app.py

# Imports
import gradio as gr
import whisper
from pytube import YouTube
from transformers import pipeline, T5Tokenizer, T5ForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM
import torch
from wordcloud import WordCloud
import re
import os
import transformers
from langchain.llms import CTransformers
from langchain import PromptTemplate, LLMChain
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler


class GradioInference:
    def __init__(self):
        
        # OpenAI's Whisper model sizes
        self.sizes = list(whisper._MODELS.keys())

        # Whisper's available languages for ASR
        self.langs = ["none"] + sorted(list(whisper.tokenizer.LANGUAGES.values()))
        
        # Default size
        self.current_size = "base"
        
        # Default model size
        self.loaded_model = whisper.load_model(self.current_size)
        
        # Initialize Pytube Object
        self.yt = None

        # Initialize summary model for English
        self.summarizer = pipeline("summarization", model="facebook/bart-large-cnn")

        # Initialize VoiceLabT5 model and tokenizer
        self.keyword_model = T5ForConditionalGeneration.from_pretrained(
            "Voicelab/vlt5-base-keywords"
        )
        self.keyword_tokenizer = T5Tokenizer.from_pretrained(
            "Voicelab/vlt5-base-keywords"
        )

        # Sentiment Classifier
        self.classifier = pipeline("text-classification", model="lxyuan/distilbert-base-multilingual-cased-sentiments-student", return_all_scores=False)

        # Initialize Multilingual summary model 
        self.tokenizer = AutoTokenizer.from_pretrained("csebuetnlp/mT5_multilingual_XLSum")
        self.model = AutoModelForSeq2SeqLM.from_pretrained("csebuetnlp/mT5_multilingual_XLSum")


        self.llm = CTransformers(model="TheBloke/Llama-2-7B-Chat-GGML", model_file = 'llama-2-7b-chat.ggmlv3.q2_K.bin', callbacks=[StreamingStdOutCallbackHandler()])

    
    def __call__(self, link, lang, size, progress=gr.Progress()):
        """
        Call the Gradio Inference python class.
        This class gets access to a YouTube video using python's library Pytube and downloads its audio.
        Then it uses the Whisper model to perform Automatic Speech Recognition (i.e Speech-to-Text).
        Once the function has the transcription of the video it proccess it to obtain:
            - Summary: using Facebook's BART transformer.
            - KeyWords: using VoiceLabT5 keyword extractor.
            - Sentiment Analysis: using Hugging Face's default sentiment classifier
            - WordCloud: using the wordcloud python library.
        """
        progress(0, desc="Starting analysis")
        if self.yt is None:
            self.yt = YouTube(link)
        
        # Pytube library to access to YouTube audio stream
        path = self.yt.streams.filter(only_audio=True)[0].download(filename="tmp.mp4")

        if lang == "none":
            lang = None

        if size != self.current_size:
            self.loaded_model = whisper.load_model(size)
            self.current_size = size

        progress(0.20, desc="Transcribing")
        
        # Transcribe the audio extracted from pytube
        results = self.loaded_model.transcribe(path, language=lang)

        progress(0.40, desc="Summarizing")
        
        # Perform summarization on the transcription
        # transcription_summary = self.summarizer(
        #     results["text"], max_length=150, min_length=30, do_sample=False
        # )

        #### Prueba
        # WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
        
        # input_ids_sum = self.tokenizer(
        #     [WHITESPACE_HANDLER(results["text"])],
        #     return_tensors="pt",
        #     padding="max_length",
        #     truncation=True,
        #     max_length=512
        # )["input_ids"]
        
        # output_ids_sum = self.model.generate(
        #     input_ids=input_ids_sum,
        #     max_length=130,
        #     no_repeat_ngram_size=2,
        #     num_beams=4
        # )[0]
        
        # summary = self.tokenizer.decode(
        #     output_ids_sum,
        #     skip_special_tokens=True,
        #     clean_up_tokenization_spaces=False
        # )
        #### Fin prueba
        
        ### Prueba con LLM ###

        template = """
        [INST] <<SYS>>
        You are a helpful, respectful and honest assistant that performs summaries of text. Write a concise summary of the following text.
        <</SYS>>
        {text}[/INST]
        """
        
        prompt = PromptTemplate(template=template, input_variables=["text"])
        llm_chain = LLMChain(prompt=prompt, llm=self.llm)
        summary2 = llm_chain.run(results["text"])

