app.py

import os
# Set environment variables for Spaces compatibility
os.environ['OMP_NUM_THREADS'] = '1'
os.environ['MKL_NUM_THREADS'] = '1'
import cv2
import yaml
import torch
import random
import gradio as gr
import numpy as np
import kagglehub
from PIL import Image
from glob import glob
import matplotlib
matplotlib.use('Agg')  # Use non-interactive backend
import matplotlib.pyplot as plt
from matplotlib import patches
from torchvision import transforms as T
from ultralytics import YOLO
import shutil
import tempfile
from pathlib import Path
import json
from io import BytesIO

# Try to import spaces for Hugging Face Spaces GPU support
try:
    import spaces
    ON_SPACES = True
except ImportError:
    ON_SPACES = False
    # Create a dummy decorator if not on Spaces
    class spaces:
        @staticmethod
        def GPU(duration=60):
            def decorator(func):
                return func
            return decorator

# Set Kaggle API credentials from environment variable
if os.getenv("KDATA_API"):
    kaggle_key = os.getenv("KDATA_API")
    # Parse the key if it's in JSON format
    if "{" in kaggle_key:
        key_data = json.loads(kaggle_key)
        os.environ["KAGGLE_USERNAME"] = key_data.get("username", "")
        os.environ["KAGGLE_KEY"] = key_data.get("key", "")

# Global variables
model = None
dataset_path = None
training_in_progress = False

class Visualization:
    def __init__(self, root, data_types, n_ims, rows, cmap=None):
        self.n_ims, self.rows = n_ims, rows
        self.cmap, self.data_types = cmap, data_types
        self.colors = ["firebrick", "darkorange", "blueviolet"]
        self.root = root
        
        self.get_cls_names()
        self.get_bboxes()
                
    def get_cls_names(self):
        yaml_path = f"{self.root}/data.yaml"
        if not os.path.exists(yaml_path):
            print(f"Warning: {yaml_path} not found")
            self.class_dict = {}
            return
            
        with open(yaml_path, 'r') as file:
            data = yaml.safe_load(file)      
        class_names = data.get('names', [])
        self.class_dict = {index: name for index, name in enumerate(class_names)}
        
        # Print class names for debugging
        if self.class_dict:
            print(f"Dataset classes: {', '.join(class_names)}")        
    
    def get_bboxes(self):
        self.vis_datas, self.analysis_datas, self.im_paths = {}, {}, {}
        for data_type in self.data_types:
            all_bboxes, all_analysis_datas = [], {}
            im_paths = glob(f"{self.root}/{data_type}/images/*")
            
            for idx, im_path in enumerate(im_paths):
                bboxes = []
                im_ext = os.path.splitext(im_path)[-1]
                lbl_path = im_path.replace(im_ext, ".txt")
                lbl_path = lbl_path.replace(f"{data_type}/images", f"{data_type}/labels")
                if not os.path.isfile(lbl_path):
                    continue                
                meta_data = open(lbl_path).readlines()
                for data in meta_data:                    
                    parts = data.strip().split()[:5]
                    cls_name = self.class_dict[int(parts[0])]
                    bboxes.append([cls_name] + [float(x) for x in parts[1:]])
                    if cls_name not in all_analysis_datas:
                        all_analysis_datas[cls_name] = 1
                    else:
                        all_analysis_datas[cls_name] += 1
                all_bboxes.append(bboxes)
                    
            self.vis_datas[data_type] = all_bboxes
            self.analysis_datas[data_type] = all_analysis_datas
            self.im_paths[data_type] = im_paths
    
    def plot_single(self, im_path, bboxes):
        or_im = np.array(Image.open(im_path).convert("RGB"))
        height, width, _ = or_im.shape

        for bbox in bboxes:
            class_id, x_center, y_center, w, h = bbox

            x_min = int((x_center - w / 2) * width)
            y_min = int((y_center - h / 2) * height)
            x_max = int((x_center + w / 2) * width)
            y_max = int((y_center + h / 2) * height)
            
            color = (random.randint(0, 255), random.randint(0, 255), random.randint(0, 255))
            cv2.rectangle(img=or_im, pt1=(x_min, y_min), pt2=(x_max, y_max), 
                         color=color, thickness=3)
        
        # Add text overlay
        cv2.putText(or_im, f"Objects: {len(bboxes)}", (10, 30), 
                   cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2, cv2.LINE_AA)
        
