app.py

import os
import time
import re
from typing import Union, AnyStr
from urllib.parse import urlparse, parse_qs
import textwrap
import streamlit as st
import openai
from openai import OpenAI
from pydub import AudioSegment
from youtube_transcript_api import YouTubeTranscriptApi
from deep_translator import GoogleTranslator
import yt_dlp as youtube_dl
from transformers import  AutoModelForCausalLM, GPT2Tokenizer
import torch
from tqdm import trange
import torch.nn.functional as F
client = OpenAI(
    api_key='sk-proj-Yzjez2g6rfAiVPpb3cfJT3BlbkFJRLU4ZQpMhyLJDf0XksF4'
)
 
def generate_response(article_text, lang ):
    messages=[
                 {"role": "system", "content": "You are an expert in summarizing text in two languages: English and Vietnamese"},
             {"role": "user", "content": f"summarize the following text professionally and return the summary according to the input language:\n{article_text}\nSummary:"}
            ]
    if lang == 'vi':
        messages=[
                 {"role": "system", "content": "Bạn là chuyên gia tóm tắt văn bản bằng hai ngôn ngữ: tiếng Anh và tiếng Việt"},
             {"role": "user", "content": f"hãy tóm tắt văn bản sau đây một cách chuyên nghiệp và trả về bản tóm tắt theo ngôn ngữ đầu vào:\n{article_text}\nBản Tóm tắt:"}
            ]
    response = client.chat.completions.create(
            model='ft:gpt-3.5-turbo-0125:personal::9eZjpJwa' ,
            messages=messages,
           max_tokens=150,         # Tăng lên để có thêm không gian cho tóm tắt
    temperature=0.3,        # Giảm xuống để tạo ra nội dung tập trung hơn
    top_p=0.95,             # Tăng nhẹ để mở rộng phạm vi từ vựng
    frequency_penalty=0.5,  # Tăng lên để khuyến khích đa dạng từ ngữ
    presence_penalty=0.5    # Tăng lên để khuyến khích đề cập đến các chủ đề mới
        )

        # Extract and return the generated summary
    summary = response.choices[0].message.content.strip()
    return summary
def cleaning_input(input_text):
    from html import unescape
    text = str(input_text)
    text = re.sub(r'\n\s*\n', '\n', text)
    text = re.sub(r'[ ]+', ' ', text)
    text = re.sub(r'\.{2,}', '.', text)
    text = re.sub(r',{2,}', ',', text)
    text = re.sub(r'-{2,}', '-', text)
    text = re.sub(r'_{2,}', '_', text)
    text = re.sub(r'!{2,}', '!', text)
    text = re.sub(r'\?{2,}', '?', text)
    text = re.sub(r'(\d)([A-Za-z])', r'\1 \2', text)
    text = re.sub(r'([A-Za-z])(\d)', r'\1 \2', text)
    text = unescape(text)
    text = re.sub(r'[^\w\s\[\]\(\)\$\\.\n\/:#<>{},_"!@\\-\\*=\\]', '', text)
    text = re.sub(r'\s+', ' ', text)
    return text

def top_k_top_p_filtering(logits, top_k, top_p, filter_value=-float('Inf')):
    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
        Args:
            logits: logits distribution shape (vocabulary size)
            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
    assert logits.dim() == 1  # batch size 1 for now - could be updated for more but the code would be less clear
    top_k = min(top_k, logits.size(-1))  # Safety check
    if top_k > 0:
        # Remove all tokens with a probability less than the last token of the top-k
        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
        logits[indices_to_remove] = filter_value

    if top_p > 0.0:
        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)

        # Remove tokens with cumulative probability above the threshold
        sorted_indices_to_remove = cumulative_probs > top_p
        # Shift the indices to the right to keep also the first token above the threshold
        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
        sorted_indices_to_remove[..., 0] = 0

        indices_to_remove = sorted_indices[sorted_indices_to_remove]
        logits[indices_to_remove] = filter_value
    return logits

def sample_seq(model, context, length, device, temperature, top_k, top_p):
    """ Generates a sequence of tokens 
        Args:
            model: gpt/gpt2 model
            context: tokenized text using gpt/gpt2 tokenizer
            length: length of generated sequence.
            device: torch.device object.
            temperature >0: used to control the randomness of predictions by scaling the logits before applying softmax.
            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
    """

    context = torch.tensor(context, dtype=torch.long, device=device)
    context = context.unsqueeze(0)
    generated = context
    with torch.no_grad():
        for _ in trange(length):
            inputs = {'input_ids': generated}
            outputs = model(
                **inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
            next_token_logits = outputs[0][0, -1, :] / temperature
            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
            next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
            generated = torch.cat((generated, next_token.unsqueeze(0)), dim=1)
    return generated
def add_special_tokens(lang):
    """ Returns GPT2 tokenizer after adding separator and padding tokens """
    token = 'gpt2'
    if lang =='vi':
        token = 'NlpHUST/gpt2-vietnamese'
    tokenizer = GPT2Tokenizer.from_pretrained(token)
    special_tokens = {'pad_token': '<|pad|>', 'sep_token': '<|sep|>'}
    tokenizer.add_special_tokens(special_tokens)
    return tokenizer

