app.py

import os
import io
import gradio as gr
import torch
import numpy as np
import re
import pronouncing  # Add this to requirements.txt for syllable counting
import functools  # Add this for lru_cache functionality
from transformers import (
    AutoModelForAudioClassification,
    AutoFeatureExtractor,
    AutoTokenizer,
    pipeline,
    AutoModelForCausalLM,
    BitsAndBytesConfig
)
from huggingface_hub import login
from utils import (
    load_audio,
    extract_audio_duration,
    extract_mfcc_features,
    format_genre_results,
    ensure_cuda_availability
)
from emotionanalysis import MusicAnalyzer
import librosa

# Login to Hugging Face Hub if token is provided
if "HF_TOKEN" in os.environ:
    login(token=os.environ["HF_TOKEN"])

# Constants
GENRE_MODEL_NAME = "dima806/music_genres_classification"
MUSIC_DETECTION_MODEL = "MIT/ast-finetuned-audioset-10-10-0.4593"
LLM_MODEL_NAME = "Qwen/Qwen3-32B"
SAMPLE_RATE = 22050  # Standard sample rate for audio processing

# Check CUDA availability (for informational purposes)
CUDA_AVAILABLE = ensure_cuda_availability()

# Load models at initialization time
print("Loading genre classification model...")
try:
    genre_feature_extractor = AutoFeatureExtractor.from_pretrained(GENRE_MODEL_NAME)
    genre_model = AutoModelForAudioClassification.from_pretrained(
        GENRE_MODEL_NAME,
        device_map="auto" if CUDA_AVAILABLE else None
    )
    # Create a convenience wrapper function with the same interface as before
    def get_genre_model():
        return genre_model, genre_feature_extractor
except Exception as e:
    print(f"Error loading genre model: {str(e)}")
    genre_model = None
    genre_feature_extractor = None

# Load LLM and tokenizer at initialization time
print("Loading Qwen LLM model with 4-bit quantization...")
try:
    # Configure 4-bit quantization for better performance
    quantization_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.float16,
        bnb_4bit_use_double_quant=True
    )
    
    llm_tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL_NAME)
    llm_model = AutoModelForCausalLM.from_pretrained(
        LLM_MODEL_NAME,
        quantization_config=quantization_config,
        device_map="auto",
        trust_remote_code=True,
        torch_dtype=torch.float16,
        use_cache=True
    )
except Exception as e:
    print(f"Error loading LLM model: {str(e)}")
    llm_tokenizer = None
    llm_model = None

# Create music analyzer instance
music_analyzer = MusicAnalyzer()

# Process uploaded audio file
def process_audio(audio_file):
    if audio_file is None:
        return "No audio file provided", None, None, None, None, None, None
    
    try:
        # Load and analyze audio
        y, sr = load_audio(audio_file, sr=SAMPLE_RATE)
        
        # Basic audio information
        duration = extract_audio_duration(y, sr)
        
        # Analyze music with MusicAnalyzer
        music_analysis = music_analyzer.analyze_music(audio_file)
        
        # Extract key information
        tempo = music_analysis["rhythm_analysis"]["tempo"]
        time_signature = music_analysis["rhythm_analysis"]["estimated_time_signature"]
        emotion = music_analysis["emotion_analysis"]["primary_emotion"]
        theme = music_analysis["theme_analysis"]["primary_theme"]
        
        # Use genre classification directly instead of pipeline
        if genre_model is not None and genre_feature_extractor is not None:
            # Resample audio to 16000 Hz for the genre model
            y_16k = librosa.resample(y, orig_sr=sr, target_sr=16000)
            
            # Extract features
            inputs = genre_feature_extractor(
                y_16k, 
                sampling_rate=16000, 
                return_tensors="pt"
            ).to(genre_model.device)
            
            # Classify genre
            with torch.no_grad():
                outputs = genre_model(**inputs)
                logits = outputs.logits
                probs = torch.nn.functional.softmax(logits, dim=-1)
                
            # Get top genres
            values, indices = torch.topk(probs[0], k=5)
            top_genres = [(genre_model.config.id2label[idx.item()], val.item()) for val, idx in zip(values, indices)]
        else:
            # Fallback if model loading failed
            top_genres = [("Unknown", 1.0)]
        
        # Format genre results for display
        genre_results_text = format_genre_results(top_genres)
        primary_genre = top_genres[0][0]
        
        # Generate lyrics using LLM
        lyrics = generate_lyrics(music_analysis, primary_genre, duration)
        
        # Prepare analysis summary
        analysis_summary = f"""
### Music Analysis Results

**Duration:** {duration:.2f} seconds
**Tempo:** {tempo:.1f} BPM
**Time Signature:** {time_signature}
**Key:** {music_analysis["tonal_analysis"]["key"]} {music_analysis["tonal_analysis"]["mode"]}
**Primary Emotion:** {emotion}
**Primary Theme:** {theme}
**Top Genre:** {primary_genre}

{genre_results_text}
        """
        
        return analysis_summary, lyrics, tempo, time_signature, emotion, theme, primary_genre
    
    except Exception as e:
        error_msg = f"Error processing audio: {str(e)}"
        print(error_msg)
        return error_msg, None, None, None, None, None, None

def generate_lyrics(music_analysis, genre, duration):
    try:
        # Extract meaningful information for context
        tempo = music_analysis["rhythm_analysis"]["tempo"]
        key = music_analysis["tonal_analysis"]["key"]
        mode = music_analysis["tonal_analysis"]["mode"]
        emotion = music_analysis["emotion_analysis"]["primary_emotion"]
        theme = music_analysis["theme_analysis"]["primary_theme"]
        
