multihead-content-moderator

Multi-Head Content Moderator

A multi-task image moderation model with two classification heads:

• NSFW Detection: Detects explicit/adult content (preserved from Falconsai)
• Violence Detection: Detects violent content (newly trained)

Architecture

• Base: ViT (Vision Transformer) from Falconsai/nsfw_image_detection
• Head 1: NSFW classifier (frozen, pretrained)
• Head 2: Violence classifier (trained on violence dataset)

Categories

NSFW Head

• nsfw
• safe

Violence Head

• safe
• violence

Performance (Violence Detection)

• Accuracy: 0.9075
• F1 Score: 0.9076

Usage

import torch
from transformers import AutoImageProcessor

# Load
checkpoint = torch.load('multihead_model.pt')
processor = AutoImageProcessor.from_pretrained('path/to/model')

# Create model class (see notebook for full class definition)
# model = MultiHeadContentModerator(...)
# model.load_state_dict(checkpoint['model_state_dict'])

# Inference
inputs = processor(images=image, return_tensors='pt')
with torch.no_grad():
    # Get both predictions
    outputs = model(inputs['pixel_values'], task='both')
    nsfw_pred = outputs['nsfw'].argmax(-1)
    violence_pred = outputs['violence'].argmax(-1)