model/predict.py

import torch
from torchvision import models, transforms
from PIL import Image

MODEL_PATH = "models/deeptrust_weights.pt"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Recreate the same model architecture
model = models.efficientnet_b0(pretrained=False)
model.classifier[1] = torch.nn.Linear(model.classifier[1].in_features, 2)

# Load checkpoint safely
state_dict = torch.load(MODEL_PATH, map_location=device)
model.load_state_dict(state_dict)
model.to(device)
model.eval()

# Transform for input images
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485,0.456,0.406],[0.229,0.224,0.225])
])

def predict(image: Image.Image):
    """
    Predicts if an image is Real or AI-generated and returns trust score.
    """
    tensor = transform(image).unsqueeze(0).to(device)
    with torch.no_grad():
        logits = model(tensor)
        probs = torch.softmax(logits, dim=1)

    # Get the predicted index
    pred_index = torch.argmax(probs, dim=1).item()
    confidence = probs[0, pred_index].item()

    # Map pred_index to Real / Fake
    # This is now automatic: higher probability class = model's predicted class
    label = "Real" if pred_index == 0 else "Fake"

    # Trust score: higher for Real images if predicted Real
    trust_score = int(confidence*100 if label=="Real" else (1-confidence)*100)

    # Debug print to verify
    print("pred_index:", pred_index, "label:", label, "confidence:", confidence)

    return label, confidence, trust_score