        # Verify LLM is loaded
        if llm_model is None or llm_tokenizer is None:
            return "Error: LLM model not properly loaded"
        
        # Construct prompt for the LLM
        prompt = f"""Write lyrics for a {genre} song with these specifications:
- Key: {key} {mode}
- Tempo: {tempo} BPM
- Emotion: {emotion}
- Theme: {theme}
- Duration: {duration:.1f} seconds
- Time signature: {music_analysis["rhythm_analysis"]["estimated_time_signature"]}

IMPORTANT INSTRUCTIONS:
- The lyrics should be in English
- Write ONLY the raw lyrics with no structural labels
- DO NOT include [verse], [chorus], [bridge], or any other section markers
- DO NOT include any explanations or thinking about the lyrics
- DO NOT number the verses or lines
- DO NOT use bullet points
- Format as simple line-by-line lyrics only
- Make sure the lyrics match the specified duration and tempo
- Keep lyrics concise enough to fit the duration when sung at the given tempo
"""

        # Generate lyrics using the LLM model directly
        # Format as chat message
        messages = [
            {"role": "user", "content": prompt}
        ]
        
        # Apply chat template
        text = llm_tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True
        )
        
        # Tokenize and move to model device
        model_inputs = llm_tokenizer([text], return_tensors="pt").to(llm_model.device)
        
        # Generate with optimized parameters
        generated_ids = llm_model.generate(
            **model_inputs,
            max_new_tokens=1024,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            repetition_penalty=1.1,
            pad_token_id=llm_tokenizer.eos_token_id
        )
        
        # Decode the output
        output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
        lyrics = llm_tokenizer.decode(output_ids, skip_special_tokens=True).strip()
        
        # Enhanced post-processing to remove ALL structural elements and thinking
        # Remove any lines with section labels using a more comprehensive pattern
        lyrics = re.sub(r'^\[.*?\].*$', '', lyrics, flags=re.MULTILINE)
        
        # Remove common prefixes and thinking text
        lyrics = re.sub(r'^(Here are|Here is|These are|This is|Let me|I will|I\'ll).*?:\s*', '', lyrics, flags=re.IGNORECASE)
        lyrics = re.sub(r'^Title:.*?$', '', lyrics, flags=re.MULTILINE).strip()
        
        # Remove all section markers in any format
        lyrics = re.sub(r'^\s*(Verse|Chorus|Bridge|Pre.?Chorus|Intro|Outro|Refrain|Hook|Breakdown)(\s*\d*|\s*[A-Z])?:?\s*$', '', lyrics, flags=re.MULTILINE|re.IGNORECASE)
        lyrics = re.sub(r'\[(Verse|Chorus|Bridge|Pre.?Chorus|Intro|Outro|Refrain|Hook|Breakdown)(\s*\d*|\s*[A-Z])?\]', '', lyrics, flags=re.IGNORECASE)
        
        # Remove any "thinking" or explanatory parts that might be at the beginning
        lyrics = re.sub(r'^.*?(Let\'s|Here\'s|I need|I want|I\'ll|First|The|This).*?:\s*', '', lyrics, flags=re.IGNORECASE)
        
        # Remove any empty lines at beginning, collapse multiple blank lines, and trim
        lyrics = re.sub(r'^\s*\n', '', lyrics)
        lyrics = re.sub(r'\n\s*\n\s*\n+', '\n\n', lyrics)
        lyrics = lyrics.strip()
        
        return lyrics
    
    except Exception as e:
        error_msg = f"Error generating lyrics: {str(e)}"
        print(error_msg)
        return error_msg

# Create Gradio interface
def create_interface():
    with gr.Blocks(title="Music Analysis & Lyrics Generator") as demo:
        gr.Markdown("# Music Analysis & Lyrics Generator")
        gr.Markdown("Upload a music file or record audio to analyze it and generate matching lyrics")
        
        with gr.Row():
            with gr.Column(scale=1):
                audio_input = gr.Audio(
                    label="Upload or Record Audio", 
                    type="filepath",
                    sources=["upload", "microphone"]
                )
                analyze_btn = gr.Button("Analyze and Generate Lyrics", variant="primary")
            
            with gr.Column(scale=2):
                with gr.Tab("Analysis"):
                    analysis_output = gr.Textbox(label="Music Analysis Results", lines=10)
                    
                    with gr.Row():
                        tempo_output = gr.Number(label="Tempo (BPM)")
                        time_sig_output = gr.Textbox(label="Time Signature")
                        emotion_output = gr.Textbox(label="Primary Emotion")
                        theme_output = gr.Textbox(label="Primary Theme")
                        genre_output = gr.Textbox(label="Primary Genre")
                
                with gr.Tab("Generated Lyrics"):
                    lyrics_output = gr.Textbox(label="Generated Lyrics", lines=20)
        
        # Set up event handlers
        analyze_btn.click(
            fn=process_audio,
            inputs=[audio_input],
            outputs=[analysis_output, lyrics_output, tempo_output, time_sig_output, 
                    emotion_output, theme_output, genre_output]
        )
        
        gr.Markdown("""
        ## How it works
        1. Upload or record a music file
        2. The system analyzes tempo, beats, time signature and other musical features
        3. It detects emotion, theme, and music genre
        4. Using this information, it generates lyrics that match the style and length of your music
        """)
    
    return demo

# Launch the app
demo = create_interface()

if __name__ == "__main__":
    demo.launch()
else:
    # For Hugging Face Spaces
    app = demo