sam_inference.py

#!/usr/bin/env python3
"""
Fixed SAM 2.1 Interface - Handles negative stride issues properly
"""

import torch
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
import gradio as gr
from transformers import Sam2Model, Sam2Processor
import warnings
import io
import base64
import os
from datetime import datetime

warnings.filterwarnings("ignore")

# Global model instance to avoid reloading
MODEL = None
PROCESSOR = None
DEVICE = None

# Global state for saving
CURRENT_MASK = None
CURRENT_IMAGE_NAME = None
CURRENT_POINTS = None

def initialize_sam(model_size="small"):
    """Initialize SAM model once"""
    global MODEL, PROCESSOR, DEVICE
    
    if MODEL is None:
        DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
        print(f"Initializing SAM 2.1 {model_size} on {DEVICE}...")
        
        model_name = f"facebook/sam2-hiera-{model_size}"
        MODEL = Sam2Model.from_pretrained(model_name).to(DEVICE)
        PROCESSOR = Sam2Processor.from_pretrained(model_name)
        
        print("✓ Model loaded successfully!")
    
    return MODEL, PROCESSOR, DEVICE

def fix_image_array(image):
    """Fix image input for SAM processing - handles filepath, numpy array, or PIL Image"""
    if isinstance(image, str):
        # Handle filepath input from Gradio
        return Image.open(image).convert("RGB")
    
    elif isinstance(image, np.ndarray):
        # Make sure array is contiguous
        if not image.flags['C_CONTIGUOUS']:
            image = np.ascontiguousarray(image)
        
        # Ensure uint8 dtype
        if image.dtype != np.uint8:
            if image.max() <= 1.0:
                image = (image * 255).astype(np.uint8)
            else:
                image = image.astype(np.uint8)
        
        # Convert to PIL Image to avoid any stride issues
        return Image.fromarray(image).convert("RGB")
    
    elif isinstance(image, Image.Image):
        return image.convert("RGB")
    
    else:
        raise ValueError(f"Unsupported image type: {type(image)}")

def apply_mask_post_processing(mask, stability_threshold=0.95):
    """Apply post-processing to refine mask size and quality"""
    import cv2
    
    # Convert to binary mask
    binary_mask = (mask > 0).astype(np.uint8)
    
    # Apply morphological operations to clean up the mask
    kernel_size = max(3, int(mask.shape[0] * 0.01))  # Adaptive kernel size
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (kernel_size, kernel_size))
    
    # Close small holes
    binary_mask = cv2.morphologyEx(binary_mask, cv2.MORPH_CLOSE, kernel)
    
    # Remove small noise
    binary_mask = cv2.morphologyEx(binary_mask, cv2.MORPH_OPEN, kernel)
    
    return binary_mask.astype(np.float32)

def apply_erosion_dilation(mask, erosion_dilation_value):
    """Apply erosion or dilation to adjust mask size"""
    import cv2
    
    binary_mask = (mask > 0).astype(np.uint8)
    
    if erosion_dilation_value == 0:
        return mask
    
    kernel_size = abs(erosion_dilation_value)
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (kernel_size, kernel_size))
    
    if erosion_dilation_value > 0:
        # Dilate (make larger)
        binary_mask = cv2.dilate(binary_mask, kernel, iterations=1)
    else:
        # Erode (make smaller)
        binary_mask = cv2.erode(binary_mask, kernel, iterations=1)
    
    return binary_mask.astype(np.float32)

def save_binary_mask(mask, image_name, points, mask_threshold, erosion_dilation, save_low_res=False, custom_folder_name=None):
    """Save binary mask to organized folder structure"""
    global CURRENT_MASK, CURRENT_IMAGE_NAME, CURRENT_POINTS
    
    try:
        # Store current state for saving
        CURRENT_MASK = mask
        CURRENT_IMAGE_NAME = image_name
        CURRENT_POINTS = points
        
        # Extract image name without extension and sanitize
        if image_name:
            base_name = os.path.splitext(os.path.basename(image_name))[0]
            # Remove any path separators and special characters
            base_name = base_name.replace('/', '_').replace('\\', '_').replace(':', '_').replace(' ', '_')
        else:
            base_name = f"image_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
        
        # Choose folder tag: user-provided name if available, else 'default'
        folder_tag = None
        if custom_folder_name and str(custom_folder_name).strip():
            folder_tag = str(custom_folder_name).strip().replace(' ', '_')
        else:
            folder_tag = "default"
        
