Spaces:
Running
on
Zero
Running
on
Zero
John6666
commited on
Commit
•
80e6c51
0
Parent(s):
Super-squash branch 'main' using huggingface_hub
Browse files- .gitattributes +35 -0
- README.md +13 -0
- app.py +118 -0
- character_series_dict.csv +0 -0
- danbooru_e621.csv +0 -0
- genimage.py +66 -0
- ja_to_danbooru/character_series_dict.json +0 -0
- ja_to_danbooru/danbooru_tagtype_dict.json +0 -0
- ja_to_danbooru/ja_danbooru_dict.json +0 -0
- ja_to_danbooru/ja_to_danbooru.py +87 -0
- llmdolphin.py +527 -0
- pre-requirements.txt +1 -0
- requirements.txt +17 -0
- tag_group.csv +0 -0
- tagger.py +506 -0
- utils.py +45 -0
.gitattributes
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
29 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
30 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
31 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
32 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
33 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
title: Natural Text to SD Prompt Translator With LLM alpha
|
3 |
+
emoji: 👀😻
|
4 |
+
colorFrom: red
|
5 |
+
colorTo: purple
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 4.38.1
|
8 |
+
app_file: app.py
|
9 |
+
pinned: false
|
10 |
+
license: apache-2.0
|
11 |
+
---
|
12 |
+
|
13 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
ADDED
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from utils import (
|
3 |
+
gradio_copy_text,
|
4 |
+
COPY_ACTION_JS,
|
5 |
+
)
|
6 |
+
from tagger import (
|
7 |
+
convert_danbooru_to_e621_prompt,
|
8 |
+
insert_recom_prompt,
|
9 |
+
)
|
10 |
+
from genimage import (
|
11 |
+
generate_image,
|
12 |
+
)
|
13 |
+
from llmdolphin import (
|
14 |
+
get_llm_formats,
|
15 |
+
get_dolphin_model_format,
|
16 |
+
get_dolphin_models,
|
17 |
+
get_dolphin_model_info,
|
18 |
+
select_dolphin_model,
|
19 |
+
select_dolphin_format,
|
20 |
+
add_dolphin_models,
|
21 |
+
get_dolphin_sysprompt,
|
22 |
+
get_dolphin_sysprompt_mode,
|
23 |
+
select_dolphin_sysprompt,
|
24 |
+
get_dolphin_languages,
|
25 |
+
select_dolphin_language,
|
26 |
+
dolphin_respond,
|
27 |
+
dolphin_parse,
|
28 |
+
)
|
29 |
+
|
30 |
+
|
31 |
+
with gr.Blocks(theme="NoCrypt/miku@>=1.2.2", css="") as app:
|
32 |
+
gr.Markdown("""# Natural Text to SD Prompt Translator With LLM alpha
|
33 |
+
Text in natural language (English, Japanese, ...) => Prompt
|
34 |
+
""")
|
35 |
+
with gr.Column(scale=1):
|
36 |
+
with gr.Group():
|
37 |
+
chatbot = gr.Chatbot(likeable=False, show_copy_button=True, show_share_button=False, layout="panel", container=True, )
|
38 |
+
with gr.Row():
|
39 |
+
chat_msg = gr.Textbox(show_label=False, placeholder="Input text in English, Japanese, or any other languages and press Enter or click Send.", scale=4)
|
40 |
+
chat_submit = gr.Button("Send", scale=1)
|
41 |
+
chat_clear = gr.Button("Clear", scale=1)
|
42 |
+
with gr.Accordion("Additional inputs", open=False):
|
43 |
+
chat_format = gr.Dropdown(choices=get_llm_formats(), value=get_dolphin_model_format(get_dolphin_models()[0][1]), label="Message format")
|
44 |
+
chat_sysmsg = gr.Textbox(value=get_dolphin_sysprompt(), label="System message")
|
45 |
+
chat_tokens = gr.Slider(minimum=1, maximum=4096, value=1024, step=1, label="Max tokens")
|
46 |
+
chat_temperature = gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature")
|
47 |
+
chat_topp = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p")
|
48 |
+
chat_topk = gr.Slider(minimum=0, maximum=100, value=40, step=1, label="Top-k")
|
49 |
+
chat_rp = gr.Slider(minimum=0.0, maximum=2.0, value=1.1, step=0.1, label="Repetition penalty")
|
50 |
+
with gr.Accordion("Add models", open=True):
|
51 |
+
chat_add_text = gr.Textbox(label="URL or Repo ID", placeholder="http://huggingface.co/.../...gguf or author/model", lines=1)
|
52 |
+
chat_add_format = gr.Dropdown(choices=get_llm_formats(), value=get_llm_formats()[0], label="Message format")
|
53 |
+
chat_add_submit = gr.Button("Update lists of models")
|
54 |
+
with gr.Accordion("Modes", open=True):
|
55 |
+
chat_model = gr.Dropdown(choices=get_dolphin_models(), value=get_dolphin_models()[0][1], allow_custom_value=True, label="Model")
|
56 |
+
chat_model_info = gr.Markdown(value=get_dolphin_model_info(get_dolphin_models()[0][1]), label="Model info")
|
57 |
+
with gr.Row():
|
58 |
+
chat_mode = gr.Dropdown(choices=get_dolphin_sysprompt_mode(), value=get_dolphin_sysprompt_mode()[0], allow_custom_value=False, label="Mode")
|
59 |
+
chat_lang = gr.Dropdown(choices=get_dolphin_languages(), value="English", allow_custom_value=True, label="Output language")
|
60 |
+
|
61 |
+
with gr.Column(scale=1):
|
62 |
+
with gr.Row():
|
63 |
+
with gr.Group():
|
64 |
+
output_text = gr.TextArea(label="Output tags", interactive=False, show_copy_button=True)
|
65 |
+
copy_btn = gr.Button(value="Copy to clipboard", size="sm", interactive=False)
|
66 |
+
elapsed_time_md = gr.Markdown(label="Elapsed time", value="", visible=False)
|
67 |
+
with gr.Group():
|
68 |
+
output_text_pony = gr.TextArea(label="Output tags (Pony e621 style)", interactive=False, show_copy_button=True)
|
69 |
+
copy_btn_pony = gr.Button(value="Copy to clipboard", size="sm", interactive=False)
|
70 |
+
with gr.Accordion(label="Advanced options", open=False, visible=False):
|
71 |
+
tag_type = gr.Radio(label="Output tag conversion", info="danbooru for Animagine, e621 for Pony.", choices=["danbooru", "e621"], value="e621", visible=False)
|
72 |
+
dummy_np = gr.Textbox(label="Negative prompt", value="", visible=False)
|
73 |
+
dummy_np_pony = gr.Textbox(label="Negative prompt", value="", visible=False)
|
74 |
+
recom_animagine = gr.Textbox(label="Animagine reccomended prompt", value="Animagine", visible=False)
|
75 |
+
recom_pony = gr.Textbox(label="Pony reccomended prompt", value="Pony", visible=False)
|
76 |
+
generate_image_btn = gr.Button(value="GENERATE IMAGE", size="lg", variant="primary")
|
77 |
+
result_image = gr.Gallery(label="Generated images", columns=1, object_fit="contain", container=True, preview=True, show_label=False, show_share_button=False, show_download_button=True, interactive=False, visible=True, format="png")
|
78 |
+
|
79 |
+
gr.on(
|
80 |
+
triggers=[chat_msg.submit, chat_submit.click],
|
81 |
+
fn=dolphin_respond,
|
82 |
+
inputs=[chat_msg, chatbot, chat_model, chat_sysmsg, chat_tokens, chat_temperature, chat_topp, chat_topk, chat_rp],
|
83 |
+
outputs=[chatbot],
|
84 |
+
queue=True,
|
85 |
+
show_progress="full",
|
86 |
+
trigger_mode="once",
|
87 |
+
).success(dolphin_parse, [chatbot], [output_text, copy_btn, copy_btn_pony]).success(
|
88 |
+
convert_danbooru_to_e621_prompt, [output_text, tag_type], [output_text_pony], queue=False,
|
89 |
+
).success(
|
90 |
+
insert_recom_prompt, [output_text, dummy_np, recom_animagine], [output_text, dummy_np], queue=False,
|
91 |
+
).success(
|
92 |
+
insert_recom_prompt, [output_text_pony, dummy_np_pony, recom_pony], [output_text_pony, dummy_np_pony], queue=False,
|
93 |
+
)
|
94 |
+
chat_clear.click(lambda: None, None, chatbot, queue=False)
|
95 |
+
chat_model.change(select_dolphin_model, [chat_model], [chat_model, chat_format, chat_model_info], queue=True, show_progress="full")\
|
96 |
+
.success(lambda: None, None, chatbot, queue=False)
|
97 |
+
chat_format.change(select_dolphin_format, [chat_format], [chat_format], queue=False)\
|
98 |
+
.success(lambda: None, None, chatbot, queue=False)
|
99 |
+
chat_mode.change(select_dolphin_sysprompt, [chat_mode], [chat_sysmsg], queue=False)
|
100 |
+
chat_lang.change(select_dolphin_language, [chat_lang], [chat_sysmsg], queue=False)
|
101 |
+
gr.on(
|
102 |
+
triggers=[chat_add_text.submit, chat_add_submit.click],
|
103 |
+
fn=add_dolphin_models,
|
104 |
+
inputs=[chat_add_text, chat_add_format],
|
105 |
+
outputs=[chat_model],
|
106 |
+
queue=False,
|
107 |
+
trigger_mode="once",
|
108 |
+
)
|
109 |
+
|
110 |
+
copy_btn.click(gradio_copy_text, [output_text], js=COPY_ACTION_JS)
|
111 |
+
copy_btn_pony.click(gradio_copy_text, [output_text_pony], js=COPY_ACTION_JS)
|
112 |
+
|
113 |
+
generate_image_btn.click(generate_image, [output_text, dummy_np], [result_image], show_progress="full")
|
114 |
+
|
115 |
+
|
116 |
+
if __name__ == "__main__":
|
117 |
+
app.queue()
|
118 |
+
app.launch()
|
character_series_dict.csv
ADDED
The diff for this file is too large to render.