        # OpenCV uses BGR, but PIL expects RGB, and we already loaded as RGB
        # So no conversion needed
        return Image.fromarray(or_im)

    def vis_samples(self, data_type, n_samples=4):
        if data_type not in self.vis_datas:
            return None
            
        indices = [random.randint(0, len(self.vis_datas[data_type]) - 1) 
                  for _ in range(min(n_samples, len(self.vis_datas[data_type])))]
        
        figs = []
        for idx in indices:
            im_path = self.im_paths[data_type][idx]
            bboxes = self.vis_datas[data_type][idx]
            fig = self.plot_single(im_path, bboxes)
            figs.append(fig)
            
        return figs

    def data_analysis(self, data_type):
        if data_type not in self.analysis_datas:
            return None
            
        plt.style.use('default')
        fig, ax = plt.subplots(figsize=(12, 6))
        
        cls_names = list(self.analysis_datas[data_type].keys())
        counts = list(self.analysis_datas[data_type].values())
        
        color_map = {"train": "firebrick", "valid": "darkorange", "test": "blueviolet"}
        color = color_map.get(data_type, "steelblue")
        
        indices = np.arange(len(counts))
        bars = ax.bar(indices, counts, 0.7, color=color)
        
        ax.set_xlabel("Class Names", fontsize=12)
        ax.set_xticks(indices)
        ax.set_xticklabels(cls_names, rotation=45, ha='right')
        ax.set_ylabel("Data Counts", fontsize=12)
        ax.set_title(f"{data_type.upper()} Dataset Class Distribution", fontsize=14)
        
        for i, (bar, v) in enumerate(zip(bars, counts)):
            ax.text(bar.get_x() + bar.get_width()/2, v + 1, str(v), 
                   ha='center', va='bottom', fontsize=10, color='navy')
        
        plt.tight_layout()
        
        # Save to BytesIO and convert to PIL Image
        buf = BytesIO()
        fig.savefig(buf, format='png', dpi=100, bbox_inches='tight')
        buf.seek(0)
        img = Image.open(buf)
        plt.close(fig)
        
        return img

def download_dataset():
    """Download the dataset using kagglehub"""
    global dataset_path
    try:
        # Create a local directory to store the dataset
        local_dir = "./xray_dataset"
        
        # Download dataset
        dataset_path = kagglehub.dataset_download("orvile/x-ray-baggage-anomaly-detection")
        
        # If the dataset is downloaded to a temporary location, copy it to our local directory
        if dataset_path != local_dir and os.path.exists(dataset_path):
            if os.path.exists(local_dir):
                shutil.rmtree(local_dir)
            shutil.copytree(dataset_path, local_dir)
            dataset_path = local_dir
        
        return f"Dataset downloaded successfully to: {dataset_path}"
    except Exception as e:
        return f"Error downloading dataset: {str(e)}\n\nPlease ensure KDATA_API environment variable is set correctly."

def visualize_data(data_type, num_samples):
    """Visualize sample images from the dataset"""
    if dataset_path is None:
        return [], "Please download the dataset first!"
    
    try:
        vis = Visualization(root=dataset_path, data_types=[data_type], 
                          n_ims=num_samples, rows=2, cmap="rgb")
        figs = vis.vis_samples(data_type, num_samples)
        if figs is None:
            return [], f"No data found for {data_type} dataset"
        return figs, f"Showing {len(figs)} samples from {data_type} dataset"
    except Exception as e:
        return [], f"Error visualizing data: {str(e)}"

def analyze_class_distribution(data_type):
    """Analyze class distribution in the dataset"""
    if dataset_path is None:
        return None, "Please download the dataset first!"
    
    try:
        vis = Visualization(root=dataset_path, data_types=[data_type], 
                          n_ims=20, rows=5, cmap="rgb")
        fig = vis.data_analysis(data_type)
        if fig is None:
            return None, f"No data found for {data_type} dataset"
        return fig, f"Class distribution for {data_type} dataset"
    except Exception as e:
        return None, f"Error analyzing data: {str(e)}"

@spaces.GPU(duration=300)  # Request GPU for 5 minutes for training
def train_model(epochs, batch_size, img_size, device_selection):
    """Train YOLOv11 model"""
    global model, training_in_progress
    
    if dataset_path is None:
        return [], "Please download the dataset first!"
    
    if training_in_progress:
        return [], "Training already in progress!"
    
