app.py

import gradio as gr
import torch
from PIL import Image
from torchvision import transforms
from transformers import AutoModelForImageSegmentation
import json
import logging
import os
from pathlib import Path
import sys
from datetime import datetime

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# ========== KONFIGURASI ==========
MODEL_ID = "mohantesting/remove_background"
MODEL_PATH = "./models/remove_background"
MAX_IMAGE_SIZE = 2048
PROCESSING_SIZE = (1024, 1024)
# ==================================

model = None
device = None
transform_image = None
stats = {
    "total_processed": 0,
    "total_errors": 0,
    "start_time": datetime.now()
}

def check_model_exists(path):
    """Cek apakah model sudah ada"""
    if not os.path.exists(path):
        return False
    
    required_files = ["config.json"]
    for file in required_files:
        if not os.path.exists(os.path.join(path, file)):
            return False
    
    has_weights = False
    for root, dirs, files in os.walk(path):
        for file in files:
            if file.endswith((".bin", ".safetensors")):
                has_weights = True
                break
        if has_weights:
            break
    
    return has_weights

def get_folder_size(folder_path):
    """Hitung total ukuran folder"""
    total_size = 0
    for dirpath, dirnames, filenames in os.walk(folder_path):
        for filename in filenames:
            filepath = os.path.join(dirpath, filename)
            if os.path.isfile(filepath):
                total_size += os.path.getsize(filepath)
    return total_size

def download_model():
    """Download model jika belum ada"""
    logger.info("="*60)
    logger.info("CHECKING BACKGROUND REMOVAL MODEL...")
    logger.info("="*60)
    
    if check_model_exists(MODEL_PATH):
        logger.info("✓ Model sudah ada di local!")
        logger.info(f"✓ Location: {MODEL_PATH}")
        
        size_bytes = get_folder_size(MODEL_PATH)
        size_mb = size_bytes / (1024 * 1024)
        logger.info(f"✓ Size: {size_mb:.2f} MB")
        logger.info("✓ Skipping download...\n")
        return True
    
    logger.info("✗ Model tidak ditemukan. Mulai download...")
    logger.info(f"Model ID: {MODEL_ID}")
    logger.info(f"Save to: {MODEL_PATH}")
    logger.info("-" * 60)
    
    try:
        os.makedirs(MODEL_PATH, exist_ok=True)
        
        logger.info("Downloading background removal model...")
        # Download langsung tanpa save - kita akan load langsung dari HF
        # karena model ini menggunakan custom code (BiRefNet)
        
        logger.info("✓ Model akan di-load langsung dari HuggingFace\n")
        return True
        
    except Exception as e:
        logger.error(f"✗ Error: {str(e)}")
        import traceback
        traceback.print_exc()
        return False

def load_model():
    """Load model ke memory"""
    global model, device, transform_image
    
    logger.info("="*60)
    logger.info("LOADING MODEL INTO MEMORY...")
    logger.info("="*60)
    
    try:
        device = 'cuda' if torch.cuda.is_available() else 'cpu'
        logger.info(f"Device: {device}")
        
        # Coba load dari local dulu, kalau gagal load dari HuggingFace
        try:
            if check_model_exists(MODEL_PATH):
                logger.info("Attempting to load from local...")
                # Add local path to sys.path for custom modules
                if MODEL_PATH not in sys.path:
                    sys.path.insert(0, MODEL_PATH)
                
                model = AutoModelForImageSegmentation.from_pretrained(
                    MODEL_PATH,
                    trust_remote_code=True,
                    local_files_only=True
                )
                logger.info("✓ Loaded from local cache")
            else:
                raise FileNotFoundError("Local model not found")
                
        except Exception as e:
            logger.info(f"Local load failed: {str(e)}")
            logger.info("Loading from HuggingFace Hub...")
            model = AutoModelForImageSegmentation.from_pretrained(
                MODEL_ID,
                trust_remote_code=True
            )
            logger.info("✓ Loaded from HuggingFace Hub")
            
            # Save untuk next time
            try:
                logger.info("Saving model to local cache...")
                model.save_pretrained(MODEL_PATH)
                logger.info(f"✓ Model saved to {MODEL_PATH}")
            except Exception as save_err:
                logger.warning(f"Could not save model: {save_err}")
        
        model.eval().to(device)
        