        # Create folder structure: masks/<image_base>/<folder_tag>/
        folder_name = f"masks/{base_name}/{folder_tag}"
        os.makedirs(folder_name, exist_ok=True)
        
        # Create binary mask (0 and 255 values)
        binary_mask = (mask > 0).astype(np.uint8) * 255
        
        # Calculate low resolution dimensions if requested
        original_height, original_width = binary_mask.shape
        if save_low_res:
            # Calculate sqrt-based resolution
            sqrt_factor = int(np.sqrt(max(original_width, original_height)))
            low_res_width = sqrt_factor
            low_res_height = sqrt_factor
            print(f"Original mask size: {original_width}x{original_height}")
            print(f"Low-res mask size: {low_res_width}x{low_res_height}")
        
        # Save binary mask
        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
        # Sanitize filename - replace problematic characters
        threshold_str = f"{mask_threshold:.2f}".replace('.', 'p')  # 0.30 -> 0p30
        adj_str = f"{erosion_dilation:+d}".replace('+', 'plus').replace('-', 'minus')  # +2 -> plus2, -2 -> minus2
        
        saved_paths = []
        
        # Save full resolution mask as JPEG with a simple filename
        mask_filename = "image.jpg"
        mask_path = os.path.join(folder_name, mask_filename)
        
        mask_image = Image.fromarray(binary_mask, mode='L')
        mask_image.save(mask_path, format="JPEG", quality=95, optimize=True)
        saved_paths.append(mask_path)

        # Save tensor mask (.pt) as float tensor (0.0/1.0)
        tensor_filename = "image.pt"
        tensor_path = os.path.join(folder_name, tensor_filename)
        torch.save(torch.from_numpy((mask > 0).astype(np.float32)), tensor_path)
        saved_paths.append(tensor_path)
        
        # Save low resolution mask if requested
        if save_low_res:
            # Resize mask to low resolution
            low_res_mask = mask_image.resize((low_res_width, low_res_height), Image.Resampling.NEAREST)
            
            low_res_filename = f"mask_lowres_{sqrt_factor}x{sqrt_factor}_t{threshold_str}_adj{adj_str}_{timestamp}.png"
            low_res_path = os.path.join(folder_name, low_res_filename)
            
            low_res_mask.save(low_res_path)
            saved_paths.append(low_res_path)
        
        # Also save metadata
        metadata = {
            "timestamp": timestamp,
            "points": points,
            "mask_threshold": mask_threshold,
            "erosion_dilation": erosion_dilation,
            "image_name": image_name,
            "original_resolution": f"{original_width}x{original_height}",
            "saved_paths": saved_paths,
            "low_resolution_saved": save_low_res
        }
        
        if save_low_res:
            metadata["low_resolution"] = f"{low_res_width}x{low_res_height}"
            metadata["sqrt_factor"] = sqrt_factor
        
        import json
        metadata_path = os.path.join(folder_name, f"metadata_{timestamp}.json")
        with open(metadata_path, 'w') as f:
            json.dump(metadata, f, indent=2)
        
        # Return appropriate message
        if save_low_res:
            return f"✅ Masks saved:\n📁 Full: {os.path.basename(mask_path)}\n📁 Low-res: {os.path.basename(low_res_path)}"
        else:
            return f"✅ Mask saved to: {os.path.basename(mask_path)}"
        
    except Exception as e:
        return f"❌ Save failed: {str(e)}"

def process_sam_segmentation(image, points_data, bbox_data, mode, image_name=None, top_k=3, mask_threshold=0.0, stability_score_threshold=0.95, erosion_dilation=0):
    """Main processing function with mask size controls - supports points and bounding boxes"""
    global CURRENT_MASK, CURRENT_IMAGE_NAME, CURRENT_POINTS
    
    if image is None:
        return None, None, "Please upload an image first."
    
    # Check input based on mode
    if mode == "Points":
        if not points_data or len(points_data) == 0:
            return None, None, "Please click on the image to select points."
    elif mode == "Bounding Box":
        if bbox_data is None:
            return None, None, "Please click two corners to define a bounding box."
    
    try:
        # Initialize model
        model, processor, device = initialize_sam()
        
        # Fix image
        pil_image = fix_image_array(image)
        
        # Prepare SAM inputs based on mode
        input_points = None
        input_labels = None
        input_boxes = None
        points = None
        
        if mode == "Points":
            # Extract points with positive/negative labels
            points = []
            labels = []
            for point_info in points_data:
                if isinstance(point_info, dict):
                    points.append([point_info.get("x", 0), point_info.get("y", 0)])
                    labels.append(1 if point_info.get("positive", True) else 0)  # 1 = positive, 0 = negative
                elif isinstance(point_info, (list, tuple)) and len(point_info) >= 2:
                    points.append([point_info[0], point_info[1]])
                    labels.append(1)  # Default to positive for old format
            
            if not points:
                return None, "No valid points found."
            