def gene(t,a):

    tokenizer = add_special_tokens(a)
    article = tokenizer.encode(t)[:900]
    # Load model directly
    model = AutoModelForCausalLM.from_pretrained("tiennlu/GPT2en_CNNen_3k")
    if a=="vi":
        model = AutoModelForCausalLM.from_pretrained("tiennlu/GPT2vi_CNNvi_3k")
    generated_text = sample_seq(model, article, 50, torch.device('cpu'), temperature=1, top_k=10, top_p=0.5)
    generated_text = generated_text[0, len(article):].tolist()
    text = tokenizer.convert_ids_to_tokens(generated_text, skip_special_tokens=True)
    text = tokenizer.convert_tokens_to_string(text)
    return text


def find_audio_files(path, extension=".mp3"):
    audio_files = []
    for root, dirs, files in os.walk(path):
        for f in files:
            if f.endswith(extension):
                audio_files.append(os.path.join(root, f))

    return audio_files


def youtube_to_mp3(youtube_url: str, output_dir: str) -> Union[AnyStr, str, bytes]:
    ydl_config = {
        "format": "bestaudio/best",
        "postprocessors": [
            {
                "key": "FFmpegExtractAudio",
                "preferredcodec": "mp3",
                "preferredquality": "192",
            }
        ],
        "outtmpl": os.path.join(output_dir, "%(title)s.%(ext)s"),
        "verbose": True,
    }

    if not os.path.exists(output_dir):
        os.makedirs(output_dir)

    with youtube_dl.YoutubeDL(ydl_config) as ydl:
        ydl.download([youtube_url])

    return find_audio_files(output_dir)[0]


def chunk_audio(filename, segment_length: int, output_dir):
    """segment lenght is in seconds"""

    # print(f"Chunking audio to {segment_length} second segments...")

    if not os.path.isdir(output_dir):
        os.mkdir(output_dir)
    # Load audio file
    audio = AudioSegment.from_mp3(filename)
    # Calculate duration in milliseconds
    duration = len(audio)

    # Calculate number of segments
    num_segments = duration // (segment_length * 1000) + 1

    print(f"Chunking {num_segments} chunks...")

    # Iterate through segments and save them
    for i in range(num_segments):
        start = i * segment_length * 1000
        end = min((i + 1) * segment_length * 1000, duration)
        segment = audio[start:end]
        segment.export(os.path.join(output_dir, f"segment_{i}.mp3"), format="mp3")

    chunked_audio_files = find_audio_files(output_dir)
    return sorted(chunked_audio_files)


def translate_text(text):
    wrapped_text = textwrap.wrap(text, 3500)
    tran_text = ""
    for line in wrapped_text:
        translation = GoogleTranslator(source='en', target='vi').translate(line)
        tran_text += translation + " "

    return tran_text


def transcribe_audio(audio_files: list, model_name="whisper-1"):
    transcripts = ""
    for audio_file in audio_files:
        audio = open(audio_file, "rb")
        try:
            response = completions_with_backoff(
                model=model_name, file=audio, response_format="text"
            )
            transcripts += response.text + " "
        except openai.OpenAIError as e:
            print(f"An error occurred: {e}")
            return None
    return transcripts


import random


# define a retry decorator
def retry_with_exponential_backoff(
        func,
        initial_delay: float = 1,
        exponential_base: float = 2,
        jitter: bool = True,
        max_retries: int = 10,
        errors: tuple = (openai.RateLimitError,),
):
    def wrapper(*args, **kwargs):
        num_retries = 0
        delay = initial_delay
        while True:
            try:
                return func(*args, **kwargs)
            except errors as e:
                print(f"Error: {e}")
                num_retries += 1
                if num_retries > max_retries:
                    raise Exception(f"Maximum number of retries ({max_retries}) exceeded.")
                delay *= exponential_base * (1 + jitter * random.random())
                time.sleep(delay)
            except Exception as e:
                raise e

    return wrapper


@retry_with_exponential_backoff
def completions_with_backoff(**kwargs):
    return client.audio.translations.create(**kwargs)


def get_video_id(youtube_url):
    """Extract video ID from YouTube URL."""
    parsed_url = urlparse(youtube_url)
    video_id = parse_qs(parsed_url.query).get("v")
    return video_id[0] if video_id else None

def get_transcript(video_id):
    tran = []
    transcript_list = YouTubeTranscriptApi.list_transcripts(video_id)
    transcript = transcript_list.find_generated_transcript(['vi', 'en'])
    translated_transcript = transcript.translate('en')   
    transcript_data = translated_transcript.fetch()
    