    training_in_progress = True
    
    try:
        # Determine device - on Spaces, always use GPU if available
        if ON_SPACES and torch.cuda.is_available():
            device = 0
        elif device_selection == "Auto":
            device = 0 if torch.cuda.is_available() else "cpu"
        elif device_selection == "CPU":
            device = "cpu"
        else:
            device = 0 if torch.cuda.is_available() else "cpu"
        
        # Read dataset info
        yaml_path = f"{dataset_path}/data.yaml"
        with open(yaml_path, 'r') as file:
            data_config = yaml.safe_load(file)
        
        class_names = data_config.get('names', [])
        print(f"Training on {len(class_names)} classes: {class_names}")
        
        # Initialize model - use yolov8n if yolo11n not available
        try:
            model = YOLO("yolo11n.pt")
        except Exception as e:
            print(f"YOLOv11 not available: {e}, falling back to YOLOv8")
            model = YOLO("yolov8n.pt")  # Fallback to YOLOv8
        
        # Create project directory
        project_dir = "./xray_detection"
        os.makedirs(project_dir, exist_ok=True)
        
        # Train model with optimized settings for X-ray detection
        results = model.train(
            data=yaml_path,
            epochs=epochs,
            imgsz=img_size,
            batch=batch_size,
            device=device,
            project=project_dir,
            name="train",
            exist_ok=True,
            verbose=True,
            patience=5,  # Reduce patience for faster training on Spaces
            save_period=5,  # Save checkpoints every 5 epochs
            workers=0,  # Important: Set to 0 to avoid multiprocessing issues
            single_cls=False,
            rect=False,
            cache=False,  # Disable caching to avoid memory issues
            amp=True,  # Use automatic mixed precision for faster training
            # Optimization settings
            optimizer='AdamW',
            lr0=0.001,  # Initial learning rate
            lrf=0.01,  # Final learning rate factor
            momentum=0.937,
            weight_decay=0.0005,
            warmup_epochs=3.0,
            warmup_momentum=0.8,
            warmup_bias_lr=0.1,
            # Loss weights
            box=7.5,
            cls=0.5,
            dfl=1.5,
            # Augmentation settings for X-ray images
            hsv_h=0.0,  # No hue augmentation for X-ray
            hsv_s=0.0,  # No saturation augmentation
            hsv_v=0.1,  # Slight value augmentation
            degrees=0.0,  # No rotation
            translate=0.1,
            scale=0.5,
            shear=0.0,
            perspective=0.0,
            flipud=0.0,  # No vertical flip for X-ray
            fliplr=0.5,  # Horizontal flip is okay
            mosaic=1.0,
            mixup=0.0,
            copy_paste=0.0
        )
        
        # Collect training result plots
        results_path = os.path.join(project_dir, "train")
        plots = []
        
        plot_files = ["results.png", "confusion_matrix.png", "val_batch0_pred.jpg", 
                     "train_batch0.jpg", "val_batch0_labels.jpg"]
        
        for plot_file in plot_files:
            plot_path = os.path.join(results_path, plot_file)
            if os.path.exists(plot_path):
                plots.append(Image.open(plot_path))
        
        # Save the model path
        model_path = os.path.join(results_path, "weights", "best.pt")
        
        # Load the trained model to ensure it's ready for inference
        model_loaded = False
        class_info = ""
        if os.path.exists(model_path):
            try:
                model = YOLO(model_path)
                model_loaded = True
                class_info = f"\n✅ Trained on {len(model.names)} classes: {', '.join(list(model.names.values()))}"
                
                # Run a test inference to ensure model works
                test_img = np.zeros((640, 640, 3), dtype=np.uint8)
                test_results = model(test_img, verbose=False)
                class_info += "\n✅ Model test passed - ready for inference!"
            except Exception as e:
                class_info = f"\n⚠️ Model loaded but test failed: {str(e)}"
        else:
            class_info = "\n❌ Model file not found!"
        
        training_in_progress = False
        
        # Provide instructions for saving the model
        save_instructions = """
        
        ✅ **Training Complete!**
        
        📥 **Next Steps:**
        1. Click "📥 Download Model (.pt)" button below to save your model
        2. Keep the downloaded file safe - you'll need it after Space restarts
        3. To reuse: Upload the model file in the "Upload & Load Model" section
        