        ### Fin prueba LLM ###
        
        progress(0.60, desc="Extracting Keywords")
        
        # Extract keywords using VoiceLabT5
        task_prefix = "Keywords: "
        input_sequence = task_prefix + results["text"]
        input_ids = self.keyword_tokenizer(
            input_sequence, return_tensors="pt", truncation=False
        ).input_ids
        output = self.keyword_model.generate(
            input_ids, no_repeat_ngram_size=3, num_beams=4
        )
        predicted = self.keyword_tokenizer.decode(output[0], skip_special_tokens=True)
        keywords = [x.strip() for x in predicted.split(",") if x.strip()]
        formatted_keywords = "\n".join([f"• {keyword}" for keyword in keywords])

        progress(0.80, desc="Extracting Sentiment")
       
        # Define a dictionary to map labels to emojis
        sentiment_emojis = {
            "positive": "Positive 👍🏼",
            "negative": "Negative 👎🏼",
            "neutral": "Neutral 😶",
        }
        
        # Sentiment label    
        label = self.classifier(summary)[0]["label"]

        # Format the label with emojis
        formatted_sentiment = sentiment_emojis.get(label, label)

        progress(0.90, desc="Generating Wordcloud")
        
        # Generate WordCloud object
        wordcloud = WordCloud(colormap = "Oranges").generate(results["text"])

        # WordCloud image to display
        wordcloud_image = wordcloud.to_image()

        if lang == "english":
            return (
                results["text"],
                summary2,
                # transcription_summary[0]["summary_text"],
                formatted_keywords,
                formatted_sentiment,
                wordcloud_image,
            )
        else:
            return (
                results["text"],
                summary2,
                formatted_keywords,
                formatted_sentiment,
                wordcloud_image,
            )


    def populate_metadata(self, link):
        """
        Access to the YouTube video title and thumbnail image to further display it
        params:
        - link: a YouTube URL.
        """
        if not link:
            return None, None
            
        self.yt = YouTube(link)
        return self.yt.thumbnail_url, self.yt.title

    def from_audio_input(self, lang, size, audio_file, progress=gr.Progress()):
        """
        Call the Gradio Inference python class.
        Uses it directly the Whisper model to perform Automatic Speech Recognition (i.e Speech-to-Text).
        Once the function has the transcription of the video it proccess it to obtain:
            - Summary: using Facebook's BART transformer.
            - KeyWords: using VoiceLabT5 keyword extractor.
            - Sentiment Analysis: using Hugging Face's default sentiment classifier
            - WordCloud: using the wordcloud python library.
        """
        progress(0, desc="Starting analysis")
        
        if lang == "none":
            lang = None

        if size != self.current_size:
            self.loaded_model = whisper.load_model(size)
            self.current_size = size

        progress(0.20, desc="Transcribing")
        
        results = self.loaded_model.transcribe(audio_file, language=lang)

        progress(0.40, desc="Summarizing")
        
        # Perform summarization on the transcription
        transcription_summary = self.summarizer(
            results["text"], max_length=150, min_length=30, do_sample=False
        )
        
        #### Prueba
        WHITESPACE_HANDLER = lambda k: re.sub('\s+', ' ', re.sub('\n+', ' ', k.strip()))
        
        input_ids_sum = self.tokenizer(
            [WHITESPACE_HANDLER(results["text"])],
            return_tensors="pt",
            padding="max_length",
            truncation=True,
            max_length=512
        )["input_ids"]
        
        output_ids_sum = self.model.generate(
            input_ids=input_ids_sum,
            max_length=130,
            no_repeat_ngram_size=2,
            num_beams=4
        )[0]
        
        summary = self.tokenizer.decode(
            output_ids_sum,
            skip_special_tokens=True,
            clean_up_tokenization_spaces=False
        )
        #### Fin prueba

        progress(0.50, desc="Extracting Keywords")
        
