src/app.py

import streamlit as st
import pandas as pd
import numpy as np
from annoy import AnnoyIndex
from sentence_transformers import SentenceTransformer
from sklearn.preprocessing import StandardScaler
import pickle
import os
from pathlib import Path

CACHE_DIR = "/app/cache"
Path(CACHE_DIR).mkdir(parents=True, exist_ok=True)

os.environ['TRANSFORMERS_CACHE'] = CACHE_DIR
os.environ['HF_HOME'] = CACHE_DIR
os.environ['XDG_CACHE_HOME'] = CACHE_DIR

TEXT_DIM = 384
AUDIO_DIM = 6

@st.cache_resource
def load_data_and_models():
    data_path = "src/dataset.csv"  
    model_dir = "src/"
    
    if not os.path.exists(data_path):
        raise FileNotFoundError(f"Dataset not found at {data_path}")
    
    data = pd.read_csv(data_path)
    
    text_model = SentenceTransformer('all-MiniLM-L6-v2', cache_folder=CACHE_DIR)
    scaler = pickle.load(open(f"{model_dir}scaler.pkl", "rb"))
    annoy_index = AnnoyIndex(TEXT_DIM + AUDIO_DIM, 'angular')
    annoy_index.load(f"{model_dir}annoy_index.ann")
    
    return data, text_model, scaler, annoy_index

def recommend(track_name, data, model, scaler, annoy_index, n=5):
    try:
        track = data[data['track_name'].str.lower() == track_name.lower()].iloc[0]
        
        text_emb = model.encode(
            f"{track['artists']} - {track['track_name']} - {track['album_name']}"
        )
        
        audio_features = track[[
            'danceability', 'energy', 'valence', 'tempo',
            'loudness', 'acousticness'
        ]].values.reshape(1, -1)
        audio_scaled = scaler.transform(audio_features)
        
        combined = np.hstack([text_emb, audio_scaled[0]])
        
        similar_indices = annoy_index.get_nns_by_vector(combined, n+1)[1:]
        similar_tracks = data.iloc[similar_indices]
        
        return similar_tracks[['artists', 'track_name', 'album_name']]
    
    except Exception as e:
        st.error(f"Ошибка: {str(e)}")
        return pd.DataFrame()

def main():
    st.set_page_config(page_title="🎧 Music Recommender", layout="wide")
    
    st.markdown(
        "<h1 style='text-align: center; color: #4A4A4A;'>🎵 Музыкальные рекомендации</h1>",
        unsafe_allow_html=True
    )
    st.markdown("<p style='text-align: center;'>Найдите треки, похожие на ваш любимый — по звучанию и настроению.</p>", unsafe_allow_html=True)

    try:
        data, text_model, scaler, annoy_index = load_data_and_models()
    except Exception as e:
        st.error(f"Ошибка загрузки данных: {str(e)}")
        return

    data['display_name'] = data['track_name'] + " — " + data['artists']

    with st.container():
        st.markdown("### 🔍 Выбор трека")
        track_display = st.selectbox(
            "Выберите трек из списка:",
            options=data['display_name'].unique(),
            index=0
        )

        selected_track_name = data[data['display_name'] == track_display]['track_name'].values[0]

        col1, col2 = st.columns([1, 3])
        with col1:
            find_button = st.button("🔎 Найти похожие", type="primary")

        if find_button:
            with st.spinner("🔬 Анализируем звучание и стиль..."):
                recommendations = recommend(
                    selected_track_name, data, text_model, scaler, annoy_index
                )

            if not recommendations.empty:
                st.markdown("### 🎧 Вам могут понравиться:")
                st.dataframe(
                    recommendations.rename(columns={
                        "artists": "👤 Исполнитель",
                        "track_name": "🎵 Трек",
                        "album_name": "💿 Альбом"
                    }),
                    use_container_width=True,
                    hide_index=True
                )
            else:
                st.warning("К сожалению, похожих треков не найдено. Попробуйте другой.")

if __name__ == "__main__":
    main()