See raw diff
|
|
danbooru_e621.csv
ADDED
The diff for this file is too large to render.
See raw diff
|
|
genimage.py
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import spaces
|
2 |
+
|
3 |
+
|
4 |
+
def load_pipeline():
|
5 |
+
from diffusers import DiffusionPipeline
|
6 |
+
import torch
|
7 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
8 |
+
pipe = DiffusionPipeline.from_pretrained(
|
9 |
+
"John6666/rae-diffusion-xl-v2-sdxl-spo-pcm",
|
10 |
+
custom_pipeline="lpw_stable_diffusion_xl",
|
11 |
+
torch_dtype=torch.float16,
|
12 |
+
)
|
13 |
+
pipe.to(device)
|
14 |
+
return pipe
|
15 |
+
|
16 |
+
|
17 |
+
def save_image(image, metadata, output_dir):
|
18 |
+
import os
|
19 |
+
import uuid
|
20 |
+
import json
|
21 |
+
from PIL import PngImagePlugin
|
22 |
+
filename = str(uuid.uuid4()) + ".png"
|
23 |
+
os.makedirs(output_dir, exist_ok=True)
|
24 |
+
filepath = os.path.join(output_dir, filename)
|
25 |
+
metadata_str = json.dumps(metadata)
|
26 |
+
info = PngImagePlugin.PngInfo()
|
27 |
+
info.add_text("metadata", metadata_str)
|
28 |
+
image.save(filepath, "PNG", pnginfo=info)
|
29 |
+
return filepath
|
30 |
+
|
31 |
+
|
32 |
+
pipe = load_pipeline()
|
33 |
+
|
34 |
+
|
35 |
+
@spaces.GPU
|
36 |
+
def generate_image(prompt, neg_prompt):
|
37 |
+
metadata = {
|
38 |
+
"prompt": prompt,
|
39 |
+
"negative_prompt": neg_prompt,
|
40 |
+
"resolution": f"{1024} x {1024}",
|
41 |
+
"guidance_scale": 7.5,
|
42 |
+
"num_inference_steps": 16,
|
43 |
+
"sampler": "Euler",
|
44 |
+
}
|
45 |
+
try:
|
46 |
+
images = pipe(
|
47 |
+
prompt=prompt,
|
48 |
+
prompt_2="anime artwork, anime style, studio anime, highly detailed, masterpiece, best quality, very aesthetic, absurdres",
|
49 |
+
negative_prompt=neg_prompt,
|
50 |
+
negative_prompt_2="lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract], photo, deformed, disfigured, low contrast, photo, deformed, disfigured, low contrast",
|
51 |
+
width=1024,
|
52 |
+
height=1024,
|
53 |
+
guidance_scale=7.5,
|
54 |
+
num_inference_steps=16,
|
55 |
+
output_type="pil",
|
56 |
+
clip_skip=1,
|
57 |
+
).images
|
58 |
+
if images:
|
59 |
+
image_paths = [
|
60 |
+
save_image(image, metadata, "./outputs")
|
61 |
+
for image in images
|
62 |
+
]
|
63 |
+
return image_paths
|
64 |
+
except Exception as e:
|
65 |
+
return []
|
66 |
+
|
ja_to_danbooru/character_series_dict.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
ja_to_danbooru/danbooru_tagtype_dict.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
ja_to_danbooru/ja_danbooru_dict.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
ja_to_danbooru/ja_to_danbooru.py
ADDED
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import re
|
3 |
+
from pathlib import Path
|
4 |
+
|
5 |
+
|
6 |
+
def load_json_dict(path: str):
|
7 |
+
import json
|
8 |
+
from pathlib import Path
|
9 |
+
dict = {}
|
10 |
+
if not Path(path).exists(): return dict
|
11 |
+
try:
|
12 |
+
with open(path, encoding='utf-8') as f:
|
13 |
+
dict = json.load(f)
|
14 |
+
except Exception:
|
15 |
+
print(f"Failed to open dictionary file: {path}")
|
16 |
+
return dict
|
17 |
+
return dict
|
18 |
+
|
19 |
+
|
20 |
+
ja_danbooru_dict = load_json_dict('ja_danbooru_dict.json')
|
21 |
+
char_series_dict = load_json_dict('character_series_dict.json')
|
22 |
+
tagtype_dict = load_json_dict('danbooru_tagtype_dict.json')
|
23 |
+
|
24 |
+
|
25 |
+
def jatags_to_danbooru_tags(jatags: list[str]):
|
26 |
+
from rapidfuzz.process import extractOne
|
27 |
+
from rapidfuzz.utils import default_process
|
28 |
+
keys = list(ja_danbooru_dict.keys())
|
29 |
+
ckeys = list(char_series_dict.keys())
|
30 |
+
tags = []
|
31 |
+
for jatag in jatags:
|
32 |
+
jatag = str(jatag).strip()
|
33 |
+
s = default_process(str(jatag))
|
34 |
+
e1 = extractOne(s, keys, processor=default_process, score_cutoff=90.0)
|
35 |
+
if e1:
|
36 |
+
tag = str(ja_danbooru_dict[e1[0]])
|
37 |
+
tags.append(tag)
|
38 |
+
if tag in tagtype_dict.keys() and tagtype_dict[tag] == "character":
|
39 |
+
cs = default_process(tag)
|
40 |
+
ce1 = extractOne(cs, ckeys, processor=default_process, score_cutoff=95.0)
|
41 |
+
if ce1:
|
42 |
+
series = str(char_series_dict[ce1[0]])
|
43 |
+
tags.append(series)
|
44 |
+
return tags
|
45 |
+
|
46 |
+
|
47 |
+
def jatags_to_danbooru(input_tag, input_file, output_file, is_append):
|
48 |
+
if input_file and Path(input_file).exists():
|
49 |
+
try:
|
50 |
+
with open(input_file, 'r', encoding='utf-8') as f:
|
51 |
+
input_tag = f.read()
|
52 |
+
except Exception:
|
53 |
+
print(f"Failed to open input file: {input_file}")
|
54 |
+
ja_tags = [tag.strip() for tag in input_tag.split(",")] if input_tag else []
|
55 |
+
tags = jatags_to_danbooru_tags(ja_tags)
|
56 |
+
output_tags = ja_tags + tags if is_append else tags
|
57 |
+
output_tag = ", ".join(output_tags)
|
58 |
+
if output_file:
|
59 |
+
try:
|
60 |
+
with open(output_file, mode='w', encoding="utf-8") as f:
|
61 |
+
f.write(output_tag)
|
62 |
+
except Exception:
|
63 |
+
print(f"Failed to write output file: {output_file}")
|
64 |
+
else:
|
65 |
+
print(output_tag)
|
66 |
+
return output_tag
|
67 |
+
|
68 |
+
|
69 |
+
if __name__ == "__main__":
|
70 |
+
parser = argparse.ArgumentParser()
|
71 |
+
parser.add_argument("--tags", default=None, type=str, required=False, help="Input tags.")
|
72 |
+
parser.add_argument("--file", default=None, type=str, required=False, help="Input tags from a text file.")
|
73 |
+
parser.add_argument("--out", default=None, type=str, help="Output to text file.")
|
74 |
+
parser.add_argument("--append", default=False, type=bool, help="Whether the output contains the input tags or not.")
|
75 |
+
|
76 |
+
args = parser.parse_args()
|
77 |
+
assert (args.tags, args.file) != (None, None), "Must provide --tags or --file!"
