This repository is inspired by ppisljar/pdf-translator and adds the following features:

• [GUI] Add support for download-save-translation process on the server (better for mobile devices)
• Support of Ollama and QWEN for translation (by api)
• Support multi threading for translation
• Make a better support of inplace translation of different languages (e.g. support Chinese translation)
• Support batch translation of PDF files without api calls (provides an example later)
• Use single process for ocr / layout model to save vram
• Use LLM for reference checking (fix the bug of supplemental translation)

• translate PDF files while preserving layout

• translation engines:

  • ollama (just added, works fine for translation)
  • openAI (best)
  • QWEN
  • google translate

• layout recognition engines:

  • UniLM DiT

• OCR engines:

  • PaddleOCR

• Render engine:

  • ReportLab

1. Clone this repository

   git clone https://github.com/poppanda/LLM_PDF_Translator.git
   cd LLM_PDF_Translator

2. Edit config.yaml and enter openai api key change type to 'openai' and enter your key under api_key if this is not changed translation engine will default to google translate

3. Build the docker image via Makefile

   make build

4. Run the docker container via Makefile

   make run

3. create venv and activte

prerequesites:

• ffmpeg, ... possibly more, check Dockerfile if you are running into issues

python3 -m venv .
source bin/activate

4. install requirements

pip3 install -r requirements.txt
pip3 install "git+https://github.com/facebookresearch/detectron2.git"

5. get models

make get_models

6. run

python3 server.py

Access to GUI via browser.

http://localhost:8765

• NVIDIA GPU (currently only support NVIDIA GPU)
• Docker

This repository does not allow commercial use.

This repository is licensed under CC BY-NC 4.0. See LICENSE for more information.

• The scene is that if you run a LLM model locally (e.g. ollama), the ocr/layout model will be loaded for nothing while translation, which is a waste of vram (for about 5GB).
• This problem is fixed by
  • Seperate the ocr / layout process from the translation process.
  • Use a single process for ocr/layout model.
  • Kill the process before translation.

• The original code checks the reference by recognizing the 'reference' keyword in the title.
• The problem is that:
  • There may be some supplemental material after the reference, and by the original code, the supplemental material will be translated as well.
  • The 'reference' keyword may not be recognized in some cases.
• The problem is fixed by:
  • Use LLM to check the reference and skip the translation.

• Support M1 Mac or CPU
• switch to VGT for layout detection
• add font detection (family/style/color/size/alignment)
• add support for translating lists
• add support for translating tables
• add support for translating text within images

# batch.py
import warnings
warnings.filterwarnings("ignore")
import server
import os
from pathlib import Path
import tempfile
from loguru import logger

pdf_dir = "" # path to the directory with pdf files

if __name__ == "__main__":
    translator = server.TranslateApi()
    files = list(os.scandir(pdf_dir))
    for file in files:
        if file.is_dir():
            files.extend(list(os.scandir(file.path)))
        elif file.is_file() and file.name.endswith(".pdf"):
            if file.name.endswith("_translated.pdf") or os.path.exists(file.path.replace(".pdf", "_translated.pdf")):
                logger.info(f"Skip {file.path}")
                continue
            logger.info(f"Translating {file.path}")
            response = translator._translate_pdf(
               file.path, 
               translator.temp_dir_name, 
               "English", 
               "Chinese", 
               translate_all=True, 
               p_from=0, 
               p_to=0, 
               side_by_side=True, 
               output_file_path=file.path.replace(".pdf", "_translated.pdf"))

• For PDF layout analysis, using DiT.

• For PDF to text conversion, using PaddlePaddle model.