app.py

# https://huggingface.co/spaces/amendolajine/OPIT

# Here are the imports
import logging
import gradio as gr
import fitz  # PyMuPDF
from transformers import BartTokenizer, BartForConditionalGeneration, pipeline
import scipy.io.wavfile
import numpy as np

# Here is the code

# Initialize logging
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s - %(levelname)s - %(message)s')

# Initialize tokenizers and models
tokenizer = BartTokenizer.from_pretrained('facebook/bart-large-cnn')
model = BartForConditionalGeneration.from_pretrained('facebook/bart-large-cnn')
synthesiser = pipeline("text-to-speech", "suno/bark")

def extract_abstract(pdf_bytes):
    try:
        doc = fitz.open(stream=pdf_bytes, filetype="pdf")
        first_page = doc[0].get_text()
        start_idx = first_page.lower().find("abstract")
        end_idx = first_page.lower().find("introduction")
        if start_idx != -1 and end_idx != -1:
            return first_page[start_idx:end_idx].strip()
        else:
            return "Abstract not found or 'Introduction' not found in the first page."
    except Exception as e:
        logging.error(f"Error extracting abstract: {e}")
        return "Error in abstract extraction"

def process_text(uploaded_file):
    logging.debug(f"Uploaded file type: {type(uploaded_file)}")
    logging.debug(f"Uploaded file content: {uploaded_file}")

    try:
        with open(uploaded_file, "rb") as file:
            pdf_bytes = file.read()
    except Exception as e:
        logging.error(f"Error reading file from path: {e}")
        return "Error reading PDF file", None

    try:
        abstract_text = extract_abstract(pdf_bytes)
        logging.info(f"Extracted abstract: {abstract_text[:200]}...")
    except Exception as e:
        logging.error(f"Error in abstract extraction: {e}")
        return "Error in processing PDF", None

    try:
        inputs = tokenizer([abstract_text], max_length=1024, return_tensors='pt', truncation=True, padding="max_length")
        summary_ids = model.generate(
            input_ids=inputs['input_ids'],
            attention_mask=inputs['attention_mask'],
            pad_token_id=model.config.pad_token_id,
            num_beams=4,
            max_length=45,
            min_length=10,
            length_penalty=2.0,
            early_stopping=True,
            no_repeat_ngram_size=2
        )
        summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

        words = summary.split()
        cleaned_summary = []
        for i, word in enumerate(words):
            if '-' in word and i < len(words) - 1:
                word = word.replace('-', '') + words[i + 1]
                words[i + 1] = ""

            if '.' in word and i != len(words) - 1:
                word = word.replace('.', '')
                cleaned_summary.append(word + ' and')
            else:
                cleaned_summary.append(word) 

        final_summary = ' '.join(cleaned_summary)
        final_summary = final_summary[0].upper() + final_summary[1:]
        final_summary = ' '.join(w[0].lower() + w[1:] if w.lower() != 'and' else w for w in final_summary.split())

        speech = synthesiser(final_summary, forward_params={"do_sample": True})
        audio_data = speech["audio"].squeeze()
        normalized_audio_data = np.int16(audio_data / np.max(np.abs(audio_data)) * 32767)

        output_file = "temp_output.wav"
        scipy.io.wavfile.write(output_file, rate=speech["sampling_rate"], data=normalized_audio_data)

        return final_summary, output_file
    except Exception as e:
        logging.error(f"Error in summary generation or TTS conversion: {e}")
        return "Error in summary or speech generation", None

iface = gr.Interface(
    fn=process_text,
    inputs=gr.components.File(label="Upload a research PDF containing an abstract"),
    outputs=["text", "audio"],
    title="Summarize an abstract and vocalize it",
    description="Upload a research paper in PDF format to extract, summarize its abstract, and convert the summarization to speech. If the upload doesn't work on the first try, refresh the page (CTRL+F5) and try again."
)

if __name__ == "__main__":
    iface.launch()