Merge pull request #17 from myhloli/master

实现spark_api，parse逻辑重构，切图逻辑重构 

demo/ocr_demo.py

@@ -115,8 +115,9 @@ def ocr_parse_pdf_core(pdf_bytes, model_output_json_list, book_name, start_page_
 if __name__ == '__main__':
     pdf_path = r"/home/cxu/workspace/Magic-PDF/ocr_demo/j.1540-627x.2006.00176.x.pdf"
     json_file_path = r"/home/cxu/workspace/Magic-PDF/ocr_demo/j.1540-627x.2006.00176.x.json"
-    ocr_local_parse(pdf_path, json_file_path)
-    # book_name = "科数网/edu_00011318"
-    # ocr_online_parse(book_name)
+    # ocr_local_parse(pdf_path, json_file_path)
+    
+    book_name = "数学新星网/edu_00001236"
+    ocr_online_parse(book_name)
     
     pass

magic_pdf/dict2md/ocr_mkcontent.py

-from magic_pdf.libs.commons import s3_image_save_path, join_path
 from magic_pdf.libs.language import detect_lang
 from magic_pdf.libs.markdown_utils import ocr_escape_special_markdown_char
 from magic_pdf.libs.ocr_content_type import ContentType
@@ -56,7 +55,7 @@ def ocr_mk_mm_markdown(pdf_info_dict: dict):
                         if not span.get('image_path'):
                             continue
                         else:
-                            content = f"![]({join_path(s3_image_save_path, span['image_path'])})"
+                            content = f"![]({span['image_path']})"
                     else:
                         content = ocr_escape_special_markdown_char(span['content'])  # 转义特殊符号
                         if span['type'] == ContentType.InlineEquation:
@@ -123,7 +122,7 @@ def ocr_mk_markdown_with_para_core(paras_of_layout, mode):
                         content = f"\n$$\n{span['content']}\n$$\n"
                     elif span_type in [ContentType.Image, ContentType.Table]:
                         if mode == 'mm':
-                            content = f"\n![]({join_path(s3_image_save_path, span['image_path'])})\n"
+                            content = f"\n![]({span['image_path']})\n"
                         elif mode == 'nlp':
                             pass
                     if content != '':
@@ -195,13 +194,13 @@ def line_to_standard_format(line):
                 if span['type'] == ContentType.Image:
                     content = {
                         'type': 'image',
-                        'img_path': join_path(s3_image_save_path, span['image_path'])
+                        'img_path': span['image_path']
                     }
                     return content
                 elif span['type'] == ContentType.Table:
                     content = {
                         'type': 'table',
-                        'img_path': join_path(s3_image_save_path, span['image_path'])
+                        'img_path': span['image_path']
                     }
                     return content
         else:

magic_pdf/libs/commons.py

@@ -24,7 +24,7 @@ error_log_path = "s3://llm-pdf-text/err_logs/"
 # json_dump_path = "s3://pdf_books_temp/json_dump/" # 这条路径仅用于临时本地测试,不能提交到main
 json_dump_path = "s3://llm-pdf-text/json_dump/"
 
-s3_image_save_path = "s3://mllm-raw-media/pdf2md_img/" # TODO 基础库不应该有这些存在的路径，应该在业务代码中定义
+# s3_image_save_path = "s3://mllm-raw-media/pdf2md_img/" # 基础库不应该有这些存在的路径，应该在业务代码中定义
 
 
 def get_top_percent_list(num_list, percent):
@@ -120,28 +120,8 @@ def read_file(pdf_path: str, s3_profile):
             return f.read()
 
