|
| 1 | +import json |
| 2 | +import os |
| 3 | +from base64 import b64encode |
| 4 | +from glob import glob |
| 5 | +from io import StringIO |
| 6 | +from typing import Tuple, Union |
| 7 | + |
| 8 | +import uvicorn |
| 9 | +from fastapi import FastAPI, UploadFile, File |
| 10 | +from fastapi.responses import JSONResponse |
| 11 | +from loguru import logger |
| 12 | +from tempfile import TemporaryDirectory |
| 13 | +from pathlib import Path |
| 14 | +import fitz # PyMuPDF |
| 15 | +import asyncio |
| 16 | +from concurrent.futures import ProcessPoolExecutor |
| 17 | +import torch |
| 18 | +import multiprocessing as mp |
| 19 | +from contextlib import asynccontextmanager |
| 20 | +import time |
| 21 | + |
| 22 | +import magic_pdf.model as model_config |
| 23 | +from magic_pdf.config.enums import SupportedPdfParseMethod |
| 24 | +from magic_pdf.data.data_reader_writer import DataWriter, FileBasedDataWriter |
| 25 | +from magic_pdf.data.dataset import PymuDocDataset |
| 26 | +from magic_pdf.model.doc_analyze_by_custom_model import doc_analyze |
| 27 | +from magic_pdf.operators.models import InferenceResult |
| 28 | +from magic_pdf.operators.pipes import PipeResult |
| 29 | + |
| 30 | +model_config.__use_inside_model__ = True |
| 31 | + |
| 32 | +app = FastAPI() |
| 33 | + |
| 34 | +process_variables = {} |
| 35 | +my_pool = None |
| 36 | + |
| 37 | +class MemoryDataWriter(DataWriter): |
| 38 | + def __init__(self): |
| 39 | + self.buffer = StringIO() |
| 40 | + |
| 41 | + def write(self, path: str, data: bytes) -> None: |
| 42 | + if isinstance(data, str): |
| 43 | + self.buffer.write(data) |
| 44 | + else: |
| 45 | + self.buffer.write(data.decode("utf-8")) |
| 46 | + |
| 47 | + def write_string(self, path: str, data: str) -> None: |
| 48 | + self.buffer.write(data) |
| 49 | + |
| 50 | + def get_value(self) -> str: |
| 51 | + return self.buffer.getvalue() # 修复：使用 getvalue() 而不是 get_value() |
| 52 | + |
| 53 | + def close(self): |
| 54 | + self.buffer.close() |
| 55 | + |
| 56 | +def worker_init(counter, lock): |
| 57 | + num_gpus = torch.cuda.device_count() |
| 58 | + processes_per_gpu = int(os.environ.get('PROCESSES_PER_GPU', 1)) |
| 59 | + with lock: |
| 60 | + worker_id = counter.value |
| 61 | + counter.value += 1 |
| 62 | + if num_gpus == 0: |
| 63 | + device = 'cpu' |
| 64 | + else: |
| 65 | + device_id = worker_id // processes_per_gpu |
| 66 | + if device_id >= num_gpus: |
| 67 | + raise ValueError(f"Worker ID {worker_id} exceeds available GPUs ({num_gpus}).") |
| 68 | + device = f'cuda:{device_id}' |
| 69 | + config = { |
| 70 | + "parse_method": "auto", |
| 71 | + "ADDITIONAL_KEY": "VALUE" |
| 72 | + } |
| 73 | + converter = init_converter(config, device_id) |
| 74 | + pid = os.getpid() |
| 75 | + process_variables[pid] = converter |
| 76 | + print(f"Worker {worker_id}: Models loaded successfully on {device}!") |
| 77 | + |
| 78 | +def init_converter(config, device_id): |
| 79 | + os.environ["CUDA_VISIBLE_DEVICES"] = str(device_id) |
| 80 | + return config |
| 81 | + |
| 82 | +def img_to_base64(img_path: str) -> str: |
| 83 | + with open(img_path, "rb") as img_file: |
| 84 | + return b64encode(img_file.read()).decode('utf-8') |
| 85 | + |
| 86 | +def embed_images_as_base64(md_content: str, image_dir: str) -> str: |
| 87 | + lines = md_content.split('\n') |
| 88 | + new_lines = [] |
| 89 | + for line in lines: |
| 90 | + if line.startswith("![") and "](" in line and ")" in line: |
| 91 | + start_idx = line.index("](") + 2 |
| 92 | + end_idx = line.index(")", start_idx) |
| 93 | + img_rel_path = line[start_idx:end_idx] |
| 94 | + img_name = os.path.basename(img_rel_path) |
| 95 | + img_path = os.path.join(image_dir, img_name) |
| 96 | + logger.info(f"Checking image: {img_path}") |
| 97 | + if os.path.exists(img_path): |
| 98 | + img_base64 = img_to_base64(img_path) |
| 99 | + new_line = f"" |