|
78 |
+
|
79 |
+
jatags_to_danbooru(args.tags, args.file, args.out, args.append)
|
80 |
+
|
81 |
+
|
82 |
+
# Usage:
|
83 |
+
# python ja_to_danbooru.py --tags "女の子, 大室櫻子"
|
84 |
+
# python danbooru_to_ja.py --file inputtag.txt
|
85 |
+
# python danbooru_to_ja.py --file inputtag.txt --append True
|
86 |
+
# Datasets: https://huggingface.co/datasets/p1atdev/danbooru-ja-tag-pair-20240715
|
87 |
+
# Datasets: https://github.com/ponapon280/danbooru-e621-converter
|
llmdolphin.py
ADDED
@@ -0,0 +1,527 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import spaces
|
3 |
+
from llama_cpp import Llama
|
4 |
+
from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
|
5 |
+
from llama_cpp_agent.providers import LlamaCppPythonProvider
|
6 |
+
from llama_cpp_agent.chat_history import BasicChatHistory
|
7 |
+
from llama_cpp_agent.chat_history.messages import Roles
|
8 |
+
from ja_to_danbooru.ja_to_danbooru import jatags_to_danbooru_tags
|
9 |
+
|
10 |
+
|
11 |
+
llm_models_dir = "./llm_models"
|
12 |
+
llm_models = {
|
13 |
+
"L3-8B-Celeste-v1-Q5_K_M.gguf": ["bartowski/L3-8B-Celeste-v1-GGUF", MessagesFormatterType.LLAMA_3],
|
14 |
+
"L3-8B-Celeste-V1.2-Q5_K_M.gguf": ["bartowski/L3-8B-Celeste-V1.2-GGUF", MessagesFormatterType.LLAMA_3],
|
15 |
+
"Llama-3-Nymeria-ELYZA-8B.i1-Q4_K_M.gguf": ["mradermacher/Llama-3-Nymeria-ELYZA-8B-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
16 |
+
"suzume-llama-3-8B-japanese.Q4_K_M.gguf": ["PrunaAI/lightblue-suzume-llama-3-8B-japanese-GGUF-smashed", MessagesFormatterType.LLAMA_3],
|
17 |
+
"suzume-llama-3-8B-multilingual-orpo-borda-top25.Q4_K_M.gguf": ["RichardErkhov/lightblue_-_suzume-llama-3-8B-multilingual-orpo-borda-top25-gguf", MessagesFormatterType.LLAMA_3],
|
18 |
+
"gemma-2-9b-it-SimPO.i1-Q4_K_M.gguf": ["mradermacher/gemma-2-9b-it-SimPO-i1-GGUF", MessagesFormatterType.ALPACA],
|
19 |
+
"Gemma-2-9B-It-SPPO-Iter3.Q4_K_M.iMatrix.gguf": ["MCZK/Gemma-2-9B-It-SPPO-Iter3-GGUF", MessagesFormatterType.ALPACA],
|
20 |
+
"Llama-3-NeuralPaca-8b.Q4_K_M.gguf": ["RichardErkhov/NeuralNovel_-_Llama-3-NeuralPaca-8b-gguf", MessagesFormatterType.ALPACA],
|
21 |
+
"SaoRPM-2x8B.i1-Q4_K_M.gguf": ["mradermacher/SaoRPM-2x8B-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
22 |
+
"L3-Hecate-8B-v1.2.Q4_K_M.gguf": ["mradermacher/L3-Hecate-8B-v1.2-GGUF", MessagesFormatterType.LLAMA_3],
|
23 |
+
"Mahou-1.3b-llama3-8B.i1-Q4_K_M.gguf": ["mradermacher/Mahou-1.3b-llama3-8B-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
24 |
+
"SwallowMaid-8B-L3-SPPO-abliterated.i1-Q5_K_M.gguf": ["mradermacher/SwallowMaid-8B-L3-SPPO-abliterated-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
25 |
+
"L3-8B-Lunar-Stheno.i1-Q5_K_M.gguf": ["mradermacher/L3-8B-Lunar-Stheno-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
26 |
+
"llama3_Loradent.Q4_K_M.gguf": ["mradermacher/llama3_Loradent-GGUF", MessagesFormatterType.LLAMA_3],
|
27 |
+
"Llama-3-8B-Stroganoff.i1-Q4_K_M.gguf": ["mradermacher/Llama-3-8B-Stroganoff-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
28 |
+
"L3-8B-EnchantedForest-v0.5.i1-Q4_K_M.gguf": ["mradermacher/L3-8B-EnchantedForest-v0.5-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
29 |
+
"gemma-radiation-rp-9b-q5_k_m.gguf": ["pegasus912/Gemma-Radiation-RP-9B-Q5_K_M-GGUF", MessagesFormatterType.MISTRAL],
|
30 |
+
"Magic-Dolphin-7b.Q4_K_M.gguf": ["mradermacher/Magic-Dolphin-7b-GGUF", MessagesFormatterType.MISTRAL],
|
31 |
+
"mathstral-7B-v0.1-Q5_K_M.gguf": ["bartowski/mathstral-7B-v0.1-GGUF", MessagesFormatterType.MISTRAL],
|
32 |
+
"Gemma2-9B-it-Boku-v1.Q5_K_M.gguf": ["mradermacher/Gemma2-9B-it-Boku-v1-GGUF", MessagesFormatterType.MISTRAL],
|
33 |
+
"Gemma-2-9B-It-SPPO-Iter3-Q5_K_M.gguf": ["grapevine-AI/Gemma-2-9B-It-SPPO-Iter3-GGUF", MessagesFormatterType.MISTRAL],
|
34 |
+
"L3-8B-Niitama-v1.i1-Q4_K_M.gguf": ["mradermacher/L3-8B-Niitama-v1-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
35 |
+
"Maidphin-Kunoichi-7B.Q5_K_M.gguf": ["RichardErkhov/nbeerbower_-_Maidphin-Kunoichi-7B-gguf", MessagesFormatterType.MISTRAL],
|
36 |
+
"L3-15B-EtherealMaid-t0.0001.i1-Q4_K_M.gguf": ["mradermacher/L3-15B-EtherealMaid-t0.0001-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
37 |
+
"L3-15B-MythicalMaid-t0.0001.i1-Q4_K_M.gguf": ["mradermacher/L3-15B-MythicalMaid-t0.0001-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
38 |
+
"llama-3-Nephilim-v3-8B.Q5_K_M.gguf": ["grimjim/llama-3-Nephilim-v3-8B-GGUF", MessagesFormatterType.LLAMA_3],
|
39 |
+
"NarutoDolphin-10B.Q5_K_M.gguf": ["RichardErkhov/FelixChao_-_NarutoDolphin-10B-gguf", MessagesFormatterType.MISTRAL],
|
40 |
+
"l3-8b-tamamo-v1-q8_0.gguf": ["Ransss/L3-8B-Tamamo-v1-Q8_0-GGUF", MessagesFormatterType.LLAMA_3],
|
41 |
+
"Tiger-Gemma-9B-v1-Q4_K_M.gguf": ["bartowski/Tiger-Gemma-9B-v1-GGUF", MessagesFormatterType.LLAMA_3],
|
42 |
+
"TooManyMixRolePlay-7B-Story_V3.5.Q4_K_M.gguf": ["mradermacher/TooManyMixRolePlay-7B-Story_V3.5-GGUF", MessagesFormatterType.LLAMA_3],
|
43 |
+
"natsumura-llama3-v1.1-8b.Q4_K_M.gguf": ["mradermacher/natsumura-llama3-v1.1-8b-GGUF", MessagesFormatterType.LLAMA_3],
|
44 |
+
"natsumura-llama3-v1-8b.i1-Q4_K_M.gguf": ["mradermacher/natsumura-llama3-v1-8b-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
45 |
+
"nephra_v1.0.Q5_K_M.gguf": ["PrunaAI/yodayo-ai-nephra_v1.0-GGUF-smashed", MessagesFormatterType.LLAMA_3],
|
46 |
+
"DPO-ONLY-Zephyr-7B.Q6_K.gguf": ["mradermacher/DPO-ONLY-Zephyr-7B-GGUF", MessagesFormatterType.LLAMA_3],
|
47 |
+
"L3-Deluxe-Scrambled-Eggs-On-Toast-8B.Q8_0.gguf": ["mradermacher/L3-Deluxe-Scrambled-Eggs-On-Toast-8B-GGUF", MessagesFormatterType.LLAMA_3],
|
48 |
+
"L3-Scrambled-Eggs-On-Toast-8B.i1-Q6_K.gguf": ["mradermacher/L3-Scrambled-Eggs-On-Toast-8B-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
49 |
+
"Llama-3-uncensored-Dare-1.Q4_K_M.gguf": ["mradermacher/Llama-3-uncensored-Dare-1-GGUF", MessagesFormatterType.LLAMA_3],
|
50 |
+
"llama3-8B-DarkIdol-2.2-Uncensored-1048K.i1-Q6_K.gguf": ["mradermacher/llama3-8B-DarkIdol-2.2-Uncensored-1048K-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
51 |
+
"dolphin-2.9.3-mistral-7b-32k-q4_k_m.gguf": ["huggingkot/dolphin-2.9.3-mistral-7B-32k-Q4_K_M-GGUF", MessagesFormatterType.MISTRAL],
|
52 |
+
"dolphin-2.9.3-mistral-7B-32k-Q5_K_M.gguf": ["bartowski/dolphin-2.9.3-mistral-7B-32k-GGUF", MessagesFormatterType.MISTRAL],
|
53 |
+
"Lexi-Llama-3-8B-Uncensored_Q5_K_M.gguf": ["Orenguteng/Llama-3-8B-Lexi-Uncensored-GGUF", MessagesFormatterType.LLAMA_3],
|
54 |
+
"Llama3-Sophie.Q8_0.gguf": ["mradermacher/Llama3-Sophie-GGUF", MessagesFormatterType.LLAMA_3],
|
55 |
+
"Aura-Uncensored-OAS-8B-L3.i1-Q4_K_M.gguf": ["mradermacher/Aura-Uncensored-OAS-8B-L3-i1-GGUF", MessagesFormatterType.LLAMA_3],
|
56 |
+
"L3-Uncen-Merger-Omelette-RP-v0.2-8B-Q5_K_S-imat.gguf": ["LWDCLS/L3-Uncen-Merger-Omelette-RP-v0.2-8B-GGUF-IQ-Imatrix-Request", MessagesFormatterType.LLAMA_3],
|
57 |
+
"qwen2-diffusion-prompter-v01-q6_k.gguf": ["trollek/Qwen2-0.5B-DiffusionPrompter-v0.1-GGUF", MessagesFormatterType.LLAMA_3],
|
58 |
+
"Smegmma-Deluxe-9B-v1-Q6_K.gguf": ["bartowski/Smegmma-Deluxe-9B-v1-GGUF", MessagesFormatterType.MISTRAL],
|
59 |
+
"Mahou-1.3c-mistral-7B.i1-Q6_K.gguf": ["mradermacher/Mahou-1.3c-mistral-7B-i1-GGUF", MessagesFormatterType.MISTRAL],
|
60 |
+
"Silicon-Maid-7B-Q8_0_X.gguf": ["duyntnet/Silicon-Maid-7B-imatrix-GGUF", MessagesFormatterType.ALPACA],
|
61 |
+
"l3-umbral-mind-rp-v3.0-8b-q5_k_m-imat.gguf": ["Casual-Autopsy/L3-Umbral-Mind-RP-v3.0-8B-Q5_K_M-GGUF", MessagesFormatterType.LLAMA_3],
|
62 |
+
"Phi-3.1-mini-128k-instruct-Q6_K_L.gguf": ["bartowski/Phi-3.1-mini-128k-instruct-GGUF", MessagesFormatterType.PHI_3],
|
63 |
+
"tifa-7b-qwen2-v0.1.q4_k_m.gguf": ["Tifa-RP/Tifa-7B-Qwen2-v0.1-GGUF", MessagesFormatterType.OPEN_CHAT],
|
64 |
+
"Llama-3-EZO-8b-Common-it.Q5_K_M.iMatrix.gguf": ["MCZK/Llama-3-EZO-8b-Common-it-GGUF", MessagesFormatterType.MISTRAL],
|
65 |
+
"EZO-Common-9B-gemma-2-it.i1-Q4_K_M.gguf": ["mradermacher/EZO-Common-9B-gemma-2-it-i1-GGUF", MessagesFormatterType.MISTRAL],
|
66 |
+
#"": ["", MessagesFormatterType.LLAMA_3],
|
67 |
+
#"": ["", MessagesFormatterType.MISTRAL],
|
68 |
+
#"": ["", MessagesFormatterType.ALPACA],
|
69 |
+
