# https://huggingface.co/spaces/varl42/audio_abstract42
# Import  libraries
import gradio as gr
import torch
import PyPDF2
from transformers import pipeline 
import numpy
import scipy
from gtts import gTTS
from io import BytesIO
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Function to extract text from PDF 
# Defines a function to extract raw text from a PDF file
def extract_text(pdf_file):
    pdfReader = PyPDF2.PdfReader(pdf_file)
    pageObj = pdfReader.pages[0]
    return pageObj.extract_text()

    
# Function to summarize text
# Defines a function to summarize the extracted text using pszemraj/led-base-book-summary    
def summarize_text(text):
    sentences = text.split(". ")
    for i, sentence in enumerate(sentences):
        if "Abstract" in sentence:
            start = i + 1
            end = start + 6
            break
    abstract = ". ".join(sentences[start:end+1]) 
    #else:
    #    return("Abstract section not found")

  # Load BART model & tokenizer
    tokenizer = AutoTokenizer.from_pretrained("pszemraj/led-base-book-summary")
    model = AutoModelForSeq2SeqLM.from_pretrained("pszemraj/led-base-book-summary")

    # Tokenize abstract 
    inputs = tokenizer(abstract,
                       max_length=1024,
                       return_tensors="pt",
                       truncation=True)

    # Generate summary
    summary_ids = model.generate(inputs['input_ids'],
                                 max_length=50,
                                 min_length=30,
                                 no_repeat_ngram_size=3,
                                 encoder_no_repeat_ngram_size=3,
                                 repetition_penalty=3.5,
                                 num_beams=4,
                                 do_sample=True,
                                 early_stopping=False)
    
    summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

    if '.' in summary:
        index = summary.rindex('.')
        if index != -1:
           summary = summary[:index+1]

    return summary

# Function to convert text to audio
# Defines a function to convert text to an audio file using Google Text-to-Speech
def text_to_audio(text):
    tts = gTTS(text, lang='en') 
    buffer = BytesIO()
    tts.write_to_fp(buffer)
    buffer.seek(0)  
    return buffer.read()

### Main function 
### The main function that ties everything together:
### extracts text, summarizes, and converts to audio.
def audio_pdf(pdf_file):
    text = extract_text(pdf_file)
    summary = summarize_text(text) 
    audio = text_to_audio(summary)
    return summary, audio

# Define Gradio interface
# Gradio web interface with a file input, text output to display the summary
# and audio output to play the audio file. # Launches the interface
inputs = gr.File() 
summary_text = gr.Text()
audio_summary = gr.Audio()


iface = gr.Interface(
    fn=audio_pdf,
    inputs=inputs,
    outputs=[summary_text,audio_summary],
    title="PDF Audio Summarizer 📻",
    description="App that converts an abstract into audio, please upload only PDFs with an abstract",
    examples=["Hidden_Technical_Debt.pdf",
              "Attention_is_all_you_need.pdf", 
              "ImageNet_Classification.pdf"
             ]
)

iface.launch() # Launch the interface