Spaces:
Sleeping
Sleeping
nihuajian
commited on
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,302 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import torch
|
| 3 |
+
from PIL import Image
|
| 4 |
+
from transformers import AutoModel, AutoTokenizer
|
| 5 |
+
import warnings
|
| 6 |
+
import os
|
| 7 |
+
|
| 8 |
+
# 禁用警告信息
|
| 9 |
+
warnings.filterwarnings("ignore")
|
| 10 |
+
|
| 11 |
+
# 全局变量存储模型
|
| 12 |
+
model = None
|
| 13 |
+
tokenizer = None
|
| 14 |
+
|
| 15 |
+
def load_model():
|
| 16 |
+
"""加载MiniCPM-o模型"""
|
| 17 |
+
global model, tokenizer
|
| 18 |
+
if model is None:
|
| 19 |
+
print("正在加载MiniCPM-o模型...")
|
| 20 |
+
model = AutoModel.from_pretrained(
|
| 21 |
+
'openbmb/MiniCPM-o-2_6',
|
| 22 |
+
trust_remote_code=True,
|
| 23 |
+
attn_implementation='sdpa',
|
| 24 |
+
torch_dtype=torch.bfloat16,
|
| 25 |
+
init_vision=True,
|
| 26 |
+
init_audio=False,
|
| 27 |
+
init_tts=False
|
| 28 |
+
)
|
| 29 |
+
model = model.eval().cuda()
|
| 30 |
+
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-o-2_6', trust_remote_code=True)
|
| 31 |
+
print("模型加载完成")
|
| 32 |
+
return model, tokenizer
|
| 33 |
+
|
| 34 |
+
def clean_markdown_output(text):
|
| 35 |
+
"""清理输出文本,只保留markdown表格"""
|
| 36 |
+
lines = text.strip().split('\n')
|
| 37 |
+
markdown_lines = []
|
| 38 |
+
|
| 39 |
+
# 查找markdown表格的开始和结束
|
| 40 |
+
in_table = False
|
| 41 |
+
for line in lines:
|
| 42 |
+
line = line.strip()
|
| 43 |
+
# 检查是否是表格行(包含|符号)
|
| 44 |
+
if '|' in line and not line.startswith('```'):
|
| 45 |
+
in_table = True
|
| 46 |
+
markdown_lines.append(line)
|
| 47 |
+
elif in_table and line == '':
|
| 48 |
+
# 空行可能表示表格结束
|
| 49 |
+
break
|
| 50 |
+
elif in_table and not line.startswith('```'):
|
| 51 |
+
# 继续收集表格相关行
|
| 52 |
+
markdown_lines.append(line)
|
| 53 |
+
|
| 54 |
+
# 如果没有找到表格,返回原始清理后的文本
|
| 55 |
+
if not markdown_lines:
|
| 56 |
+
# 移除代码块标记和多余的说明文字
|
| 57 |
+
cleaned_text = text.replace('```markdown', '').replace('```', '').strip()
|
| 58 |
+
# 移除常见的解释性文字
|
| 59 |
+
lines = cleaned_text.split('\n')
|
| 60 |
+
result_lines = []
|
| 61 |
+
for line in lines:
|
| 62 |
+
line = line.strip()
|
| 63 |
+
if line and not line.startswith('这个表格') and not line.startswith('该表格') and not line.startswith('表格显示'):
|
| 64 |
+
result_lines.append(line)
|
| 65 |
+
return '\n'.join(result_lines)
|
| 66 |
+
|
| 67 |
+
return '\n'.join(markdown_lines)
|
| 68 |
+
|
| 69 |
+
def clean_formula_output(text):
|
| 70 |
+
"""清理输出文本,只保留LaTeX公式"""
|
| 71 |
+
lines = text.strip().split('\n')
|
| 72 |
+
formula_lines = []
|
| 73 |
+
|
| 74 |
+
for line in lines:
|
| 75 |
+
line = line.strip()
|
| 76 |
+
# 跳过解释性文字
|
| 77 |
+
if line and not any(line.startswith(prefix) for prefix in [
|
| 78 |
