【nougat推理】pdf转markdown文件代码demo示例&web_demo示例

模型介绍

Nougat是一个名为Donut的模型，它经过训练，可以将PDF文档转录成Markdown格式文档。该模型由Swin Transformer作为视觉编码器，以及mBART模型作为文本解码器组成。

该模型被训练成在只给出PDF图像像素作为输入的情况下，自回归地预测Markdown格式。
https://huggingface.co/facebook/nougat-base

安装依赖

pip install nltk python-Levenshtein
pip install spaces
pip install git+https://github.com/facebookresearch/nougat

直接使用

from huggingface_hub import hf_hub_download
import re
from PIL import Image
from pdf2image import convert_from_path
from transformers import NougatProcessor, VisionEncoderDecoderModel
from datasets import load_dataset
import torch
import os

processor = NougatProcessor.from_pretrained("/your/model/path/nougat-base")
model = VisionEncoderDecoderModel.from_pretrained("/your/model/path/nougat-base")

device = "cuda"
model.to(device)

# prepare PDF image for the model
pdf_file = 'your/pdf_data/path/xxx.pdf'
pages = convert_from_path(pdf_file, 300)  # 300 是输出图片的 DPI（每英寸点数）
# 创建保存图片的目录
output_dir = os.path.join(os.path.dirname(pdf_file), 'images')
os.makedirs(output_dir, exist_ok=True)

for i, page in enumerate(pages):
    image_path = os.path.join(output_dir, f'page_{i + 1}.jpg')
    page.save(image_path, 'JPEG')
    image = Image.open(image_path)
    print("*"*30,f"第{i}页","*"*30)
    pixel_values = processor(image, return_tensors="pt").pixel_values

    # generate transcription (here we only generate 30 tokens)
    outputs = model.generate(
        pixel_values.to(device),
        min_length=1,
        max_new_tokens=3000,
        bad_words_ids=[[processor.tokenizer.unk_token_id]],
    )

    sequences = processor.batch_decode(outputs, skip_special_tokens=True)
    for sequence in sequences:
        sequence = processor.post_process_generation(sequence, fix_markdown=False)
        # note: we're using repr here such for the sake of printing the \n characters, feel free to just print the sequence
        print(repr(sequence)）

搭建web并生成.md文件

from huggingface_hub import hf_hub_download
import re
from PIL import Image
import requests
from nougat.dataset.rasterize import rasterize_paper
from transformers import NougatProcessor, VisionEncoderDecoderModel
import torch
import gradio as gr
import uuid
import os
import spaces

processor = NougatProcessor.from_pretrained("/your/model/path/nougat-base")
model = VisionEncoderDecoderModel.from_pretrained("/your/model/path/nougat-base")

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device) 


def get_pdf(pdf_link):
  unique_filename = f"{os.getcwd()}/downloaded_paper_{uuid.uuid4().hex}.pdf"

  response = requests.get(pdf_link)

  if response.status_code == 200:
      with open(unique_filename, 'wb') as pdf_file:
          pdf_file.write(response.content)
      print("PDF downloaded successfully.")
  else:
      print("Failed to download the PDF.")
  return unique_filename


@spaces.GPU
def predict(image):
  #为模型准备PDF图像
  image = Image.open(image)
  pixel_values = processor(image, return_tensors="pt").pixel_values

  outputs = model.generate(
      pixel_values.to(device),
      min_length=1,
      max_new_tokens=3000,
      bad_words_ids=[[processor.tokenizer.unk_token_id]],
  )

  page_sequence = processor.batch_decode(outputs, skip_special_tokens=True)[0]
  page_sequence = processor.post_process_generation(page_sequence, fix_markdown=False)
  return page_sequence



def inference(pdf_file, pdf_link):
  if pdf_file is None:
    if pdf_link == '':
      print("未上传任何文件，也未提供任何链接")
      return "未提供任何数据。上传一个pdf文件或提供一个pdf链接，然后重试！"
    else:
      file_name = get_pdf(pdf_link)
  else:
    file_name = pdf_file.name
    pdf_name = pdf_file.name.split('/')[-1].split('.')[0]

  images = rasterize_paper(file_name, return_pil=True)
  sequence = ""
  # 推理每一页并合并
  for image in images:
    sequence += predict(image)


  content = sequence.replace(r'\(', '$').replace(r'\)', '$').replace(r'\[', '$$').replace(r'\]', '$$')
  with open(f"{os.getcwd()}/output.md","w+") as f:
      f.write(content)
      f.close()

      
  return content, f"{os.getcwd()}/output.md"


css = """
  #mkd {
    height: 500px; 
    overflow: auto; 
    border: 1px solid #ccc; 
  }
"""

with gr.Blocks(css=css) as demo:
  gr.HTML("<h1><center>Nougat模型推理PDF转Markdown 🍫<center><h1>")

  with gr.Row():
    mkd = gr.Markdown('<h4><center>上传PDF</center></h4>')
    mkd = gr.Markdown('<h4><center>提供PDF链接</center></h4>')
  
  
  with gr.Row(equal_height=True):
    pdf_file = gr.File(label='PDF 📑', file_count='single')
    pdf_link = gr.Textbox(placeholder='在此处输入arxiv链接', label='Link to Paper🔗')
  with gr.Row():
    btn = gr.Button('运行 🍫')
  with gr.Row():
    clr = gr.Button('清除输入和输出 🧼')

  output_headline = gr.Markdown("## Markdown展示👇")
  with gr.Row():
      parsed_output = gr.Markdown(elem_id='mkd', value='输出文本 📝')
      output_file = gr.File(file_types = ["txt"], label="输出文件 📑")
  
  btn.click(inference, [pdf_file, pdf_link], [parsed_output, output_file])
  clr.click(lambda : (gr.update(value=None), 
                      gr.update(value=None),
                      gr.update(value=None), 
                      gr.update(value=None)), 
             [], 
             [pdf_file, pdf_link, parsed_output, output_file]
            )
  gr.Examples(
      [["/your/pdf_data/path/xxx.pdf", ""], [None, "https://arxiv.org/pdf/2308.08316.pdf"]],
      inputs = [pdf_file, pdf_link],
      outputs = [parsed_output, output_file],
      fn=inference,
      cache_examples=True,
      label='点击下面的任何示例，快速获取 Nougat OCR 🍫结果:'
  )
    
demo.queue()
demo.launch(debug=True,share=True, server_name="192.168.10.60", server_port=7899)

测试结果

正常的arxiv上的论文识别效果很好，但是这个只限于英文
对于特殊pdf的识别效果很差，nougat自己的rasterize_paper函数没有pdf2image好
另外文档的排版也会干扰识别结果