app.py

import streamlit as st
import numpy as np
import pandas as pd
import tensorflow as tf
from transformers import RobertaTokenizer, TFRobertaForSequenceClassification
from sentence_transformers import SentenceTransformer, util
import spotipy
from spotipy.oauth2 import SpotifyClientCredentials
import os 

sim_model = SentenceTransformer("sentence-transformers/all-mpnet-base-v2")

# Function to analyze sentiment of the user's input
def analyze_user_input(user_input, tokenizer, model):
    encoded_input = tokenizer(user_input, return_tensors="tf", truncation=True, padding=True, max_length=512)
    outputs = model(encoded_input)
    scores = tf.nn.softmax(outputs.logits, axis=-1).numpy()[0]
    predicted_class_idx = tf.argmax(outputs.logits, axis=-1).numpy()[0]
    sentiment_label = model.config.id2label[predicted_class_idx]
    sentiment_score = scores[predicted_class_idx]
    return sentiment_label, sentiment_score

# Function to match songs from the dataset with the user's sentiment
def match_songs_with_sentiment(user_sentiment_label, user_sentiment_score,inputVector, score_range,songs_df):

    # Filter songs with the same sentiment label
    matched_songs = songs_df[songs_df['sentiment'] == user_sentiment_label]

    # Calculate the score range
    score_min = max(0, user_sentiment_score - score_range)
    score_max = min(1, user_sentiment_score + score_range)

    # Further filter songs whose scores fall within the specified range
    matched_songs = matched_songs[(matched_songs['score'] >= score_min) & (matched_songs['score'] <= score_max)]

    # Shuffle the matched songs to get a random order
    matched_songs = matched_songs.sample(frac=1).reset_index(drop=True)

    matched_songs['similarity'] = matched_songs['seq'].apply(lambda x: util.pytorch_cos_sim(sim_model.encode(x), inputVector))

    top_5 = matched_songs['similarity'].sort_values(ascending=False).head(5)

    # Sort the songs by how close their score is to the user's sentiment score
    # matched_songs['score_diff'] = abs(matched_songs['score'] - user_sentiment_score)
    # matched_songs = matched_songs.sort_values(by='score_diff')

    # Select the top five songs and return
    return matched_songs.loc[top_5.index, ['song','artist','seq','similarity','sentiment','score']]


client_id = os.getenv('client_id')
client_secret = os.getenv('client_secret')

client_credentials_manager = SpotifyClientCredentials(client_id=client_id, client_secret=client_secret)
sp = spotipy.Spotify(client_credentials_manager=client_credentials_manager)

def get_track_id(song_name):
    # Search for the track ID using the song name
    results = sp.search(q=song_name, type='track', limit=1)
    if results['tracks']['items']:
        track_id = results['tracks']['items'][0]['id']
        return track_id
    else:
        print(f"No results found for {song_name}")
        return None

def get_track_preview_url(track_id):
    # Get the 30-second preview URL for the track
    track_info = sp.track(track_id)
    preview_url = track_info['preview_url']
    return preview_url

# Initialize the tokenizer and model outside of the functions to speed up repeated calls
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
model = TFRobertaForSequenceClassification.from_pretrained('arpanghoshal/EmoRoBERTa')

# Streamlit app layout
st.set_page_config(page_title="MODUS MUSIC", layout="wide")  # New: Setting page title and layout

# Custom CSS for background and text color
st.markdown("""
    <style>
    .stApp {
        background: rgb(0,0,0);
        background-size: cover;
        color: white;  /* Sets global text color to white */
    }
    /* General rule for all labels */
    label {
        color: white !important;
    }
    /* Specific color for the main title */
    h1 {
        color: red !important;  /* Making the MODUS MUSIC title red */
    }
    /* Additional specific styling */
    .stTextInput > label, .stButton > button, .css-10trblm, .css-1yjuwjr, .intro {
        color: white !important;
    }
    </style>
    """, unsafe_allow_html=True)


image_path = './MODUSMUSIC.png'  # Replace with the actual path to your image

st.image(image_path, use_column_width=False, width=250)  # Adjust the width as needed
# Custom gradient background using CSS
st.markdown("""
    <style>
    .stApp {
        background: rgb(0,0,0);
        background-size: cover;
    }
    </style>
    """, unsafe_allow_html=True)

# Custom HTML for the main title
st.markdown("<h1 style='text-align: center; font-weight: bold;'>MODUS MUSIC</h1>", unsafe_allow_html=True)

st.title('Music Suggestion Based on Your Feeling')  # Existing Title

# New: Introduction Section
with st.container():
    st.markdown("""
        <style>
        .intro {
            font-size:18px;
        }
        </style>
        <div class='intro'>
        Welcome to Modus Music! Share your vibe, and let's find the perfect songs to match your mood.
        Just type in your thoughts, and we'll do the rest.
        </div>
        """, unsafe_allow_html=True)

# User input text area
with st.container():
    user_input = st.text_area("What's your vibe? Tell me about it:", key="123", height=150, max_chars=500)
m = st.markdown("""
<style>
div.stButton > button:first-child {
    background-color: rgb(204, 49, 49);

}
</style>""", unsafe_allow_html=True)
# Use the custom style for the button
submit_button = st.button("Generate music")


# Processing and Displaying Results
if submit_button and len(user_input.split()) > 5:
    # New: Define inputVector here
    inputVector = sim_model.encode(user_input)

    # Run sentiment analysis on the user input
    sentiment_label, sentiment_score = analyze_user_input(user_input, tokenizer, model)
    st.write(f"Sentiment: {sentiment_label}, Score: {sentiment_score:.2f}")

    # Load songs dataframe
    songs_df = pd.read_csv('./music_mental_health.csv')

    # Suggest songs
    suggested_songs = match_songs_with_sentiment(sentiment_label, sentiment_score, inputVector, 0.0003, songs_df)
    suggested_songs['similarity'] = suggested_songs['similarity'].apply(lambda x: x.numpy()[0][0])

    # Styling for the suggested songs display
    with st.container():
        st.markdown("<div class='song-list'>", unsafe_allow_html=True)
        st.write("Based on your vibe, you might like these songs:")
        for index, row in suggested_songs.iterrows():
          song = row['song']
          artist = row['artist']
          track_id = get_track_id(song)
          if track_id.strip():
            preview_url = get_track_preview_url(track_id)
            #st.write(f"Similarity: {row['similarity']}")
            st.write(f"{song} by {artist}")
            with st.expander(f"Show Lyrics for {song} by {artist}", expanded=False):
                st.write(f"Lyrics: {row['seq']}")

            if preview_url:
              st.audio(preview_url)
            else:
              st.write("No Preview Available")
        st.markdown("</div>", unsafe_allow_html=True)
        st.dataframe(suggested_songs[['song','artist','seq','similarity','sentiment','score']])
elif submit_button and not len(user_input.split()) > 5:
    st.warning("Please provide a longer response with 5 words or more.")
    st.rerun()