VarunGumma
commited on
Commit
•
53ee433
1
Parent(s):
684893f
Upload tokenization_indictrans.py with huggingface_hub
Browse files- tokenization_indictrans.py +261 -0
tokenization_indictrans.py
ADDED
@@ -0,0 +1,261 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
|
4 |
+
from typing import Dict, List, Optional, Union, Tuple
|
5 |
+
|
6 |
+
from transformers.utils import logging
|
7 |
+
from sentencepiece import SentencePieceProcessor
|
8 |
+
from transformers.tokenization_utils import PreTrainedTokenizer
|
9 |
+
|
10 |
+
|
11 |
+
logger = logging.get_logger(__name__)
|
12 |
+
|
13 |
+
SPIECE_UNDERLINE = "▁"
|
14 |
+
|
15 |
+
SPECIAL_TAGS = {
|
16 |
+
"_bt_",
|
17 |
+
"_ft_",
|
18 |
+
"asm_Beng",
|
19 |
+
"awa_Deva",
|
20 |
+
"ben_Beng",
|
21 |
+
"bho_Deva",
|
22 |
+
"brx_Deva",
|
23 |
+
"doi_Deva",
|
24 |
+
"eng_Latn",
|
25 |
+
"gom_Deva",
|
26 |
+
"gon_Deva",
|
27 |
+
"guj_Gujr",
|
28 |
+
"hin_Deva",
|
29 |
+
"hne_Deva",
|
30 |
+
"kan_Knda",
|
31 |
+
"kas_Arab",
|
32 |
+
"kas_Deva",
|
33 |
+
"kha_Latn",
|
34 |
+
"lus_Latn",
|
35 |
+
"mag_Deva",
|
36 |
+
"mai_Deva",
|
37 |
+
"mal_Mlym",
|
38 |
+
"mar_Deva",
|
39 |
+
"mni_Beng",
|
40 |
+
"mni_Mtei",
|
41 |
+
"npi_Deva",
|
42 |
+
"ory_Orya",
|
43 |
+
"pan_Guru",
|
44 |
+
"san_Deva",
|
45 |
+
"sat_Olck",
|
46 |
+
"snd_Arab",
|
47 |
+
"snd_Deva",
|
48 |
+
"tam_Taml",
|
49 |
+
"tel_Telu",
|
50 |
+
"urd_Arab",
|
51 |
+
"unr_Deva",
|
52 |
+
}
|
53 |
+
|
54 |
+
VOCAB_FILES_NAMES = {
|
55 |
+
"src_vocab_fp": "dict.SRC.json",
|
56 |
+
"tgt_vocab_fp": "dict.TGT.json",
|
57 |
+
"src_spm_fp": "model.SRC",
|
58 |
+
"tgt_spm_fp": "model.TGT",
|
59 |
+
}
|
60 |
+
|
61 |
+
|
62 |
+
class IndicTransTokenizer(PreTrainedTokenizer):
|
63 |
+
_added_tokens_encoder = {}
|
64 |
+
_added_tokens_decoder = {}
|
65 |
+
|
66 |
+
vocab_files_names = VOCAB_FILES_NAMES
|
67 |
+
model_input_names = ["input_ids", "attention_mask"]
|
68 |
+
|
69 |
+
def __init__(
|
70 |
+
self,
|
71 |
+
src_vocab_fp=None,
|
72 |
+
tgt_vocab_fp=None,
|
73 |
+
src_spm_fp=None,
|
74 |
+
tgt_spm_fp=None,
|
75 |
+
unk_token="<unk>",
|
76 |
+
bos_token="<s>",
|
77 |
+
eos_token="</s>",
|
78 |
+
pad_token="<pad>",
|
79 |
+
do_lower_case=False,
|
80 |
+
**kwargs,
|
81 |
+
):
|
82 |
+
|
83 |
+
self.src = True
|
84 |
+
|
85 |
+
self.src_vocab_fp = src_vocab_fp
|
86 |
+
self.tgt_vocab_fp = tgt_vocab_fp
|
87 |
+
self.src_spm_fp = src_spm_fp
|
88 |
+
self.tgt_spm_fp = tgt_spm_fp
|
89 |
+
|
90 |
+
self.unk_token = unk_token
|
91 |
+
self.pad_token = pad_token
|
92 |
+
self.eos_token = eos_token
|
93 |
+
self.bos_token = bos_token
|
94 |
+
|
95 |
+
self.encoder = self._load_json(self.src_vocab_fp)
|
96 |
+
if self.unk_token not in self.encoder:
|
97 |
+
raise KeyError("<unk> token must be in vocab")
|
98 |
+
assert self.pad_token in self.encoder
|
99 |
+
self.encoder_rev = {v: k for k, v in self.encoder.items()}
|
100 |
+
|
101 |
+
self.decoder = self._load_json(self.tgt_vocab_fp)
|
102 |
+
if self.unk_token not in self.encoder:
|
103 |
+
raise KeyError("<unk> token must be in vocab")
|
104 |
+
assert self.pad_token in self.encoder
|
105 |
+
self.decoder_rev = {v: k for k, v in self.decoder.items()}
|
106 |
+
|
107 |
+
# load SentencePiece model for pre-processing
|
108 |
+
self.src_spm = self._load_spm(self.src_spm_fp)
|
109 |
+
