import torch
import torch.nn as nn
import numpy as np
from PIL import Image, ImageDraw
from torchvision import transforms
from transformers import ViTImageProcessor, ViTModel
from facenet_pytorch import MTCNN
import gradio as gr
import time

# Define the Vision Transformer (ViT) architecture
class ViT(nn.Module):
    def __init__(self, base_model):
        super(ViT, self).__init__()
        self.base_model = base_model
        self.dropout = nn.Dropout(p=0.2)
        self.fc = nn.Linear(base_model.config.hidden_size, 512)
        self.dropout2 = nn.Dropout(p=0.2)
        self.l2_norm = nn.functional.normalize

    def forward(self, x):
        x = self.base_model(x).pooler_output
        x = self.dropout(x)
        x = self.fc(x)
        x = self.dropout2(x)
        x = self.l2_norm(x, p=2, dim=1)  # Apply L2 normalization
        return x

# Load the pre-trained ViT model and processor
model_name = "google/vit-base-patch16-224"
processor = ViTImageProcessor.from_pretrained(model_name) 
base_model = ViTModel.from_pretrained(model_name)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

model = ViT(base_model).to(device)
model_path = r'best_vit11.pth'
model.load_state_dict(torch.load(model_path))


# Initialize MTCNN for face detection
mtcnn = MTCNN(keep_all=True, post_process=False, min_face_size=12, device=device)

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 euclidean_distance(embedding1, embedding2):
    return np.linalg.norm(embedding1 - embedding2)

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."
    
    # Use processor to preprocess the images
    face1 = processor(images=face1, return_tensors="pt").pixel_values.to('cuda' if torch.cuda.is_available() else 'cpu')
    face2 = processor(images=face2, return_tensors="pt").pixel_values.to('cuda' if torch.cuda.is_available() else 'cpu')
    
    with torch.no_grad():
        embedding1 = model(face1).cpu().numpy()
        embedding2 = model(face2).cpu().numpy()
    
    # Flatten the embeddings if necessary (ensuring they are 1D)
    embedding1 = embedding1.flatten()
    embedding2 = embedding2.flatten()
    
    euclidean_dist = euclidean_distance(embedding1, embedding2)
    is_match = euclidean_dist < 0.8
    
    # Calculate confidence
    confidence = max(0.0, 1.0 - euclidean_dist / 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"Euclidean Distance: {euclidean_dist:.2f}\n"
    result += f"Match: {is_match}\n"
    result += f"Inference 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 Vision Transformer",
    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)