            print(f"Processing {len(points)} points: {points} with labels: {labels}")
            input_points = [[points]]
            input_labels = [[labels]]
            
        elif mode == "Bounding Box":
            # Use bounding box
            bbox = bbox_data  # [x1, y1, x2, y2]
            print(f"Processing bounding box: {bbox}")
            input_boxes = [[bbox]]
            # For visualization, store the bbox corners as points
            points = [[bbox[0], bbox[1]], [bbox[2], bbox[3]]]
        
        # Process with SAM
        processor_inputs = {
            "images": pil_image,
            "return_tensors": "pt"
        }
        
        # Add points or boxes based on mode
        if mode == "Points":
            processor_inputs["input_points"] = input_points
            processor_inputs["input_labels"] = input_labels
        elif mode == "Bounding Box":
            processor_inputs["input_boxes"] = input_boxes
        
        inputs = processor(**processor_inputs).to(device)
        
        # Generate masks with multiple outputs for better control
        with torch.no_grad():
            outputs = model(**inputs, multimask_output=True)
        
        # Get masks and scores
        masks = processor.post_process_masks(
            outputs.pred_masks.cpu(),
            inputs["original_sizes"]
        )[0]
        
        scores = outputs.iou_scores.cpu().numpy().flatten()
        
        # Get top-k masks
        top_indices = np.argsort(scores)[::-1][:top_k]
        
        # Apply mask threshold to control size
        best_mask = masks[0, top_indices[0]].numpy()
        best_score = scores[top_indices[0]]
        
        # Apply threshold to control mask size
        if mask_threshold > 0:
            best_mask = (best_mask > mask_threshold).astype(np.float32)
        
        # Additional mask processing for size control
        best_mask = apply_mask_post_processing(best_mask, stability_score_threshold)
        
        # Apply erosion/dilation for fine size control
        if erosion_dilation != 0:
            best_mask = apply_erosion_dilation(best_mask, erosion_dilation)
        
        # Store current state for saving
        CURRENT_MASK = best_mask
        CURRENT_IMAGE_NAME = image_name
        CURRENT_POINTS = points
        
        # Create dual visualizations
        original_with_input = create_original_with_input_visualization(pil_image, points, bbox_data, mode)
        mask_result = create_mask_visualization(pil_image, best_mask, best_score, mask_threshold)
        
        status = f"✓ Generated mask with score: {float(best_score):.3f}\n🔄 Ready to save!"
        return original_with_input, mask_result, status
        
    except Exception as e:
        print(f"Error in processing: {e}")
        return None, None, f"Error: {str(e)}"

def create_original_with_input_visualization(pil_image, points, bbox, mode, negative_points=None):
    """Create visualization of original image with input points/bbox overlay"""
    # Convert PIL to numpy for matplotlib
    img_array = np.array(pil_image)
    
    fig, ax = plt.subplots(1, 1, figsize=(8, 6))
    
    # Show original image only
    ax.imshow(img_array)
    
    # Show input visualization based on mode
    if mode == "Points":
        total_points = 0
        # Show positive points (green)
        if points:
            for point in points:
                ax.plot(point[0], point[1], 'go', markersize=12, markeredgewidth=3, markerfacecolor='lime')
            total_points += len(points)
        
        # Show negative points (red)
        if negative_points:
            for point in negative_points:
                ax.plot(point[0], point[1], 'ro', markersize=12, markeredgewidth=3, markerfacecolor='red')
            total_points += len(negative_points)
            
        pos_count = len(points) if points else 0
        neg_count = len(negative_points) if negative_points else 0
        title_suffix = f"Points: {pos_count}+ {neg_count}-" if neg_count > 0 else f"Points: {pos_count}"
    elif mode == "Bounding Box" and bbox:
        # Show bounding box
        x1, y1, x2, y2 = bbox
        width = x2 - x1
        height = y2 - y1
        
        # Draw bounding box rectangle
        from matplotlib.patches import Rectangle
        rect = Rectangle((x1, y1), width, height, linewidth=3, edgecolor='lime', facecolor='none')
        ax.add_patch(rect)
        
