuple(str)`: Paths to the files saved.
150
+ """
151
+ if not os.path.isdir(save_directory):
152
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory")
153
+ return
154
+ out_vocab_file = os.path.join(
155
+ save_directory,
156
+ (filename_prefix + "-" if filename_prefix else "")
157
+ + VOCAB_FILES_NAMES["vocab_file"],
158
+ )
159
+
160
+ if os.path.abspath(self.vocab_file) != os.path.abspath(
161
+ out_vocab_file
162
+ ) and os.path.isfile(self.vocab_file):
163
+ copyfile(self.vocab_file, out_vocab_file)
164
+ elif not os.path.isfile(self.vocab_file):
165
+ with open(out_vocab_file, "wb") as fi:
166
+ content_spiece_model = self.sp_model.serialized_model_proto()
167
+ fi.write(content_spiece_model)
168
+
169
+ return (out_vocab_file,)
170
+
171
+ def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
172
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
173
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
174
+
175
+ output = bos_token_id + token_ids_0 + eos_token_id
176
+
177
+ if token_ids_1 is not None:
178
+ output = output + bos_token_id + token_ids_1 + eos_token_id
179
+
180
+ return output
181
+
182
+ def get_special_tokens_mask(
183
+ self,
184
+ token_ids_0: List[int],
185
+ token_ids_1: Optional[List[int]] = None,
186
+ already_has_special_tokens: bool = False,
187
+ ) -> List[int]:
188
+ """
189
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
190
+ special tokens using the tokenizer `prepare_for_model` method.
191
+
192
+ Args:
193
+ token_ids_0 (`List[int]`):
194
+ List of IDs.
195
+ token_ids_1 (`List[int]`, *optional*):
196
+ Optional second list of IDs for sequence pairs.
197
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
198
+ Whether or not the token list is already formatted with special tokens for the model.
199
+
200
+ Returns:
201
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
202
+ """
203
+ if already_has_special_tokens:
204
+ return super().get_special_tokens_mask(
205
+ token_ids_0=token_ids_0,
206
+ token_ids_1=token_ids_1,
207
+ already_has_special_tokens=True,
208
+ )
209
+
210
+ bos_token_id = [1] if self.add_bos_token else []
211
+ eos_token_id = [1] if self.add_eos_token else []
212
+
213
+ if token_ids_1 is None:
214
+ return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
215
+ return (
216
+ bos_token_id
217
+ + ([0] * len(token_ids_0))
218
+ + eos_token_id
219
+ + bos_token_id
220
+ + ([0] * len(token_ids_1))
221
+ + eos_token_id
222
+ )
223
+
224
+ def create_token_type_ids_from_sequences(
225
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
226
+ ) -> List[int]:
227
+ """
228
+ Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
229
+ sequence pair mask has the following format:
230
+
231
+ ```
232
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
233
+ | first sequence | second sequence |
234
+ ```
235
+
236
+ if token_ids_1 is None, only returns the first portion of the mask (0s).
237
+
238
+ Args:
239
+ token_ids_0 (`List[int]`):
240
+ List of ids.
241
+ token_ids_1 (`List[int]`, *optional*):
242
+ Optional second list of IDs for sequence pairs.
243
+
244
+ Returns:
245
+ `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
246
+ """
247
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
248
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
249
+
250
+ output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
251
+
252
+ if token_ids_1 is not None:
253
+ output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
254
+
255
+ return output