+
'这个公式', '该公式', '公式表示', '根据图片', '图片中的', '识别结果'
|
| 79 |
+
]):
|
| 80 |
+
# 保留包含LaTeX语法的行
|
| 81 |
+
if any(symbol in line for symbol in ['$', '\\', '{', '}', '^', '_']) or '=' in line:
|
| 82 |
+
formula_lines.append(line)
|
| 83 |
+
# 或者保留纯数学表达式
|
| 84 |
+
elif any(char.isdigit() or char in '+-*/=()[]{}^_' for char in line):
|
| 85 |
+
formula_lines.append(line)
|
| 86 |
+
|
| 87 |
+
# 如果没有找到公式,返回原始清理后的文本
|
| 88 |
+
if not formula_lines:
|
| 89 |
+
cleaned_text = text.replace('```latex', '').replace('```', '').strip()
|
| 90 |
+
lines = cleaned_text.split('\n')
|
| 91 |
+
result_lines = []
|
| 92 |
+
for line in lines:
|
| 93 |
+
line = line.strip()
|
| 94 |
+
if line and not any(line.startswith(prefix) for prefix in [
|
| 95 |
+
'这个公式', '该公式', '公式表示', '根据图片', '图片中的'
|
| 96 |
+
]):
|
| 97 |
+
result_lines.append(line)
|
| 98 |
+
return '\n'.join(result_lines)
|
| 99 |
+
|
| 100 |
+
return '\n'.join(formula_lines)
|
| 101 |
+
|
| 102 |
+
def clean_text_output(text):
|
| 103 |
+
"""清理输出文本,只保留识别的文字内容"""
|
| 104 |
+
lines = text.strip().split('\n')
|
| 105 |
+
text_lines = []
|
| 106 |
+
|
| 107 |
+
# 移除代码块标记
|
| 108 |
+
cleaned_text = text.replace('```text', '').replace('```', '').strip()
|
| 109 |
+
lines = cleaned_text.split('\n')
|
| 110 |
+
|
| 111 |
+
for line in lines:
|
| 112 |
+
line = line.strip()
|
| 113 |
+
# 跳过解释性文字
|
| 114 |
+
if line and not any(line.startswith(prefix) for prefix in [
|
| 115 |
+
'图片中的文字', '识别结果', '文字内容', '根据图片', '这张图片', '该图片'
|
| 116 |
+
]):
|
| 117 |
+
text_lines.append(line)
|
| 118 |
+
|
| 119 |
+
return '\n'.join(text_lines)
|
| 120 |
+
|
| 121 |
+
def parse_image(image, parse_type):
|
| 122 |
+
"""解析图片内容为指定格式"""
|
| 123 |
+
try:
|
| 124 |
+
# 确保模型已加载
|
| 125 |
+
model, tokenizer = load_model()
|
| 126 |
+
|
| 127 |
+
if image is None:
|
| 128 |
+
return "请上传一张图片", ""
|
| 129 |
+
|
| 130 |
+
# 转换图片格式
|
| 131 |
+
if isinstance(image, str):
|
| 132 |
+
image = Image.open(image).convert('RGB')
|
| 133 |
+
elif hasattr(image, 'convert'):
|
| 134 |
+
image = image.convert('RGB')
|
| 135 |
+
|
| 136 |
+
# 根据解析类型设置不同的提示词
|
| 137 |
+
questions = {
|
| 138 |
+
"表格解析": "解析一下这个表格为markdown格式,不需要任何解释和思考,直接输出markdown格式",
|
| 139 |
+
"公式解析": "识别并提取���片中的数学公式,用LaTeX格式输出,不需要任何解释,直接输出公式",
|
| 140 |
+
"文本解析": "识别并提取图片中的所有文字内容,保持原有格式,不需要任何解释,直接输出文字内容"
|
| 141 |
+
}
|
| 142 |
+
|
| 143 |
+
question = questions.get(parse_type, questions["表格解析"])
|
| 144 |
+
msgs = [{'role': 'user', 'content': [image, question]}]
|
| 145 |
+
|
| 146 |
+
# 使用流式输出获取结果
|
| 147 |
+
res = model.chat(
|
| 148 |
+
msgs=msgs,
|
| 149 |
+
tokenizer=tokenizer,
|
| 150 |
+
sampling=True,
|
| 151 |
+
stream=True
|
| 152 |
+
)
|
| 153 |
+
|
| 154 |
+
# 收集所有输出文本
|
| 155 |
+
generated_text = ""
