model_class.py

import torch
import torch.nn as nn
from transformers import AutoModelForImageClassification, AutoImageProcessor

class MultiHeadContentModerator(nn.Module):
    """
    Multi-task model with two classification heads:
    - Head 1: NSFW detection (frozen, pretrained)
    - Head 2: Violence detection (trainable)
    """
    def __init__(self, base_model_name="Falconsai/nsfw_image_detection", num_violence_labels=2):
        super().__init__()
        
        # Load base model
        original_model = AutoModelForImageClassification.from_pretrained(base_model_name)
        hidden_size = original_model.config.hidden_size
        
        # ViT backbone (shared)
        self.vit = original_model.vit
        
        # Head 1: Original NSFW classifier
        self.nsfw_classifier = original_model.classifier
        
        # Head 2: Violence classifier
        self.violence_classifier = nn.Linear(hidden_size, num_violence_labels)
        
        # Label mappings - use actual Falconsai config
        self.nsfw_id2label = original_model.config.id2label  # {0: 'normal', 1: 'nsfw'}
        self.violence_id2label = {0: 'safe', 1: 'violence'}  # Will be overwritten from checkpoint
        
    def forward(self, pixel_values, task='both'):
        outputs = self.vit(pixel_values=pixel_values)
        pooled_output = outputs.last_hidden_state[:, 0]
        
        if task == 'nsfw':
            return self.nsfw_classifier(pooled_output)
        elif task == 'violence':
            return self.violence_classifier(pooled_output)
        elif task == 'both':
            return {
                'nsfw': self.nsfw_classifier(pooled_output),
                'violence': self.violence_classifier(pooled_output)
            }
        return self.violence_classifier(pooled_output)

def load_multihead_model(checkpoint_path, device='cuda'):
    """Load trained multi-head model"""
    checkpoint = torch.load(checkpoint_path, map_location=device)
    
    model = MultiHeadContentModerator(
        base_model_name=checkpoint['base_model'],
        num_violence_labels=checkpoint['num_violence_labels']
    )
    model.load_state_dict(checkpoint['model_state_dict'])
    model.violence_id2label = checkpoint['violence_id2label']
    model.nsfw_id2label = checkpoint['nsfw_id2label']
    
    return model.to(device)