app.py

# Create a Gradio demo for the Multi-Head model

import gradio as gr
import torch
import torch.nn as nn
from transformers import AutoModelForImageClassification, AutoImageProcessor
from huggingface_hub import hf_hub_download

# ========================================
# MODEL DEFINITION
# ========================================
class MultiHeadContentModerator(nn.Module):
    def __init__(self, base_model_name="Falconsai/nsfw_image_detection", num_violence_labels=2):
        super().__init__()
        original_model = AutoModelForImageClassification.from_pretrained(base_model_name)
        hidden_size = original_model.config.hidden_size
        
        self.vit = original_model.vit
        self.nsfw_classifier = original_model.classifier
        self.violence_classifier = nn.Linear(hidden_size, num_violence_labels)
        
        # Falconsai uses: {0: 'normal', 1: 'nsfw'}
        self.nsfw_id2label = {0: 'normal', 1: 'nsfw'}
        self.violence_id2label = {0: 'safe', 1: 'violence'}
        
    def forward(self, pixel_values, task='both'):
        outputs = self.vit(pixel_values=pixel_values)
        pooled_output = outputs.last_hidden_state[:, 0]
        
        if task == 'both':
            return {
                'nsfw': self.nsfw_classifier(pooled_output),
                'violence': self.violence_classifier(pooled_output)
            }
        elif task == 'nsfw':
            return self.nsfw_classifier(pooled_output)
        elif task == 'violence':
            return self.violence_classifier(pooled_output)

# ========================================
# LOAD MODEL
# ========================================
MODEL_ID = "Ali7880/multihead-content-moderator"  # Change this!

# Download model files
checkpoint_path = hf_hub_download(MODEL_ID, "multihead_model.pt")
checkpoint = torch.load(checkpoint_path, map_location='cpu', weights_only=False)

# Create and load model
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']
model.eval()

processor = AutoImageProcessor.from_pretrained(MODEL_ID)

# ========================================
# INFERENCE FUNCTION
# ========================================
def moderate_image(image):
    if image is None:
        return None, None, "Please upload an image"
    
    # Preprocess
    inputs = processor(images=image, return_tensors="pt")
    
    # Predict
    with torch.no_grad():
        outputs = model(inputs['pixel_values'], task='both')
        nsfw_probs = torch.softmax(outputs['nsfw'], dim=-1).numpy()[0]
        violence_probs = torch.softmax(outputs['violence'], dim=-1).numpy()[0]
    
    # Format results
    nsfw_results = {model.nsfw_id2label[i]: float(p) for i, p in enumerate(nsfw_probs)}
    violence_results = {model.violence_id2label[i]: float(p) for i, p in enumerate(violence_probs)}
    
    # Falconsai: {0: 'normal', 1: 'nsfw'}, Violence: {0: 'safe', 1: 'violence'}
    is_nsfw = nsfw_probs.argmax() == 1  # 1 = nsfw
    is_violent = violence_probs.argmax() == 1  # 1 = violence
    
    flags = []
    if is_nsfw:
        flags.append(f"NSFW ({nsfw_results.get('nsfw', 0):.0%})")
    if is_violent:
        flags.append(f"Violence ({violence_results.get('violence', 0):.0%})")
    
    if flags:
        verdict = "❌ UNSAFE - " + ", ".join(flags)
    else:
        verdict = f"✅ SAFE (Normal: {nsfw_results.get('normal', 0):.0%}, Safe: {violence_results.get('safe', 0):.0%})"
    
    return nsfw_results, violence_results, verdict

# ========================================
# GRADIO INTERFACE
# ========================================
demo = gr.Interface(
    fn=moderate_image,
    inputs=gr.Image(type="pil", label="Upload Image"),
    outputs=[
        gr.Label(label="NSFW Detection", num_top_classes=2),
        gr.Label(label="Violence Detection", num_top_classes=2),
        gr.Textbox(label="Overall Verdict")
    ],
    title="🛡️ Multi-Head Content Moderator",
    description="Upload an image to check for NSFW and Violence content simultaneously."
)

if __name__ == "__main__":
    demo.launch()