|
| 156 |
+
for new_text in res:
|
| 157 |
+
generated_text += new_text
|
| 158 |
+
|
| 159 |
+
# 根据类型清理输出
|
| 160 |
+
if parse_type == "表格解析":
|
| 161 |
+
result = clean_markdown_output(generated_text)
|
| 162 |
+
output_format = "Markdown"
|
| 163 |
+
elif parse_type == "公式解析":
|
| 164 |
+
result = clean_formula_output(generated_text)
|
| 165 |
+
output_format = "LaTeX"
|
| 166 |
+
elif parse_type == "文本解析":
|
| 167 |
+
result = clean_text_output(generated_text)
|
| 168 |
+
output_format = "纯文本"
|
| 169 |
+
else:
|
| 170 |
+
result = generated_text.strip()
|
| 171 |
+
output_format = "原始输出"
|
| 172 |
+
|
| 173 |
+
return result, f"解析完成 - 输出格式: {output_format}"
|
| 174 |
+
|
| 175 |
+
except Exception as e:
|
| 176 |
+
return f"解析失败: {str(e)}", "错误"
|
| 177 |
+
|
| 178 |
+
def create_interface():
|
| 179 |
+
"""创建Gradio界面"""
|
| 180 |
+
|
| 181 |
+
# 自定义CSS样式
|
| 182 |
+
css = """
|
| 183 |
+
.gradio-container {
|
| 184 |
+
font-family: 'Helvetica Neue', Arial, sans-serif;
|
| 185 |
+
}
|
| 186 |
+
.output-text {
|
| 187 |
+
font-family: 'Courier New', monospace;
|
| 188 |
+
font-size: 14px;
|
| 189 |
+
}
|
| 190 |
+
"""
|
| 191 |
+
|
| 192 |
+
with gr.Blocks(css=css, title="MiniCPM 多模态内容解析工具") as interface:
|
| 193 |
+
gr.Markdown("""
|
| 194 |
+
# 🚀 MiniCPM 多模态内容解析工具
|
| 195 |
+
|
| 196 |
+
基于MiniCPM-o多模态模型的智能图片内容解析工具,支持表格、公式、文本三种解析模式。
|
| 197 |
+
|
| 198 |
+
## 📋 使用说明
|
| 199 |
+
1. **上传图片**: 支持 PNG、JPG、JPEG 等格式
|
| 200 |
+
2. **选择解析类型**: 根据图片内容选择相应的解析模式
|
| 201 |
+
3. **获取结果**: 自动清理输出,获得纯净的解析结果
|
| 202 |
+
|
| 203 |
+
## 🎯 解析类型说明
|
| 204 |
+
- **📊 表格解析**: 将表格图片转换为Markdown格式
|
| 205 |
+
- **🧮 公式解析**: 识别数学公式并输出LaTeX格式
|
| 206 |
+
- **📝 文本解析**: 提取图片中的所有文字内容
|
| 207 |
+
""")
|
| 208 |
+
|
| 209 |
+
with gr.Row():
|
| 210 |
+
with gr.Column(scale=1):
|
| 211 |
+
# 输入组件
|
| 212 |
+
image_input = gr.Image(
|
| 213 |
+
label="📷 上传图片",
|
| 214 |
+
type="pil",
|
| 215 |
+
height=400
|
| 216 |
+
)
|
| 217 |
+
|
| 218 |
+
parse_type = gr.Radio(
|
| 219 |
+
choices=["表格解析", "公式解析", "文本解析"],
|
| 220 |
+
value="表格解析",
|
| 221 |
+
label="🎛️ 选择解析类型",
|
| 222 |
+
info="根据图片内容选择合适的解析模式"
|
| 223 |
+
)
|
| 224 |
+
|
| 225 |
+
parse_button = gr.Button(
|
| 226 |
+
"🔍 开始解析",
|
| 227 |
+
variant="primary",
|
| 228 |
+
size="lg"
|
| 229 |
+
)
|
| 230 |
+
|
| 231 |
+
with gr.Column(scale=1):
|
| 232 |
+
# 输出组件
|
| 233 |
+
status_output = gr.Textbox(
|
| 234 |
+
label="📊 解析状态",
|
| 235 |
+
value="等待上传图片...",
|
| 236 |
+
interactive=False
|
| 237 |
+
)
|
| 238 |
+
|
| 239 |
+
result_output = gr.Textbox(
|
| 240 |
+
label="📄 解析结果",
|
| 241 |
+
lines=20,
|
| 242 |
+
max_lines=30,
|
| 243 |
+
show_copy_button=True,
|
| 244 |
+
elem_classes=["output-text"],
|
| 245 |
+
placeholder="解析结果将在这里显示..."
