src/classifier.py

import torch
import torchaudio
import librosa
import numpy as np
from transformers import pipeline
from typing import Union, Tuple, List

class MusicGenreClassifier:
    def __init__(self):
        try:
            # Initialize both audio and text classification pipelines with auto device mapping
            self.text_classifier = pipeline(
                "zero-shot-classification",
                model="facebook/bart-large-mnli",
                device_map="auto"
            )
            
            # For audio classification, we'll use MIT's music classification model
            self.audio_classifier = pipeline(
                "audio-classification",
                model="mit/ast-finetuned-audioset-10-10-0.4593",
                device_map="auto"
            )
        except Exception as e:
            print(f"Warning: GPU initialization failed, falling back to CPU. Error: {str(e)}")
            # Fall back to CPU if GPU initialization fails
            self.text_classifier = pipeline(
                "zero-shot-classification",
                model="facebook/bart-large-mnli",
                device="cpu"
            )
            
            self.audio_classifier = pipeline(
                "audio-classification",
                model="mit/ast-finetuned-audioset-10-10-0.4593",
                device="cpu"
            )
        
        # Define standard genres for classification
        self.genres = [
            "rock", "pop", "hip hop", "country", "jazz",
            "classical", "electronic", "blues", "reggae", "metal"
        ]
        
        # Mapping from model output labels to our standard genres
        self.label_mapping = {
            "Music": "pop",  # Default mapping
            "Rock music": "rock",
            "Pop music": "pop",
            "Hip hop music": "hip hop",
            "Country": "country",
            "Jazz": "jazz",
            "Classical music": "classical",
            "Electronic music": "electronic",
            "Blues": "blues",
            "Reggae": "reggae",
            "Heavy metal": "metal"
        }
    
    def process_audio(self, audio_path: str) -> torch.Tensor:
        """Process audio file to match model requirements."""
        try:
            # Load audio using librosa (handles more formats)
            waveform, sample_rate = librosa.load(audio_path, sr=16000)
            # Convert to torch tensor and ensure proper shape
            waveform = torch.from_numpy(waveform).float()
            if len(waveform.shape) == 1:
                waveform = waveform.unsqueeze(0)
            return waveform
        except Exception as e:
            raise ValueError(f"Error processing audio file: {str(e)}")

    def map_label_to_genre(self, label: str) -> str:
        """Map model output label to standard genre."""
        return self.label_mapping.get(label, "pop")  # Default to pop if unknown

    def classify_audio(self, audio_path: str) -> Tuple[str, float]:
        """Classify genre from audio file."""
        try:
            waveform = self.process_audio(audio_path)
            predictions = self.audio_classifier(waveform, top_k=3)
            
            # Process predictions
            if isinstance(predictions, list):
                predictions = predictions[0]
            
            # Find the highest scoring music-related prediction
            music_preds = [
                (self.map_label_to_genre(p['label']), p['score'])
                for p in predictions
                if p['label'] in self.label_mapping
            ]
            
            if not music_preds:
                # If no music genres found, return default
                return "pop", 0.5
            
            # Get the highest scoring genre
            genre, score = max(music_preds, key=lambda x: x[1])
            return genre, score
            
        except Exception as e:
            raise ValueError(f"Audio classification failed: {str(e)}")

    def classify_text(self, lyrics: str) -> Tuple[str, float]:
        """Classify genre from lyrics text."""
        try:
            # Prepare the hypothesis template for zero-shot classification
            hypothesis_template = "This text contains {} music lyrics."
            
            result = self.text_classifier(
                lyrics,
                candidate_labels=self.genres,
                hypothesis_template=hypothesis_template
            )
            
            return result['labels'][0], result['scores'][0]
        except Exception as e:
            raise ValueError(f"Text classification failed: {str(e)}")

    def predict(self, input_data: str, input_type: str = None) -> dict:
        """
        Main prediction method that handles both audio and text inputs.
        
        Args:
            input_data: Path to audio file or lyrics text
            input_type: Optional, 'audio' or 'text'. If None, will try to auto-detect
            
        Returns:
            dict containing predicted genre and confidence score
        """
        # Try to auto-detect input type if not specified
        if input_type is None:
            input_type = 'audio' if input_data.lower().endswith(('.mp3', '.wav', '.ogg', '.flac')) else 'text'
        
        try:
            if input_type == 'audio':
                genre, confidence = self.classify_audio(input_data)
            else:
                genre, confidence = self.classify_text(input_data)
                
            return {
                'genre': genre,
                'confidence': float(confidence),
                'input_type': input_type
            }
        except Exception as e:
            raise ValueError(f"Prediction failed: {str(e)}")