app.py

import gradio as gr
import torch
from transformers import AutoProcessor, AutoModel
from PIL import Image, ImageDraw, ImageFont
import numpy as np
import random
import os
import wget
import traceback

# --- Configuration & Model Loading ---

# Device Selection with fallback
DEVICE = "cuda" if torch.cuda.is_available() else "cpu" # Simplified check
print(f"Using device: {DEVICE}")

# --- CLIP Setup ---
CLIP_MODEL_ID = "openai/clip-vit-base-patch32"
clip_processor = None
clip_model = None

def load_clip_model():
    global clip_processor, clip_model
    if clip_processor is None:
        try:
            print(f"Loading CLIP processor: {CLIP_MODEL_ID}...")
            clip_processor = AutoProcessor.from_pretrained(CLIP_MODEL_ID)
            print("CLIP processor loaded.")
        except Exception as e:
            print(f"Error loading CLIP processor: {e}")
            traceback.print_exc() # Print traceback
            return False
    if clip_model is None:
        try:
            print(f"Loading CLIP model: {CLIP_MODEL_ID}...")
            clip_model = AutoModel.from_pretrained(CLIP_MODEL_ID).to(DEVICE)
            print(f"CLIP model loaded to {DEVICE}.")
        except Exception as e:
            print(f"Error loading CLIP model: {e}")
            traceback.print_exc() # Print traceback
            return False
    return True

# --- FastSAM Setup ---
FASTSAM_CHECKPOINT = "FastSAM-s.pt"
FASTSAM_CHECKPOINT_URL = f"https://huggingface.co/CASIA-IVA-Lab/FastSAM-s/resolve/main/{FASTSAM_CHECKPOINT}"

fastsam_model = None
fastsam_lib_imported = False
FastSAM = None # Define placeholders
FastSAMPrompt = None # Define placeholders

def check_and_import_fastsam():
    global fastsam_lib_imported, FastSAM, FastSAMPrompt # Make sure globals are modified
    if not fastsam_lib_imported:
        try:
            from fastsam import FastSAM as FastSAM_lib, FastSAMPrompt as FastSAMPrompt_lib # Use temp names
            FastSAM = FastSAM_lib # Assign to global
            FastSAMPrompt = FastSAMPrompt_lib # Assign to global
            fastsam_lib_imported = True
            print("fastsam library imported successfully.")
        except ImportError as e:
            print(f"Error: 'fastsam' library not found. Please install it: pip install git+https://github.com/CASIA-IVA-Lab/FastSAM.git")
            print(f"ImportError: {e}")
            fastsam_lib_imported = False
        except Exception as e:
            print(f"Unexpected error during fastsam import: {e}")
            traceback.print_exc()
            fastsam_lib_imported = False
    return fastsam_lib_imported

def download_fastsam_weights(retries=3):
    if not os.path.exists(FASTSAM_CHECKPOINT):
        print(f"Downloading FastSAM weights: {FASTSAM_CHECKPOINT} from {FASTSAM_CHECKPOINT_URL}...")
        for attempt in range(retries):
            try:
                # Ensure the directory exists if FASTSAM_CHECKPOINT includes a path
                os.makedirs(os.path.dirname(FASTSAM_CHECKPOINT) or '.', exist_ok=True)
                wget.download(FASTSAM_CHECKPOINT_URL, FASTSAM_CHECKPOINT)
                print("FastSAM weights downloaded.")
                return True # Return True on successful download
            except Exception as e:
                print(f"Attempt {attempt + 1}/{retries} failed to download FastSAM weights: {e}")
                if os.path.exists(FASTSAM_CHECKPOINT): # Cleanup partial download
                    try:
                        os.remove(FASTSAM_CHECKPOINT)
                    except OSError:
                        pass
                if attempt + 1 == retries:
                    print("Failed to download weights after all attempts.")
                    return False
        return False # Should not be reached if loop completes, but added for clarity
    else:
        print("FastSAM weights already exist.")
        return True # Weights exist

def load_fastsam_model():
    global fastsam_model
    if fastsam_model is None:
        if not check_and_import_fastsam():
            print("Cannot load FastSAM model due to library import failure.")
            return False
        if download_fastsam_weights():
            # Ensure FastSAM class is available (might fail if import failed earlier but file exists)
            if FastSAM is None:
                 print("FastSAM class not available, check import status.")
                 return False
            try:
                print(f"Loading FastSAM model: {FASTSAM_CHECKPOINT}...")
                # Instantiate the imported class
                fastsam_model = FastSAM(FASTSAM_CHECKPOINT)
                # Move model to device *after* initialization (common practice)
                # Note: Check FastSAM docs if it needs explicit .to(DEVICE) or handles it internally
                # fastsam_model.model.to(DEVICE) # Example if needed, adjust based on FastSAM structure
                print("FastSAM model loaded.")
                return True
            except Exception as e:
                print(f"Error loading FastSAM model weights or initializing: {e}")
                traceback.print_exc()
                return False
        else:
            print("FastSAM weights not found or download failed.")
            return False
    # Model already loaded
    return True