#"": ["", MessagesFormatterType.OPEN_CHAT],
|
70 |
+
}
|
71 |
+
llm_formats = {
|
72 |
+
"MISTRAL": MessagesFormatterType.MISTRAL,
|
73 |
+
"CHATML": MessagesFormatterType.CHATML,
|
74 |
+
"VICUNA": MessagesFormatterType.VICUNA,
|
75 |
+
"LLAMA 2": MessagesFormatterType.LLAMA_2,
|
76 |
+
"SYNTHIA": MessagesFormatterType.SYNTHIA,
|
77 |
+
"NEURAL CHAT": MessagesFormatterType.NEURAL_CHAT,
|
78 |
+
"SOLAR": MessagesFormatterType.SOLAR,
|
79 |
+
"OPEN CHAT": MessagesFormatterType.OPEN_CHAT,
|
80 |
+
"ALPACA": MessagesFormatterType.ALPACA,
|
81 |
+
"CODE DS": MessagesFormatterType.CODE_DS,
|
82 |
+
"B22": MessagesFormatterType.B22,
|
83 |
+
"LLAMA 3": MessagesFormatterType.LLAMA_3,
|
84 |
+
"PHI 3": MessagesFormatterType.PHI_3,
|
85 |
+
"Autocoder": MessagesFormatterType.AUTOCODER,
|
86 |
+
"DeepSeek Coder v2": MessagesFormatterType.DEEP_SEEK_CODER_2,
|
87 |
+
"Gemma 2": MessagesFormatterType.ALPACA,
|
88 |
+
"Qwen2": MessagesFormatterType.OPEN_CHAT,
|
89 |
+
}
|
90 |
+
# https://github.com/Maximilian-Winter/llama-cpp-agent
|
91 |
+
llm_languages = ["English", "Japanese", "Chinese"]
|
92 |
+
llm_models_tupled_list = []
|
93 |
+
default_llm_model_filename = list(llm_models.keys())[0]
|
94 |
+
override_llm_format = None
|
95 |
+
|
96 |
+
|
97 |
+
def to_list(s):
|
98 |
+
return [x.strip() for x in s.split(",") if not s == ""]
|
99 |
+
|
100 |
+
|
101 |
+
def list_uniq(l):
|
102 |
+
return sorted(set(l), key=l.index)
|
103 |
+
|
104 |
+
|
105 |
+
def to_list_ja(s):
|
106 |
+
import re
|
107 |
+
s = re.sub(r'[、。]', ',', s)
|
108 |
+
return [x.strip() for x in s.split(",") if not s == ""]
|
109 |
+
|
110 |
+
|
111 |
+
def is_japanese(s):
|
112 |
+
import unicodedata
|
113 |
+
for ch in s:
|
114 |
+
name = unicodedata.name(ch, "")
|
115 |
+
if "CJK UNIFIED" in name or "HIRAGANA" in name or "KATAKANA" in name:
|
116 |
+
return True
|
117 |
+
return False
|
118 |
+
|
119 |
+
|
120 |
+
def update_llm_model_tupled_list():
|
121 |
+
from pathlib import Path
|
122 |
+
global llm_models_tupled_list
|
123 |
+
llm_models_tupled_list = []
|
124 |
+
for k, v in llm_models.items():
|
125 |
+
name = k
|
126 |
+
value = k
|
127 |
+
llm_models_tupled_list.append((name, value))
|
128 |
+
model_files = Path(llm_models_dir).glob('*.gguf')
|
129 |
+
for path in model_files:
|
130 |
+
name = path.name
|
131 |
+
value = path.name
|
132 |
+
llm_models_tupled_list.append((name, value))
|
133 |
+
llm_models_tupled_list = list_uniq(llm_models_tupled_list)
|
134 |
+
return llm_models_tupled_list
|
135 |
+
|
136 |
+
|
137 |
+
def download_llm_models():
|
138 |
+
from huggingface_hub import hf_hub_download
|
139 |
+
global llm_models_tupled_list
|
140 |
+
llm_models_tupled_list = []
|
141 |
+
for k, v in llm_models.items():
|
142 |
+
try:
|
143 |
+
hf_hub_download(repo_id = v[0], filename = k, local_dir = llm_models_dir)
|
144 |
+
except Exception:
|
145 |
+
continue
|
146 |
+
name = k
|
147 |
+
value = k
|
148 |
+
llm_models_tupled_list.append((name, value))
|
149 |
+
|
150 |
+
|
151 |
+
def download_llm_model(filename):
|
152 |
+
from huggingface_hub import hf_hub_download
|
153 |
+
if not filename in llm_models.keys(): return default_llm_model_filename
|
154 |
+
try:
|
155 |
+
hf_hub_download(repo_id = llm_models[filename][0], filename = filename, local_dir = llm_models_dir)
|
156 |
+
except Exception:
|
157 |
+
return default_llm_model_filename
|
158 |
+
update_llm_model_tupled_list()
|
159 |
+
return filename
|
160 |
+
|
161 |
+
|
162 |
+
def get_dolphin_model_info(filename):
|
163 |
+
md = "None"
|
164 |
+
items = llm_models.get(filename, None)
|
165 |
+
if items:
|
166 |
+
md = f'Repo: [{items[0]}](https://huggingface.co/{items[0]})'
|
167 |
+
return md
|
168 |
+
|
169 |
+
|
170 |
+
def select_dolphin_model(filename, progress=gr.Progress(track_tqdm=True)):
|
171 |
+
global override_llm_format
|
172 |
+
override_llm_format = None
|
173 |
+
progress(0, desc="Loading model...")
|
174 |
+
value = download_llm_model(filename)
|
175 |
+
progress(1, desc="Model loaded.")
|
176 |
+
md = get_dolphin_model_info(filename)
|
177 |
+
return gr.update(value=value, choices=get_dolphin_models()), gr.update(value=get_dolphin_model_format(value)), gr.update(value=md)
|
178 |
+
|
179 |
+
|
180 |
+
def select_dolphin_format(format_name):
|
181 |
+
global override_llm_format
|
182 |
+
override_llm_format = llm_formats[format_name]
|
183 |
+
return gr.update(value=format_name)
|
184 |
+
|
185 |
+
|
186 |
+
#download_llm_models()
|
187 |
+
download_llm_model(default_llm_model_filename)
|
188 |
+
|
189 |
+
|
190 |
+
def get_dolphin_models():
|
191 |
+
return update_llm_model_tupled_list()
|
192 |
+
|
193 |
+
|
194 |
+
def get_llm_formats():
|
195 |
+
return list(llm_formats.keys())
|
196 |
+
|
197 |
+
|
198 |
+
def get_key_from_value(d, val):
|
199 |
+
keys = [k for k, v in d.items() if v == val]
|
200 |
+
if keys:
|
201 |
+
return keys[0]
|
202 |
+
return None
|
203 |
+
|
204 |
+
|
205 |
+
def get_dolphin_model_format(filename):
|
206 |
+
if not filename in llm_models.keys(): filename = default_llm_model_filename
|
207 |
+
format = llm_models[filename][1]
|
208 |
+
format_name = get_key_from_value(llm_formats, format)
|
209 |
+
return format_name
|
210 |
+
|
211 |
+
|
212 |
+
def add_dolphin_models(query, format_name):
|
213 |
+
import re
|
214 |
+
from huggingface_hub import HfApi
|
215 |
+
global llm_models
|
216 |
+
api = HfApi()
|
217 |
+
add_models = {}
|
218 |
+
format = llm_formats[format_name]
|
219 |
+
filename = ""
|
220 |
+
repo = ""
|
221 |
+
try:
|
222 |
+
s = list(re.findall(r'^(?:https?://huggingface.co/)?(.+?/.+?)(?:/.*/(.+?.gguf).*?)?$', query)[0])
|
223 |
+
if s and "" in s: s.remove("")
|
224 |
+
if len(s) == 1:
|
225 |
+
repo = s[0]
|
226 |
+
if not api.repo_exists(repo_id = repo): return gr.update(visible=True)
|
227 |
+
files = api.list_repo_files(repo_id = repo)
|
228 |
+
for file in files:
|
229 |
+
if str(file).endswith(".gguf"): add_models[filename] = [repo, format]
|
230 |
+
elif len(s) >= 2:
|
231 |
+
repo = s[0]
|
232 |
+
filename = s[1]
|
233 |
+
if not api.repo_exists(repo_id = repo) or not api.file_exists(repo_id = repo, filename = filename): return gr.update(visible=True)
|
234 |
+
add_models[filename] = [repo, format]
|
235 |
+
else: return gr.update(visible=True)
|
236 |
+
except Exception:
|
237 |
+
return gr.update(visible=True)
|
238 |
+
print(add_models)
|
239 |
+
llm_models = (llm_models | add_models).copy()
|
240 |
+
return gr.update(choices=get_dolphin_models())
|
241 |
+
|
242 |
+
|
243 |
+
dolphin_output_language = "English"
|
244 |
+
dolphin_sysprompt_mode = "Default"
|
245 |
+
dolphin_system_prompt = {"Default": r'''You are a helpful AI assistant to generate messages for AI that outputs an image when I enter a message.
|
246 |
+
The message must have the following [Tags] generated in strict accordance with the following [Rules]:
|
247 |
+
```
|
248 |
+
[Tags]
|
249 |
+
- Words to describe full names of characters and names of series in which they appear.
|
250 |
+
- Words to describe names of the people there and their numbers, such as 2girls, 1boy.
|
251 |
+
- Words to describe their hair color, hairstyle, hair length, hair accessory, eye color, eye shape, facial expression, breast size, and clothing of them in detail, such as long hair.
|
252 |
+
- Words to describe their external features, ornaments and belongings (also specify colors, patterns, shapes) in detail.
|
253 |
+
- Words to describe their stance from head to toe in detail.
|
254 |
+
- Words to describe their acting, especially with sexual activity in detail.
|
255 |
+
- Words to describe their surroundings in detail.
|
256 |
+
- Words to describe background details, such as inside room, forest, starry sky.
|
257 |
+
[Rules]
|
258 |
+
- Any output should be plain text in English and don't use line breaks.
|
259 |
+
- Output only composed of Tags in 1 line, separated by commas with spaces between Tags, in lower case English.
|
260 |
+
- Output should be in the format: "//GENBEGIN//://1girl, Tag, Tag, ..., Tag//://GENEND//".
|
261 |
+
- Preferably refer to and describe the information obtained from Danbooru. If not, describe it in own way.
