app.py

import json
import importlib.util
import sys
from pathlib import Path
import torch
import torch.nn as nn
import gradio as gr
from PIL import Image
from torchvision import transforms
from huggingface_hub import hf_hub_download

MODEL_REPO = "OhMyYuwan/face-forgery-detection"

with open(hf_hub_download(MODEL_REPO, "registry.json")) as f:
    REGISTRY = json.load(f)["models"]

with open(hf_hub_download(MODEL_REPO, "optimal_thresholds.json")) as f:
    THRESHOLDS = json.load(f)

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
_model_cache = {}


def load_model_class(model_name):
    """Dynamically load model class from HuggingFace"""
    model_file = hf_hub_download(MODEL_REPO, f"{model_name}/model.py")

    spec = importlib.util.spec_from_file_location(f"{model_name}_model", model_file)
    module = importlib.util.module_from_spec(spec)
    sys.modules[f"{model_name}_model"] = module
    spec.loader.exec_module(module)

    return module.OurNet


def load_model(model_name):
    if model_name in _model_cache:
        return _model_cache[model_name]

    config_path = hf_hub_download(MODEL_REPO, REGISTRY[model_name]["config_file"])
    model_path = hf_hub_download(MODEL_REPO, REGISTRY[model_name]["model_file"])

    with open(config_path) as f:
        config = json.load(f)

    # Dynamically load model class
    OurNet = load_model_class(model_name)
    model = OurNet(config)

    state = torch.load(model_path, map_location=DEVICE, weights_only=False)
    model.load_state_dict(state, strict=False)
    model.to(DEVICE).eval()
    _model_cache[model_name] = model
    return model


def predict(image, model_name):
    if image is None:
        return "", ""

    try:
        model = load_model(model_name)
        threshold = THRESHOLDS[model_name]["threshold"]

        x = transform(image.convert("RGB")).unsqueeze(0).to(DEVICE)
        with torch.no_grad():
            _, det = model.forward_det(x)
            score = torch.sigmoid(det).item()

        label = "🔴 Forged" if score > threshold else "✅ Real"
        confidence = score if score > threshold else 1 - score

        result_text = f"**Prediction:** {label}\n**Confidence:** {confidence*100:.2f}%"
        details = f"Raw Score: {score:.4f}\nThreshold: {threshold:.4f}\nModel: {model_name}"

        return result_text, details
    except Exception as e:
        return f"❌ Error: {str(e)}", ""


# Only support models that don't require local files
SUPPORTED_MODELS = [
    "convnext_base",
    "inceptionnext_base",
    "maxvit_base",
]
MODEL_NAMES = [m for m in REGISTRY.keys() if m in SUPPORTED_MODELS]

with gr.Blocks(title="Face Forgery Detection") as demo:
    gr.Markdown("# 🔍 Face Forgery Detection\nDetect whether a face image is real or forged using state-of-the-art models.")

    model_selector = gr.Dropdown(choices=MODEL_NAMES, value=MODEL_NAMES[0], label="Select Model")

    with gr.Row():
        with gr.Column():
            img_input = gr.Image(type="pil", label="Upload Image")
            detect_btn = gr.Button("🔍 Detect", variant="primary", size="lg")
        with gr.Column():
            result_output = gr.Markdown(label="Result")
            details_output = gr.Textbox(label="Details", lines=3, interactive=False)

    detect_btn.click(
        predict,
        inputs=[img_input, model_selector],
        outputs=[result_output, details_output],
        api_name=False,
    )

demo.launch()