 
-def get_docx_model_output(pdf_model_output, pdf_model_s3_profile, page_id):
-    if isinstance(pdf_model_output, str):
-        model_output_json_path = join_path(pdf_model_output, f"page_{page_id + 1}.json")  # 模型输出的页面编号从1开始的
-        if os.path.exists(model_output_json_path):
-            json_from_docx = read_file(model_output_json_path, pdf_model_s3_profile)
-            model_output_json = json.loads(json_from_docx)
-        else:
-            try:
-                model_output_json_path = join_path(pdf_model_output, "model.json")
-                with open(model_output_json_path, "r", encoding="utf-8") as f:
-                    model_output_json = json.load(f)
-                    model_output_json = model_output_json["doc_layout_result"][page_id]
-            except:
-                s3_model_output_json_path = join_path(pdf_model_output, f"page_{page_id + 1}.json")
-                s3_model_output_json_path = join_path(pdf_model_output, f"{page_id}.json")
-                #s3_model_output_json_path = join_path(pdf_model_output, f"page_{page_id }.json")
-                # logger.warning(f"model_output_json_path: {model_output_json_path} not found. try to load from s3: {s3_model_output_json_path}")
-
-                s = read_file(s3_model_output_json_path, pdf_model_s3_profile)
-                return json.loads(s)
-
-    elif isinstance(pdf_model_output, list):
+def get_docx_model_output(pdf_model_output, page_id):
+
     model_output_json = pdf_model_output[page_id]
 
     return model_output_json

magic_pdf/libs/hash_utils.py

+import hashlib
+
+
+def compute_md5(file_bytes):
+    hasher = hashlib.md5()
+    hasher.update(file_bytes)
+    return hasher.hexdigest().upper()
+
+
+def compute_sha256(input_string):
+    hasher = hashlib.sha256()
+    # 在Python3中，需要将字符串转化为字节对象才能被哈希函数处理
+    input_bytes = input_string.encode('utf-8')
+    hasher.update(input_bytes)
+    return hasher.hexdigest()

magic_pdf/libs/pdf_image_tools.py

-import os
-from pathlib import Path
-from typing import Tuple
-import io
 
-# from app.common.s3 import get_s3_client
 from magic_pdf.libs.commons import fitz
 from loguru import logger
-from magic_pdf.libs.commons import parse_bucket_key, join_path
+from magic_pdf.libs.commons import join_path
+from magic_pdf.libs.hash_utils import compute_sha256
 
 
-def cut_image(bbox: Tuple, page_num: int, page: fitz.Page, save_parent_path: str, s3_return_path=None, img_s3_client=None, upload_switch=True):
+def cut_image(bbox: tuple, page_num: int, page: fitz.Page, return_path, imageWriter):
     """
     从第page_num页的page中，根据bbox进行裁剪出一张jpg图片，返回图片路径
     save_path：需要同时支持s3和本地, 图片存放在save_path下，文件名是: {page_num}_{bbox[0]}_{bbox[1]}_{bbox[2]}_{bbox[3]}.jpg , bbox内数字取整。
     """
     # 拼接文件名
-    filename = f"{page_num}_{int(bbox[0])}_{int(bbox[1])}_{int(bbox[2])}_{int(bbox[3])}.jpg"
-    # 拼接路径
-    image_save_path = join_path(save_parent_path, filename)
-    s3_img_path = join_path(s3_return_path, filename) if s3_return_path is not None else None
-    # 打印图片文件名
-    # print(f"Saved {image_save_path}")
-
-    #检查坐标
-    # x_check = int(bbox[2]) - int(bbox[0])
-    # y_check = int(bbox[3]) - int(bbox[1])
-    # if x_check <= 0 or y_check <= 0:
-    #
-    #     if image_save_path.startswith("s3://"):
-    #         logger.exception(f"传入图片坐标有误，x1<x0或y1<y0,{s3_img_path}")
-    #         return s3_img_path
-    #     else:
-    #         logger.exception(f"传入图片坐标有误，x1<x0或y1<y0,{image_save_path}")
-    #         return image_save_path
+    filename = f"{page_num}_{int(bbox[0])}_{int(bbox[1])}_{int(bbox[2])}_{int(bbox[3])}"
 
+    # 老版本返回不带bucket的路径
+    img_path = join_path(return_path, filename) if return_path is not None else None
+
+    # 新版本生成平铺路径
+    img_hash256_path = f"{compute_sha256(img_path)}.jpg"
 
     # 将坐标转换为fitz.Rect对象
     rect = fitz.Rect(*bbox)
@@ -42,39 +26,17 @@ def cut_image(bbox: Tuple, page_num: int, page: fitz.Page, save_parent_path: str
     # 截取图片
     pix = page.get_pixmap(clip=rect, matrix=zoom)
 