        # Extract keywords using VoiceLabT5
        task_prefix = "Keywords: "
        input_sequence = task_prefix + results["text"]
        input_ids = self.keyword_tokenizer(
            input_sequence, return_tensors="pt", truncation=False
        ).input_ids
        output = self.keyword_model.generate(
            input_ids, no_repeat_ngram_size=3, num_beams=4
        )
        predicted = self.keyword_tokenizer.decode(output[0], skip_special_tokens=True)
        keywords = [x.strip() for x in predicted.split(",") if x.strip()]
        formatted_keywords = "\n".join([f"• {keyword}" for keyword in keywords])

        progress(0.80, desc="Extracting Sentiment")

        # Define a dictionary to map labels to emojis
        sentiment_emojis = {
            "positive": "Positive 👍🏼",
            "negative": "Negative 👎🏼",
            "neutral": "Neutral 😶",
        }
        
        # Sentiment label    
        label = self.classifier(summary)[0]["label"]

        # Format the label with emojis
        formatted_sentiment = sentiment_emojis.get(label, label)
        
        progress(0.90, desc="Generating Wordcloud")
        # WordCloud object
        wordcloud = WordCloud(colormap = "Oranges").generate(
            results["text"]
        )
        wordcloud_image = wordcloud.to_image()

        if lang == "english":
            return (
                results["text"],
                # summ,
                transcription_summary[0]["summary_text"],
                formatted_keywords,
                formatted_sentiment,
                wordcloud_image,
            )
        else:
            return (
                results["text"],
                # summ,
                summary,
                formatted_keywords,
                formatted_sentiment,
                wordcloud_image,
            )


gio = GradioInference()
title = "YouTube Insights"
description = "Your AI-powered video analytics tool"

block = gr.Blocks()

with block as demo:
    gr.HTML(
        """
        <div style="text-align: center; max-width: 500px; margin: 0 auto;">
          <div>
            <h1>YouTube <span style="color: #FFA500;">Insights</span> 💡</h1>
          </div>
          <h4 style="margin-bottom: 10px; font-size: 95%">
            Your AI-powered video analytics tool ✨
          </h4>
        </div>
        """
    )
    with gr.Group():
        with gr.Tab("From YouTube 📹"):
            with gr.Box():
                
                with gr.Row().style(equal_height=True):
                    size = gr.Dropdown(
                        label="Speech-to-text Model Size", choices=gio.sizes, value="base"
                    )
                    lang = gr.Dropdown(
                        label="Language (Optional)", choices=gio.langs, value="none"
                    )
                link = gr.Textbox(
                    label="YouTube Link", placeholder="Enter YouTube link..."
                )
                title = gr.Label(label="Video Title")
                
                with gr.Row().style(equal_height=True):
                    img = gr.Image(label="Thumbnail")
                    text = gr.Textbox(
                        label="Transcription",
                        placeholder="Transcription Output...",
                        lines=10,
                    ).style(show_copy_button=True, container=True)
                    
                with gr.Row().style(equal_height=True):
                    summary = gr.Textbox(
                        label="Summary", placeholder="Summary Output...", lines=5
                    ).style(show_copy_button=True, container=True)
                    keywords = gr.Textbox(
                        label="Keywords", placeholder="Keywords Output...", lines=5
                    ).style(show_copy_button=True, container=True)
                    label = gr.Label(label="Sentiment Analysis")
                    wordcloud_image = gr.Image(label="WordCloud")
                    
                with gr.Row().style(equal_height=True):
                    clear = gr.ClearButton(
                        [link, title, img, text, summary, keywords, label, wordcloud_image], scale=1, value="Clear 🗑️"
                    )
                    btn = gr.Button("Get video insights 🔎", variant="primary", scale=1)
                btn.click(
                    gio,
                    inputs=[link, lang, size],
                    outputs=[text, summary, keywords, label, wordcloud_image],
                )
                link.change(gio.populate_metadata, inputs=[link], outputs=[img, title])

        with gr.Tab("From Audio file 🎙️"):
            with gr.Box():
                
                with gr.Row().style(equal_height=True):
                    size = gr.Dropdown(
                        label="Model Size", choices=gio.sizes, value="base"
                    )
                    lang = gr.Dropdown(
                        label="Language (Optional)", choices=gio.langs, value="none"
                    )
                audio_file = gr.Audio(type="filepath")
                