        ⚠️ **Important**: This model will be lost when the Space restarts!
        """
        
        return plots, f"Model saved to {model_path}{class_info}{save_instructions}"
        
    except Exception as e:
        training_in_progress = False
        return [], f"Error during training: {str(e)}"

# 🔍 Inference (Modified to highlight bomb, pistol, spring, grenade, eod_gear, battery)
@spaces.GPU(duration=60)  # Request GPU for 1 minute for inference
def run_inference(input_image, conf_threshold):
    """Run inference on a single image and print detected item names."""
    global model

    # Try to load the trained model if not already loaded
    if model is None:
        trained_model_path = "./xray_detection/train/weights/best.pt"
        if os.path.exists(trained_model_path):
            try:
                model = YOLO(trained_model_path)
                print(f"Loaded trained model from {trained_model_path}")
            except Exception:
                pass

        # If still no model, try default
        if model is None:
            for fallback in ("yolo11n.pt", "yolov8n.pt"):
                try:
                    model = YOLO(fallback)
                    print(f"Loaded fallback model: {fallback}")
                    break
                except Exception:
                    continue
            if model is None:
                return None, "Please train the model first or load a pre-trained model!"

    if input_image is None:
        return None, "Please upload an image!"

    try:
        # Save the input image temporarily with proper format
        temp_path = "temp_inference.jpg"
        if input_image.mode != 'RGB':
            input_image = input_image.convert('RGB')
        input_image.save(temp_path, format='JPEG', quality=95)

        # Run inference
        imgsz = 640
        results = model(
            temp_path,
            conf=conf_threshold,
            verbose=False,
            device=0 if torch.cuda.is_available() else 'cpu',
            imgsz=imgsz,
            augment=False,
            agnostic_nms=False,
            max_det=300
        )

        # Draw annotated image
        annotated_image = results[0].plot(
            conf=True,
            labels=True,
            boxes=True,
            masks=False,
            probs=False
        )

        # Prepare detection information
        detections = []
        detection_count = 0
        danger_set = {'bomb', 'pistol', 'spring', 'grenade', 'eod_gear', 'battery'}

        if results[0].boxes is not None:
            detection_count = len(results[0].boxes)
            for idx, box in enumerate(results[0].boxes):
                cls = int(box.cls)
                conf_val = float(box.conf)
                xyxy = list(map(int, box.xyxy[0].tolist()))
                cls_name = model.names.get(cls, f"Class {cls}")

                # Highlight dangerous items
                prefix = "‼️ " if cls_name in danger_set else ""
                detections.append(
                    f"{idx + 1}. {prefix}{cls_name}: {conf_val:.3f} "
                    f"| Box: [{xyxy[0]}, {xyxy[1]}, {xyxy[2]}, {xyxy[3]}]"
                )

        # Clean up temp file
        if os.path.exists(temp_path):
            os.remove(temp_path)

        # Assemble detection text
        det_text_header = (
            f"Model classes ({len(model.names)}): {', '.join(list(model.names.values())[:10])}...\n"
            f"Confidence threshold: {conf_threshold}\n\n"
        )
        if detections:
            detection_text = (
                det_text_header +
                f"✅ Found {detection_count} object(s):\n\n" +
                "\n".join(detections)
            )
        else:
            detection_text = det_text_header + "❌ No objects detected."

        return Image.fromarray(annotated_image), detection_text

    except Exception as e:
        import traceback
        traceback.print_exc()
        return None, f"Error during inference: {str(e)}"


@spaces.GPU(duration=60)  # Request GPU for batch inference
def batch_inference(data_type, num_images):
    """Run inference on multiple images from test set"""
    global model
    
    # Try to load the trained model if not already loaded
    if model is None:
        trained_model_path = "./xray_detection/train/weights/best.pt"
        if os.path.exists(trained_model_path):
            try:
                model = YOLO(trained_model_path)
                print(f"Loaded trained model for batch inference")
            except:
                try:
                    model = YOLO("yolo11n.pt")
                    print("Loaded default model for batch inference")
                except:
                    try:
                        model = YOLO("yolov8n.pt")
                        print("Loaded YOLOv8 model as fallback for batch inference")
                    except:
                        return [], "Please train the model first!"
        else:
            return [], "No trained model found. Please train the model first!"
    
    if dataset_path is None:
        return [], "Please download the dataset first!"
    
    try:
        image_dir = f"{dataset_path}/{data_type}/images"
        if not os.path.exists(image_dir):
            return [], f"Directory {image_dir} not found!"
            