-    if image_save_path.startswith("s3://"):
-        if not upload_switch:
-            pass
-        else:
-            # 图片保存到s3
-            bucket_name, bucket_key = parse_bucket_key(image_save_path)
-            # 将字节流上传到s3
     byte_data = pix.tobytes(output='jpeg', jpg_quality=95)
-            file_obj = io.BytesIO(byte_data)
-            if img_s3_client is not None:
-                img_s3_client.upload_fileobj(file_obj, bucket_name, bucket_key)
-                # 每个图片上传任务都创建一个新的client
-                # img_s3_client_once = get_s3_client(image_save_path)
-                # img_s3_client_once.upload_fileobj(file_obj, bucket_name, bucket_key)
-            else:
-                logger.exception("must input img_s3_client")
-        return s3_img_path
-    else:
-        # 保存图片到本地
-        # 先检查一下image_save_path的父目录是否存在，如果不存在，就创建
-        parent_dir = os.path.dirname(image_save_path)
-        if not os.path.exists(parent_dir):
-            os.makedirs(parent_dir)
-        pix.save(image_save_path, jpg_quality=95)
-        # 为了直接能在markdown里看，这里把地址改为相对于mardown的地址
-        pth = Path(image_save_path)
-        image_save_path = f"{pth.parent.name}/{pth.name}"
-        return image_save_path
-
-
-def save_images_by_bboxes(book_name: str, page_num: int, page: fitz.Page, save_path: str,
-                            image_bboxes: list, images_overlap_backup:list, table_bboxes: list, equation_inline_bboxes: list,
-                            equation_interline_bboxes: list, img_s3_client) -> dict:
+
+    imageWriter.write(data=byte_data, path=img_hash256_path, mode="binary")
+
+    return img_hash256_path
+
+
+def save_images_by_bboxes(page_num: int, page: fitz.Page, pdf_bytes_md5: str,
+                          image_bboxes: list, images_overlap_backup: list, table_bboxes: list,
+                          equation_inline_bboxes: list,
+                          equation_interline_bboxes: list, imageWriter) -> dict:
     """
     返回一个dict, key为bbox, 值是图片地址
     """
@@ -85,53 +47,30 @@ def save_images_by_bboxes(book_name: str, page_num: int, page: fitz.Page, save_p
     interline_eq_info = []
 
     # 图片的保存路径组成是这样的： {s3_or_local_path}/{book_name}/{images|tables|equations}/{page_num}_{bbox[0]}_{bbox[1]}_{bbox[2]}_{bbox[3]}.jpg
-    s3_return_image_path = join_path(book_name, "images")
-    image_save_path = join_path(save_path, s3_return_image_path)
-
-    s3_return_table_path = join_path(book_name, "tables")
-    table_save_path = join_path(save_path, s3_return_table_path)
-
-    s3_return_equations_inline_path = join_path(book_name, "equations_inline")
-    equation_inline_save_path = join_path(save_path, s3_return_equations_inline_path)
-
-    s3_return_equation_interline_path = join_path(book_name, "equation_interline")
-    equation_interline_save_path = join_path(save_path, s3_return_equation_interline_path)
 
+    def return_path(type):
+        return join_path(pdf_bytes_md5, type)
 
     for bbox in image_bboxes:
-        if any([bbox[0]>=bbox[2], bbox[1]>=bbox[3]]):
+        if any([bbox[0] >= bbox[2], bbox[1] >= bbox[3]]):
             logger.warning(f"image_bboxes: 错误的box, {bbox}")
             continue
 
-        image_path = cut_image(bbox, page_num, page, image_save_path, s3_return_image_path, img_s3_client)
+        image_path = cut_image(bbox, page_num, page, return_path("images"), imageWriter)
         image_info.append({"bbox": bbox, "image_path": image_path})
 
     for bbox in images_overlap_backup:
-        if any([bbox[0]>=bbox[2], bbox[1]>=bbox[3]]):
+        if any([bbox[0] >= bbox[2], bbox[1] >= bbox[3]]):
             logger.warning(f"images_overlap_backup: 错误的box, {bbox}")
             continue
-        image_path = cut_image(bbox, page_num, page, image_save_path, s3_return_image_path, img_s3_client)
+        image_path = cut_image(bbox, page_num, page, return_path("images"), imageWriter)
         image_backup_info.append({"bbox": bbox, "image_path": image_path})
 