| 100 | + new_lines.append(new_line) |
| 101 | + else: |
| 102 | + logger.warning(f"Image not found: {img_path}") |
| 103 | + new_lines.append(line) |
| 104 | + else: |
| 105 | + new_lines.append(line) |
| 106 | + return '\n'.join(new_lines) |
| 107 | + |
| 108 | +def process_pdf(pdf_path, output_dir): |
| 109 | + try: |
| 110 | + pid = os.getpid() |
| 111 | + config = process_variables.get(pid, "No variable") |
| 112 | + parse_method = config["parse_method"] |
| 113 | + |
| 114 | + with open(str(pdf_path), "rb") as f: |
| 115 | + pdf_bytes = f.read() |
| 116 | + |
| 117 | + output_path = Path(output_dir) / f"{Path(pdf_path).stem}_output" |
| 118 | + os.makedirs(str(output_path), exist_ok=True) |
| 119 | + image_dir = os.path.join(str(output_path), "images") |
| 120 | + os.makedirs(image_dir, exist_ok=True) |
| 121 | + image_writer = FileBasedDataWriter(str(output_path)) |
| 122 | + |
| 123 | + # 处理 PDF |
| 124 | + infer_result, pipe_result = process_pdf_content(pdf_bytes, parse_method, image_writer) |
| 125 | + |
| 126 | + md_content_writer = MemoryDataWriter() |
| 127 | + pipe_result.dump_md(md_content_writer, "", "images") |
| 128 | + md_content = md_content_writer.get_value() |
| 129 | + md_content_writer.close() |
| 130 | + |
| 131 | + # 获取保存的图片路径 |
| 132 | + image_paths = glob(os.path.join(image_dir, "*.jpg")) |
| 133 | + logger.info(f"Saved images by magic_pdf: {image_paths}") |
| 134 | + |
| 135 | + # 如果 magic_pdf 未保存足够图片，使用 fitz 提取 |
| 136 | + if not image_paths or len(image_paths) < 3: # 假设至少 3 张图片 |
| 137 | + logger.warning("Insufficient images saved by magic_pdf, falling back to fitz extraction") |
| 138 | + image_map = {} |
| 139 | + original_names = [] |
| 140 | + # 收集 Markdown 中的所有图片文件名 |
| 141 | + for line in md_content.split('\n'): |
| 142 | + if line.startswith("![") and "](" in line and ")" in line: |
| 143 | + start_idx = line.index("](") + 2 |
| 144 | + end_idx = line.index(")", start_idx) |
| 145 | + img_rel_path = line[start_idx:end_idx] |
| 146 | + original_names.append(os.path.basename(img_rel_path)) |
| 147 | + |
| 148 | + # 提取图片并映射 |
| 149 | + with fitz.open(pdf_path) as doc: |
| 150 | + img_counter = 0 |
| 151 | + for page_num, page in enumerate(doc): |
| 152 | + for img_index, img in enumerate(page.get_images(full=True)): |
| 153 | + xref = img[0] |
| 154 | + base = doc.extract_image(xref) |
| 155 | + if img_counter < len(original_names): |
| 156 | + img_name = original_names[img_counter] # 使用 Markdown 中的原始文件名 |
| 157 | + else: |
| 158 | + img_name = f"page_{page_num}_img_{img_index}.jpg" |
| 159 | + img_path = os.path.join(image_dir, img_name) |
| 160 | + with open(img_path, "wb") as f: |
| 161 | + f.write(base["image"]) |
| 162 | + if img_counter < len(original_names): |
| 163 | + image_map[original_names[img_counter]] = img_name |
| 164 | + img_counter += 1 |
| 165 | + |
| 166 | + image_paths = glob(os.path.join(image_dir, "*.jpg")) |
| 167 | + logger.info(f"Images extracted by fitz: {image_paths}") |
| 168 | + |
| 169 | + # 更新 Markdown（仅在必要时替换） |
| 170 | + for original_name, new_name in image_map.items(): |
| 171 | + if original_name != new_name: |
| 172 | + md_content = md_content.replace(f"images/{original_name}", f"images/{new_name}") |
| 173 | + |
| 174 | + return { |
| 175 | + "status": "success", |
| 176 | + "text": md_content, |
| 177 | + "output_path": str(output_path), |
| 178 | + "images": image_paths |
| 179 | + } |
| 180 | + except Exception as e: |
| 181 | + logger.error(f"Error processing PDF: {str(e)}") |
| 182 | + return { |
| 183 | + "status": "error", |
| 184 | + "message": str(e), |
| 185 | + "file": str(pdf_path) |
| 186 | + } |
| 187 | + |
| 188 | +def process_pdf_content(pdf_bytes, parse_method, image_writer): |
| 189 | + ds = PymuDocDataset(pdf_bytes) |
| 190 | + infer_result: InferenceResult = None |
| 191 | + pipe_result: PipeResult = None |