        # Show corner points
        ax.plot([x1, x2], [y1, y2], 'go', markersize=8, markeredgewidth=2, markerfacecolor='lime')
        title_suffix = f"BBox: {int(width)}×{int(height)}"
    else:
        title_suffix = "No input"
    
    ax.set_title(f"Input Selection ({title_suffix})", fontsize=14)
    ax.axis('off')
    
    # Convert to numpy array
    fig.canvas.draw()
    buf = fig.canvas.buffer_rgba()
    result_array = np.asarray(buf)
    # Convert RGBA to RGB
    result_array = result_array[:, :, :3]
    
    plt.close(fig)
    return result_array

def create_mask_visualization(pil_image, mask, score, mask_threshold=0.0):
    """Create clean mask visualization without input overlays"""
    # Convert PIL to numpy for matplotlib
    img_array = np.array(pil_image)
    
    fig, ax = plt.subplots(1, 1, figsize=(8, 6))
    
    # Show original image
    ax.imshow(img_array)
    
    # Overlay mask in red
    mask_overlay = np.zeros((*mask.shape, 4))
    mask_overlay[mask > 0] = [1, 0, 0, 0.6]  # Red with transparency
    ax.imshow(mask_overlay)
    
    ax.set_title(f"Generated Mask (Score: {float(score):.3f}, Threshold: {mask_threshold:.2f})", fontsize=14)
    ax.axis('off')
    
    # Convert to numpy array
    fig.canvas.draw()
    buf = fig.canvas.buffer_rgba()
    result_array = np.asarray(buf)
    # Convert RGBA to RGB
    result_array = result_array[:, :, :3]
    
    plt.close(fig)
    return result_array

def create_interface():
    """Create a simplified single-image annotator interface."""
    
    with gr.Blocks(title="SAM 2.1 - Simple Annotator", theme=gr.themes.Soft(), css="""
        .negative-mode-checkbox label {
            color: #d00000 !important;
            font-weight: 800 !important;
            font-size: 16px !important;
        }
        """) as interface:
        gr.HTML("""
        <div style="text-align: center;">
            <h1>🎯 SAM 2.1 Simple Annotator</h1>
            <p>Upload one image, click to add positive/negative points, generate mask, and save.</p>
        </div>
        """)

        # Image input (single image) - directly annotate; this serves as uploader too
        # Users can upload by clicking the annotatable image component below.
        image_input = gr.Image(
            label=None,
            type="filepath",
            height=0,
            visible=False
        )

        # Main layout: Selected Points on the left, annotatable image in the center, preview on the right
        with gr.Row():
            with gr.Column(scale=1):
                points_display = gr.JSON(label="📍 Selected Points", value=[], visible=True)
            with gr.Column(scale=3):
                # Negative mode toggle with clear red styling
                negative_point_mode = gr.Checkbox(
                    label="➖ NEGATIVE POINT MODE",
                    value=False,
                    info="🔴 Enable to add negative points (shown in red)",
                    interactive=True,
                    elem_classes="negative-mode-checkbox"
                )
                original_with_input = gr.Image(
                    label="📍 Click to Annotate (toggle negative mode to exclude)",
                    height=640,
                    interactive=True
                )
            with gr.Column(scale=1):
                points_overlay = gr.Image(label="📍 Points Preview (green=positive, red=negative)", height=720, interactive=False)

        # Action buttons
        with gr.Row():
            generate_btn = gr.Button("🎯 Generate Mask", variant="primary", size="lg")
            clear_btn = gr.Button("🗑️ Clear Points", variant="secondary", size="lg")

        # Mask result under buttons
        with gr.Row():
            mask_result = gr.Image(label="🎭 Generated Mask", height=512)

        # Save controls under mask
        with gr.Row():
            mask_name_input = gr.Textbox(label="Folder name (optional)", placeholder="e.g., michael_phelps_bottom_left")
            save_btn = gr.Button("💾 Save Mask", variant="stop", size="lg")