        # Setup transform
        transform_image = transforms.Compose([
            transforms.Resize(PROCESSING_SIZE),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])
        
        logger.info("="*60)
        logger.info("✓ MODEL READY!")
        logger.info(f"  Model: {MODEL_ID}")
        logger.info(f"  Device: {device}")
        logger.info(f"  Processing Size: {PROCESSING_SIZE}")
        logger.info("="*60 + "\n")
        
        return True
        
    except Exception as e:
        logger.error(f"✗ Failed to load model: {str(e)}")
        import traceback
        traceback.print_exc()
        return False

# ========== STARTUP SEQUENCE ==========
logger.info("\n" + "="*60)
logger.info(f"  APPLICATION STARTUP - {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
logger.info("="*60 + "\n")

if not download_model():
    raise Exception("Failed to prepare model")

if not load_model():
    raise Exception("Failed to load model into memory")

# ========================================

def remove_background(input_image):
    """Remove background dari image"""
    
    try:
        if model is None or transform_image is None:
            return None, None, json.dumps({
                "success": False,
                "error": "Model belum siap"
            }, indent=2, ensure_ascii=False)
        
        if input_image is None:
            return None, None, json.dumps({
                "success": False,
                "error": "Image tidak boleh kosong"
            }, indent=2, ensure_ascii=False)
        
        # Convert to PIL Image
        if not isinstance(input_image, Image.Image):
            input_image = Image.fromarray(input_image).convert("RGB")
        else:
            input_image = input_image.convert("RGB")
        
        original_size = input_image.size
        logger.info(f"Processing image... Size: {original_size[0]}x{original_size[1]}")
        
        # Check if image is too large
        max_dim = max(original_size)
        if max_dim > MAX_IMAGE_SIZE:
            scale = MAX_IMAGE_SIZE / max_dim
            new_size = (int(original_size[0] * scale), int(original_size[1] * scale))
            logger.info(f"Resizing large image to {new_size[0]}x{new_size[1]}")
            input_image = input_image.resize(new_size, Image.Resampling.LANCZOS)
        
        # Transform image
        input_tensor = transform_image(input_image).unsqueeze(0).to(device)
        
        # Prediction
        with torch.no_grad():
            preds = model(input_tensor)[-1].sigmoid().cpu()
        
        pred = preds[0].squeeze()
        pred_pil = transforms.ToPILImage()(pred)
        mask = pred_pil.resize(input_image.size, Image.Resampling.LANCZOS)
        
        # Create output with alpha channel
        output_image = input_image.copy()
        output_image.putalpha(mask)
        
        # Update stats
        stats["total_processed"] += 1
        
        logger.info(f"✓ Background removed. Output: {output_image.width}x{output_image.height}")
        
        # JSON result
        result = {
            "success": True,
            "input_size": f"{input_image.width}x{input_image.height}",
            "output_size": f"{output_image.width}x{output_image.height}",
            "output_format": "PNG with alpha channel",
            "model": MODEL_ID,
            "device": device,
            "processing_time": "~1-3 seconds"
        }
        
        return output_image, mask, json.dumps(result, indent=2, ensure_ascii=False)
    
    except Exception as e:
        stats["total_errors"] += 1
        logger.error(f"Error removing background: {str(e)}", exc_info=True)
        return None, None, json.dumps({
            "success": False,
            "error": str(e)
        }, indent=2, ensure_ascii=False)

def get_model_info():
    """Return model info sebagai JSON"""
    try:
        uptime = datetime.now() - stats["start_time"]
        
        info = {
            "model_name": "Background Removal Model (BiRefNet)",
            "model_id": MODEL_ID,
            "model_path": MODEL_PATH,
            "model_type": "Image Segmentation",
            "architecture": "BiRefNet (Bilateral Reference Network)",
            "device": device if device else "unknown",
            "model_loaded": model is not None,
            "processing_size": f"{PROCESSING_SIZE[0]}x{PROCESSING_SIZE[1]}",
            "max_input_size": f"{MAX_IMAGE_SIZE}x{MAX_IMAGE_SIZE}",
            "output_format": "PNG with transparency (alpha channel)",
            "statistics": {
                "total_processed": stats["total_processed"],
                "total_errors": stats["total_errors"],
                "uptime": str(uptime).split('.')[0],
                "success_rate": f"{((stats['total_processed'] - stats['total_errors']) / max(stats['total_processed'], 1) * 100):.1f}%"
            },
            "capabilities": [
                "Automatic background removal",
                "High-quality segmentation",
                "Preserve original image resolution",
                "Generate alpha mask",
                "Handles complex backgrounds"
            ],
            "use_cases": [
                "Product photography",
                "Portrait editing",
                "E-commerce images",
                "Graphic design",
                "Social media content",
                "Profile pictures"
            ],
            "technical_details": {
                "framework": "PyTorch + Transformers",
                "trust_remote_code": True,
                "normalization": "ImageNet stats",
                "model_size": "~840MB"
            }
        }
        
        return json.dumps(info, indent=2, ensure_ascii=False)
    except Exception as e:
        return json.dumps({"error": str(e)}, indent=2, ensure_ascii=False)