|
262 |
+
- It's preferable that each Tag is a plain phrase, word, caption, Danbooru tag, or E621 tag.
|
263 |
+
- Convert any nicknames to full names first.
|
264 |
+
- If a sexual theme is given, priority should be given to specific and rich descriptions of sexual activity, especially about genitals, fluids.
|
265 |
+
- Assemble a short story internally which is developed from the themes provided, then describe a scene into an detailed English sentences based on the central character internally.
|
266 |
+
- Split sentences into short phrases or words, and then convert them to Tags.
|
267 |
+
- Use associated Danbooru tags, E621 tags.
|
268 |
+
- Same Tags should be used only once per output.
|
269 |
+
- Anyway, keep processing until you've finished outputting message.
|
270 |
+
```
|
271 |
+
Based on these Rules, please tell me message within 40 Tags that can generate an image for the following themes:
|
272 |
+
''',
|
273 |
+
"With dialogue and description": r'''You are a helpful AI assistant to generate messages for AI that outputs an image when I enter a message.
|
274 |
+
The message must have the following [Tags] generated in strict accordance with the following [Rules]:
|
275 |
+
```
|
276 |
+
[Tags]
|
277 |
+
- Words to describe full names of characters and names of series in which they appear.
|
278 |
+
- Words to describe names of the people there and their numbers, such as 2girls, 1boy.
|
279 |
+
- Words to describe their hair color, hairstyle, hair length, hair accessory, eye color, eye shape, facial expression, breast size, and clothing of them in detail, such as long hair.
|
280 |
+
- Words to describe their external features, ornaments and belongings (also specify colors, patterns, shapes) in detail.
|
281 |
+
- Words to describe their stance from head to toe in detail.
|
282 |
+
- Words to describe their acting, especially with sexual activity in detail.
|
283 |
+
- Words to describe their surroundings in detail.
|
284 |
+
- Words to describe background details, such as inside room, forest, starry sky.
|
285 |
+
[Rules]
|
286 |
+
- Any Tags should be plain text in English and don't use line breaks.
|
287 |
+
- Message is only composed of Tags in 1 line, separated by commas with spaces between Tags, in lower case English.
|
288 |
+
- Message should be in the format: "//GENBEGIN//://1girl, Tag, Tag, ..., Tag//://GENEND//".
|
289 |
+
- Preferably refer to and describe the information obtained from Danbooru. If not, describe it in own way.
|
290 |
+
- It's preferable that each Tag is a plain phrase, word, caption, Danbooru tag, or E621 tag.
|
291 |
+
- Convert any nicknames to full names first.
|
292 |
+
- If a sexual theme is given, priority should be given to specific and rich descriptions of sexual activity, especially about genitals, fluids.
|
293 |
+
- Assemble a short story internally which is developed from the themes provided, then describe a scene into an detailed English sentences based on the central character internally.
|
294 |
+
- Split sentences into short phrases or words, and then convert them to Tags.
|
295 |
+
- Use associated Danbooru tags, E621 tags.
|
296 |
+
- Same Tags should be used only once per output.
|
297 |
+
- Anyway, keep processing until you've finished outputting message.
|
298 |
+
```
|
299 |
+
Based on these Rules, please tell me message within 40 Tags that can generate an image for the following themes,
|
300 |
+
then write the character's long actor's line composed of one's voices and moaning and voices in thought, based on the story you have assembled, in <LANGUAGE> only,
|
301 |
+
enclosed in //VOICEBEGIN//:// and //://VOICEEND//, then describe the message you've generated in short, in <LANGUAGE> only.:
|
302 |
+
''', "Japanese to Danbooru Dictionary": r"""You are a helpful AI assistant.
|
303 |
+
Extract Japanese words from the following sentences and output them separated by commas. Convert words in their original forms.
|
304 |
+
Output should be enclosed in //GENBEGIN//:// and //://GENEND//. The text to be given is as follows:""",
|
305 |
+
"Chat with LLM": r"You are a helpful AI assistant. Respond in <LANGUAGE>."}
|
306 |
+
|
307 |
+
|
308 |
+
def get_dolphin_sysprompt():
|
309 |
+
import re
|
310 |
+
prompt = re.sub('<LANGUAGE>', dolphin_output_language, dolphin_system_prompt.get(dolphin_sysprompt_mode, ""))
|
311 |
+
return prompt
|
312 |
+
|
313 |
+
|
314 |
+
def get_dolphin_sysprompt_mode():
|
315 |
+
return list(dolphin_system_prompt.keys())
|
316 |
+
|
317 |
+
|
318 |
+
def select_dolphin_sysprompt(key: str):
|
319 |
+
global dolphin_sysprompt_mode
|
320 |
+
if not key in dolphin_system_prompt.keys():
|
321 |
+
dolphin_sysprompt_mode = "Default"
|
322 |
+
else:
|
323 |
+
dolphin_sysprompt_mode = key
|
324 |
+
return gr.update(value=get_dolphin_sysprompt())
|
325 |
+
|
326 |
+
|
327 |
+
def get_dolphin_languages():
|
328 |
+
return llm_languages
|
329 |
+
|
330 |
+
|
331 |
+
def select_dolphin_language(lang: str):
|
332 |
+
global dolphin_output_language
|
333 |
+
dolphin_output_language = lang
|
334 |
+
return gr.update(value=get_dolphin_sysprompt())
|
335 |
+
|
336 |
+
|
337 |
+
@spaces.GPU
|
338 |
+
def dolphin_respond(
|
339 |
+
message: str,
|
340 |
+
history: list[tuple[str, str]],
|
341 |
+
model: str = default_llm_model_filename,
|
342 |
+
system_message: str = get_dolphin_sysprompt(),
|
343 |
+
max_tokens: int = 1024,
|
344 |
+
temperature: float = 0.7,
|
345 |
+
top_p: float = 0.95,
|
346 |
+
top_k: int = 40,
|
347 |
+
repeat_penalty: float = 1.1,
|
348 |
+
progress=gr.Progress(track_tqdm=True),
|
349 |
+
):
|
350 |
+
from pathlib import Path
|
351 |
+
progress(0, desc="Processing...")
|
352 |
+
|
353 |
+
if override_llm_format:
|
354 |
+
chat_template = override_llm_format
|
355 |
+
else:
|
356 |
+
chat_template = llm_models[model][1]
|
357 |
+
|
358 |
+
llm = Llama(
|
359 |
+
model_path=str(Path(f"{llm_models_dir}/{model}")),
|
360 |
+
flash_attn=True,
|
361 |
+
n_gpu_layers=81,
|
362 |
+
n_batch=1024,
|
363 |
+
n_ctx=8192,
|
364 |
+
)
|
365 |
+
provider = LlamaCppPythonProvider(llm)
|
366 |
+
|
367 |
+
agent = LlamaCppAgent(
|
368 |
+
provider,
|
369 |
+
system_prompt=f"{system_message}",
|
370 |
+
predefined_messages_formatter_type=chat_template,
|
371 |
+
debug_output=False
|
372 |
+
)
|
373 |
+
|
374 |
+
settings = provider.get_provider_default_settings()
|
375 |
+
settings.temperature = temperature
|
376 |
+
settings.top_k = top_k
|
377 |
+
settings.top_p = top_p
|
378 |
+
settings.max_tokens = max_tokens
|
379 |
+
settings.repeat_penalty = repeat_penalty
|
380 |
+
settings.stream = True
|
381 |
+
|
382 |
+
messages = BasicChatHistory()
|
383 |
+
|
384 |
+
for msn in history:
|
385 |
+
user = {
|
386 |
+
'role': Roles.user,
|
387 |
+
'content': msn[0]
|
388 |
+
}
|
389 |
+
assistant = {
|
390 |
+
'role': Roles.assistant,
|
391 |
+
'content': msn[1]
|
392 |
+
}
|
393 |
+
messages.add_message(user)
|
394 |
+
messages.add_message(assistant)
|
395 |
+
|
396 |
+
stream = agent.get_chat_response(
|
397 |
+
message,
|
398 |
+
llm_sampling_settings=settings,
|
399 |
+
chat_history=messages,
|
400 |
+
returns_streaming_generator=True,
|
401 |
+
print_output=False
|
402 |
+
)
|
403 |
+
|
404 |
+
progress(0.5, desc="Processing...")
|
405 |
+
|
406 |
+
outputs = ""
|
407 |
+
for output in stream:
|
408 |
+
outputs += output
|
409 |
+
yield [(outputs, None)]
|
410 |
+
|
411 |
+
|
412 |
+
def dolphin_parse(
|
413 |
+
history: list[tuple[str, str]],
|
414 |
+
):
|
415 |
+
import re
|
416 |
+
if dolphin_sysprompt_mode == "Chat with LLM" or not history or len(history) < 1: "", gr.update(visible=True), gr.update(visible=True)
|
417 |
+
try:
|
418 |
+
msg = history[-1][0]
|
419 |
+
except Exception:
|
420 |
+
return ""
|
421 |
+
m = re.findall(r'/GENBEGIN/((?:.|\s)+?)/GENEND/', msg)
|
422 |
+
raw_prompt = re.sub(r'[*/:_"#]|\n', ' ', ", ".join(m)).lower() if m else ""
|
423 |
+
prompts = []
|
424 |
+
if dolphin_sysprompt_mode == "Japanese to Danbooru Dictionary" and is_japanese(raw_prompt):
|
425 |
+
prompts = list_uniq(jatags_to_danbooru_tags(to_list_ja(raw_prompt)) + ["nsfw", "explicit"])
|
426 |
+
else:
|
427 |
+
prompts = list_uniq(to_list(raw_prompt) + ["nsfw", "explicit"])
|
428 |
+
return ", ".join(prompts), gr.update(interactive=True), gr.update(interactive=True)
|
429 |
+
|
430 |
+
|
431 |
+
@spaces.GPU
|
432 |
+
def dolphin_respond_auto(
|
433 |
+
message: str,
|
434 |
+
history: list[tuple[str, str]],
|
435 |
+
model: str = default_llm_model_filename,
|
436 |
+
system_message: str = get_dolphin_sysprompt(),
|
437 |
+
max_tokens: int = 1024,
|
438 |
+
temperature: float = 0.7,
|
439 |
+
top_p: float = 0.95,
|
440 |
+
top_k: int = 40,
|
441 |
+
repeat_penalty: float = 1.1,
|
442 |
+
progress=gr.Progress(track_tqdm=True),
|
443 |
+
):
|
444 |
+
#if not is_japanese(message): return [(None, None)]
|
445 |
+
|
446 |
+
from pathlib import Path
|
447 |
+
progress(0, desc="Processing...")