    words_to_remove = ['[music]', '[clause]', '[smile]', '[laugh]', '[cry]', '[sigh]', '[uh]', '[um]', '[uh-huh]', '[sob]', '[giggle]', '[hmm]']

    
    for t in transcript_data:
        if t['text'].lower() not in words_to_remove:
            tran.append(t['text'])
    
    return ' '.join(tran)



def chunk_text(text, chunk_size=1000, overlap_size=24):
    encoder = RecursiveCharacterTextSplitter().from_tiktoken_encoder(model_name="gpt-3.5-turbo", chunk_size=chunk_size,
                                                                     chunk_overlap=overlap_size)
    return encoder.split_text(text=text)


def summarize_youtube_video(youtube_url, outputs_dir):
    # Tạo đường dẫn đầy đủ cho thư mục đầu ra
    video_id = get_video_id(youtube_url)
    en_transcript = get_transcript(video_id)
    if not os.path.exists(outputs_dir):
        os.makedirs(outputs_dir)
    if not en_transcript:
        outputs_dir = f"{outputs_dir}\\{video_id}"
        raw_audio_dir = f"{outputs_dir}\\raw_audio\\"
        chunks_dir = f"{outputs_dir}\\chunks"
        segment_length = 10 * 60  # chunk to 10 minute segments
        if not os.path.exists(outputs_dir):
            os.makedirs(outputs_dir)
        audio_filename = youtube_to_mp3(youtube_url, output_dir=raw_audio_dir)
        chunked_audio_files = chunk_audio(
            audio_filename, segment_length=segment_length, output_dir=chunks_dir
        )
        en_transcript = transcribe_audio(chunked_audio_files)
    en_transcript = cleaning_input(en_transcript)
    vi_transcript = translate_text(en_transcript)
    summ_en = summary(en_transcript, 'en')
    summ_vi = summary(vi_transcript, 'vi')
    return tuple(summ_en), tuple(summ_vi)


def main():
    st.set_page_config(layout="wide")

    st.title("YouTube Video Summarizer 🎥")
    st.markdown('<style>h1{color: orange; text-align: center;}</style>', unsafe_allow_html=True)
    st.subheader('Built with the GPT2, Streamlit and ❤️')
    st.markdown('<style>h3{color: pink;  text-align: center;}</style>', unsafe_allow_html=True)

    # Expander for app details
    with st.expander("About the App"):
        st.write("This app allows you to summarize while watching a YouTube video.")
        st.write(
            "Enter a YouTube URL in the input box below and click 'Submit' to start. This app is built by AI Anytime.")

    # Input box for YouTube URL
    youtube_url = st.text_input("Enter YouTube URL")
    # Submit button
    if st.button("Submit") and youtube_url:
        start_time = time.time()  # Start the timer
        summ, tran = summarize_youtube_video(youtube_url, "./outputs")

        sum = summ[0]
        script = summ[1]
        sum_tran = tran[0]
        script_tran = tran[1]

        end_time = time.time()  # End the timer
        elapsed_time = end_time - start_time

        # Centering the video and elapsed time
        st.markdown("""
        <div style="display: flex; justify-content: center; flex-direction: column; align-items: center;">
            <div style="width: 60%; max-width: 720px;">
                <iframe width="100%" height="315" src="{youtube_url}" frameborder="0" allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
            </div>
            <h2>Summarization of YouTube Video</h2>
            <p>Time taken: {elapsed_time:.2f} seconds</p>
        </div>
    """.format(youtube_url=youtube_url.replace("watch?v=", "embed/"), elapsed_time=elapsed_time),
                    unsafe_allow_html=True)

        col1, col2 = st.columns(2)

        with col1:
            st.subheader("Transcript english")
            st.markdown(
                f'<div style="height: 400px; overflow-y: auto; border: 1px solid #ccc; padding: 10px;">{script}</div>',
                unsafe_allow_html=True)
            st.subheader("Summary english")
            st.write(sum)

        with col2:
            st.subheader("Transcript vietnamese")
            st.markdown(
                f'<div style="height: 400px; overflow-y: auto; border: 1px solid #ccc; padding: 10px;">{script_tran}</div>',
                unsafe_allow_html=True)
            st.subheader("Summary vietnamese")
            st.write(sum_tran)


from langchain.text_splitter import RecursiveCharacterTextSplitter


def chunk_overlap_text(text, chunk_size=1000, overlap_size=24):
    return RecursiveCharacterTextSplitter().from_tiktoken_encoder(model_name="gpt-3.5-turbo", chunk_size=chunk_size,
                                                                  chunk_overlap=overlap_size).split_text(text=text)


def summary(text, lang):
    chunks = chunk_overlap_text(text)
    rs = ""
    print(len(chunks[0]))
    print(f"Number of chunks: {len(chunks)}")

    for t in chunks:
        generated_summary = generate_response(t, lang)
        rs += generated_summary + " "
    text = ""
    for t in chunks:
        text += t + " "
    return rs, text


if __name__ == "__main__":
    main()