backend/src/model.py

"""Lyrical emotion classification model."""

import torch
from transformers import AutoModelForSequenceClassification, AutoConfig
from pathlib import Path

from .config import config


class LyricEmotionClassifier:
    """RoBERTa-based emotion classifier for song lyrics."""

    def __init__(self, model_name: str = None, num_labels: int = None):
        self.model_name = model_name or config.model_name
        self.num_labels = num_labels or config.num_labels
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model = None

    def load_pretrained(self) -> "LyricEmotionClassifier":
        """Load pretrained model for fine-tuning."""
        model_config = AutoConfig.from_pretrained(
            self.model_name,
            num_labels=self.num_labels,
            id2label={i: label for i, label in enumerate(config.label_names)},
            label2id={label: i for i, label in enumerate(config.label_names)},
        )

        self.model = AutoModelForSequenceClassification.from_pretrained(
            self.model_name,
            config=model_config
        )
        self.model.to(self.device)
        return self

    def load_finetuned(self, path: Path) -> "LyricEmotionClassifier":
        """Load a fine-tuned model from disk."""
        self.model = AutoModelForSequenceClassification.from_pretrained(path)
        self.model.to(self.device)
        return self

    def save(self, path: Path):
        """Save model to disk."""
        path.mkdir(parents=True, exist_ok=True)
        self.model.save_pretrained(path)

    def predict(self, inputs: dict) -> dict:
        """Run inference on tokenized inputs."""
        self.model.eval()

        # Move inputs to device
        inputs = {k: v.to(self.device) for k, v in inputs.items()}

        with torch.no_grad():
            outputs = self.model(**inputs)

        probs = torch.softmax(outputs.logits, dim=-1)
        predictions = torch.argmax(probs, dim=-1)

        return {
            "predictions": predictions.cpu().numpy(),
            "probabilities": probs.cpu().numpy(),
            "labels": [config.label_names[p] for p in predictions.cpu().numpy()]
        }


def get_model() -> LyricEmotionClassifier:
    """Factory function to get classifier instance."""
    return LyricEmotionClassifier().load_pretrained()