|
| 246 |
+
)
|
| 247 |
+
|
| 248 |
+
# 示例图片
|
| 249 |
+
gr.Markdown("## 📖 示例图片")
|
| 250 |
+
with gr.Row():
|
| 251 |
+
gr.Examples(
|
| 252 |
+
examples=[
|
| 253 |
+
["table.png", "表格解析"],
|
| 254 |
+
["formulas.png", "公式解析"],
|
| 255 |
+
["text.png", "文本解析"]
|
| 256 |
+
],
|
| 257 |
+
inputs=[image_input, parse_type],
|
| 258 |
+
label="点击示例快速体验"
|
| 259 |
+
)
|
| 260 |
+
|
| 261 |
+
# 绑定事件
|
| 262 |
+
parse_button.click(
|
| 263 |
+
fn=parse_image,
|
| 264 |
+
inputs=[image_input, parse_type],
|
| 265 |
+
outputs=[result_output, status_output]
|
| 266 |
+
)
|
| 267 |
+
|
| 268 |
+
# 添加页脚信息
|
| 269 |
+
gr.Markdown("""
|
| 270 |
+
---
|
| 271 |
+
### 💡 使用提示
|
| 272 |
+
- 确保图片清晰,内容结构明显
|
| 273 |
+
- 复杂表格建议分段处理
|
| 274 |
+
- 公式图片建议使用高分辨率
|
| 275 |
+
- 文字图片避免模糊、倾斜或光线不足
|
| 276 |
+
|
| 277 |
+
### 🔧 技术支持
|
| 278 |
+
- 模���: MiniCPM-o-2.6
|
| 279 |
+
- 框架: Gradio + Transformers
|
| 280 |
+
- GPU: CUDA加速推理
|
| 281 |
+
""")
|
| 282 |
+
|
| 283 |
+
return interface
|
| 284 |
+
|
| 285 |
+
if __name__ == "__main__":
|
| 286 |
+
# 预加载模型(可选,在启动时加载以减少首次使用延迟)
|
| 287 |
+
try:
|
| 288 |
+
load_model()
|
| 289 |
+
print("✅ 模型预加载完成")
|
| 290 |
+
except Exception as e:
|
| 291 |
+
print(f"⚠️ 模型预加载失败: {e}")
|
| 292 |
+
print("模型将在首次使用时加载")
|
| 293 |
+
|
| 294 |
+
# 创建并启动界面
|
| 295 |
+
interface = create_interface()
|
| 296 |
+
interface.launch(
|
| 297 |
+
server_name="0.0.0.0", # 允许外部访问
|
| 298 |
+
server_port=7860, # Hugging Face Spaces默认端口
|
| 299 |
+
share=False, # 在Hugging Face上部署时设为False
|
| 300 |
+
show_error=True, # 显示详细错误信息
|
| 301 |
+
quiet=False # 显示启动信息
|
| 302 |
+
)
|