        image_files = glob(f"{image_dir}/*")[:num_images]
        
        if not image_files:
            return [], f"No images found in {image_dir}"
        
        results_images = []
        detection_counts = []
        
        for img_path in image_files:
            results = model(img_path, verbose=False, conf=0.25, imgsz=640)
            annotated = results[0].plot()
            results_images.append(Image.fromarray(annotated))
            
            # Count detections
            if results[0].boxes is not None:
                detection_counts.append(len(results[0].boxes))
            else:
                detection_counts.append(0)
        
        # Check model type
        model_type = "X-ray detection model" if len(model.names) != 80 else "General COCO model"
        avg_detections = sum(detection_counts) / len(detection_counts) if detection_counts else 0
        
        return results_images, f"Processed {len(results_images)} images using {model_type}\nAverage detections per image: {avg_detections:.1f}"
        
    except Exception as e:
        return [], f"Error during batch inference: {str(e)}"

def get_dataset_info():
    """Get information about the X-ray dataset classes"""
    if dataset_path is None:
        return "Dataset not downloaded yet."
    
    try:
        yaml_path = f"{dataset_path}/data.yaml"
        if not os.path.exists(yaml_path):
            return "Dataset configuration file not found."
        
        with open(yaml_path, 'r') as file:
            data = yaml.safe_load(file)
        
        class_names = data.get('names', [])
        num_classes = len(class_names)
        
        # Count images
        train_images = len(glob(f"{dataset_path}/train/images/*")) if os.path.exists(f"{dataset_path}/train/images") else 0
        valid_images = len(glob(f"{dataset_path}/valid/images/*")) if os.path.exists(f"{dataset_path}/valid/images") else 0
        test_images = len(glob(f"{dataset_path}/test/images/*")) if os.path.exists(f"{dataset_path}/test/images") else 0
        
        info = f"### 📊 X-ray Baggage Dataset Info\n\n"
        info += f"**Classes ({num_classes}):** {', '.join(class_names)}\n\n"
        info += f"**Dataset Split:**\n"
        info += f"- Training: {train_images} images\n"
        info += f"- Validation: {valid_images} images\n"
        info += f"- Test: {test_images} images\n"
        info += f"- Total: {train_images + valid_images + test_images} images\n\n"
        info += f"**What to expect:** The model will learn to detect these prohibited items in X-ray scans."
        
        return info
    except Exception as e:
        return f"Error reading dataset info: {str(e)}"
    """Load a pre-trained model"""
    global model
    try:
        # Check if it's a HuggingFace model path
        if model_path.startswith("hf://") or "/" in model_path and not os.path.exists(model_path):
            # Load from HuggingFace Hub
            model = YOLO(model_path)
            return f"Model loaded successfully from HuggingFace: {model_path}"
        
        if not os.path.exists(model_path):
            # Try default paths
            default_paths = [
                "./xray_detection/train/weights/best.pt",
                "./xray_detection/train/weights/last.pt",
                "yolo11n.pt",
                "yolov8n.pt"
            ]
            for path in default_paths:
                if os.path.exists(path):
                    model_path = path
                    break
        
        if os.path.exists(model_path):
            model = YOLO(model_path)
            # Check if it's a trained model by looking at class names
            try:
                if hasattr(model, 'names') and len(model.names) > 0:
                    class_names = ", ".join([f"{i}: {name}" for i, name in model.names.items()][:5])
                    if len(model.names) > 5:
                        class_names += f"... (총 {len(model.names)} 클래스)"
                    return f"Model loaded successfully from {model_path}\n클래스: {class_names}"
            except:
                pass
            return f"Model loaded successfully from {model_path}"
        else:
            return "Model file not found. Please train a model first or provide a valid path."
    except Exception as e:
        return f"Error loading model: {str(e)}"

def load_pretrained_model(model_file):
    """Load a pre-trained model from uploaded file"""
    global model
    
    if model_file is None:
        return "Please upload a model file (.pt)"
    
    try:
        # model_file is already a filepath string when type="filepath"
        temp_path = model_file
        
