Upload tokenization_dart.py
Browse files- tokenization_dart.py +45 -60
tokenization_dart.py
CHANGED
|
@@ -1,26 +1,51 @@
|
|
| 1 |
import logging
|
| 2 |
-
import os
|
| 3 |
import json
|
| 4 |
-
from typing import
|
| 5 |
from pydantic.dataclasses import dataclass
|
| 6 |
|
| 7 |
-
import numpy as np
|
| 8 |
-
from numpy.typing import NDArray
|
| 9 |
-
|
| 10 |
from transformers import PreTrainedTokenizerFast
|
| 11 |
from tokenizers.decoders import Decoder
|
| 12 |
|
| 13 |
logger = logging.getLogger(__name__)
|
| 14 |
|
| 15 |
-
VOCAB_FILES_NAMES = {
|
| 16 |
-
"category_config": "category_config.json",
|
| 17 |
-
}
|
| 18 |
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
}
|
| 23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
|
| 26 |
@dataclass
|
|
@@ -71,57 +96,17 @@ class DartDecoder:
|
|
| 71 |
class DartTokenizer(PreTrainedTokenizerFast):
|
| 72 |
"""Dart tokenizer"""
|
| 73 |
|
| 74 |
-
|
| 75 |
-
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
| 76 |
-
|
| 77 |
-
def __init__(self, category_config, **kwargs):
|
| 78 |
super().__init__(**kwargs)
|
| 79 |
|
| 80 |
self._tokenizer.decoder = Decoder.custom( # type: ignore
|
| 81 |
DartDecoder(list(self.get_added_vocab().keys()))
|
| 82 |
)
|
| 83 |
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
tokens,
|
| 90 |
-
) in self.category_config.category_to_token_ids.items():
|
| 91 |
-
self._id_to_category_map[tokens] = int(category_id)
|
| 92 |
-
|
| 93 |
-
def create_vocab_mask(self, value: int = 1):
|
| 94 |
-
"""Create an array of vocab size filled with specified value"""
|
| 95 |
-
return np.full(self.vocab_size, value).astype("uint8")
|
| 96 |
-
|
| 97 |
-
def get_token_ids_in_category(self, category_id: Union[int, str]):
|
| 98 |
-
"""Get token ids in the specified category"""
|
| 99 |
-
return self.category_config.category_to_token_ids[str(category_id)]
|
| 100 |
-
|
| 101 |
-
def get_category(self, category_id: Union[int, str]):
|
| 102 |
-
"""Get the specified category config"""
|
| 103 |
-
return self.category_config.categories[str(category_id)]
|
| 104 |
-
|
| 105 |
-
def convert_ids_to_category_ids(self, token_ids: Union[int, List[int]]):
|
| 106 |
-
"""Get the category ids of specified tokens"""
|
| 107 |
-
return self._id_to_category_map[token_ids]
|
| 108 |
-
|
| 109 |
-
def get_banned_tokens_mask(self, tokens: Union[str, List[str], int, List[int]]):
|
| 110 |
-
if isinstance(tokens, str):
|
| 111 |
-
tokens = [tokens]
|
| 112 |
-
elif isinstance(tokens, int):
|
| 113 |
-
tokens = [tokens]
|
| 114 |
-
elif isinstance(tokens, list):
|
| 115 |
-
tokens = [ # type: ignore
|
| 116 |
-
self.convert_tokens_to_ids(token) if isinstance(token, str) else token
|
| 117 |
-
for token in tokens
|
| 118 |
-
]
|
| 119 |
-
|
| 120 |
-
assert isinstance(tokens, list) and all(
|
| 121 |
-
[isinstance(token, int) for token in tokens]
|
| 122 |
-
)
|
| 123 |
-
|
| 124 |
-
mask = self.create_vocab_mask(value=1)
|
| 125 |
-
mask[tokens] = 0
|
| 126 |
|
| 127 |
-
return
|
|
|
|
| 1 |
import logging
|
|
|
|
| 2 |
import json
|
| 3 |
+
from typing import Dict, List
|
| 4 |
from pydantic.dataclasses import dataclass
|
| 5 |
|
|
|
|
|
|
|
|
|
|
| 6 |
from transformers import PreTrainedTokenizerFast
|
| 7 |
from tokenizers.decoders import Decoder
|
| 8 |
|
| 9 |
logger = logging.getLogger(__name__)
|
| 10 |
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
+
# fmt: off
|
| 13 |
+
# https://huggingface.co/docs/transformers/main/en/chat_templating
|
| 14 |
+
PROMPT_TEMPLATE = (
|
| 15 |
+
"{{ '<|bos|>' }}"
|
| 16 |
+
|
| 17 |
+
"{{ '<rating>' }}"
|
| 18 |
+
"{% if 'rating' not in messages or messages['rating'] is none %}"
|
| 19 |
+
"{{ 'rating:sfw, rating:general' }}"
|
| 20 |
+
"{% else %}"
|
| 21 |
+
"{{ messages['rating'] }}"
|
| 22 |
+
"{% endif %}"
|
| 23 |
+
"{{ '</rating>' }}"
|
| 24 |
+
|
| 25 |
+
"{{ '<copyright>' }}"
|
| 26 |
+
"{% if 'copyright' not in messages or messages['copyright'] is none %}"
|
| 27 |
+
"{{ '' }}"
|
| 28 |
+
"{% else %}"
|
| 29 |
+
"{{ messages['copyright'] }}"
|
| 30 |
+
"{% endif %}"
|
| 31 |
+
"{{ '</copyright>' }}"
|
| 32 |
+
|
| 33 |
+
"{{ '<character>' }}"
|
| 34 |
+
"{% if 'character' not in messages or messages['character'] is none %}"
|
| 35 |
+
"{{ '' }}"
|
| 36 |
+
"{% else %}"
|
| 37 |
+
"{{ messages['character'] }}"
|
| 38 |
+
"{% endif %}"
|
| 39 |
+
"{{ '</character>' }}"
|
| 40 |
+
|
| 41 |
+
"{{ '<general>' }}"
|
| 42 |
+
"{% if 'general' not in messages or messages['general'] is none %}"
|
| 43 |
+
"{{ '' }}"
|
| 44 |
+
"{% else %}"
|
| 45 |
+
"{{ messages['general'] }}"
|
| 46 |
+
"{% endif %}"
|
| 47 |
+
).strip()
|
| 48 |
+
# fmt: on
|
| 49 |
|
| 50 |
|
| 51 |
@dataclass
|
|
|
|
| 96 |
class DartTokenizer(PreTrainedTokenizerFast):
|
| 97 |
"""Dart tokenizer"""
|
| 98 |
|
| 99 |
+
def __init__(self, **kwargs):
|
|
|
|
|
|
|
|
|
|
| 100 |
super().__init__(**kwargs)
|
| 101 |
|
| 102 |
self._tokenizer.decoder = Decoder.custom( # type: ignore
|
| 103 |
DartDecoder(list(self.get_added_vocab().keys()))
|
| 104 |
)
|
| 105 |
|
| 106 |
+
@property
|
| 107 |
+
def default_chat_template(self):
|
| 108 |
+
"""
|
| 109 |
+
Danbooru Tags Transformer uses special format prompt to generate danbooru tags.
|
| 110 |
+
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
|
| 112 |
+
return PROMPT_TEMPLATE
|