     for bbox in table_bboxes:
-        if any([bbox[0]>=bbox[2], bbox[1]>=bbox[3]]):
+        if any([bbox[0] >= bbox[2], bbox[1] >= bbox[3]]):
             logger.warning(f"table_bboxes: 错误的box, {bbox}")
             continue
-        image_path = cut_image(bbox, page_num, page, table_save_path, s3_return_table_path, img_s3_client)
+        image_path = cut_image(bbox, page_num, page, return_path("tables"), imageWriter)
         table_info.append({"bbox": bbox, "image_path": image_path})
 
-    for bbox in equation_inline_bboxes:
-        if any([bbox[0]>=bbox[2], bbox[1]>=bbox[3]]):
-            logger.warning(f"equation_inline_bboxes: 错误的box, {bbox}")
-            continue
-        image_path = cut_image(bbox[:4], page_num, page, equation_inline_save_path, s3_return_equations_inline_path, img_s3_client, upload_switch=False)
-        inline_eq_info.append({'bbox':bbox[:4], "image_path":image_path, "latex_text":bbox[4]})
-
-    for bbox in equation_interline_bboxes:
-        if any([bbox[0]>=bbox[2], bbox[1]>=bbox[3]]):
-            logger.warning(f"equation_interline_bboxes: 错误的box, {bbox}")
-            continue
-        image_path = cut_image(bbox[:4], page_num, page, equation_interline_save_path, s3_return_equation_interline_path, img_s3_client, upload_switch=False)
-        interline_eq_info.append({"bbox":bbox[:4], "image_path":image_path, "latex_text":bbox[4]})
-
     return image_info, image_backup_info, table_info, inline_eq_info, interline_eq_info
\ No newline at end of file

magic_pdf/pdf_parse_by_ocr.py

-import json
-import os
 import time
-
 from loguru import logger
-
-from magic_pdf.libs.draw_bbox import draw_layout_bbox, draw_text_bbox
 from magic_pdf.libs.commons import (
-    read_file,
-    join_path,
     fitz,
-    get_img_s3_client,
     get_delta_time,
     get_docx_model_output,
 )
 from magic_pdf.libs.coordinate_transform import get_scale_ratio
 from magic_pdf.libs.drop_tag import DropTag
+from magic_pdf.libs.hash_utils import compute_md5
 from magic_pdf.libs.ocr_content_type import ContentType
-from magic_pdf.libs.safe_filename import sanitize_filename
 from magic_pdf.para.para_split import para_split
 from magic_pdf.pre_proc.construct_page_dict import ocr_construct_page_component
 from magic_pdf.pre_proc.detect_footer_by_model import parse_footers
@@ -38,38 +30,16 @@ from magic_pdf.pre_proc.remove_bbox_overlap import remove_overlap_between_bbox
 def parse_pdf_by_ocr(
         pdf_bytes,
         pdf_model_output,
-        save_path,
-        book_name,
-        pdf_model_profile=None,
-        image_s3_config=None,
+        imageWriter,
         start_page_id=0,
         end_page_id=None,
         debug_mode=False,
 ):
-
-    save_tmp_path = os.path.join(os.path.dirname(__file__), "../..", "tmp", "unittest")
-    book_name = sanitize_filename(book_name)
-    md_bookname_save_path = ""
-    if debug_mode:
-        save_path = join_path(save_tmp_path, "md")
-        pdf_local_path = join_path(save_tmp_path, "download-pdfs", book_name)
-
-        if not os.path.exists(os.path.dirname(pdf_local_path)):
-            # 如果目录不存在，创建它
-            os.makedirs(os.path.dirname(pdf_local_path))
-
-        md_bookname_save_path = join_path(save_tmp_path, "md", book_name)
-        if not os.path.exists(md_bookname_save_path):
-            # 如果目录不存在，创建它
-            os.makedirs(md_bookname_save_path)
-
-        with open(pdf_local_path + ".pdf", "wb") as pdf_file:
-            pdf_file.write(pdf_bytes)
+    pdf_bytes_md5 = compute_md5(pdf_bytes)
 
     pdf_docs = fitz.open("pdf", pdf_bytes)
     # 初始化空的pdf_info_dict
     pdf_info_dict = {}
-    img_s3_client = get_img_s3_client(save_path, image_s3_config)
 
     start_time = time.time()
 