self.tgt_spm = self._load_spm(self.tgt_spm_fp)
|
110 |
+
|
111 |
+
self.current_spm = self.src_spm
|
112 |
+
self.current_encoder = self.encoder
|
113 |
+
self.current_encoder_rev = self.encoder_rev
|
114 |
+
|
115 |
+
self.unk_token_id = self.encoder[self.unk_token]
|
116 |
+
self.pad_token_id = self.encoder[self.pad_token]
|
117 |
+
self.eos_token_id = self.encoder[self.eos_token]
|
118 |
+
self.bos_token_id = self.encoder[self.bos_token]
|
119 |
+
|
120 |
+
super().__init__(
|
121 |
+
src_vocab_file=self.src_vocab_fp,
|
122 |
+
tgt_vocab_file=self.src_vocab_fp,
|
123 |
+
do_lower_case=do_lower_case,
|
124 |
+
unk_token=unk_token,
|
125 |
+
bos_token=bos_token,
|
126 |
+
eos_token=eos_token,
|
127 |
+
pad_token=pad_token,
|
128 |
+
**kwargs,
|
129 |
+
)
|
130 |
+
|
131 |
+
def add_new_special_tags(self, new_tags: List[str]):
|
132 |
+
SPECIAL_TAGS.update(new_tags)
|
133 |
+
|
134 |
+
def _switch_to_input_mode(self):
|
135 |
+
self.src = True
|
136 |
+
self.padding_side = "left"
|
137 |
+
self.current_spm = self.src_spm
|
138 |
+
self.current_encoder = self.encoder
|
139 |
+
self.current_encoder_rev = self.encoder_rev
|
140 |
+
|
141 |
+
def _switch_to_target_mode(self):
|
142 |
+
self.src = False
|
143 |
+
self.padding_side = "right"
|
144 |
+
self.current_spm = self.tgt_spm
|
145 |
+
self.current_encoder = self.decoder
|
146 |
+
self.current_encoder_rev = self.decoder_rev
|
147 |
+
|
148 |
+
def _load_spm(self, path: str) -> SentencePieceProcessor:
|
149 |
+
return SentencePieceProcessor(model_file=path)
|
150 |
+
|
151 |
+
def _save_json(self, data, path: str) -> None:
|
152 |
+
with open(path, "w", encoding="utf-8") as f:
|
153 |
+
json.dump(data, f, indent=2)
|
154 |
+
|
155 |
+
def _load_json(self, path: str) -> Union[Dict, List]:
|
156 |
+
with open(path, "r", encoding="utf-8") as f:
|
157 |
+
return json.load(f)
|
158 |
+
|
159 |
+
def _split_tags(self, tokens: List[str]) -> Tuple[List[str], List[str]]:
|
160 |
+
tags = [token for token in tokens if token in SPECIAL_TAGS]
|
161 |
+
tokens = [token for token in tokens if token not in SPECIAL_TAGS]
|
162 |
+
return tags, tokens
|
163 |
+
|
164 |
+
def _split_pads(self, tokens: List[str]) -> Tuple[List[str], List[str]]:
|
165 |
+
pads = [token for token in tokens if token == self.pad_token]
|
166 |
+
tokens = [token for token in tokens if token != self.pad_token]
|
167 |
+
return pads, tokens
|
168 |
+
|
169 |
+
@property
|
170 |
+
def src_vocab_size(self) -> int:
|
171 |
+
return len(self.encoder)
|
172 |
+
|
173 |
+
@property
|
174 |
+
def tgt_vocab_size(self) -> int:
|
175 |
+
return len(self.decoder)
|
176 |
+
|
177 |
+
def get_src_vocab(self) -> Dict[str, int]:
|
178 |
+
return dict(self.encoder, **self.added_tokens_encoder)
|
179 |
+
|
180 |
+
def get_tgt_vocab(self) -> Dict[str, int]:
|
181 |
+
return dict(self.decoder, **self.added_tokens_decoder)
|
182 |
+
|
183 |
+
# hack override
|
184 |
+
def get_vocab(self) -> Dict[str, int]:
|
185 |
+
return self.get_src_vocab()
|
186 |
+
|
187 |
+
# hack override
|
188 |
+
@property
|
189 |
+
def vocab_size(self) -> int:
|
190 |
+
return self.src_vocab_size
|
191 |
+
|
192 |
+
def _convert_token_to_id(self, token: str) -> int:
|
193 |
+
"""Converts an token (str) into an index (integer) using the source/target vocabulary map."""
