Update app.py

app.py

@@ -1,198 +1 @@
-##########################################
-# Step 0: Import required libraries
-##########################################
-import streamlit as st  # For building the web application interface
-from transformers import (
-    pipeline,
-    SpeechT5Processor,
-    SpeechT5ForTextToSpeech,
-    SpeechT5HifiGan,
-    AutoModelForCausalLM,
-    AutoTokenizer
-)  # For sentiment analysis, text-to-speech, and text generation
-from datasets import load_dataset  # For loading datasets (e.g., speaker embeddings)
-import torch  # For tensor operations
-import soundfile as sf  # For saving audio as .wav files
-import sentencepiece  # Required by SpeechT5Processor for tokenization
-
-##########################################
-# Streamlit application title and input
-##########################################
-# Display a colorful, large title in a visually appealing font
-st.markdown(
-    "<h1 style='text-align: center; color: #FF5720; font-size: 50px;'>Just Comment</h1>",
-    unsafe_allow_html=True
-)  # Use HTML and CSS for a custom title design
-
-# Display a smaller, gentle subtitle below the title
-st.markdown(
-    "<h3 style='text-align: center; color: #5D6D7E; font-style: italic;'>I'm listening to you, my friend</h3>",
-    unsafe_allow_html=True
-)  # Use HTML for a friendly and soft-styled subtitle
-
-# Add a well-designed text area for user input
-text = st.text_area(
-    "Enter your comment",
-    placeholder="Type something here...",
-    height=150,
-    help="Write a comment you would like us to analyze and respond to!"  # Provide a helpful tooltip
-)
-
-##########################################
-# Step 1: Sentiment Analysis Function
-##########################################
-def analyze_dominant_emotion(user_review):
-    """
-    Analyze the dominant emotion in the user's comment using a fine-tuned text classification model.
-    """
-    emotion_classifier = pipeline(
-        "text-classification",
-        model="Thea231/jhartmann_emotion_finetuning",
-        return_all_scores=True
-    )  # Load the fine-tuned text classification model from Hugging Face
-
-    emotion_results = emotion_classifier(user_review)[0]  # Perform sentiment analysis on the input text
-    dominant_emotion = max(emotion_results, key=lambda x: x['score'])  # Identify the emotion with the highest confidence
-    return dominant_emotion  # Return the dominant emotion (label and score)
-
-##########################################
-# Step 2: Response Generation Function
-##########################################
-
-
-def response_gen(user_review):
-    """
-    Generate a concise and logical response based on the sentiment of the user's comment.
-    """
-    dominant_emotion = analyze_dominant_emotion(user_review)  # Get the dominant emotion of the user's comment
-    emotion_label = dominant_emotion['label'].lower()  # Extract the emotion label in lowercase format
-    
-    # Define response templates for each emotion
-    emotion_prompts = {
-        "anger": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft a professional response that:\n"
-            "- Begins with sincere apology and acknowledgment\n"
-            "- Clearly explains solution process with concrete steps\n"
-            "- Offers appropriate compensation/redemption\n"
-            "- Keeps tone humble and solution-focused (3-4 sentences)\n\n"
-            "Response:"
-        ),
-        "disgust": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft a response that:\n"
-            "- Immediately acknowledges the product issue\n"
-            "- Explains quality control measures being taken\n"
-            "- Provides clear return/replacement instructions\n"
-            "- Offers goodwill gesture (3-4 sentences)\n\n"
-            "Response:"
-        ),
-        "fear": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft a reassuring response that:\n"
-            "- Directly addresses the safety worries\n"
-            "- References relevant certifications/standards\n"
-            "- Offers dedicated support contact\n"
-            "- Provides satisfaction guarantee (3-4 sentences)\n\n"
-            "Response:"
-        ),
-        "joy": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft a concise and enthusiastic response that:\n"
-            "- Thanks the customer for their feedback\n"
-            "- Acknowledges both positive and constructive comments\n"
-            "- Invites them to explore loyalty programs\n\n"
-            "Response:"
-        ),
-        "neutral": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft a balanced response that:\n"
-            "- Provides additional relevant product information\n"
-            "- Highlights key service features\n"
-            "- Politely requests more detailed feedback\n"
-            "- Maintains professional tone (3-4 sentences)\n\n"
-            "Response:"
-        ),
-        "sadness": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft an empathetic response that:\n"
-            "- Shows genuine understanding of the issue\n"
-            "- Proposes personalized recovery solution\n"
-            "- Offers extended support options\n"
-            "- Maintains positive outlook (3-4 sentences)\n\n"
-            "Response:"
-        ),
-        "surprise": (
-            f"'{user_review}'\n\n"
-            "As a customer service representative, craft a response that:\n"
-            "- Matches customer's positive energy appropriately\n"
-            "- Highlights unexpected product benefits\n"
-            "- Invites to user community/events\n"
-            "- Maintains brand voice (3-4 sentences)\n\n"
-            "Response:"
-        )
-    }
-    
-    prompt = emotion_prompts.get(
-        emotion_label,
-        f"Neutral feedback: '{user_review}'\n\nWrite a professional and concise response (50-200 words max).\n\nResponse:"
-    )  # Default to neutral if emotion is not found
-
-    # Load the tokenizer and language model for response generation
-    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B")  # Load tokenizer for processing text inputs
-    model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-0.5B")  # Load language model for response generation
-
-    inputs = tokenizer(prompt, return_tensors="pt")  # Tokenize the input prompt
-    outputs = model.generate(
-        **inputs,
-        max_new_tokens=300,
-        min_length=75,  # Ensure concise and complete responses
-        no_repeat_ngram_size=2,  # Avoid repetitive phrases
-        temperature=0.7  # Add randomness for more natural responses
-    )
-    response = tokenizer.decode(outputs[0], skip_special_tokens=True)  # Decode the generated response
-    return response  # Return the response
-
-##########################################
-# Step 3: Text-to-Speech Conversion Function
-##########################################
-def sound_gen(response):
-    """
-    Convert the generated response to speech and save it as a .wav file.
