Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,186 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gradio as gr
|
3 |
+
import huggingface_hub
|
4 |
+
import numpy as np
|
5 |
+
import onnxruntime as rt
|
6 |
+
import pandas as pd
|
7 |
+
from PIL import Image
|
8 |
+
|
9 |
+
# 模型配置
|
10 |
+
MODEL_REPO = "SmilingWolf/wd-swinv2-tagger-v3" # 默认模型
|
11 |
+
MODEL_FILENAME = "model.onnx"
|
12 |
+
LABEL_FILENAME = "selected_tags.csv"
|
13 |
+
HF_TOKEN = os.environ.get("HF_TOKEN", "")
|
14 |
+
|
15 |
+
# 标签处理配置
|
16 |
+
kaomojis = [
|
17 |
+
"0_0",
|
18 |
+
"(o)_(o)",
|
19 |
+
"+_+",
|
20 |
+
"+_-",
|
21 |
+
"._.",
|
22 |
+
"<o>_<o>",
|
23 |
+
"<|>_<|>",
|
24 |
+
"=_=",
|
25 |
+
">_<",
|
26 |
+
"3_3",
|
27 |
+
"6_9",
|
28 |
+
">_o",
|
29 |
+
"@_@",
|
30 |
+
"^_^",
|
31 |
+
"o_o",
|
32 |
+
"u_u",
|
33 |
+
"x_x",
|
34 |
+
"|_|",
|
35 |
+
"||_||",
|
36 |
+
]
|
37 |
+
|
38 |
+
class Tagger:
|
39 |
+
def __init__(self):
|
40 |
+
self.model = None
|
41 |
+
self.tag_names = []
|
42 |
+
self.model_size = None
|
43 |
+
self._init_model()
|
44 |
+
|
45 |
+
def _init_model(self):
|
46 |
+
"""初始化模型和标签"""
|
47 |
+
# 下载模型文件
|
48 |
+
label_path = huggingface_hub.hf_hub_download(
|
49 |
+
MODEL_REPO,
|
50 |
+
LABEL_FILENAME,
|
51 |
+
token=HF_TOKEN
|
52 |
+
)
|
53 |
+
model_path = huggingface_hub.hf_hub_download(
|
54 |
+
MODEL_REPO,
|
55 |
+
MODEL_FILENAME,
|
56 |
+
token=HF_TOKEN
|
57 |
+
)
|
58 |
+
|
59 |
+
# 加载标签
|
60 |
+
tags_df = pd.read_csv(label_path)
|
61 |
+
self.tag_names = tags_df["name"].tolist()
|
62 |
+
self.categories = {
|
63 |
+
"rating": np.where(tags_df["category"] == 9)[0],
|
64 |
+
"general": np.where(tags_df["category"] == 0)[0],
|
65 |
+
"character": np.where(tags_df["category"] == 4)[0]
|
66 |
+
}
|
67 |
+
|
68 |
+
# 加载ONNX模型
|
69 |
+
self.model = rt.InferenceSession(model_path)
|
70 |
+
self.model_size = self.model.get_inputs()[0].shape[1]
|
71 |
+
|
72 |
+
def _preprocess(self, img):
|
73 |
+
"""图像预处理"""
|
74 |
+
# 转换为RGB
|
75 |
+
if img.mode != "RGB":
|
76 |
+
img = img.convert("RGB")
|
77 |
+
|
78 |
+
# 填充为正方形
|
79 |
+
size = max(img.size)
|
80 |
+
padded = Image.new("RGB", (size, size), (255, 255, 255))
|
81 |
+
padded.paste(img, ((size - img.width)//2, (size - img.height)//2))
|
82 |
+
|
83 |
+
# 调整尺寸
|
84 |
+
if size != self.model_size:
|
85 |
+
padded = padded.resize((self.model_size, self.model_size), Image.BICUBIC)
|
86 |
+
|
87 |
+
# 转换为BGR格式
|
88 |
+
return np.array(padded)[:, :, ::-1].astype(np.float32)
|
89 |
+
|
90 |
+
def predict(self, img, general_thresh=0.35, character_thresh=0.85):
|
91 |
+
"""执行预测"""
|
92 |
+
# 预处理
|
93 |
+
img_data = self._preprocess(img)[np.newaxis]
|
94 |
+
|
95 |
+
# 运行模型
|
96 |
+
input_name = self.model.get_inputs()[0].name
|
97 |
+
outputs = self.model.run(None, {input_name: img_data})[0][0]
|
98 |
+
|
99 |
+
# 组织结果
|
100 |
+
results = {
|
101 |
+
"ratings": {},
|
102 |
+
"general": {},
