Spaces:
Runtime error
Runtime error
update: nougat gpu
Browse files
app.py
CHANGED
|
@@ -87,43 +87,107 @@ except:
|
|
| 87 |
if not terminators:
|
| 88 |
terminators = [2] # 使用常见的</s>标记ID作为默认值
|
| 89 |
|
| 90 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
@spaces.GPU(stateless=True)
|
| 92 |
-
def
|
| 93 |
-
"""使用Nougat处理PDF文件
|
| 94 |
try:
|
| 95 |
-
#
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
except Exception as e:
|
| 122 |
import traceback
|
| 123 |
-
|
| 124 |
-
|
|
|
|
| 125 |
|
| 126 |
-
#
|
| 127 |
def convert_pdf_to_markdown(pdf_file):
|
| 128 |
"""使用Nougat将PDF转换为Markdown (GPU优化版)"""
|
| 129 |
if pdf_file is None:
|
|
@@ -141,73 +205,34 @@ def convert_pdf_to_markdown(pdf_file):
|
|
| 141 |
with open(temp_pdf_path, "wb") as f:
|
| 142 |
f.write(pdf_file)
|
| 143 |
|
| 144 |
-
#
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
# 设置GPU环境变量
|
| 149 |
-
env = os.environ.copy()
|
| 150 |
-
env["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
|
| 151 |
|
| 152 |
-
|
| 153 |
-
result = subprocess.run(
|
| 154 |
-
["nougat", temp_pdf_path, "-o", output_dir],
|
| 155 |
-
stdout=subprocess.PIPE,
|
| 156 |
-
stderr=subprocess.PIPE,
|
| 157 |
-
text=True,
|
| 158 |
-
timeout=180,
|
| 159 |
-
env=env
|
| 160 |
-
)
|
| 161 |
-
|
| 162 |
-
# 检查命令行转换是否成功
|
| 163 |
-
if result.returncode == 0:
|
| 164 |
-
# 读取生成的Markdown文件
|
| 165 |
-
markdown_path = os.path.join(output_dir, "temp.mmd")
|
| 166 |
-
if os.path.exists(markdown_path):
|
| 167 |
-
with open(markdown_path, "r", encoding="utf-8") as f:
|
| 168 |
-
markdown_content = f.read()
|
| 169 |
-
|
| 170 |
-
# 限制文本长度
|
| 171 |
-
if len(markdown_content) > 20000:
|
| 172 |
-
markdown_content = markdown_content[:20000] + "\n\n...(Markdown内容已截断)"
|
| 173 |
-
|
| 174 |
-
status = f"PDF已成功转换为Markdown (GPU命令行): 生成了{len(markdown_content)}个字符"
|
| 175 |
-
return markdown_content, status
|
| 176 |
-
|
| 177 |
-
# 如果命令行方式失败,尝试空间GPU API
|
| 178 |
-
print("命令行转换失败,正在尝试使用GPU API方式处理PDF...")
|
| 179 |
-
markdown_content = process_pdf_with_nougat(temp_pdf_path)
|
| 180 |
-
|
| 181 |
-
# 限制文本长度
|
| 182 |
-
if len(markdown_content) > 20000:
|
| 183 |
-
markdown_content = markdown_content[:20000] + "\n\n...(Markdown内容已截断)"
|
| 184 |
-
|
| 185 |
-
status = f"PDF已成功转换为Markdown (GPU API): 生成了{len(markdown_content)}个字符"
|
| 186 |
-
return markdown_content, status
|
| 187 |
-
|
| 188 |
-
except subprocess.TimeoutExpired:
|
| 189 |
-
print("命令行处理超时,尝试使用GPU API...")
