app.py

import yt_dlp
import os
import streamlit as st
import transformers
from transformers import pipeline
from transformers import AutoTokenizer
import nltk
from PIL import Image
import torch

icon = Image.open("Traçado laranja #f1863d.png")

st.set_page_config(
    page_title = "Turing Videos",
    page_icon = icon,
    layout = "wide",
    initial_sidebar_state = "auto",  
)

#Download youtube video
@st.cache_data
def download_audio(link):
    with yt_dlp.YoutubeDL({'extract_audio': True, 'format': 'bestaudio', 'outtmpl': 'video.mp3'}) as video:
        video.download(link)

#Load Whisper pipeline via HuggingFace
@st.cache_resource
def load_whisper(seconds):
    return pipeline("automatic-speech-recognition",
                    model="openai/whisper-tiny",
                    chunk_length_s=seconds,
                   )

#Load Extractive Summarizer pipeline via HuggingFace
@st.cache_resource
def load_extractive():
    return pipeline("summarization",
               model = "NotXia/longformer-bio-ext-summ",
               tokenizer = AutoTokenizer.from_pretrained("NotXia/longformer-bio-ext-summ"),
               trust_remote_code = True,
               )

#Load QA pipeline via HuggingFace
@st.cache_resource
def load_qa():
    return pipeline("question-answering",
                    model='distilbert-base-uncased-distilled-squad'
                   )

#Download punkt function from nltk
@st.cache_data
def load_nltk():
    nltk.download("punkt")

#Make the ASR task
@st.cache_data
def audio_speech_recognition(model_pipeline, video="video.mp3"):
    return model_pipeline(video, batch_size=8)["text"].strip()

#Make the Summarization task
@st.cache_data
def text_summarization(model_pipeline, full_text, ratio):
    sentences = nltk.sent_tokenize(full_text)
    extractive_sentences = model_pipeline({"sentences": sentences}, strategy="ratio", strategy_args=ratio)
    extractive_text = " ".join(extractive_sentences[0])
    return extractive_text.strip()

#Make the QA task
@st.cache_data
def answer_questions(model_pipeline, full_text, questionings):
    answers = []
    for question in questionings:
        result = model_pipeline(question=question, context=full_text)
        answers.append(result["answer"])
    return answers
    
def main():

    header = st.container()
    model = st.container()
    model_1, model_2 = st.columns(2)
    
    with st.sidebar:
        
        st.title(":red[Turing]Videos")

        with st.form("data_collection"):

            language = st.selectbox('Qual a linguagem do seu modelo?',
                                    ('Português (pt)', 'Inglês (en)', 'Outra')
                                   )

            link = st.text_area(label="Coloque o link do seu vídeo do YouTube:",
                                height=25, placeholder="Digite seu link...")

            compression_rate = st.slider(label="Selecione a taxa de compressão:",
                                         min_value=0.1, max_value=0.9,
                                         value=0.25, step=0.05
                                        )

            questions = st.text_area(label="Coloque suas perguntas separadas por vírgula!",
                                     height=50, placeholder="Digite suas perguntas..."
                                    ).split(",")

            seconds = st.select_slider(label="Digite a duração do seu vídeo para otimização:",
                                       options = ["5 min", "15 min", "30 min", "45 min", "60 min"],
                                       value = "15 min", 
                                      )

            seconds = int(seconds.replace(" min", "")) * 60

            submitted = st.form_submit_button("Submit")
            
            if submitted:
                st.success('Dados coletados!', icon="✅")
            else:
                st.error('Dados ainda não coletados!', icon="🚨")
    
    with header:
        st.title(":red[Turing]Videos")
        st.subheader("Este projeto inovador utiliza técnicas avançadas de inteligência artificial para simplificar e acelerar a compreensão de conteúdo audiovisual.",
                     divider = "red"
                    )

    with model:
        if submitted:
            with st.spinner("Carregando modelos..."):

                if language == "Inglês (en)":
                    download_audio(link)
                    load_nltk()
                    whisper = load_whisper(seconds)
                    extractive = load_extractive()
                    qa_model = load_qa()

                elif language == "Português (pt)":
                    st.header("Modelo ainda não implementado.")

                else:
                    st.header("Erro na seleção de linguagem.")

            with st.spinner("Transcrevendo texto..."):
                transcript_text = audio_speech_recognition(whisper)

            with model_1:
                st.header("Texto Sumarizado:")
                with st.spinner("Carregando sumarização..."):
                    summary = text_summarization(extractive, transcript_text, compression_rate)
                st.subheader(summary)

            with model_2:
                st.header("Resposta das perguntas:")
                with st.spinner("Carregando respostas..."):
                    answers = answer_questions(qa_model, transcript_text, questions)
                
                for i in range(len(answers)):
                    st.subheader(questions[i])
                    st.subheader(answers[i])
                    st.write("\n")

main()