Upload lora-scripts/sd-scripts/finetune/clean_captions_and_tags.py with huggingface_hub
Browse files
lora-scripts/sd-scripts/finetune/clean_captions_and_tags.py
ADDED
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# このスクリプトのライセンスは、Apache License 2.0とします
|
2 |
+
# (c) 2022 Kohya S. @kohya_ss
|
3 |
+
|
4 |
+
import argparse
|
5 |
+
import glob
|
6 |
+
import os
|
7 |
+
import json
|
8 |
+
import re
|
9 |
+
|
10 |
+
from tqdm import tqdm
|
11 |
+
from library.utils import setup_logging
|
12 |
+
setup_logging()
|
13 |
+
import logging
|
14 |
+
logger = logging.getLogger(__name__)
|
15 |
+
|
16 |
+
PATTERN_HAIR_LENGTH = re.compile(r', (long|short|medium) hair, ')
|
17 |
+
PATTERN_HAIR_CUT = re.compile(r', (bob|hime) cut, ')
|
18 |
+
PATTERN_HAIR = re.compile(r', ([\w\-]+) hair, ')
|
19 |
+
PATTERN_WORD = re.compile(r', ([\w\-]+|hair ornament), ')
|
20 |
+
|
21 |
+
# 複数人がいるとき、複数の髪色や目の色が定義されていれば削除する
|
22 |
+
PATTERNS_REMOVE_IN_MULTI = [
|
23 |
+
PATTERN_HAIR_LENGTH,
|
24 |
+
PATTERN_HAIR_CUT,
|
25 |
+
re.compile(r', [\w\-]+ eyes, '),
|
26 |
+
re.compile(r', ([\w\-]+ sleeves|sleeveless), '),
|
27 |
+
# 複数の髪型定義がある場合は削除する
|
28 |
+
re.compile(
|
29 |
+
r', (ponytail|braid|ahoge|twintails|[\w\-]+ bun|single hair bun|single side bun|two side up|two tails|[\w\-]+ braid|sidelocks), '),
|
30 |
+
]
|
31 |
+
|
32 |
+
|
33 |
+
def clean_tags(image_key, tags):
|
34 |
+
# replace '_' to ' '
|
35 |
+
tags = tags.replace('^_^', '^@@@^')
|
36 |
+
tags = tags.replace('_', ' ')
|
37 |
+
tags = tags.replace('^@@@^', '^_^')
|
38 |
+
|
39 |
+
# remove rating: deepdanbooruのみ
|
40 |
+
tokens = tags.split(", rating")
|
41 |
+
if len(tokens) == 1:
|
42 |
+
# WD14 taggerのときはこちらになるのでメッセージは出さない
|
43 |
+
# logger.info("no rating:")
|
44 |
+
# logger.info(f"{image_key} {tags}")
|
45 |
+
pass
|
46 |
+
else:
|
47 |
+
if len(tokens) > 2:
|
48 |
+
logger.info("multiple ratings:")
|
49 |
+
logger.info(f"{image_key} {tags}")
|
50 |
+
tags = tokens[0]
|
51 |
+
|
52 |
+
tags = ", " + tags.replace(", ", ", , ") + ", " # カンマ付きで検索をするための身も蓋もない対策
|
53 |
+
|
54 |
+
# 複数の人物がいる場合は髪色等のタグを削除する
|
55 |
+
if 'girls' in tags or 'boys' in tags:
|
56 |
+
for pat in PATTERNS_REMOVE_IN_MULTI:
|
57 |
+
found = pat.findall(tags)
|
58 |
+
if len(found) > 1: # 二つ以上、タグがある
|
59 |
+
tags = pat.sub("", tags)
|
60 |
+
|
61 |
+
# 髪の特殊対応
|
62 |
+
srch_hair_len = PATTERN_HAIR_LENGTH.search(tags) # 髪の長さタグは例外なので避けておく(全員が同じ髪の長さの場合)
|
63 |
+
if srch_hair_len:
|
64 |
+
org = srch_hair_len.group()
|
65 |
+
tags = PATTERN_HAIR_LENGTH.sub(", @@@, ", tags)
|
66 |
+
|
67 |
+
found = PATTERN_HAIR.findall(tags)
|
68 |
+
if len(found) > 1:
|
69 |
+
tags = PATTERN_HAIR.sub("", tags)
|
70 |
+
|
71 |
+
if srch_hair_len:
|
72 |
+
tags = tags.replace(", @@@, ", org) # 戻す
|
73 |
+
|
74 |
+
# white shirtとshirtみたいな重複タグの削除
|
75 |
+
found = PATTERN_WORD.findall(tags)
|
76 |
+
for word in found:
|
77 |
+
if re.search(f", ((\w+) )+{word}, ", tags):
|
78 |
+
tags = tags.replace(f", {word}, ", "")
|
79 |
+
|
80 |
+
tags = tags.replace(", , ", ", ")
|
81 |
+
assert tags.startswith(", ") and tags.endswith(", ")
|
82 |
+
tags = tags[2:-2]
|
83 |
+
return tags
|
84 |
+
|
85 |
+
|
86 |
+
# 上から順に検索、置換される
|
87 |
+
# ('置換元文字列', '置換後文字列')
|
88 |
+
CAPTION_REPLACEMENTS = [
|
89 |
+
('anime anime', 'anime'),
|
90 |
+
('young ', ''),
|
91 |
+
('anime girl', 'girl'),
|
92 |
+
('cartoon female', 'girl'),
|
93 |
+
('cartoon lady', 'girl'),
|
94 |
+
('cartoon character', 'girl'), # a or ~s
|
95 |
+
('cartoon woman', 'girl'),
|
96 |
+
('cartoon women', 'girls'),
|
97 |
+
('cartoon girl', 'girl'),
|
98 |
+
('anime female', 'girl'),
|
99 |
+
('anime lady', 'girl'),
|
100 |
+
('anime character', 'girl'), # a or ~s
|
101 |
+
('anime woman', 'girl'),
|