|
194 |
+
return self.current_encoder.get(token, self.current_encoder[self.unk_token])
|
195 |
+
|
196 |
+
def _convert_id_to_token(self, index: int) -> str:
|
197 |
+
"""Converts an index (integer) into a token (str) using the source/target vocabulary map."""
|
198 |
+
return self.current_encoder_rev.get(index, self.unk_token)
|
199 |
+
|
200 |
+
def convert_tokens_to_string(self, tokens: List[str]) -> str:
|
201 |
+
"""Uses sentencepiece model for detokenization"""
|
202 |
+
pads, tokens = self._split_pads(tokens)
|
203 |
+
|
204 |
+
if self.src:
|
205 |
+
|
206 |
+
tags, non_tags = self._split_tags(tokens)
|
207 |
+
|
208 |
+
return (
|
209 |
+
" ".join(pads)
|
210 |
+
+ " "
|
211 |
+
+ " ".join(tags)
|
212 |
+
+ " "
|
213 |
+
+ "".join(non_tags).replace(SPIECE_UNDERLINE, " ").strip()
|
214 |
+
)
|
215 |
+
|
216 |
+
return (
|
217 |
+
"".join(tokens).replace(SPIECE_UNDERLINE, " ").strip()
|
218 |
+
+ " "
|
219 |
+
+ " ".join(pads)
|
220 |
+
)
|
221 |
+
|
222 |
+
def _tokenize(self, text) -> List[str]:
|
223 |
+
if self.src:
|
224 |
+
tokens = text.split(" ")
|
225 |
+
tags, non_tags = self._split_tags(tokens)
|
226 |
+
text = " ".join(non_tags)
|
227 |
+
tokens = self.current_spm.EncodeAsPieces(text)
|
228 |
+
return tags + tokens
|
229 |
+
else:
|
230 |
+
return self.current_spm.EncodeAsPieces(text)
|
231 |
+
|
232 |
+
def build_inputs_with_special_tokens(
|
233 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
234 |
+
) -> List[int]:
|
235 |
+
if token_ids_1 is None:
|
236 |
+
return token_ids_0 + [self.eos_token_id]
|
237 |
+
# We don't expect to process pairs, but leave the pair logic for API consistency
|
238 |
+
return token_ids_0 + [self.eos_token_id] + token_ids_1 + [self.eos_token_id]
|
239 |
+
|
240 |
+
def save_vocabulary(
|
241 |
+
self, save_directory: str, filename_prefix: Optional[str] = None
|
242 |
+
) -> Tuple[str]:
|
243 |
+
if not os.path.isdir(save_directory):
|
244 |
+
logger.error(f"Vocabulary path ({save_directory}) should be a directory")
|
245 |
+
return
|
246 |
+
|
247 |
+
src_spm_fp = os.path.join(save_directory, "model.SRC")
|
248 |
+
tgt_spm_fp = os.path.join(save_directory, "model.TGT")
|
249 |
+
src_vocab_fp = os.path.join(save_directory, "dict.SRC.json")
|
250 |
+
tgt_vocab_fp = os.path.join(save_directory, "dict.TGT.json")
|
251 |
+
|
252 |
+
self._save_json(self.encoder, src_vocab_fp)
|
253 |
+
self._save_json(self.decoder, tgt_vocab_fp)
|
254 |
+
|
255 |
+
with open(src_spm_fp, "wb") as f:
|
256 |
+
f.write(self.src_spm.serialized_model_proto())
|
257 |
+
|
258 |
+
with open(tgt_spm_fp, "wb") as f:
|
259 |
+
f.write(self.tgt_spm.serialized_model_proto())
|
260 |
+
|
261 |
+
return src_vocab_fp, tgt_vocab_fp, src_spm_fp, tgt_spm_fp
|