inference.py

import torch
import torch.nn.functional as F
import numpy as np
import cv2
import io
import matplotlib.pyplot as plt
from PIL import Image
from torchvision import transforms
from model import DeepfakeClassifier

# Device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load model
model = DeepfakeClassifier()
state_dict = torch.load("model_weights/model_87_acc_20_frames_final_data.pt", map_location=device)
model.load_state_dict(state_dict, strict=False)
model.to(device)
model.eval()

# ✅ Image transform (ImageNet normalization restored)
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406],
                         [0.229, 0.224, 0.225])
])

# Softmax with temperature scaling
def softmax_with_temperature(logits, T=0.5):
    return torch.softmax(logits / T, dim=1)

# Grad-CAM logic
def apply_gradcam(image_tensor, target_layer):
    gradients, activations = [], []

    def backward_hook(module, grad_input, grad_output):
        gradients.append(grad_output[0])

    def forward_hook(module, input, output):
        activations.append(output)

    handle_fwd = target_layer.register_forward_hook(forward_hook)
    handle_bwd = target_layer.register_backward_hook(backward_hook)

    output = model(image_tensor)
    class_idx = output.argmax().item()
    score = output[0, class_idx]
    model.zero_grad()
    score.backward()

    grads = gradients[0]
    acts = activations[0]
    weights = grads.mean(dim=[2, 3], keepdim=True)
    cam = (weights * acts).sum(dim=1).squeeze().cpu().detach().numpy()
    cam = np.maximum(cam, 0)
    cam = cv2.resize(cam, (224, 224))
    cam -= cam.min()
    cam /= (cam.max() + 1e-8)

    handle_fwd.remove()
    handle_bwd.remove()
    return cam, class_idx

# Grad-CAM overlay
def overlay_cam(image, cam):
    heatmap = cv2.applyColorMap(np.uint8(255 * cam), cv2.COLORMAP_JET)
    image = np.array(image.resize((224, 224)).convert("RGB"))
    overlay = cv2.addWeighted(image, 0.6, heatmap, 0.4, 0)
    return overlay

# ✅ Predict image (with normalization and T=1.0)
def predict_image(pil_image):
    image_tensor = transform(pil_image).unsqueeze(0).to(device)
    cam, _ = apply_gradcam(image_tensor, model.model[6][2].conv3)

    prob = softmax_with_temperature(model(image_tensor), T=0.5)[0]
    print("Image Probabilities:", prob.tolist())  # Optional debug

    fake_confidence = prob[0].item()
    label = "Fake" if fake_confidence > 0.7 else "Real"
    confidence = fake_confidence if label == "Fake" else prob[1].item()

    overlay = overlay_cam(pil_image, cam)
    return f"{label} ({confidence:.2f} confidence)", Image.fromarray(overlay)

# Predict video
def predict_video(video_path):
    cap = cv2.VideoCapture(video_path)
    frames = []

    while len(frames) < 20:
        ret, frame = cap.read()
        if not ret:
            break
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        pil_img = Image.fromarray(frame)
        frames.append(transform(pil_img))
    cap.release()

    if len(frames) == 0:
        return "No valid frames found.", [], None

    # Per-frame prediction + Grad-CAM
    frame_probs = []
    cam_images = []

    for frame in frames:
        frame_tensor = frame.unsqueeze(0).to(device)

        with torch.no_grad():
            out = model(frame_tensor)
            prob = softmax_with_temperature(out, T=2.0)[0][0].item()
        frame_probs.append(prob)

        cam, _ = apply_gradcam(frame_tensor, model.model[6][2].conv3)
        frame_np = (frame.permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8)
        frame_np = np.clip(frame_np, 0, 255)
        pil_frame = Image.fromarray(frame_np)
        overlay = overlay_cam(pil_frame, cam)
        cam_images.append(Image.fromarray(overlay))

    # Sequence prediction (LSTM)
    if len(frames) < 20:
        frames += [frames[-1]] * (20 - len(frames))
    video_tensor = torch.stack(frames).unsqueeze(0).to(device)

    with torch.no_grad():
        output = model(video_tensor)
        seq_prob = softmax_with_temperature(output, T=2.0)[0][0].item()
    seq_label = "Fake" if seq_prob > 0.8 else "Real"

    # Confidence chart
    fig, ax = plt.subplots()
    ax.plot(frame_probs, marker='o', label='Fake Probability')
    ax.axhline(0.5, color='red', linestyle='--', label='Threshold')
    ax.set_title('Per-frame Fake Confidence')
    ax.set_xlabel('Frame Index')
    ax.set_ylabel('Probability')
    ax.legend()
    ax.grid(True)

    buf = io.BytesIO()
    plt.savefig(buf, format='png')
    plt.close(fig)
    buf.seek(0)
    chart_img = Image.open(buf)

    # Final decision
    label_text = f"Final Decision (Sequence): {seq_label} ({seq_prob:.2f})"
    return label_text, cam_images, chart_img