@@ -91,16 +61,14 @@ def parse_pdf_by_ocr(
 
         # 获取当前页的模型数据
         ocr_page_info = get_docx_model_output(
-            pdf_model_output, pdf_model_profile, page_id
+            pdf_model_output, page_id
         )
 
         """从json中获取每页的页码、页眉、页脚的bbox"""
         page_no_bboxes = parse_pageNos(page_id, page, ocr_page_info)
         header_bboxes = parse_headers(page_id, page, ocr_page_info)
         footer_bboxes = parse_footers(page_id, page, ocr_page_info)
-        footnote_bboxes = parse_footnotes_by_model(
-            page_id, page, ocr_page_info, md_bookname_save_path, debug_mode=debug_mode
-        )
+        footnote_bboxes = parse_footnotes_by_model(page_id, page, ocr_page_info, debug_mode=debug_mode)
 
         # 构建需要remove的bbox字典
         need_remove_spans_bboxes_dict = {
@@ -179,7 +147,7 @@ def parse_pdf_by_ocr(
         spans, dropped_spans_by_removed_bboxes = remove_spans_by_bboxes_dict(spans, need_remove_spans_bboxes_dict)
 
         '''对image和table截图'''
-        spans = cut_image_and_table(spans, page, page_id, book_name, save_path, img_s3_client)
+        spans = cut_image_and_table(spans, page, page_id, pdf_bytes_md5, imageWriter)
 
         '''行内公式调整, 高度调整至与同行文字高度一致(优先左侧, 其次右侧)'''
         displayed_list = []
@@ -242,16 +210,4 @@ def parse_pdf_by_ocr(
     """分段"""
     para_split(pdf_info_dict, debug_mode=debug_mode)
 
-    '''在测试时,保存调试信息'''
-    if debug_mode:
-        params_file_save_path = join_path(
-            save_tmp_path, "md", book_name, "preproc_out.json"
-        )
-        with open(params_file_save_path, "w", encoding="utf-8") as f:
-            json.dump(pdf_info_dict, f, ensure_ascii=False, indent=4)
-
-        # drow_bbox
-        draw_layout_bbox(pdf_info_dict, pdf_bytes, md_bookname_save_path)
-        draw_text_bbox(pdf_info_dict, pdf_bytes, md_bookname_save_path)
-
     return pdf_info_dict

magic_pdf/pdf_parse_for_train.py

@@ -112,7 +112,6 @@ def parse_pdf_for_train(
     pdf_model_output,
     save_path,
     book_name,
-    pdf_model_profile=None,
     image_s3_config=None,
     start_page_id=0,
     end_page_id=None,
@@ -200,7 +199,7 @@ def parse_pdf_for_train(
             flags=fitz.TEXTFLAGS_TEXT,
         )["blocks"]
         model_output_json = get_docx_model_output(
-            pdf_model_output, pdf_model_profile, page_id
+            pdf_model_output, page_id
         )
 
         # 解析图片

magic_pdf/pre_proc/detect_footnote.py

@@ -3,7 +3,7 @@ from magic_pdf.libs.commons import fitz             # pyMuPDF库
 from magic_pdf.libs.coordinate_transform import get_scale_ratio
 
 
-def parse_footnotes_by_model(page_ID: int, page: fitz.Page, json_from_DocXchain_obj: dict, md_bookname_save_path, debug_mode=False):
+def parse_footnotes_by_model(page_ID: int, page: fitz.Page, json_from_DocXchain_obj: dict, md_bookname_save_path=None, debug_mode=False):
     """
     :param page_ID: int类型，当前page在当前pdf文档中是第page_D页。
     :param page :fitz读取的当前页的内容

magic_pdf/pre_proc/ocr_cut_image.py

@@ -3,18 +3,16 @@ from magic_pdf.libs.ocr_content_type import ContentType
 from magic_pdf.libs.pdf_image_tools import cut_image
 