-    """
-    processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")  # Pre-trained processor for TTS
-    model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")  # Pre-trained TTS model
-    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")  # Vocoder for generating waveforms
-
-    # Create speaker embedding to match text input
-    embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")  # Load speaker embeddings
-    speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)  # Use a default embedding
-
-    inputs = processor(text=response, return_tensors="pt")  # Process text for spectrogram generation
-    inputs["input_ids"] = inputs["input_ids"].to(torch.int32)  # Match tensor format (fix runtime error)
-    spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings)  # Generate the spectrogram
-
-    with torch.no_grad():
-        speech = vocoder(spectrogram)  # Convert spectrogram to waveform
-
-    sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000)  # Save as .wav file
-    st.audio("customer_service_response.wav", start_time=0)  # Embed an auto-playing audio player
-
-##########################################
-# Main Function
-##########################################
-def main():
-    """
-    Main function to handle sentiment analysis, response generation, and text-to-speech functionalities.
-    """
-    if text:  # Check if the user has entered a comment
-        response = response_gen(text)  # Generate the response
-        st.markdown(
-            f"<p style='color:#3498DB; font-size:20px;'>{response}</p>",
-            unsafe_allow_html=True
-        )  # Display the response with styled formatting
-        sound_gen(response)  # Convert the response to speech and play it
-
-# Execute the main function
-if __name__ == "__main__":
-    main()
+########################################## # Step 0: Import required libraries ########################################## import streamlit as st # For building the web application from transformers import ( pipeline, SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan, AutoModelForCausalLM, AutoTokenizer ) # For emotion analysis, text-to-speech, and text generation from datasets import load_dataset # For loading datasets (e.g., speaker embeddings) import torch # For tensor operations import soundfile as sf # For saving audio as .wav files ########################################## # Streamlit application title and input ########################################## st.title("Comment Reply for You") # Application title st.write("Generate automatic replies for user comments") # Application description text = st.text_area("Enter your comment", "") # Text input for user to enter comments ########################################## # Step 1: Sentiment Analysis Function ########################################## def analyze_dominant_emotion(user_review): """ Analyze the dominant emotion in the user's review using a text classification model. """ emotion_classifier = pipeline( "text-classification", model="Thea231/jhartmann_emotion_finetuning", return_all_scores=True ) # Load pre-trained emotion classification model emotion_results = emotion_classifier(user_review)[0] # Get emotion scores for the review dominant_emotion = max(emotion_results, key=lambda x: x['score']) # Find the emotion with the highest confidence return dominant_emotion ########################################## # Step 2: Response Generation Function ########################################## def response_gen(user_review): """ Generate a response based on the sentiment of the user's review. """ # Use Llama-based model to create a response based on a generated prompt dominant_emotion = analyze_dominant_emotion(user_review) # Get the dominant emotion emotion_label = dominant_emotion['label'].lower() # Extract emotion label # Define response templates for each emotion emotion_prompts = { "anger": ( "Customer complaint: '{review}'\n\n" "As a customer service representative, write a response that:\n" "- Sincerely apologizes for the issue\n" "- Explains how the issue will be resolved\n" "- Offers compensation where appropriate\n\n" "Response:" ), "joy": ( "Customer review: '{review}'\n\n" "As a customer service representative, write a positive response that:\n" "- Thanks the customer for their feedback\n" "- Acknowledges both positive and constructive comments\n" "- Invites them to explore loyalty programs\n\n" "Response:" ), # Add other emotions as needed... } # Format the prompt with the user's review prompt = emotion_prompts.get(emotion_label, "Neutral").format(review=user_review) # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B") model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-0.5B") inputs = tokenizer(prompt, return_tensors="pt") # Tokenize the prompt outputs = model.generate(**inputs, max_new_tokens=100) # Generate a response input_length = inputs.input_ids.shape[1] # Length of the input text response = tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True) # Decode the generated text return response ########################################## # Step 3: Text-to-Speech Conversion Function ########################################## def sound_gen(response): """ Convert the generated response to speech and save as a .wav file. """ # Load the pre-trained TTS models processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts") vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan") # Load speaker embeddings (e.g., neutral female voice) embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation") speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0) # Process the input text and generate a spectrogram inputs = processor(text=response, return_tensors="pt") spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings) # Use the vocoder to generate a waveform with torch.no_grad(): speech = vocoder(spectrogram) # Save the generated speech as a .wav file sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000) st.audio("customer_service_response.wav") # Play the audio in Streamlit ########################################## # Main Function ########################################## def main(): """ Main function to orchestrate the workflow of sentiment analysis, response generation, and text-to-speech. """ if text: # Check if the user entered a comment response = response_gen(text) # Generate a response st.write(f"Generated response: {response}") # Display the generated response sound_gen(response) # Convert the response to speech and play it # Run the main function if __name__ == "__main__": main()