        # Load the model
        model = YOLO(temp_path)
        
        # Check model info
        try:
            if hasattr(model, 'names') and len(model.names) > 0:
                num_classes = len(model.names)
                class_names = ", ".join([f"{name}" for name in list(model.names.values())[:5]])
                if len(model.names) > 5:
                    class_names += f"... (총 {num_classes} 클래스)"
                
                if num_classes == 80:
                    return f"⚠️ Loaded COCO model with {num_classes} classes. This is not trained for X-ray detection.\nClasses: {class_names}"
                else:
                    return f"✅ Model loaded successfully!\nClasses ({num_classes}): {class_names}"
            else:
                return "✅ Model loaded successfully!"
        except:
            return "✅ Model loaded successfully!"
            
    except Exception as e:
        return f"Error loading model: {str(e)}"

def check_model_status():
    """Check current model status"""
    global model
    if model is None:
        # Try to load trained model
        trained_path = "./xray_detection/train/weights/best.pt"
        if os.path.exists(trained_path):
            try:
                model = YOLO(trained_path)
                num_classes = len(model.names)
                class_names = ', '.join(list(model.names.values()))
                return f"✅ Trained model loaded: {num_classes} classes\n📋 Classes: {class_names}"
            except:
                return "❌ No model loaded. Please train or load a model first."
        return "❌ No model loaded. Please train or load a model first."
    else:
        try:
            num_classes = len(model.names)
            class_names = ', '.join(list(model.names.values()))
            
            if num_classes == 80:
                return f"⚠️ Default COCO model loaded ({num_classes} classes). For X-ray detection, please train on the X-ray dataset."
            else:
                return f"✅ Model loaded: {num_classes} classes\n📋 Classes: {class_names}"
        except:
            return "✅ Model loaded"

# Create Gradio interface
with gr.Blocks(title="X-ray Baggage Anomaly Detection", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🎯 X-ray Baggage Anomaly Detection with YOLO
    
    This application allows you to:
    1. Download and visualize the X-ray baggage dataset
    2. Analyze class distributions
    3. Train a YOLO model for object detection
    4. Run inference on new images
    
    **Note:** GPU will be automatically allocated when needed for training and inference.
    """)
    
    # Check if there's a pre-existing model
    initial_model_status = "🔍 Checking for existing models..."
    if os.path.exists("./xray_detection/train/weights/best.pt"):
        try:
            model = YOLO("./xray_detection/train/weights/best.pt")
            initial_model_status = "✅ Found previously trained model! Ready to use."
        except:
            initial_model_status = "❌ No model loaded. Please train or upload a model."
    else:
        initial_model_status = "❌ No model loaded. Please train or upload a model."
    
    gr.Markdown(f"**Model Status:** {initial_model_status}")
    
    # Add instructions for Kaggle API setup
    with gr.Accordion("📝 Setup Instructions", open=False):
        gr.Markdown("""
        ### Kaggle API Setup
        1. Get your Kaggle API credentials from https://www.kaggle.com/settings
        2. Set the KDATA_API environment variable in Hugging Face Spaces settings:
           ```
           KDATA_API={"username":"your_username","key":"your_api_key"}
           ```
           
        ### Model Persistence on Hugging Face Spaces
        - Models trained on Spaces are **temporary** and will be lost when the Space restarts
        - After training, download your model using the "📥 Download Model" button
        - Upload the downloaded model file to reuse it after Space restarts
        - No need for HuggingFace Hub or complex setups!
        """)
    
    with gr.Tab("📊 Dataset"):
        with gr.Row():
            download_btn = gr.Button("Download Dataset", variant="primary", scale=1)
            download_status = gr.Textbox(label="Status", interactive=False, scale=3)
        
        download_btn.click(download_dataset, outputs=download_status)
        
        # Dataset info section
        with gr.Row():
            dataset_info = gr.Markdown(value="Dataset not downloaded yet.")
            info_btn = gr.Button("🔄 Refresh Dataset Info", scale=0)
        
        def update_dataset_info():
            return get_dataset_info()
        
        info_btn.click(update_dataset_info, outputs=dataset_info)
        
        gr.Markdown("### Visualize Dataset Samples")
        with gr.Row():
            data_type_viz = gr.Dropdown(["train", "valid", "test"], value="train", label="Dataset Type")
            num_samples = gr.Slider(1, 8, 4, step=1, label="Number of Samples")
            viz_btn = gr.Button("Visualize Samples")
        
        viz_gallery = gr.Gallery(label="Sample Images", columns=2, height="auto")
        viz_status = gr.Textbox(label="Status", interactive=False)
        
        viz_btn.click(visualize_data, inputs=[data_type_viz, num_samples], 
                     outputs=[viz_gallery, viz_status])
        