|
103 |
+
"characters": {}
|
104 |
+
}
|
105 |
+
|
106 |
+
# 处理评分标签
|
107 |
+
for idx in self.categories["rating"]:
|
108 |
+
tag = self.tag_names[idx].replace("_", " ")
|
109 |
+
results["ratings"][tag] = float(outputs[idx])
|
110 |
+
|
111 |
+
# 处理通用标签
|
112 |
+
for idx in self.categories["general"]:
|
113 |
+
if outputs[idx] > general_thresh:
|
114 |
+
tag = self.tag_names[idx].replace("_", " ")
|
115 |
+
results["general"][tag] = float(outputs[idx])
|
116 |
+
|
117 |
+
# 处理角色标签
|
118 |
+
for idx in self.categories["character"]:
|
119 |
+
if outputs[idx] > character_thresh:
|
120 |
+
tag = self.tag_names[idx].replace("_", " ")
|
121 |
+
results["characters"][tag] = float(outputs[idx])
|
122 |
+
|
123 |
+
# 排序结果
|
124 |
+
results["general"] = dict(sorted(
|
125 |
+
results["general"].items(),
|
126 |
+
key=lambda x: x[1],
|
127 |
+
reverse=True
|
128 |
+
))
|
129 |
+
|
130 |
+
return results
|
131 |
+
|
132 |
+
# 创建Gradio界面
|
133 |
+
with gr.Blocks(theme=gr.themes.Soft(), title="AI图像标签分析器") as demo:
|
134 |
+
gr.Markdown("# 🖼️ AI图像标签分析器")
|
135 |
+
gr.Markdown("上传图片自动分析图像内容标签")
|
136 |
+
|
137 |
+
with gr.Row():
|
138 |
+
with gr.Column(scale=1):
|
139 |
+
img_input = gr.Image(type="pil", label="上传图片")
|
140 |
+
with gr.Accordion("高级设置", open=False):
|
141 |
+
general_slider = gr.Slider(0, 1, 0.35,
|
142 |
+
label="通用标签阈值",
|
143 |
+
info="值越高标签越少但更准确")
|
144 |
+
char_slider = gr.Slider(0, 1, 0.85,
|
145 |
+
label="角色标签阈值",
|
146 |
+
info="推荐保持较高阈值")
|
147 |
+
analyze_btn = gr.Button("开始分析", variant="primary")
|
148 |
+
|
149 |
+
with gr.Column(scale=2):
|
150 |
+
with gr.Tabs():
|
151 |
+
with gr.TabItem("🏷️ 通用标签"):
|
152 |
+
general_tags = gr.Label(label="检测到的通用标签")
|
153 |
+
with gr.TabItem("👤 角色标签"):
|
154 |
+
char_tags = gr.Label(label="检测到的角色标签")
|
155 |
+
with gr.TabItem("⭐ 评分标签"):
|
156 |
+
rating_tags = gr.Label(label="图像评级标签")
|
157 |
+
|
158 |
+
output_text = gr.Textbox(label="标签文本",
|
159 |
+
placeholder="生成的标签文本将显示在这里...")
|
160 |
+
|
161 |
+
# 处理逻辑
|
162 |
+
def process_image(img, gen_thresh, char_thresh):
|
163 |
+
tagger = Tagger()
|
164 |
+
results = tagger.predict(img, gen_thresh, char_thresh)
|
165 |
+
|
166 |
+
# 格式化文本输出
|
167 |
+
tag_text = ", ".join(results["general"].keys())
|
168 |
+
if results["characters"]:
|
169 |
+
tag_text += ", " + ", ".join(results["characters"].keys())
|
170 |
+
|
171 |
+
return {
|
172 |
+
general_tags: results["general"],
|
173 |
+
char_tags: results["characters"],
|
174 |
+
rating_tags: results["ratings"],
|
175 |
+
output_text: tag_text
|
176 |
+
}
|
177 |
+
|
178 |
+
analyze_btn.click(
|
179 |
+
process_image,
|
180 |
+
inputs=[img_input, general_slider, char_slider],
|
181 |
+
outputs=[general_tags, char_tags, rating_tags, output_text]
|
182 |
+
)
|
183 |
+
|
184 |
+
# 启动应用
|
185 |
+
if __name__ == "__main__":
|
186 |
+
demo.launch(server_name="0.0.0.0", server_port=7860)
|