app.py

from text_extractor import TextExtractor
from tqdm import tqdm
from transformers import PegasusForConditionalGeneration, PegasusTokenizer
from transformers import pipeline
from mdutils.mdutils import MdUtils
from pathlib import Path

import gradio as gr
import fitz
import torch
import copy
import os

FILENAME = ""

preprocess = TextExtractor()
model_name = "sshleifer/distill-pegasus-cnn-16-4"
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = PegasusTokenizer.from_pretrained(model_name, max_length=500)
model = PegasusForConditionalGeneration.from_pretrained(model_name).to(device)

def summarize(slides):
    generated_slides = copy.deepcopy(slides)
    for page, contents in tqdm(generated_slides.items()):
        for idx, (tag, content) in enumerate(contents):
            if tag.startswith('p'): 
                try:
                    input = tokenizer(content, truncation=True, padding="longest", return_tensors="pt").to(device)
                    tensor = model.generate(**input)
                    summary = tokenizer.batch_decode(tensor, skip_special_tokens=True)[0]
                    contents[idx] = (tag, summary)
                except Exception as e:
                    print(e)
                    print("Summarization Fails")
    return generated_slides


def convert2markdown(generate_slides):
    mdFile = MdUtils(file_name=FILENAME, title=f'{FILENAME} Presentation')
    for k, v in generate_slides.items():
        mdFile.new_paragraph('---')
        for section in v:
            tag = section[0]
            content = section[1]
            if tag.startswith('h'):
                mdFile.new_header(level=int(tag[1]), title=content)
            if tag == 'p':
                contents = content.split('<n>')
                for content in contents:
                    mdFile.new_paragraph(content)
    mdFile.create_md_file()
    return f"{FILENAME}.md"

def inference(document):
    global FILENAME
    print(document)
    doc = fitz.open(document)
    FILENAME = Path(doc.name).stem
    font_counts, styles = preprocess.get_font_info(doc, granularity=False)
    size_tag = preprocess.get_font_tags(font_counts, styles)
    texts = preprocess.assign_tags(doc, size_tag)
    slides = preprocess.get_slides(texts)
    generated_slides = summarize(slides)
    markdown_path = convert2markdown(generated_slides)
    # with open(markdown_path, 'rt') as f:
    #     markdown_str = f.read()
    return markdown_path


with gr.Blocks() as demo:
    inp = gr.File(file_types=['pdf'])
    out = gr.File(label="Markdown File")
    # out = gr.Textbox(label="Markdown Content")
    inference_btn = gr.Button("Summarized PDF")
    inference_btn.click(fn=inference, inputs=inp, outputs=out, show_progress=True, api_name="summarize")
    
demo.launch()