|
448 |
+
|
449 |
+
if override_llm_format:
|
450 |
+
chat_template = override_llm_format
|
451 |
+
else:
|
452 |
+
chat_template = llm_models[model][1]
|
453 |
+
|
454 |
+
llm = Llama(
|
455 |
+
model_path=str(Path(f"{llm_models_dir}/{model}")),
|
456 |
+
flash_attn=True,
|
457 |
+
n_gpu_layers=81,
|
458 |
+
n_batch=1024,
|
459 |
+
n_ctx=8192,
|
460 |
+
)
|
461 |
+
provider = LlamaCppPythonProvider(llm)
|
462 |
+
|
463 |
+
agent = LlamaCppAgent(
|
464 |
+
provider,
|
465 |
+
system_prompt=f"{system_message}",
|
466 |
+
predefined_messages_formatter_type=chat_template,
|
467 |
+
debug_output=False
|
468 |
+
)
|
469 |
+
|
470 |
+
settings = provider.get_provider_default_settings()
|
471 |
+
settings.temperature = temperature
|
472 |
+
settings.top_k = top_k
|
473 |
+
settings.top_p = top_p
|
474 |
+
settings.max_tokens = max_tokens
|
475 |
+
settings.repeat_penalty = repeat_penalty
|
476 |
+
settings.stream = True
|
477 |
+
|
478 |
+
messages = BasicChatHistory()
|
479 |
+
|
480 |
+
for msn in history:
|
481 |
+
user = {
|
482 |
+
'role': Roles.user,
|
483 |
+
'content': msn[0]
|
484 |
+
}
|
485 |
+
assistant = {
|
486 |
+
'role': Roles.assistant,
|
487 |
+
'content': msn[1]
|
488 |
+
}
|
489 |
+
messages.add_message(user)
|
490 |
+
messages.add_message(assistant)
|
491 |
+
|
492 |
+
progress(0, desc="Translating...")
|
493 |
+
stream = agent.get_chat_response(
|
494 |
+
message,
|
495 |
+
llm_sampling_settings=settings,
|
496 |
+
chat_history=messages,
|
497 |
+
returns_streaming_generator=True,
|
498 |
+
print_output=False
|
499 |
+
)
|
500 |
+
|
501 |
+
progress(0.5, desc="Processing...")
|
502 |
+
|
503 |
+
outputs = ""
|
504 |
+
for output in stream:
|
505 |
+
outputs += output
|
506 |
+
yield [(outputs, None)]
|
507 |
+
|
508 |
+
|
509 |
+
def dolphin_parse_simple(
|
510 |
+
message: str,
|
511 |
+
history: list[tuple[str, str]],
|
512 |
+
):
|
513 |
+
import re
|
514 |
+
#if not is_japanese(message) or not history or len(history) < 1: return message
|
515 |
+
if dolphin_sysprompt_mode == "Chat with LLM" or not history or len(history) < 1: return message
|
516 |
+
try:
|
517 |
+
msg = history[-1][0]
|
518 |
+
except Exception:
|
519 |
+
return ""
|
520 |
+
m = re.findall(r'/GENBEGIN/((?:.|\s)+?)/GENEND/', msg)
|
521 |
+
raw_prompt = re.sub(r'[*/:_"#]|\n', ' ', ", ".join(m)).lower() if m else ""
|
522 |
+
prompts = []
|
523 |
+
if dolphin_sysprompt_mode == "Japanese to Danbooru Dictionary" and is_japanese(raw_prompt):
|
524 |
+
prompts = list_uniq(jatags_to_danbooru_tags(to_list_ja(raw_prompt)) + ["nsfw", "explicit"])
|
525 |
+
else:
|
526 |
+
prompts = list_uniq(to_list(raw_prompt) + ["nsfw", "explicit"])
|
527 |
+
return ", ".join(prompts)
|
pre-requirements.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
pip>=23.0.0
|
requirements.txt
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
huggingface_hub
|
2 |
+
scikit-build-core
|
3 |
+
https://github.com/abetlen/llama-cpp-python/releases/download/v0.2.82-cu124/llama_cpp_python-0.2.82-cp310-cp310-linux_x86_64.whl
|
4 |
+
git+https://github.com/Maximilian-Winter/llama-cpp-agent
|
5 |
+
pybind11>=2.12
|
6 |
+
torch
|
7 |
+
torchvision
|
8 |
+
accelerate
|
9 |
+
transformers
|
10 |
+
optimum[onnxruntime]
|
11 |
+
spaces
|
12 |
+
dartrs
|
13 |
+
httpx==0.13.3
|
14 |
+
httpcore
|
15 |
+
googletrans==4.0.0rc1
|
16 |
+
git+https://github.com/huggingface/diffusers
|
17 |
+
rapidfuzz
|
tag_group.csv
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tagger.py
ADDED
@@ -0,0 +1,506 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from PIL import Image
|
2 |
+
import torch
|
3 |
+
import gradio as gr
|
4 |
+
import spaces # ZERO GPU
|
5 |
+
|
6 |
+
from transformers import (
|
7 |
+
AutoImageProcessor,
|
8 |
+
AutoModelForImageClassification,
|
9 |
+
)
|
10 |
+
|
11 |
+
WD_MODEL_NAMES = ["p1atdev/wd-swinv2-tagger-v3-hf"]
|
12 |
+
WD_MODEL_NAME = WD_MODEL_NAMES[0]
|
13 |
+
|
14 |
+
wd_model = AutoModelForImageClassification.from_pretrained(WD_MODEL_NAME, trust_remote_code=True)
|
15 |
+
wd_model.to("cuda" if torch.cuda.is_available() else "cpu")
|
16 |
+
wd_processor = AutoImageProcessor.from_pretrained(WD_MODEL_NAME, trust_remote_code=True)
|
17 |
+
|
18 |
+
|
19 |
+
def _people_tag(noun: str, minimum: int = 1, maximum: int = 5):
|
20 |
+
return (
|
21 |
+
[f"1{noun}"]
|
22 |
+
+ [f"{num}{noun}s" for num in range(minimum + 1, maximum + 1)]
|
23 |
+
+ [f"{maximum+1}+{noun}s"]
|
24 |
+
)
|
25 |
+
|
26 |
+
|
27 |
+
PEOPLE_TAGS = (
|
28 |
+
_people_tag("girl") + _people_tag("boy") + _people_tag("other") + ["no humans"]
|
29 |
+
)
|
30 |
+
|
31 |
+
|
32 |
+
RATING_MAP = {
|
33 |
+
"general": "safe",
|
34 |
+
"sensitive": "sensitive",
|
35 |
+
"questionable": "nsfw",
|
36 |
+
"explicit": "explicit, nsfw",
|
37 |
+
}
|
38 |
+
DANBOORU_TO_E621_RATING_MAP = {
|
39 |
+
"safe": "rating_safe",
|
40 |
+
"sensitive": "rating_safe",
|
41 |
+
"nsfw": "rating_explicit",
|
42 |
+
"explicit, nsfw": "rating_explicit",
|
43 |
+
"explicit": "rating_explicit",
|
44 |
+
"rating:safe": "rating_safe",
|
45 |
+
"rating:general": "rating_safe",
|
46 |
+
"rating:sensitive": "rating_safe",
|
47 |
+
"rating:questionable, nsfw": "rating_explicit",
|
48 |
+
"rating:explicit, nsfw": "rating_explicit",
|
49 |
+
}
|
50 |
+
|
51 |
+
|
52 |
+
def to_list(s):
|
53 |
+
return [x.strip() for x in s.split(",") if not s == ""]
|
54 |
+
|
55 |
+
|
56 |
+
def list_sub(a, b):
|
57 |
+
return [e for e in a if e not in b]
|
58 |
+
|
59 |
+
|
60 |
+
def list_uniq(l):
|
61 |
+
return sorted(set(l), key=l.index)
|
62 |
+
|
63 |
+
|
64 |
+
def load_dict_from_csv(filename):
|
65 |
+
with open(filename, 'r', encoding="utf-8") as f:
|
66 |
+
lines = f.readlines()
|
67 |
+
dict = {}
|
68 |
+
for line in lines:
|
69 |
+
parts = line.strip().split(',')
|
70 |
+
dict[parts[0]] = parts[1]
|
71 |
+
return dict
|
72 |
+
|
73 |
+
|
74 |
+
anime_series_dict = load_dict_from_csv('character_series_dict.csv')
|
75 |
+
|
76 |
+
|
77 |
+
def character_list_to_series_list(character_list):
|
78 |
+
output_series_tag = []
|
79 |
+
series_tag = ""
|
80 |
+
series_dict = anime_series_dict
|
81 |
+
for tag in character_list:
|
82 |
+
series_tag = series_dict.get(tag, "")
|
83 |
+
if tag.endswith(")"):
|
84 |
+
tags = tag.split("(")
|
85 |
+
character_tag = "(".join(tags[:-1])
|
86 |
+
if character_tag.endswith(" "):
|
87 |
+
character_tag = character_tag[:-1]
|
88 |
+
series_tag = tags[-1].replace(")", "")
|
89 |
+
|
90 |
+
if series_tag:
|
91 |
+
output_series_tag.append(series_tag)
|
92 |
+
|
93 |
+
return output_series_tag
|
94 |
+
|
95 |
+
|
96 |
+
def select_random_character(series: str, character: str):
|
97 |
+
from random import randrange
|
98 |
+
character_list = list(anime_series_dict.keys())
|
99 |
+
character = character_list[randrange(len(character_list) - 1)]
|
100 |
+
series = anime_series_dict.get(character.split(",")[0].strip(), "")
|
101 |
+
return series, character
|
102 |
+
|
103 |
+
|
104 |
+
def danbooru_to_e621(dtag, e621_dict):
|
105 |
+
def d_to_e(match, e621_dict):
|
106 |
+
dtag = match.group(0)
|
107 |
+
etag = e621_dict.get(dtag.strip().replace("_", " "), "")
|
108 |
+
if etag:
|
109 |
+
return etag
|
110 |
+
else:
|
111 |
+
return dtag
|
112 |
+
|
113 |
+
import re
|
114 |
+
tag = re.sub(r'[\w ]+', lambda wrapper: d_to_e(wrapper, e621_dict), dtag, 2)
|
115 |
+
|
116 |
+
return tag
|
117 |
+
|
118 |
+
|
119 |
+
danbooru_to_e621_dict = load_dict_from_csv('danbooru_e621.csv')
|