        # Status
        with gr.Row():
            status_text = gr.Textbox(label="📊 Status", interactive=False, lines=3)

        # State to store points only
        points_state = gr.State([])

        # Event handlers
        def on_image_click(image, current_points, negative_mode, evt: gr.SelectData):
            """Handle clicks on the image for point annotations only."""
            if evt.index is not None and image is not None:
                x, y = evt.index
                try:
                    pil_image = fix_image_array(image)
                    is_negative = negative_mode
                    new_point = {"x": int(x), "y": int(y), "positive": not is_negative}
                    updated_points = current_points + [new_point]

                    positive_points = [[p["x"], p["y"]] for p in updated_points if p.get("positive", True)]
                    negative_points = [[p["x"], p["y"]] for p in updated_points if not p.get("positive", True)]

                    updated_visualization = create_original_with_input_visualization(
                        pil_image, positive_points, None, "Points", negative_points
                    )

                    point_type = "positive" if not is_negative else "negative"
                    pos_count = len(positive_points)
                    neg_count = len(negative_points)
                    return updated_points, updated_points, updated_visualization, (
                        f"Added {point_type} point at ({x}, {y}). Total: {pos_count} positive, {neg_count} negative points."
                    )
                except Exception as e:
                    print(f"Error in visualization: {e}")
                    return current_points, current_points, None, f"Error updating visualization: {str(e)}"
            return current_points, current_points, None, "Click on the image to add points."

        def on_image_upload(image):
            """Handle image upload and show it for annotation."""
            if image is not None:
                try:
                    pil_image = fix_image_array(image)
                    img_array = np.array(pil_image)
                    # Populate both the annotation image (left) and the points preview (right)
                    return img_array, img_array, [], [], "Image uploaded. Click on the left image to add points (enable negative mode for exclusion)."
                except Exception as e:
                    return None, None, [], [], f"Error loading image: {str(e)}"
            return None, None, [], [], "No image uploaded."

        def clear_all_points(image):
            """Clear points and keep the image visible for annotation."""
            try:
                if image is not None:
                    pil_image = fix_image_array(image)
                    img_array = np.array(pil_image)
                    return [], [], img_array, img_array, None, "All points cleared. You can continue annotating."
            except Exception:
                pass
            return [], [], None, None, None, "All points cleared."

        def generate_segmentation(image, points):
            """Generate a single segmentation mask using points only."""
            # Determine image name
            if isinstance(image, str):
                image_name = os.path.basename(image)
            else:
                # Prefer an explicit friendly default if metadata lacks a good name
                image_name = None
                if hasattr(image, 'orig_name'):
                    image_name = image.orig_name
                elif isinstance(image, dict) and 'orig_name' in image:
                    image_name = image['orig_name']
                elif hasattr(image, 'name'):
                    image_name = image.name
                if not image_name or 'tmp' in str(image_name).lower() or 'uploaded_image' in str(image_name).lower():
                    image_name = "michael_phelps_bottom_left.jpg"

            # Run segmentation (points mode)
            _, mask_img, status = process_sam_segmentation(
                image, points, None, "Points", image_name, 1, 0.0, 0.95, 0
            )
            if mask_img is not None:
                status += f"\n📁 Image: {os.path.basename(image_name)}"
            return mask_img, status

        def save_current_mask(custom_folder_name):
            """Save the currently generated mask."""
            global CURRENT_MASK, CURRENT_IMAGE_NAME, CURRENT_POINTS
            if CURRENT_MASK is None:
                return "❌ No mask to save. Generate a mask first."
            if CURRENT_POINTS is None:
                return "❌ No points available. Generate a mask first."
            return save_binary_mask(CURRENT_MASK, CURRENT_IMAGE_NAME, CURRENT_POINTS, 0.0, 0, False, custom_folder_name=(custom_folder_name or None))

        # Wire events
        # Let the annotatable image also handle image uploads (drag & drop / click upload)
        original_with_input.upload(
            on_image_upload,
            inputs=[original_with_input],
            outputs=[original_with_input, points_overlay, points_state, points_display, status_text]
        )

        original_with_input.select(
            on_image_click,
            inputs=[original_with_input, points_state, negative_point_mode],
            outputs=[points_state, points_display, points_overlay, status_text]
        )

        generate_btn.click(
            generate_segmentation,
            inputs=[original_with_input, points_state],
            outputs=[mask_result, status_text]
        )

        clear_btn.click(
            clear_all_points,
            inputs=[original_with_input],
            outputs=[points_state, points_display, points_overlay, original_with_input, mask_result, status_text]
        )

        save_btn.click(
            save_current_mask,
            inputs=[mask_name_input],
            outputs=[status_text]
        )
    
    return interface

def main():
    """Main function"""
    print("🚀 Starting Fixed SAM 2.1 Interface...")
    
    interface = create_interface()
    
    print("🌐 Launching web interface...")
    print("📍 Click on objects in images to segment them!")
    
    interface.launch(
        server_port=int(os.environ.get("GRADIO_SERVER_PORT", 7860)),
        share=False,
        inbrowser=False,  # Don't auto-open browser in server environment
        show_error=True
    )

if __name__ == "__main__":
    main()