102 |
+
('anime women', 'girls'),
|
103 |
+
('lady', 'girl'),
|
104 |
+
('female', 'girl'),
|
105 |
+
('woman', 'girl'),
|
106 |
+
('women', 'girls'),
|
107 |
+
('people', 'girls'),
|
108 |
+
('person', 'girl'),
|
109 |
+
('a cartoon figure', 'a figure'),
|
110 |
+
('a cartoon image', 'an image'),
|
111 |
+
('a cartoon picture', 'a picture'),
|
112 |
+
('an anime cartoon image', 'an image'),
|
113 |
+
('a cartoon anime drawing', 'a drawing'),
|
114 |
+
('a cartoon drawing', 'a drawing'),
|
115 |
+
('girl girl', 'girl'),
|
116 |
+
]
|
117 |
+
|
118 |
+
|
119 |
+
def clean_caption(caption):
|
120 |
+
for rf, rt in CAPTION_REPLACEMENTS:
|
121 |
+
replaced = True
|
122 |
+
while replaced:
|
123 |
+
bef = caption
|
124 |
+
caption = caption.replace(rf, rt)
|
125 |
+
replaced = bef != caption
|
126 |
+
return caption
|
127 |
+
|
128 |
+
|
129 |
+
def main(args):
|
130 |
+
if os.path.exists(args.in_json):
|
131 |
+
logger.info(f"loading existing metadata: {args.in_json}")
|
132 |
+
with open(args.in_json, "rt", encoding='utf-8') as f:
|
133 |
+
metadata = json.load(f)
|
134 |
+
else:
|
135 |
+
logger.error("no metadata / メタデータファイルがありません")
|
136 |
+
return
|
137 |
+
|
138 |
+
logger.info("cleaning captions and tags.")
|
139 |
+
image_keys = list(metadata.keys())
|
140 |
+
for image_key in tqdm(image_keys):
|
141 |
+
tags = metadata[image_key].get('tags')
|
142 |
+
if tags is None:
|
143 |
+
logger.error(f"image does not have tags / メタデータにタグがありません: {image_key}")
|
144 |
+
else:
|
145 |
+
org = tags
|
146 |
+
tags = clean_tags(image_key, tags)
|
147 |
+
metadata[image_key]['tags'] = tags
|
148 |
+
if args.debug and org != tags:
|
149 |
+
logger.info("FROM: " + org)
|
150 |
+
logger.info("TO: " + tags)
|
151 |
+
|
152 |
+
caption = metadata[image_key].get('caption')
|
153 |
+
if caption is None:
|
154 |
+
logger.error(f"image does not have caption / メタデータにキャプションがありません: {image_key}")
|
155 |
+
else:
|
156 |
+
org = caption
|
157 |
+
caption = clean_caption(caption)
|
158 |
+
metadata[image_key]['caption'] = caption
|
159 |
+
if args.debug and org != caption:
|
160 |
+
logger.info("FROM: " + org)
|
161 |
+
logger.info("TO: " + caption)
|
162 |
+
|
163 |
+
# metadataを書き出して終わり
|
164 |
+
logger.info(f"writing metadata: {args.out_json}")
|
165 |
+
with open(args.out_json, "wt", encoding='utf-8') as f:
|
166 |
+
json.dump(metadata, f, indent=2)
|
167 |
+
logger.info("done!")
|
168 |
+
|
169 |
+
|
170 |
+
def setup_parser() -> argparse.ArgumentParser:
|
171 |
+
parser = argparse.ArgumentParser()
|
172 |
+
# parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
|
173 |
+
parser.add_argument("in_json", type=str, help="metadata file to input / 読み込むメタデータファイル")
|
174 |
+
parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
|
175 |
+
parser.add_argument("--debug", action="store_true", help="debug mode")
|
176 |
+
|
177 |
+
return parser
|
178 |
+
|
179 |
+
|
180 |
+
if __name__ == '__main__':
|
181 |
+
parser = setup_parser()
|
182 |
+
|
183 |
+
args, unknown = parser.parse_known_args()
|
184 |
+
if len(unknown) == 1:
|
185 |
+
logger.warning("WARNING: train_data_dir argument is removed. This script will not work with three arguments in future. Please specify two arguments: in_json and out_json.")
|
186 |
+
logger.warning("All captions and tags in the metadata are processed.")
|
187 |
+
logger.warning("警告: train_data_dir引数は不要になりました。将来的には三つの引数を指定すると動かなくなる予定です。読み込み元のメタデータと書き出し先の二つの引数だけ指定してください。")
|
188 |
+
logger.warning("メタデータ内のすべてのキャプションとタグが処理されます。")
|
189 |
+
args.in_json = args.out_json
|
190 |
+
args.out_json = unknown[0]
|
191 |
+
elif len(unknown) > 0:
|
192 |
+
raise ValueError(f"error: unrecognized arguments: {unknown}")
|
193 |
+
|
194 |
+
main(args)
|