# Custom CSS
custom_css = """
#output_json {
    font-family: 'Courier New', monospace;
    font-size: 14px;
}
.gradio-container {
    max-width: 1600px !important;
}
.tab-nav button {
    font-size: 16px;
    font-weight: 500;
}
"""

# Gradio Interface
with gr.Blocks(css=custom_css, theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🎨 Background Removal API
    ### AI-powered automatic background removal using BiRefNet
    Remove backgrounds from images with high-quality segmentation
    """)
    
    with gr.Tabs():
        # Tab Background Removal
        with gr.Tab("✂️ Remove Background"):
            with gr.Row():
                with gr.Column(scale=1):
                    input_image = gr.Image(
                        label="📸 Input Image",
                        type="pil",
                        height=450
                    )
                    
                    with gr.Row():
                        remove_btn = gr.Button(
                            "✂️ Remove Background",
                            variant="primary",
                            size="lg",
                            scale=2
                        )
                        clear_btn = gr.ClearButton(
                            components=[input_image],
                            value="🗑️ Clear",
                            size="lg",
                            scale=1
                        )
                
                with gr.Column(scale=1):
                    output_image = gr.Image(
                        label="🖼️ Output (No Background)",
                        type="pil",
                        height=450
                    )
                    
                    output_mask = gr.Image(
                        label="🎭 Alpha Mask",
                        type="pil",
                        height=200
                    )
            
            output_json = gr.Code(
                label="📄 JSON Response",
                language="json",
                lines=10,
                elem_id="output_json"
            )
            
            gr.Markdown("""
            ### 💡 Tips for Best Results
            - Use images with **clear subject-background separation**
            - **Good lighting** improves accuracy
            - **Higher resolution** = better edge quality
            - Images are automatically resized if too large (max 2048px)
            - Save as **PNG** to preserve transparency
            """)
            
            remove_btn.click(
                fn=remove_background,
                inputs=[input_image],
                outputs=[output_image, output_mask, output_json]
            )
        
        # Tab Model Info
        with gr.Tab("ℹ️ Model Info"):
            model_info_output = gr.Code(
                label="Model Information & Statistics",
                language="json",
                lines=35
            )
            info_btn = gr.Button("🔍 Get Model Info & Stats", variant="secondary", size="lg")
            
            gr.Markdown("""
            ### About BiRefNet
            
            **BiRefNet** (Bilateral Reference Network) is a state-of-the-art image segmentation model
            specifically designed for high-quality background removal. It uses bilateral reference 
            mechanisms to achieve precise object segmentation with clean edges.
            
            **Key Features:**
            - Advanced bilateral architecture for precise segmentation
            - Handles complex backgrounds and fine details
            - Preserves hair, fur, and transparent objects
            - Production-ready quality
            """)
            
            info_btn.click(
                fn=get_model_info,
                inputs=[],
                outputs=model_info_output
            )
        
        # Tab API Documentation
        with gr.Tab("📚 API Usage"):
            gr.Markdown("""
            ## 🚀 API Usage Guide
            
            ### 1. Python Example with Requests
            ```python
            import requests
            import base64
            from PIL import Image
            from io import BytesIO
            import json
            
            # Load and encode image
            with open("input.jpg", "rb") as f:
                img_data = base64.b64encode(f.read()).decode()
            
            # API endpoint
            url = "https://YOUR-SPACE-URL/api/predict"
            
            payload = {
                "data": [f"data:image/jpeg;base64,{img_data}"]
            }
            
            # Make request
            response = requests.post(url, json=payload)
            result = response.json()
            
            # Get output image (PNG with transparency)
            output_image_data = result['data'][0]
            output_json = json.loads(result['data'][2])
            
            # Decode and save
            img_bytes = base64.b64decode(output_image_data.split(',')[1])
            img = Image.open(BytesIO(img_bytes))
            img.save('output_no_bg.png')
            
            print(json.dumps(output_json, indent=2))
            ```
            
            ### 2. Using Gradio Client
            ```python
            from gradio_client import Client
            from PIL import Image
            
            client = Client("YOUR-SPACE-URL")
            
            # Process image
            result = client.predict(
                input_image="path/to/image.jpg",
                api_name="/predict"
            )
            
