app.py

import gradio as gr
import torch
import torch.nn as nn
from torchvision import models
from ultralytics import YOLO
import cv2
import numpy as np

# ==========================================
# 1. SETUP & MODEL LOADING
# ==========================================
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"--- System Boot: Using {device} ---")

# --- LOAD VISUAL SYSTEM (YOLO) ---
try:
    yolo_model = YOLO("best.pt")
    print("✅ Visual System: Custom EAGLE A7 Model Loaded")
except:
    yolo_model = YOLO("yolo11n.pt")

# --- LOAD THERMAL SYSTEM (ResNet-18) ---
def get_thermal_model():
    model = models.resnet18(weights=None)
    model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
    model.fc = nn.Linear(model.fc.in_features, 1)
    return model

thermal_model = get_thermal_model().to(device)
MODEL_PATH = "thermal_landmine_scanner.pth"

try:
    thermal_model.load_state_dict(torch.load(MODEL_PATH, map_location=device))
    thermal_model.eval()
    print(f"✅ Thermal System: Loaded {MODEL_PATH}")
except Exception as e:
    print(f"❌ CRITICAL ERROR: Could not load thermal model. {e}")

# ==========================================
# 2. SETUP GRAD-CAM (THE "X-RAY" HOOK)
# ==========================================
# We need to steal the features from inside the model while it thinks
features_blob = []
def hook_feature(module, input, output):
    features_blob.clear() # Clear old data
    features_blob.append(output.data.cpu().numpy())

# Attach the spy hook to the last layer (Layer 4)
thermal_model.layer4.register_forward_hook(hook_feature)

# Get weights from the final decision layer
params = list(thermal_model.parameters())
weight_softmax = params[-2].data.cpu().numpy() # The weights connecting features to "Mine/Safe"

# ==========================================
# 3. PROCESSING FUNCTIONS
# ==========================================

def run_visual_detection(image):
    if image is None: return None, "Waiting for feed..."
    results = yolo_model.predict(image, conf=0.65)
    return results[0].plot(), f"Objects Detected: {len(results[0].boxes)}"

def run_thermal_scan(image):
    if image is None: return None, "No Signal", "N/A"

    # --- PREPROCESSING (Standard) ---
    if len(image.shape) == 3:
        gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
    else:
        gray = image

    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
    enhanced_img = clahe.apply(gray)
    resized = cv2.resize(enhanced_img, (224, 224))
    normalized_img = resized.astype(np.float32) / 255.0
    
    tensor = torch.from_numpy(normalized_img).float().unsqueeze(0).unsqueeze(0)
    tensor = tensor.to(device)
    
    # --- INFERENCE ---
    with torch.no_grad():
        output = thermal_model(tensor)
        prob = torch.sigmoid(output).item()
        
    # --- GENERATE HEATMAP (Explainable AI) ---
    # 1. Get the features captured by our hook [1, 512, 7, 7]
    feature_data = features_blob[0]
    
    # 2. Calculate the "Attention Map"
    cam = np.zeros((7, 7), dtype=np.float32)
    # Use the weights for the "Mine" class to weight the features
    for i, w in enumerate(weight_softmax[0]): 
        cam += w * feature_data[0, i, :, :]
        
    # 3. Process the CAM
    cam = np.maximum(cam, 0) # ReLU (Remove negative influence)
    cam = cv2.resize(cam, (224, 224)) # Resize to image size
    cam = cam - np.min(cam)
    if np.max(cam) != 0:
        cam = cam / np.max(cam) # Normalize 0-1
        
    # 4. Colorize
    heatmap = cv2.applyColorMap(np.uint8(255*cam), cv2.COLORMAP_JET)
    
    # 5. Overlay on original (Grayscale -> RGB)
    orig_rgb = cv2.cvtColor(resized, cv2.COLOR_GRAY2RGB)
    
    # Mix: 60% Original Image + 40% Heatmap
    # If prob is low (Safe), we show less heatmap so it doesn't look scary
    intensity = 0.5 if prob > 0.5 else 0.2 
    overlay = cv2.addWeighted(orig_rgb, 1-intensity, heatmap, intensity, 0)
    
    # --- DECISION ---
    is_mine = prob > 0.50
    if is_mine:
        label = f"<h2 style='color: red; text-align: center;'>🔴 MINE DETECTED</h2>"
        conf_text = f"CONFIDENCE: {prob*100:.1f}%"
    else:
        label = f"<h2 style='color: green; text-align: center;'>🟢 SAFE SOIL</h2>"
        conf_text = f"Risk Level: {prob*100:.1f}%"

    return overlay, label, conf_text

# ==========================================
# 4. DASHBOARD UI
# ==========================================
custom_css = ".gradio-container {background-color: #1e1e1e; color: white}"

with gr.Blocks(css=custom_css, title="EAGLE A7 Mission Control") as demo:
    gr.Markdown("# 🦅 EAGLE A7: Autonomous Demining Interface")
    
    with gr.Tabs():
        with gr.TabItem("☀️ Daytime Vision"):
            with gr.Row():
                vis_input = gr.Image(label="Input", type="numpy")
                vis_output = gr.Image(label="YOLO Detections")
            vis_btn = gr.Button("SCAN", variant="primary")
            vis_status = gr.Textbox(label="Status")
            vis_btn.click(run_visual_detection, inputs=vis_input, outputs=[vis_output, vis_status])

        with gr.TabItem("🌙 Night Vision (X-Ray)"):
            gr.Markdown("### Thermal Anomaly Localization")
            with gr.Row():
                with gr.Column():
                    therm_input = gr.Image(label="Thermal Feed", type="numpy")
                    therm_btn = gr.Button("ANALYZE & LOCATE", variant="stop")
                with gr.Column():
                    therm_output = gr.Image(label="Target Localization (Heatmap)")
                    therm_label = gr.HTML(label="Result")
                    therm_conf = gr.Textbox(label="Telemetry")
            
            therm_btn.click(run_thermal_scan, inputs=therm_input, outputs=[therm_output, therm_label, therm_conf])

print("--- Launching Dashboard with X-Ray Vision ---")
demo.launch(server_name="0.0.0.0", share=True)