import gradio as gr
import torch
import torch.nn as nn
from torchvision import models, transforms
from PIL import Image
import cv2
import numpy as np

# 1. Download OpenCV Haarcascades for eye tracking
import urllib.request
import os
cascade_path = 'haarcascade_eye.xml'
if not os.path.exists(cascade_path):
    urllib.request.urlretrieve(
        'https://raw.githubusercontent.com/opencv/opencv/master/data/haarcascades/haarcascade_eye.xml', 
        cascade_path
    )
eye_cascade = cv2.CascadeClassifier(cascade_path)

# 2. Re-initialize and load the model (Using your weights)
model = models.mobilenet_v2(weights=None)
# Ensure this matches exactly how you defined it in the Masterpiece training step
model.classifier = nn.Sequential(
    nn.Dropout(p=0.5),
    nn.Linear(model.last_channel, 2)
)
model.load_state_dict(torch.load('ddobj_model.pth', map_location=torch.device('cpu')))
model.eval()

# 3. Transforms (Grayscale is key to matching the MRL dataset!)
transform = transforms.Compose([
    transforms.Grayscale(num_output_channels=3), # Convert to 3-channel grayscale
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

# 4. The Smart Prediction Function
def predict_drowsiness(image):
    # Convert Gradio image to OpenCV format
    img_cv = np.array(image)
    gray = cv2.cvtColor(img_cv, cv2.COLOR_RGB2GRAY)
    
    # Detect eyes in the image
    eyes = eye_cascade.detectMultiScale(gray, scaleFactor=1.1, minNeighbors=5, minSize=(30, 30))
    
    if len(eyes) == 0:
        return "ERROR: Could not detect any eyes in the image. Please upload a clear face photo.", None
    
    # Take the first detected eye (largest/clearest)
    (x, y, w, h) = eyes[0]
    
    # Crop the eye from the original image
    eye_crop = img_cv[y:y+h, x:x+w]
    
    # Convert the cropped eye back to PIL for PyTorch
    eye_pil = Image.fromarray(eye_crop)
    input_tensor = transform(eye_pil).unsqueeze(0)
    
    # Run the model
    with torch.no_grad():
        outputs = model(input_tensor)
        _, predicted = torch.max(outputs, 1)
        
    classes = ["DROWSY ALERT! 🚨 (Eyes Closed)", "NOT DROWSY ✅ (Eyes Open)"]
    result = classes[predicted.item()]
    
    # Return the prediction AND show the user the exact crop the model looked at
    return result, eye_pil

# 5. Build the UI
interface = gr.Interface(
    fn=predict_drowsiness,
    inputs=gr.Image(label="Upload Full Face Photo"),
    outputs=[
        gr.Textbox(label="DDobj System Status"), 
        gr.Image(label="What the AI saw (Eye Crop)")
    ],
    title="DDobj: Driver Drowsiness Detection",
    description="Upload a photo. The system will automatically locate the eyes, isolate them, and analyze them for fatigue.",
    theme="default"
)

if __name__ == "__main__":
    interface.launch()