            # result contains: [output_image, mask, json_response]
            output_path, mask_path, json_data = result
            
            # Load and use
            output = Image.open(output_path)
            output.save("no_background.png")
            ```
            
            ### 3. Response Format
            ```json
            {
              "success": true,
              "input_size": "1200x1600",
              "output_size": "1200x1600",
              "output_format": "PNG with alpha channel",
              "model": "mohantesting/remove_background",
              "device": "cuda",
              "processing_time": "~1-3 seconds"
            }
            ```
            
            ### 4. Batch Processing Script
            ```python
            import os
            from pathlib import Path
            from gradio_client import Client
            from PIL import Image
            
            client = Client("YOUR-SPACE-URL")
            
            input_dir = 'input_images'
            output_dir = 'output_images'
            os.makedirs(output_dir, exist_ok=True)
            
            for img_file in Path(input_dir).glob('*.jpg'):
                print(f"Processing: {img_file.name}")
                
                result = client.predict(
                    input_image=str(img_file),
                    api_name="/predict"
                )
                
                output_path = result[0]
                img = Image.open(output_path)
                
                save_path = Path(output_dir) / f"{img_file.stem}_no_bg.png"
                img.save(save_path)
                
                print(f"✓ Saved: {save_path}")
            ```
            
            ### 5. Output Format Details
            
            **Image Format:**
            - Format: PNG with full alpha transparency
            - Resolution: Same as input (up to 2048x2048)
            - Background: Completely transparent (alpha = 0)
            - Foreground: Fully preserved with smooth edges
            
            **Alpha Mask:**
            - Grayscale image showing segmentation confidence
            - White (255) = foreground
            - Black (0) = background
            - Gray values = edge transitions
            
            ### 6. Best Practices
            
            ✅ **DO:**
            - Use high-resolution images (1000px+ recommended)
            - Ensure good contrast between subject and background
            - Use well-lit, sharp images
            - Save output as PNG to preserve transparency
            - Test with sample images first
            
            ❌ **DON'T:**
            - Don't use extremely large images (>4K) - they'll be auto-resized
            - Don't expect perfect results on very complex backgrounds
            - Don't save as JPEG (loses transparency!)
            - Don't use blurry or low-quality input images
            
            ### 7. Common Use Cases
            
            **E-Commerce Product Photos:**
            ```python
            # Remove background for clean product shots
            result = remove_background('product.jpg')
            result.save('product_transparent.png')
            # Upload to Shopify, Amazon, etc.
            ```
            
            **Portrait Photography:**
            ```python
            # Create professional headshots
            result = remove_background('portrait.jpg')
            # Composite on professional backgrounds
            ```
            
            **Social Media Content:**
            ```python
            # Create stickers, cutouts, graphics
            result = remove_background('subject.jpg')
            # Use in Instagram, TikTok, YouTube thumbnails
            ```
            
            **Graphic Design:**
            ```python
            # Create design elements
            result = remove_background('object.jpg')
            # Import into Photoshop, Illustrator, Canva
            ```
            
            ### 8. Performance Metrics
            
            - **Processing Time**: 1-3 seconds per image (GPU) / 5-10 seconds (CPU)
            - **Max Resolution**: 2048x2048 (auto-resized if larger)
            - **Model Size**: ~840MB
            - **GPU Memory**: ~2GB recommended
            - **Accuracy**: High-quality segmentation with clean edges
            
            ### 9. Error Handling
            
            ```python
            try:
                result = client.predict(input_image="image.jpg")
                output_data = json.loads(result[2])
                
                if output_data['success']:
                    print("Success!")
                else:
                    print(f"Error: {output_data['error']}")
                    
            except Exception as e:
                print(f"Request failed: {e}")
            ```
            
            ### 10. Rate Limits & Quotas
            
            - No built-in rate limits (depends on hosting)
            - For HuggingFace Spaces: Check your space tier
            - For self-hosted: Limited by GPU/CPU resources
            - Recommended: Process images sequentially for stability
            
            ---
            
            **Model:** mohantesting/remove_background (BiRefNet)  
            **Framework:** PyTorch + Transformers + Gradio  
            **License:** Check model repository for licensing details
            """)
    
    gr.Markdown("""
    ---
    <div style="text-align: center; padding: 20px; background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); border-radius: 10px; color: white;">
        <h3 style="margin: 0; color: white;">🚀 Ready to integrate background removal into your app?</h3>
        <p style="margin: 10px 0 0 0; opacity: 0.9;">Use the API documentation above to get started!</p>
    </div>
    """)

# Launch
if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )