import torch
import torchvision.transforms as transforms
import numpy as np
import gradio as gr
from PIL import Image, ImageDraw
from facenet_pytorch import MTCNN, InceptionResnetV1
import time
# Initialize MTCNN for face detection with smaller face size detection
mtcnn = MTCNN(keep_all=True, device='cuda' if torch.cuda.is_available() else 'cpu', min_face_size=20)
# Load the pre-trained FaceNet model
facenet = InceptionResnetV1(pretrained='vggface2').eval().to('cuda' if torch.cuda.is_available() else 'cpu')
model_path = r'faceNet_update_transformation.pth'
model_state_dict = torch.load(model_path)
facenet.load_state_dict(model_state_dict)
facenet.eval()  # Set the model to evaluation mode
# Define the transformation with normalization
val_test_transform = transforms.Compose([
    transforms.Resize((160, 160)),  # FaceNet expects 160x160 input
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
def compare_faces(embedding1, embedding2, threshold=0.2):  # Adjusted threshold
    dist = np.linalg.norm(embedding1 - embedding2)
    return dist, dist < threshold
def align_face(frame):
    # Convert the frame to a PIL image if it's a numpy array
    if isinstance(frame, np.ndarray):
        frame = Image.fromarray(frame)
        boxes, _ = mtcnn.detect(frame)
    if boxes is not None and len(boxes) > 0:
        faces = mtcnn(frame)
        if faces is not None and len(faces) > 0:
            face = faces[0]
            # Convert the face tensor to PIL Image
            face = transforms.ToPILImage()(face)
            return face, boxes[0]
    return None, None
def draw_bounding_box(image, box):
    draw = ImageDraw.Draw(image)
    draw.rectangle(box.tolist(), outline="red", width=3)
    return image
def l2_normalize(tensor):
    norm = np.linalg.norm(tensor, ord=2, axis=1, keepdims=True)
    return tensor / norm
def process_images(image1, image2):
    start_time = time.time()
    
    frame1 = np.array(image1)
    frame2 = np.array(image2)
    
    face1, box1 = align_face(frame1)
    face2, box2 = align_face(frame2)
    
    if face1 is None or face2 is None:
        return None, "Face not detected in one or both images."
    
    face1 = val_test_transform(face1).unsqueeze(0).to('cuda' if torch.cuda.is_available() else 'cpu')
    face2 = val_test_transform(face2).unsqueeze(0).to('cuda' if torch.cuda.is_available() else 'cpu')
    
    with torch.no_grad():
        embedding1 = facenet(face1).cpu().numpy()
        embedding2 = facenet(face2).cpu().numpy()
    
    embedding1 = l2_normalize(embedding1)
    embedding2 = l2_normalize(embedding2)
    
    distance, is_match = compare_faces(embedding1, embedding2, threshold=0.2)
    
    # Calculate confidence
    confidence = max(0.0, 1.0 - distance / 1.0)  # Ensure confidence is between 0 and 1
    print(f'confidence={confidence}')
    end_time = time.time()
    inference_time = end_time - start_time
    
    # Draw bounding boxes on the original images
    image1_with_box = draw_bounding_box(image1, box1)
    image2_with_box = draw_bounding_box(image2, box2)
    
    result = f"Distance: {distance:.2f}\nMatch: {is_match}\nInference time: {inference_time:.2f} seconds"
    
    return [image1_with_box, image2_with_box], result
# Create the Gradio interface
iface = gr.Interface(
    fn=process_images,
    inputs=[gr.Image(type="pil"), gr.Image(type="pil")],
    outputs=[gr.Gallery(), gr.Textbox()],
    title="Face Verification with FaceNet",
    description="Upload two images and the model will verify if the faces in both images are of the same person."
)
# Launch the interface
iface.launch(share=True, debug=True)