 
-def cut_image_and_table(spans, page, page_id, book_name, save_path, img_s3_client):
-    def s3_return_path(type):
-        return join_path(book_name, type)
+def cut_image_and_table(spans, page, page_id, pdf_bytes_md5, imageWriter):
 
-    def img_save_path(type):
-        return join_path(save_path, s3_return_path(type))
+    def return_path(type):
+        return join_path(pdf_bytes_md5, type)
 
     for span in spans:
         span_type = span['type']
         if span_type == ContentType.Image:
-            span['image_path'] = cut_image(span['bbox'], page_id, page, img_save_path('images'), s3_return_path=s3_return_path('images'), img_s3_client=img_s3_client)
+            span['image_path'] = cut_image(span['bbox'], page_id, page, return_path=return_path('images'), imageWriter=imageWriter)
         elif span_type == ContentType.Table:
-            span['image_path'] = cut_image(span['bbox'], page_id, page, img_save_path('tables'), s3_return_path=s3_return_path('tables'), img_s3_client=img_s3_client)
+            span['image_path'] = cut_image(span['bbox'], page_id, page, return_path=return_path('tables'), imageWriter=imageWriter)
 
     return spans

magic_pdf/spark/spark_api.py

@@ -12,27 +12,78 @@
 其余部分至于构造s3cli, 获取ak,sk都在code-clean里写代码完成。不要反向依赖！！！
 
 """
-
+from loguru import logger
 
 from magic_pdf.io import AbsReaderWriter
+from magic_pdf.pdf_parse_by_ocr import parse_pdf_by_ocr
+from magic_pdf.pdf_parse_by_txt import parse_pdf_by_txt
 
 
 def parse_txt_pdf(pdf_bytes:bytes, pdf_models:list, imageWriter: AbsReaderWriter, is_debug=False, start_page=0, *args, **kwargs):
     """
     解析文本类pdf
     """
-    pass
+    pdf_info_dict = parse_pdf_by_txt(
+        pdf_bytes,
+        pdf_models,
+        imageWriter,
+        start_page_id=start_page,
+        debug_mode=is_debug,
+    )
+
+    pdf_info_dict["parse_type"] = "txt"
+
+    return pdf_info_dict
 
 
 def parse_ocr_pdf(pdf_bytes:bytes,  pdf_models:list, imageWriter: AbsReaderWriter, is_debug=False, start_page=0, *args, **kwargs):
     """
     解析ocr类pdf
     """
-    pass
+    pdf_info_dict = parse_pdf_by_ocr(
+        pdf_bytes,
+        pdf_models,
+        imageWriter,
+        start_page_id=start_page,
+        debug_mode=is_debug,
+    )
+
+    pdf_info_dict["parse_type"] = "ocr"
+
+    return pdf_info_dict
 
 
 def parse_union_pdf(pdf_bytes:bytes,  pdf_models:list, imageWriter: AbsReaderWriter, is_debug=False, start_page=0,  *args, **kwargs):
     """
     ocr和文本混合的pdf，全部解析出来
     """
+    def parse_pdf(method):
+        try:
+            return method(
+                pdf_bytes,
+                pdf_models,
+                imageWriter,
+                start_page_id=start_page,
+                debug_mode=is_debug,
+            )
+        except Exception as e:
+            logger.error(f"{method.__name__} error: {e}")
+            return None
+
+    pdf_info_dict = parse_pdf(parse_pdf_by_txt)
+
+    if pdf_info_dict is None or pdf_info_dict.get("need_drop", False):
+        logger.warning(f"parse_pdf_by_txt drop or error, switch to parse_pdf_by_ocr")
+        pdf_info_dict = parse_pdf(parse_pdf_by_ocr)
+        if pdf_info_dict is None:
+            raise Exception("Both parse_pdf_by_txt and parse_pdf_by_ocr failed.")
+        else:
+            pdf_info_dict["parse_type"] = "ocr"
+    else:
+        pdf_info_dict["parse_type"] = "txt"
+
+    return pdf_info_dict
+
+
+def spark_json_extractor(jso:dict):
     pass