                with gr.Row().style(equal_height=True):
                    text = gr.Textbox(
                        label="Transcription",
                        placeholder="Transcription Output...",
                        lines=10,
                    ).style(show_copy_button=True, container=False)
                    
                with gr.Row().style(equal_height=True):
                    summary = gr.Textbox(
                        label="Summary", placeholder="Summary Output", lines=5
                    )
                    keywords = gr.Textbox(
                        label="Keywords", placeholder="Keywords Output", lines=5
                    )
                    label = gr.Label(label="Sentiment Analysis")
                    wordcloud_image = gr.Image(label="WordCloud")
                    
                with gr.Row().style(equal_height=True):
                    clear = gr.ClearButton([audio_file,text, summary, keywords, label, wordcloud_image], scale=1, value="Clear 🗑️")
                    btn = gr.Button(
                        "Get audio insights 🔎", variant="primary", scale=1
                    )
                btn.click(
                    gio.from_audio_input,
                    inputs=[lang, size, audio_file],
                    outputs=[text, summary, keywords, label, wordcloud_image],
                )


with block:
    gr.Markdown("### Video Examples")
    gr.Examples(["https://www.youtube.com/shorts/xDNzz8yAH7I","https://www.youtube.com/watch?v=kib6uXQsxBA&pp=ygURc3RldmUgam9icyBzcGVlY2g%3D"], inputs=link)

    gr.Markdown("### Audio Examples")
    gr.Examples(
        [[os.path.join(os.path.dirname(__file__),"audios/TED_lagrange_point.wav")],[os.path.join(os.path.dirname(__file__),"audios/TED_platon.wav")]], 
        inputs=audio_file)
    
    gr.Markdown("### About the app:")

    with gr.Accordion("What is YouTube Insights?", open=False):
        gr.Markdown(
            "YouTube Insights is a tool developed for academic purposes that allows you to analyze YouTube videos or audio files. It provides features like transcription, summarization, keyword extraction, sentiment analysis, and word cloud generation for multimedia content."
        )

    with gr.Accordion("How does YouTube Insights work?", open=False):
        gr.Markdown(
            "YouTube Insights leverages several powerful AI models and libraries. It uses OpenAI's Whisper for Automatic Speech Recognition (ASR) to transcribe audio content. It summarizes the transcribed text using Facebook's BART model, extracts keywords with VoiceLabT5, performs sentiment analysis with DistilBERT, and generates word clouds."
        )

    with gr.Accordion("What languages are supported for the analysis?", open=False):
        gr.Markdown(
            "YouTube Insights supports multiple languages for transcription and analysis. You can select your preferred language from the available options when using the app."
        )

    with gr.Accordion("Can I analyze audio files instead of YouTube videos?", open=False):
        gr.Markdown(
            "Yes, you can analyze audio files directly. Simply upload your audio file to the app, and it will provide the same transcription, summarization, keyword extraction, sentiment analysis, and word cloud generation features."
        )

    with gr.Accordion("What are the different model sizes available for transcription?", open=False):
        gr.Markdown(
            "The app uses a Speech-to-text model that has different training sizes, from tiny to large. Hence, the bigger the model the accurate the transcription."
        )

    with gr.Accordion("How long does it take to analyze a video or audio file?", open=False):
        gr.Markdown(
            "The time taken for analysis may vary based on the duration of the video or audio file and the selected model size. Shorter content will be processed more quickly."
        )
    
    with gr.Accordion("Who developed YouTube Insights?" ,open=False):
        gr.Markdown(
            "YouTube Insights was developed by students as part of the 2022/23 Master's in Big Data & Data Science program at Universidad Complutense de Madrid for academic purposes (Trabajo de Fin de Master)."
        )
    
    gr.HTML(
        """
        <div style="text-align: center; max-width: 500px; margin: 0 auto;">
          <p style="margin-bottom: 10px; font-size: 96%">
            Trabajo de Fin de Máster - Grupo 3
          </p>
          <p style="margin-bottom: 10px; font-size: 90%">
            2023 Master in Big Data & Data Science - Universidad Complutense de Madrid
          </p>
        </div>
        """
    )

demo.launch()