| 192 | + |
| 193 | + if parse_method == "ocr": |
| 194 | + infer_result = ds.apply(doc_analyze, ocr=True) |
| 195 | + pipe_result = infer_result.pipe_ocr_mode(image_writer) |
| 196 | + elif parse_method == "txt": |
| 197 | + infer_result = ds.apply(doc_analyze, ocr=False) |
| 198 | + pipe_result = infer_result.pipe_txt_mode(image_writer) |
| 199 | + else: # auto |
| 200 | + if ds.classify() == SupportedPdfParseMethod.OCR: |
| 201 | + infer_result = ds.apply(doc_analyze, ocr=True) |
| 202 | + pipe_result = infer_result.pipe_ocr_mode(image_writer) |
| 203 | + else: |
| 204 | + infer_result = ds.apply(doc_analyze, ocr=False) |
| 205 | + pipe_result = infer_result.pipe_txt_mode(image_writer) |
| 206 | + |
| 207 | + return infer_result, pipe_result |
| 208 | + |
| 209 | +@asynccontextmanager |
| 210 | +async def lifespan(app: FastAPI): |
| 211 | + try: |
| 212 | + mp.set_start_method('spawn') |
| 213 | + except RuntimeError: |
| 214 | + raise RuntimeError("Set start method to spawn twice. This may be a temporary issue with the script. Please try running it again.") |
| 215 | + global my_pool |
| 216 | + manager = mp.Manager() |
| 217 | + worker_counter = manager.Value('i', 0) |
| 218 | + worker_lock = manager.Lock() |
| 219 | + gpu_count = torch.cuda.device_count() |
| 220 | + my_pool = ProcessPoolExecutor(max_workers=gpu_count * int(os.environ.get('PROCESSES_PER_GPU', 1)), |
| 221 | + initializer=worker_init, initargs=(worker_counter, worker_lock)) |
| 222 | + yield |
| 223 | + if my_pool: |
| 224 | + my_pool.shutdown(wait=True) |
| 225 | + print("Application shutdown, cleaning up...") |
| 226 | + |
| 227 | +app.router.lifespan_context = lifespan |
| 228 | + |
| 229 | +@app.post("/v2/parse/file") |
| 230 | +async def process_pdfs(file: UploadFile = File(...)): |
| 231 | + s_time = time.time() |
| 232 | + with TemporaryDirectory() as temp_dir: |
| 233 | + temp_path = Path(temp_dir) / file.filename |
| 234 | + with open(str(temp_path), "wb") as buffer: |
| 235 | + buffer.write(await file.read()) |
| 236 | + |
| 237 | + # 验证 PDF 文件 |
| 238 | + try: |
| 239 | + with fitz.open(str(temp_path)) as pdf_document: |
| 240 | + total_pages = pdf_document.page_count |
| 241 | + except fitz.fitz.FileDataError: |
| 242 | + return JSONResponse(content={"success": False, "message": "", "error": "Invalid PDF file"}, status_code=400) |
| 243 | + except Exception as e: |
| 244 | + logger.error(f"Error opening PDF: {str(e)}") |
| 245 | + return JSONResponse(content={"success": False, "message": "", "error": f"Internal server error: {str(e)}"}, status_code=500) |
| 246 | + |
| 247 | + try: |
| 248 | + loop = asyncio.get_running_loop() |
| 249 | + results = await loop.run_in_executor( |
| 250 | + my_pool, |
| 251 | + process_pdf, |
| 252 | + str(temp_path), |
| 253 | + str(temp_dir) |
| 254 | + ) |
| 255 | + |
| 256 | + if results.get("status") == "error": |
| 257 | + return JSONResponse(content={ |
| 258 | + "success": False, |
| 259 | + "message": "", |
| 260 | + "error": results.get("message") |
| 261 | + }, status_code=500) |
| 262 | + |
| 263 | + # 嵌入 Base64 |
| 264 | + image_dir = os.path.join(results.get("output_path"), "images") |
| 265 | + md_content_with_base64 = embed_images_as_base64(results.get("text"), image_dir) |
| 266 | + |
| 267 | + return { |
| 268 | + "success": True, |
| 269 | + "message": "", |
| 270 | + "markdown": md_content_with_base64, |
| 271 | + "pages": total_pages |
| 272 | + } |
| 273 | + except Exception as e: |
| 274 | + logger.error(f"Error in process_pdfs: {str(e)}") |
| 275 | + return JSONResponse(content={ |
| 276 | + "success": False, |
| 277 | + "message": "", |
| 278 | + "error": f"Internal server error: {str(e)}" |
| 279 | + }, status_code=500) |
| 280 | + |
| 281 | +if __name__ == "__main__": |
| 282 | + uvicorn.run(app, host="0.0.0.0", port=7231) |
0 commit comments