|
| 190 |
-
# 尝试使用GPU API
|
| 191 |
-
markdown_content = process_pdf_with_nougat(temp_pdf_path)
|
| 192 |
-
|
| 193 |
# 限制文本长度
|
| 194 |
if len(markdown_content) > 20000:
|
| 195 |
markdown_content = markdown_content[:20000] + "\n\n...(Markdown内容已截断)"
|
| 196 |
|
| 197 |
-
status = f"PDF已成功转换为Markdown (GPU
|
| 198 |
return markdown_content, status
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
|
|
|
|
|
|
| 205 |
# 限制文本长度
|
| 206 |
if len(markdown_content) > 20000:
|
| 207 |
markdown_content = markdown_content[:20000] + "\n\n...(Markdown内容已截断)"
|
| 208 |
|
| 209 |
status = f"PDF已成功转换为Markdown (GPU API): 生成了{len(markdown_content)}个字符"
|
| 210 |
return markdown_content, status
|
|
|
|
|
|
|
|
|
|
| 211 |
|
| 212 |
except Exception as e:
|
| 213 |
import traceback
|
|
|
|
| 87 |
if not terminators:
|
| 88 |
terminators = [2] # 使用常见的</s>标记ID作为默认值
|
| 89 |
|
| 90 |
+
# 使用CUDA运行Nougat的PDF处理函数
|
| 91 |
+
def process_pdf_with_nougat_gpu(pdf_path, output_dir=None):
|
| 92 |
+
"""使用GPU运行Nougat处理PDF文件"""
|
| 93 |
+
try:
|
| 94 |
+
# 如果未指定输出目录,使用PDF所在目录
|
| 95 |
+
if output_dir is None:
|
| 96 |
+
output_dir = os.path.dirname(pdf_path)
|
| 97 |
+
|
| 98 |
+
# 设置CUDA环境变量
|
| 99 |
+
env = os.environ.copy()
|
| 100 |
+
env["CUDA_VISIBLE_DEVICES"] = "0" # 使用第一个GPU
|
| 101 |
+
env["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
|
| 102 |
+
|
| 103 |
+
# 执行带有GPU支持的Nougat命令
|
| 104 |
+
print(f"使用GPU运行Nougat: {pdf_path}")
|
| 105 |
+
cmd = ["nougat", pdf_path, "-o", output_dir, "--device", "cuda"]
|
| 106 |
+
|
| 107 |
+
# 执行命令并捕获输出
|
| 108 |
+
result = subprocess.run(
|
| 109 |
+
cmd,
|
| 110 |
+
stdout=subprocess.PIPE,
|
| 111 |
+
stderr=subprocess.PIPE,
|
| 112 |
+
text=True,
|
| 113 |
+
env=env,
|
| 114 |
+
timeout=300 # 5分钟超时
|
| 115 |
+
)
|
| 116 |
+
|
| 117 |
+
# 检查命令执行结果
|
| 118 |
+
if result.returncode != 0:
|
| 119 |
+
print(f"Nougat GPU处理失败: {result.stderr}")
|
| 120 |
+
return None, result.stderr
|
| 121 |
+
|
| 122 |
+
# 获取生成的markdown文件路径
|
| 123 |
+
base_name = os.path.basename(pdf_path)
|
| 124 |
+
name_without_ext = os.path.splitext(base_name)[0]
|
| 125 |
+
markdown_path = os.path.join(output_dir, f"{name_without_ext}.mmd")
|
| 126 |
+
|
| 127 |
+
# 检查markdown文件是否生成
|
| 128 |
+
if not os.path.exists(markdown_path):
|
| 129 |
+
return None, "Nougat处理完成,但未找到生成的Markdown文件"
|
| 130 |
+
|
| 131 |
+
# 读取markdown内容
|
| 132 |
+
with open(markdown_path, "r", encoding="utf-8") as f:
|
| 133 |
+
markdown_content = f.read()
|
| 134 |
+
|
| 135 |
+
return markdown_content, None
|
| 136 |
+
|
| 137 |
+
except subprocess.TimeoutExpired:
|
| 138 |
+
return None, "Nougat处理超时"
|
| 139 |
+
|
| 140 |
+
except Exception as e:
|
| 141 |
+
import traceback
|
| 142 |
+
error = f"Nougat处理异常: {str(e)}\n{traceback.format_exc()}"
|
| 143 |
+
print(error)
|
| 144 |
+
return None, error
|
| 145 |
+
|
| 146 |
+
# 使用Python API的GPU处理方式
|
| 147 |
@spaces.GPU(stateless=True)
|
| 148 |
+
def process_pdf_with_nougat_api(pdf_path):
|
| 149 |
+
"""使用Nougat Python API与GPU处理PDF文件"""
|
| 150 |
try:
|
| 151 |
+
# 导入必要的库
|
| 152 |