        gr.Markdown("### Analyze Class Distribution")
        with gr.Row():
            data_type_analysis = gr.Dropdown(["train", "valid", "test"], value="train", label="Dataset Type")
            analyze_btn = gr.Button("Analyze Distribution")
        
        distribution_plot = gr.Image(label="Class Distribution", type="pil")
        analysis_status = gr.Textbox(label="Status", interactive=False)
        
        analyze_btn.click(analyze_class_distribution, inputs=data_type_analysis, 
                         outputs=[distribution_plot, analysis_status])
        
        gr.Markdown("### Visualize Dataset Samples")
        with gr.Row():
            data_type_viz = gr.Dropdown(["train", "valid", "test"], value="train", label="Dataset Type")
            num_samples = gr.Slider(1, 8, 4, step=1, label="Number of Samples")
            viz_btn = gr.Button("Visualize Samples")
        
        viz_gallery = gr.Gallery(label="Sample Images", columns=2, height="auto")
        viz_status = gr.Textbox(label="Status", interactive=False)
        
        viz_btn.click(visualize_data, inputs=[data_type_viz, num_samples], 
                     outputs=[viz_gallery, viz_status])
        
        gr.Markdown("### Analyze Class Distribution")
        with gr.Row():
            data_type_analysis = gr.Dropdown(["train", "valid", "test"], value="train", label="Dataset Type")
            analyze_btn = gr.Button("Analyze Distribution")
        
        distribution_plot = gr.Image(label="Class Distribution", type="pil")
        analysis_status = gr.Textbox(label="Status", interactive=False)
        
        analyze_btn.click(analyze_class_distribution, inputs=data_type_analysis, 
                         outputs=[distribution_plot, analysis_status])
    
    with gr.Tab("🚀 Training"):
        gr.Markdown("### Train YOLO Model")
        gr.Markdown("""
        **Note:** Training will automatically use GPU if available. This may take several minutes.
        
        **Recommended Settings for X-ray Detection:**
        - **Epochs:** 20-30 for good results
        - **Batch Size:** 2-4 for better convergence
        - **Image Size:** 640 for best quality
        - **Expected time:** ~2-5 minutes for 20 epochs
        
        ⚠️ **Important**: Models are temporary on Spaces! Download your model after training.
        """)
        
        with gr.Row():
            epochs_input = gr.Slider(1, 50, 20, step=1, label="Epochs (20+ recommended)")
            batch_size_input = gr.Slider(2, 16, 4, step=2, label="Batch Size (lower for better results)")
            img_size_input = gr.Slider(320, 640, 640, step=32, label="Image Size (640 recommended)")
            device_input = gr.Radio(["Auto", "GPU", "CPU"], value="Auto", label="Device")
        
        train_btn = gr.Button("Start Training", variant="primary")
        
        training_gallery = gr.Gallery(label="Training Results", columns=3, height="auto")
        training_status = gr.Textbox(label="Training Status", interactive=False)
        
        train_btn.click(train_model, 
                       inputs=[epochs_input, batch_size_input, img_size_input, device_input],
                       outputs=[training_gallery, training_status])
        
        gr.Markdown("### 📥 Model Management")
        
        with gr.Row():
            with gr.Column():
                gr.Markdown("#### 1️⃣ Download Trained Model")
                gr.Markdown("After training, download your model to save it permanently.")
                
                # Function to prepare model for download
                def prepare_model_download():
                    model_path = "./xray_detection/train/weights/best.pt"
                    if os.path.exists(model_path):
                        return gr.update(value=model_path, visible=True), "✅ Model ready for download!"
                    else:
                        return gr.update(value=None, visible=False), "❌ No trained model found. Please train a model first."
                
                download_btn = gr.Button("📥 Download Model (.pt)", variant="secondary")
                download_file = gr.File(label="Download Model File", visible=False)
                download_status = gr.Textbox(label="Download Status", interactive=False)
                
                download_btn.click(prepare_model_download, outputs=[download_file, download_status])
                
            with gr.Column():
                gr.Markdown("#### 2️⃣ Upload & Load Model")
                gr.Markdown("Upload a previously trained model file to continue using it.")
                
                model_upload = gr.File(
                    label="Upload Model File (.pt)",
                    file_types=[".pt"],
                    type="filepath"
                )
                load_btn = gr.Button("📤 Load Uploaded Model", variant="secondary")
                load_status = gr.Textbox(label="Load Status", interactive=False)
                
                load_btn.click(load_pretrained_model, inputs=model_upload, outputs=load_status)
                