120 |
+
|
121 |
+
|
122 |
+
def convert_danbooru_to_e621_prompt(input_prompt: str = "", prompt_type: str = "danbooru"):
|
123 |
+
if prompt_type == "danbooru": return input_prompt
|
124 |
+
tags = input_prompt.split(",") if input_prompt else []
|
125 |
+
people_tags: list[str] = []
|
126 |
+
other_tags: list[str] = []
|
127 |
+
rating_tags: list[str] = []
|
128 |
+
|
129 |
+
e621_dict = danbooru_to_e621_dict
|
130 |
+
for tag in tags:
|
131 |
+
tag = tag.strip().replace("_", " ")
|
132 |
+
tag = danbooru_to_e621(tag, e621_dict)
|
133 |
+
if tag in PEOPLE_TAGS:
|
134 |
+
people_tags.append(tag)
|
135 |
+
elif tag in DANBOORU_TO_E621_RATING_MAP.keys():
|
136 |
+
rating_tags.append(DANBOORU_TO_E621_RATING_MAP.get(tag.replace(" ",""), ""))
|
137 |
+
else:
|
138 |
+
other_tags.append(tag)
|
139 |
+
|
140 |
+
rating_tags = sorted(set(rating_tags), key=rating_tags.index)
|
141 |
+
rating_tags = [rating_tags[0]] if rating_tags else []
|
142 |
+
rating_tags = ["explicit, nsfw"] if rating_tags and rating_tags[0] == "explicit" else rating_tags
|
143 |
+
|
144 |
+
output_prompt = ", ".join(people_tags + other_tags + rating_tags)
|
145 |
+
|
146 |
+
return output_prompt
|
147 |
+
|
148 |
+
|
149 |
+
def translate_prompt(prompt: str = ""):
|
150 |
+
def translate_to_english(prompt):
|
151 |
+
import httpcore
|
152 |
+
setattr(httpcore, 'SyncHTTPTransport', 'AsyncHTTPProxy')
|
153 |
+
from googletrans import Translator
|
154 |
+
translator = Translator()
|
155 |
+
try:
|
156 |
+
translated_prompt = translator.translate(prompt, src='auto', dest='en').text
|
157 |
+
return translated_prompt
|
158 |
+
except Exception as e:
|
159 |
+
return prompt
|
160 |
+
|
161 |
+
def is_japanese(s):
|
162 |
+
import unicodedata
|
163 |
+
for ch in s:
|
164 |
+
name = unicodedata.name(ch, "")
|
165 |
+
if "CJK UNIFIED" in name or "HIRAGANA" in name or "KATAKANA" in name:
|
166 |
+
return True
|
167 |
+
return False
|
168 |
+
|
169 |
+
def to_list(s):
|
170 |
+
return [x.strip() for x in s.split(",")]
|
171 |
+
|
172 |
+
prompts = to_list(prompt)
|
173 |
+
outputs = []
|
174 |
+
for p in prompts:
|
175 |
+
p = translate_to_english(p) if is_japanese(p) else p
|
176 |
+
outputs.append(p)
|
177 |
+
|
178 |
+
return ", ".join(outputs)
|
179 |
+
|
180 |
+
|
181 |
+
def translate_prompt_to_ja(prompt: str = ""):
|
182 |
+
def translate_to_japanese(prompt):
|
183 |
+
import httpcore
|
184 |
+
setattr(httpcore, 'SyncHTTPTransport', 'AsyncHTTPProxy')
|
185 |
+
from googletrans import Translator
|
186 |
+
translator = Translator()
|
187 |
+
try:
|
188 |
+
translated_prompt = translator.translate(prompt, src='en', dest='ja').text
|
189 |
+
return translated_prompt
|
190 |
+
except Exception as e:
|
191 |
+
return prompt
|
192 |
+
|
193 |
+
def is_japanese(s):
|
194 |
+
import unicodedata
|
195 |
+
for ch in s:
|
196 |
+
name = unicodedata.name(ch, "")
|
197 |
+
if "CJK UNIFIED" in name or "HIRAGANA" in name or "KATAKANA" in name:
|
198 |
+
return True
|
199 |
+
return False
|
200 |
+
|
201 |
+
def to_list(s):
|
202 |
+
return [x.strip() for x in s.split(",")]
|
203 |
+
|
204 |
+
prompts = to_list(prompt)
|
205 |
+
outputs = []
|
206 |
+
for p in prompts:
|
207 |
+
p = translate_to_japanese(p) if not is_japanese(p) else p
|
208 |
+
outputs.append(p)
|
209 |
+
|
210 |
+
return ", ".join(outputs)
|
211 |
+
|
212 |
+
|
213 |
+
def tags_to_ja(itag, dict):
|
214 |
+
def t_to_j(match, dict):
|
215 |
+
tag = match.group(0)
|
216 |
+
ja = dict.get(tag.strip().replace("_", " "), "")
|
217 |
+
if ja:
|
218 |
+
return ja
|
219 |
+
else:
|
220 |
+
return tag
|
221 |
+
|
222 |
+
import re
|
223 |
+
tag = re.sub(r'[\w ]+', lambda wrapper: t_to_j(wrapper, dict), itag, 2)
|
224 |
+
|
225 |
+
return tag
|
226 |
+
|
227 |
+
|
228 |
+
def convert_tags_to_ja(input_prompt: str = ""):
|
229 |
+
tags = input_prompt.split(",") if input_prompt else []
|
230 |
+
out_tags = []
|
231 |
+
|
232 |
+
tags_to_ja_dict = load_dict_from_csv('all_tags_ja_ext.csv')
|
233 |
+
dict = tags_to_ja_dict
|
234 |
+
for tag in tags:
|
235 |
+
tag = tag.strip().replace("_", " ")
|
236 |
+
tag = tags_to_ja(tag, dict)
|
237 |
+
out_tags.append(tag)
|
238 |
+
|
239 |
+
return ", ".join(out_tags)
|
240 |
+
|
241 |
+
|
242 |
+
enable_auto_recom_prompt = True
|
243 |
+
|
244 |
+
|
245 |
+
animagine_ps = to_list("anime artwork, anime style, studio anime, highly detailed, masterpiece, best quality, very aesthetic, absurdres")
|
246 |
+
animagine_nps = to_list("lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]")
|
247 |
+
pony_ps = to_list("source_anime, score_9, score_8_up, score_7_up, masterpiece, best quality, very aesthetic, absurdres")
|
248 |
+
pony_nps = to_list("source_pony, source_furry, source_cartoon, score_6, score_5, score_4, busty, ugly face, mutated hands, low res, blurry face, black and white, the simpsons, overwatch, apex legends")
|
249 |
+
other_ps = to_list("anime artwork, anime style, studio anime, highly detailed, cinematic photo, 35mm photograph, film, bokeh, professional, 4k, highly detailed")
|
250 |
+
other_nps = to_list("photo, deformed, black and white, realism, disfigured, low contrast, drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly")
|
251 |
+
default_ps = to_list("score_9, score_8_up, score_7_up, highly detailed, masterpiece, best quality, very aesthetic, absurdres")
|
252 |
+
default_nps = to_list("score_6, score_5, score_4, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]")
|
253 |
+
def insert_recom_prompt(prompt: str = "", neg_prompt: str = "", type: str = "None"):
|
254 |
+
global enable_auto_recom_prompt
|
255 |
+
prompts = to_list(prompt)
|
256 |
+
neg_prompts = to_list(neg_prompt)
|
257 |
+
|
258 |
+
prompts = list_sub(prompts, animagine_ps + pony_ps)
|
259 |
+
neg_prompts = list_sub(neg_prompts, animagine_nps + pony_nps)
|
260 |
+
|
261 |
+
last_empty_p = [""] if not prompts and type != "None" else []
|
262 |
+
last_empty_np = [""] if not neg_prompts and type != "None" else []
|
263 |
+
|
264 |
+
if type == "Auto":
|
265 |
+
enable_auto_recom_prompt = True
|
266 |
+
else:
|
267 |
+
enable_auto_recom_prompt = False
|
268 |
+
if type == "Animagine":
|
269 |
+
prompts = prompts + animagine_ps
|
270 |
+
neg_prompts = neg_prompts + animagine_nps
|
271 |
+
elif type == "Pony":
|
272 |
+
prompts = prompts + pony_ps
|
273 |
+
neg_prompts = neg_prompts + pony_nps
|
274 |
+
|
275 |
+
prompt = ", ".join(list_uniq(prompts) + last_empty_p)
|
276 |
+
neg_prompt = ", ".join(list_uniq(neg_prompts) + last_empty_np)
|
277 |
+
|
278 |
+
return prompt, neg_prompt
|
279 |
+
|
280 |
+
|
281 |
+
def load_model_prompt_dict():
|
282 |
+
import json
|
283 |
+
dict = {}
|
284 |
+
try:
|
285 |
+
with open('model_dict.json', encoding='utf-8') as f:
|
286 |
+
dict = json.load(f)
|
287 |
+
except Exception:
|
288 |
+
pass
|
289 |
+
return dict
|
290 |
+
|
291 |
+
|
292 |
+
model_prompt_dict = load_model_prompt_dict()
|
293 |
+
|
294 |
+
|
295 |
+
def insert_model_recom_prompt(prompt: str = "", neg_prompt: str = "", model_name: str = "None"):
|
296 |
+
if not model_name or not enable_auto_recom_prompt: return prompt, neg_prompt
|
297 |
+
prompts = to_list(prompt)
|
298 |
+
neg_prompts = to_list(neg_prompt)
|
299 |
+
prompts = list_sub(prompts, animagine_ps + pony_ps + other_ps)
|
300 |
+
neg_prompts = list_sub(neg_prompts, animagine_nps + pony_nps + other_nps)
|
301 |
+
last_empty_p = [""] if not prompts and type != "None" else []
|
302 |
+