+
from nougat import NougatModel
|
| 153 |
+
from nougat.utils.checkpoint import get_checkpoint
|
| 154 |
+
from nougat.dataset.rasterize import rasterize_paper
|
| 155 |
+
import torch
|
| 156 |
+
|
| 157 |
+
# 确保GPU可用
|
| 158 |
+
if not torch.cuda.is_available():
|
| 159 |
+
return None, "GPU不可用,无法使用Nougat API处理PDF"
|
| 160 |
+
|
| 161 |
+
# 显示GPU信息
|
| 162 |
+
device_count = torch.cuda.device_count()
|
| 163 |
+
device_name = torch.cuda.get_device_name(0) if device_count > 0 else "Unknown"
|
| 164 |
+
print(f"使用GPU: {device_name}, 可用GPU数量: {device_count}")
|
| 165 |
+
|
| 166 |
+
# 初始化模型并移至GPU
|
| 167 |
+
ckpt = get_checkpoint()
|
| 168 |
+
model = NougatModel.from_pretrained(ckpt)
|
| 169 |
+
device = torch.device("cuda")
|
| 170 |
+
model = model.to(device)
|
| 171 |
+
|
| 172 |
+
# 处理PDF
|
| 173 |
+
markdown_content = ""
|
| 174 |
+
pages = list(rasterize_paper(pdf_path))
|
| 175 |
+
|
| 176 |
+
# 使用tqdm显示进度
|
| 177 |
+
for page_idx, page in enumerate(tqdm(pages, desc="处理PDF页面")):
|
| 178 |
+
page = page.to(device)
|
| 179 |
+
markdown = model.inference(page)
|
| 180 |
+
markdown_content += f"--- Page {page_idx+1} ---\n{markdown}\n\n"
|
| 181 |
+
|
| 182 |
+
return markdown_content, None
|
| 183 |
+
|
| 184 |
except Exception as e:
|
| 185 |
import traceback
|
| 186 |
+
error = f"Nougat API处理异常: {str(e)}\n{traceback.format_exc()}"
|
| 187 |
+
print(error)
|
| 188 |
+
return None, error
|
| 189 |
|
| 190 |
+
# 添加PDF转换为Markdown函数
|
| 191 |
def convert_pdf_to_markdown(pdf_file):
|
| 192 |
"""使用Nougat将PDF转换为Markdown (GPU优化版)"""
|
| 193 |
if pdf_file is None:
|
|
|
|
| 205 |
with open(temp_pdf_path, "wb") as f:
|
| 206 |
f.write(pdf_file)
|
| 207 |
|
| 208 |
+
# 方法1: 首先尝试使用命令行GPU方式
|
| 209 |
+
print("方法1: 尝试使用命令行GPU方式处理PDF...")
|
| 210 |
+
markdown_content, error = process_pdf_with_nougat_gpu(temp_pdf_path, temp_dir)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 211 |
|
| 212 |
+
if markdown_content is not None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 213 |
# 限制文本长度
|
| 214 |
if len(markdown_content) > 20000:
|
| 215 |
markdown_content = markdown_content[:20000] + "\n\n...(Markdown内容已截断)"
|
| 216 |
|
| 217 |
+
status = f"PDF已成功转换为Markdown (GPU命令行): 生成了{len(markdown_content)}个字符"
|
| 218 |
return markdown_content, status
|
| 219 |
+
|
| 220 |
+
# 方法2: 如果命令行方式失败,尝试使用Python API方式
|
| 221 |
+
print(f"方法1失败: {error}")
|
| 222 |
+
print("方法2: 尝试使用Python API GPU方式处理PDF...")
|
| 223 |
+
|
| 224 |
+
markdown_content, api_error = process_pdf_with_nougat_api(temp_pdf_path)
|
| 225 |
+
|
| 226 |
+
if markdown_content is not None:
|
| 227 |
# 限制文本长度
|
| 228 |
if len(markdown_content) > 20000:
|
| 229 |
markdown_content = markdown_content[:20000] + "\n\n...(Markdown内容已截断)"
|
| 230 |
|
| 231 |
status = f"PDF已成功转换为Markdown (GPU API): 生成了{len(markdown_content)}个字符"
|
| 232 |
return markdown_content, status
|
| 233 |
+
|
| 234 |
+
# 所有方法都失败
|
| 235 |
+
return "", f"PDF转换失败: 所有GPU方法都失败了\n命令行错误: {error}\nAPI错误: {api_error}"
|
| 236 |
|
| 237 |
except Exception as e:
|
| 238 |
import traceback
|