                # Auto-load when file is uploaded
                model_upload.change(load_pretrained_model, inputs=model_upload, outputs=load_status)
    
    with gr.Tab("🔍 Inference"):
        # Model status check
        with gr.Row():
            model_status = gr.Textbox(label="Model Status", value=check_model_status(), interactive=False)
            refresh_status_btn = gr.Button("🔄 Refresh Status", scale=0)
        
        refresh_status_btn.click(check_model_status, outputs=model_status)
        
        gr.Markdown("""
        ## 🎯 모델이 객체를 감지하지 못하나요?
        
        **권장 학습 설정:**
        - **Epochs: 30** (최소 20 이상)
        - **Batch Size: 2 또는 4**
        - **Image Size: 640**
        
        **체크리스트:**
        1. ✅ X-ray 이미지인가? (일반 사진은 작동 안 함)
        2. ✅ 충분히 학습했나? (20+ epochs)
        3. ✅ Confidence threshold를 0.01로 낮춰봤나?
        4. ✅ 모델이 제대로 로드되었나? (상태 확인)
        
        **성공적인 학습 후 예상 결과:**
        - Firearm (총기류) 감지
        - Knife (칼) 감지
        - Pliers (펜치) 감지
        - Scissors (가위) 감지
        - Wrench (렌치) 감지
        """)
        
        gr.Markdown("### Single Image Inference")
        gr.Markdown("Upload an X-ray baggage image to detect prohibited items.")
        
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(type="pil", label="Upload X-ray Image")
                conf_threshold = gr.Slider(0.01, 0.9, 0.25, step=0.01, label="Confidence Threshold (낮을수록 더 많이 감지)")
                
                # Debug options
                with gr.Row():
                    inference_btn = gr.Button("Run Detection", variant="primary")
                    test_btn = gr.Button("Test with 0.01 threshold", variant="secondary", scale=0)
                
                # Add example images if dataset is available
                example_images = []
                if dataset_path and os.path.exists(f"{dataset_path}/test/images"):
                    test_images = glob(f"{dataset_path}/test/images/*")[:5]
                    example_images.extend(test_images)
                
                if example_images:
                    gr.Examples(
                        examples=[[img] for img in example_images],
                        inputs=input_image,
                        label="Example X-ray Images (Click to load)"
                    )
            
            with gr.Column():
                output_image = gr.Image(type="pil", label="Detection Result")
                detection_info = gr.Textbox(label="Detection Info", lines=8)
        
        inference_btn.click(run_inference, 
                          inputs=[input_image, conf_threshold],
                          outputs=[output_image, detection_info])
        
        # Test with very low threshold
        test_btn.click(
            lambda img: run_inference(img, 0.01),
            inputs=[input_image],
            outputs=[output_image, detection_info]
        )
        
        # Auto-refresh model status after inference
        inference_btn.click(check_model_status, outputs=model_status)
        
        gr.Markdown("### Batch Inference")
        gr.Markdown("Run detection on multiple images from the test dataset.")
        
        with gr.Row():
            batch_data_type = gr.Dropdown(["test", "valid"], value="test", label="Dataset Type")
            batch_num_images = gr.Slider(1, 10, 5, step=1, label="Number of Images")
            batch_btn = gr.Button("Run Batch Inference")
        
        batch_gallery = gr.Gallery(label="Batch Results", columns=3, height="auto")
        batch_status = gr.Textbox(label="Status", interactive=False)
        
        batch_btn.click(batch_inference,
                       inputs=[batch_data_type, batch_num_images],
                       outputs=[batch_gallery, batch_status])

    # Footer
    gr.Markdown("---")
    gr.Markdown("""
    <div style='text-align: center; font-size: 14px; color: #666;'>
        💡 <b>Quick Start:</b> Download Dataset → Train Model (20+ epochs) → Run Inference<br>
        🔍 <b>No detections?</b> Try lowering threshold to 0.01 or train for more epochs<br>
        🚀 Built with Gradio, YOLOv8, and ❤️ for X-ray security
    </div>
    """)

# Launch the app
if __name__ == "__main__":
    # Check if running on Hugging Face Spaces
    if ON_SPACES:
        demo.launch(ssr_mode=False)
    else:
        demo.launch(share=True, ssr_mode=False)