last_empty_np = [""] if not neg_prompts and type != "None" else []
|
303 |
+
ps = []
|
304 |
+
nps = []
|
305 |
+
if model_name in model_prompt_dict.keys():
|
306 |
+
ps = to_list(model_prompt_dict[model_name]["prompt"])
|
307 |
+
nps = to_list(model_prompt_dict[model_name]["negative_prompt"])
|
308 |
+
else:
|
309 |
+
ps = default_ps
|
310 |
+
nps = default_nps
|
311 |
+
prompts = prompts + ps
|
312 |
+
neg_prompts = neg_prompts + nps
|
313 |
+
prompt = ", ".join(list_uniq(prompts) + last_empty_p)
|
314 |
+
neg_prompt = ", ".join(list_uniq(neg_prompts) + last_empty_np)
|
315 |
+
return prompt, neg_prompt
|
316 |
+
|
317 |
+
|
318 |
+
tag_group_dict = load_dict_from_csv('tag_group.csv')
|
319 |
+
|
320 |
+
|
321 |
+
def remove_specific_prompt(input_prompt: str = "", keep_tags: str = "all"):
|
322 |
+
def is_dressed(tag):
|
323 |
+
import re
|
324 |
+
p = re.compile(r'dress|cloth|uniform|costume|vest|sweater|coat|shirt|jacket|blazer|apron|leotard|hood|sleeve|skirt|shorts|pant|loafer|ribbon|necktie|bow|collar|glove|sock|shoe|boots|wear|emblem')
|
325 |
+
return p.search(tag)
|
326 |
+
|
327 |
+
def is_background(tag):
|
328 |
+
import re
|
329 |
+
p = re.compile(r'background|outline|light|sky|build|day|screen|tree|city')
|
330 |
+
return p.search(tag)
|
331 |
+
|
332 |
+
un_tags = ['solo']
|
333 |
+
group_list = ['groups', 'body_parts', 'attire', 'posture', 'objects', 'creatures', 'locations', 'disambiguation_pages', 'commonly_misused_tags', 'phrases', 'verbs_and_gerunds', 'subjective', 'nudity', 'sex_objects', 'sex', 'sex_acts', 'image_composition', 'artistic_license', 'text', 'year_tags', 'metatags']
|
334 |
+
keep_group_dict = {
|
335 |
+
"body": ['groups', 'body_parts'],
|
336 |
+
"dress": ['groups', 'body_parts', 'attire'],
|
337 |
+
"all": group_list,
|
338 |
+
}
|
339 |
+
|
340 |
+
def is_necessary(tag, keep_tags, group_dict):
|
341 |
+
if keep_tags == "all":
|
342 |
+
return True
|
343 |
+
elif tag in un_tags or group_dict.get(tag, "") in explicit_group:
|
344 |
+
return False
|
345 |
+
elif keep_tags == "body" and is_dressed(tag):
|
346 |
+
return False
|
347 |
+
elif is_background(tag):
|
348 |
+
return False
|
349 |
+
else:
|
350 |
+
return True
|
351 |
+
|
352 |
+
if keep_tags == "all": return input_prompt
|
353 |
+
keep_group = keep_group_dict.get(keep_tags, keep_group_dict["body"])
|
354 |
+
explicit_group = list(set(group_list) ^ set(keep_group))
|
355 |
+
|
356 |
+
tags = input_prompt.split(",") if input_prompt else []
|
357 |
+
people_tags: list[str] = []
|
358 |
+
other_tags: list[str] = []
|
359 |
+
|
360 |
+
group_dict = tag_group_dict
|
361 |
+
for tag in tags:
|
362 |
+
tag = tag.strip().replace("_", " ")
|
363 |
+
if tag in PEOPLE_TAGS:
|
364 |
+
people_tags.append(tag)
|
365 |
+
elif is_necessary(tag, keep_tags, group_dict):
|
366 |
+
other_tags.append(tag)
|
367 |
+
|
368 |
+
output_prompt = ", ".join(people_tags + other_tags)
|
369 |
+
|
370 |
+
return output_prompt
|
371 |
+
|
372 |
+
|
373 |
+
def sort_taglist(tags: list[str]):
|
374 |
+
if not tags: return []
|
375 |
+
character_tags: list[str] = []
|
376 |
+
series_tags: list[str] = []
|
377 |
+
people_tags: list[str] = []
|
378 |
+
group_list = ['groups', 'body_parts', 'attire', 'posture', 'objects', 'creatures', 'locations', 'disambiguation_pages', 'commonly_misused_tags', 'phrases', 'verbs_and_gerunds', 'subjective', 'nudity', 'sex_objects', 'sex', 'sex_acts', 'image_composition', 'artistic_license', 'text', 'year_tags', 'metatags']
|
379 |
+
group_tags = {}
|
380 |
+
other_tags: list[str] = []
|
381 |
+
rating_tags: list[str] = []
|
382 |
+
|
383 |
+
group_dict = tag_group_dict
|
384 |
+
group_set = set(group_dict.keys())
|
385 |
+
character_set = set(anime_series_dict.keys())
|
386 |
+
series_set = set(anime_series_dict.values())
|
387 |
+
rating_set = set(DANBOORU_TO_E621_RATING_MAP.keys()) | set(DANBOORU_TO_E621_RATING_MAP.values())
|
388 |
+
|
389 |
+
for tag in tags:
|
390 |
+
tag = tag.strip().replace("_", " ")
|
391 |
+
if tag in PEOPLE_TAGS:
|
392 |
+
people_tags.append(tag)
|
393 |
+
elif tag in rating_set:
|
394 |
+
rating_tags.append(tag)
|
395 |
+
elif tag in group_set:
|
396 |
+
elem = group_dict[tag]
|
397 |
+
group_tags[elem] = group_tags[elem] + [tag] if elem in group_tags else [tag]
|
398 |
+
elif tag in character_set:
|
399 |
+
character_tags.append(tag)
|
400 |
+
elif tag in series_set:
|
401 |
+
series_tags.append(tag)
|
402 |
+
else:
|
403 |
+
other_tags.append(tag)
|
404 |
+
|
405 |
+
output_group_tags: list[str] = []
|
406 |
+
for k in group_list:
|
407 |
+
output_group_tags.extend(group_tags.get(k, []))
|
408 |
+
|
409 |
+
rating_tags = [rating_tags[0]] if rating_tags else []
|
410 |
+
rating_tags = ["explicit, nsfw"] if rating_tags and rating_tags[0] == "explicit" else rating_tags
|
411 |
+
|
412 |
+
output_tags = character_tags + series_tags + people_tags + output_group_tags + other_tags + rating_tags
|
413 |
+
|
414 |
+
return output_tags
|
415 |
+
|
416 |
+
|
417 |
+
def sort_tags(tags: str):
|
418 |
+
if not tags: return ""
|
419 |
+
taglist: list[str] = []
|
420 |
+
for tag in tags.split(","):
|
421 |
+
taglist.append(tag.strip())
|
422 |
+
taglist = list(filter(lambda x: x != "", taglist))
|
423 |
+
return ", ".join(sort_taglist(taglist))
|
424 |
+
|
425 |
+
|
426 |
+
def postprocess_results(results: dict[str, float], general_threshold: float, character_threshold: float):
|
427 |
+
results = {
|
428 |
+
k: v for k, v in sorted(results.items(), key=lambda item: item[1], reverse=True)
|
429 |
+
}
|
430 |
+
|
431 |
+
rating = {}
|
432 |
+
character = {}
|
433 |
+
general = {}
|
434 |
+
|
435 |
+
for k, v in results.items():
|
436 |
+
if k.startswith("rating:"):
|
437 |
+
rating[k.replace("rating:", "")] = v
|
438 |
+
continue
|
439 |
+
elif k.startswith("character:"):
|
440 |
+
character[k.replace("character:", "")] = v
|
441 |
+
continue
|
442 |
+
|
443 |
+
general[k] = v
|
444 |
+
|
445 |
+
character = {k: v for k, v in character.items() if v >= character_threshold}
|
446 |
+
general = {k: v for k, v in general.items() if v >= general_threshold}
|
447 |
+
|
448 |
+
return rating, character, general
|
449 |
+
|
450 |
+
|
451 |
+
def gen_prompt(rating: list[str], character: list[str], general: list[str]):
|
452 |
+
people_tags: list[str] = []
|
453 |
+
other_tags: list[str] = []
|
454 |
+
rating_tag = RATING_MAP[rating[0]]
|
455 |
+
|
456 |
+
for tag in general:
|
457 |
+
if tag in PEOPLE_TAGS:
|
458 |
+
people_tags.append(tag)
|
459 |
+
else:
|
460 |
+
other_tags.append(tag)
|
461 |
+
|
462 |
+
all_tags = people_tags + other_tags
|
463 |
+
|
464 |
+
return ", ".join(all_tags)
|
465 |
+
|
466 |
+
|
467 |
+
@spaces.GPU()
|
468 |
+
def predict_tags(image: Image.Image, general_threshold: float = 0.3, character_threshold: float = 0.8):
|
469 |
+
inputs = wd_processor.preprocess(image, return_tensors="pt")
|
470 |
+
|
471 |
+
outputs = wd_model(**inputs.to(wd_model.device, wd_model.dtype))
|
472 |
+
logits = torch.sigmoid(outputs.logits[0]) # take the first logits
|
473 |
+
|
474 |
+
# get probabilities
|
475 |
+
results = {
|
476 |
+
wd_model.config.id2label[i]: float(logit.float()) for i, logit in enumerate(logits)
|
477 |
+
}
|
478 |
+
# rating, character, general
|
479 |
+
rating, character, general = postprocess_results(
|
480 |
+
results, general_threshold, character_threshold
|
481 |
+
)
|
482 |
+
prompt = gen_prompt(
|
483 |
+
list(rating.keys()), list(character.keys()), list(general.keys())
< |