demo/inference.py

"""
Inference module for AI-Generated Image Detection using ARNIQA model.
Simplified for demo purposes - CPU-only, no training dependencies.
"""
import os
import sys
import torch
import torch.nn.functional as F
from torchvision import transforms
from PIL import Image
import numpy as np

# Add parent directory to path to import existing modules
parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, parent_dir)

import features.ARNIQA as ARNIQA
import functions.networks as networks
import defaults


def load_model(device='cpu'):
    """
    Load ARNIQA feature extractor and classifier.

    Args:
        device (str): Device to load model on ('cpu' or 'cuda')

    Returns:
        tuple: (feature_extractor, classifier) - both models ready for inference
    """
    print("Loading ARNIQA model...")

    # Load ARNIQA feature extractor
    feature_extractor = ARNIQA.Compute_ARNIQA(device=device)
    feature_extractor.eval()

    # Load classifier
    classifier = networks.Classifier_Arch2(
        input_dim=4096,  # ARNIQA produces 4096-dim features
        hidden_layers=[1024]
    )

    # Load checkpoint
    checkpoint_path = os.path.join(
        defaults.main_checkpoints_dir,
        "GenImage/extensive/MarginContrastiveLoss_CrossEntropy/arniqa/best_model.ckpt"
    )

    if not os.path.exists(checkpoint_path):
        raise FileNotFoundError(
            f"Checkpoint not found at: {checkpoint_path}\n"
            f"Please ensure the trained model checkpoint exists."
        )

    # Load checkpoint (PyTorch Lightning format)
    checkpoint = torch.load(checkpoint_path, map_location=device)

    # Extract state dict from Lightning checkpoint
    state_dict = checkpoint['state_dict']

    # Remove 'classifier.' prefix from keys if present
    new_state_dict = {}
    for key, value in state_dict.items():
        if key.startswith('classifier.'):
            new_state_dict[key.replace('classifier.', '')] = value
        else:
            new_state_dict[key] = value

    classifier.load_state_dict(new_state_dict)
    classifier = classifier.to(device)
    classifier.eval()

    print("Model loaded successfully!")
    return feature_extractor, classifier


def preprocess_image(image_pil):
    """
    Preprocess image for ARNIQA feature extraction.
    Creates two scales: original size and 224x224.

    Args:
        image_pil (PIL.Image): Input image

    Returns:
        tuple: (img1_tensor, img2_tensor) - two scales of the image
    """
    # Normalization (ImageNet stats)
    normalize = transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )

    # Transform for original size
    transform_original = transforms.Compose([
        transforms.ToTensor(),
        normalize
    ])

    # Transform for 224x224
    transform_224 = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        normalize
    ])

    # Convert to RGB if needed
    if image_pil.mode != 'RGB':
        image_pil = image_pil.convert('RGB')

    # Apply transforms
    img1 = transform_224(image_pil).unsqueeze(0)  # Add batch dimension
    img2 = transform_original(image_pil).unsqueeze(0)

    return img1, img2


def predict(image_pil, feature_extractor, classifier, device='cpu'):
    """
    Run inference on a single image.

    Args:
        image_pil (PIL.Image): Input image
        feature_extractor: ARNIQA feature extractor
        classifier: Trained classifier
        device (str): Device to run inference on

    Returns:
        tuple: (prediction, confidence, (prob_real, prob_fake))
            - prediction (str): "Real" or "Fake"
            - confidence (float): Confidence percentage (0-100)
            - prob_real (float): Probability of real (0-100)
            - prob_fake (float): Probability of fake (0-100)
    """
    # Preprocess
    img1, img2 = preprocess_image(image_pil)
    img1 = img1.to(device)
    img2 = img2.to(device)

    # Extract features
    with torch.no_grad():
        features = feature_extractor(img1, img2)
        features = torch.flatten(features, start_dim=1).to(torch.float32)

        # Classify
        _, logits = classifier(features)

        # Get probabilities
        probs = F.softmax(logits, dim=1)
        prob_real = probs[0, 0].item() * 100  # Class 0 = Real
        prob_fake = probs[0, 1].item() * 100  # Class 1 = Fake

        # Determine prediction (threshold = 0.5)
        if prob_fake > 50.0:
            prediction = "Fake"
            confidence = prob_fake
        else:
            prediction = "Real"
            confidence = prob_real

    return prediction, confidence, (prob_real, prob_fake)


# Test function
if __name__ == "__main__":
    # Test the model loading
    feature_extractor, classifier = load_model()
    print("Model test passed!")