# --- Processing Functions ---

def run_clip_zero_shot(image: Image.Image, text_labels: str):
    # Keep CLIP as is, seems less likely to be the primary issue
    if not isinstance(image, Image.Image):
         print(f"CLIP input is not a PIL Image, type: {type(image)}")
         # Try to convert if it's a numpy array (common from Gradio)
         if isinstance(image, np.ndarray):
             try:
                 image = Image.fromarray(image)
                 print("Converted numpy input to PIL Image for CLIP.")
             except Exception as e:
                 print(f"Failed to convert numpy array to PIL Image: {e}")
                 return "Error: Invalid image input format.", None
         else:
             return "Error: Please provide a valid image.", None

    if clip_model is None or clip_processor is None:
        if not load_clip_model():
            # Return None for the image part on critical error
            return "Error: CLIP Model could not be loaded.", None
    if not text_labels:
        # Return empty dict and original image if no labels
        return {}, image

    labels = [label.strip() for label in text_labels.split(',') if label.strip()]
    if not labels:
        # Return empty dict and original image if no valid labels
        return {}, image

    print(f"Running CLIP zero-shot classification with labels: {labels}")
    try:
        # Ensure image is RGB
        if image.mode != "RGB":
            print(f"Converting image from {image.mode} to RGB for CLIP.")
            image = image.convert("RGB")

        inputs = clip_processor(text=labels, images=image, return_tensors="pt", padding=True).to(DEVICE)
        with torch.no_grad():
            outputs = clip_model(**inputs)
            # Calculate probabilities
            logits_per_image = outputs.logits_per_image # B x N_labels
            probs = logits_per_image.softmax(dim=1) # Softmax over labels

        # Create confidences dictionary
        confidences = {labels[i]: float(probs[0, i].item()) for i in range(len(labels))}
        print(f"CLIP Confidences: {confidences}")
        # Return confidences and the original (potentially converted) image
        return confidences, image
    except Exception as e:
        print(f"Error during CLIP processing: {e}")
        traceback.print_exc()
        # Return error message and None for image
        return f"Error during CLIP processing: {e}", None


def run_fastsam_segmentation(image_pil: Image.Image, conf_threshold: float = 0.4, iou_threshold: float = 0.9):
    # Add input type check
    if not isinstance(image_pil, Image.Image):
         print(f"FastSAM input is not a PIL Image, type: {type(image_pil)}")
         if isinstance(image_pil, np.ndarray):
             try:
                 image_pil = Image.fromarray(image_pil)
                 print("Converted numpy input to PIL Image for FastSAM.")
             except Exception as e:
                 print(f"Failed to convert numpy array to PIL Image: {e}")
                 # Return None for image on error
                 return None, "Error: Invalid image input format." # Return tuple for consistency
         else:
             # Return None for image on error
             return None, "Error: Please provide a valid image." # Return tuple

    # Ensure model is loaded
    if not load_fastsam_model() or not fastsam_lib_imported or FastSAMPrompt is None:
        # Return None for image on critical error
        return None, "Error: FastSAM not loaded or library unavailable."

    print(f"Running FastSAM 'segment everything' with conf={conf_threshold}, iou={iou_threshold}...")
    output_image = None # Initialize output image
    status_message = "Processing..." # Initial status

    try:
        # Ensure image is RGB
        if image_pil.mode != "RGB":
            print(f"Converting image from {image_pil.mode} to RGB for FastSAM.")
            image_pil_rgb = image_pil.convert("RGB")
        else:
            image_pil_rgb = image_pil

        # Convert PIL Image to NumPy array (RGB)
        image_np_rgb = np.array(image_pil_rgb)
        print(f"Input image shape for FastSAM: {image_np_rgb.shape}")

        # Run FastSAM model
        # Make sure the arguments match what FastSAM expects
        everything_results = fastsam_model(
            image_np_rgb,
            device=DEVICE,
            retina_masks=True,
            imgsz=640, # Or another size FastSAM supports
            conf=conf_threshold,
            iou=iou_threshold,
            verbose=True # Keep verbose for debugging
        )

        # Check if results are valid before creating prompt
        if everything_results is None or not isinstance(everything_results, list) or len(everything_results) == 0:
            print("FastSAM model returned None or empty results.")
            # Return original image and status
            return image_pil, "FastSAM did not return valid results."

        # Results might be in a different format, inspect 'everything_results'
        print(f"Type of everything_results: {type(everything_results)}")
        print(f"Length of everything_results: {len(everything_results)}")
        if len(everything_results) > 0:
            print(f"Type of first element: {type(everything_results[0])}")
             # Try to access potential attributes like 'masks' if it's an object
            if hasattr(everything_results[0], 'masks') and everything_results[0].masks is not None:
                 print(f"Masks found in results object, shape: {everything_results[0].masks.data.shape}")
            else:
                 print("First result element does not have 'masks' attribute or it's None.")


        # Process results with FastSAMPrompt
        # Ensure FastSAMPrompt class is available
        if FastSAMPrompt is None:
            print("FastSAMPrompt class is not available.")
            return image_pil, "Error: FastSAMPrompt class not loaded."

        prompt_process = FastSAMPrompt(image_np_rgb, everything_results, device=DEVICE)
        ann = prompt_process.everything_prompt() # Get all annotations

        # Check annotation format - Adjust based on actual FastSAM output structure
        # Assuming 'ann' is a list and the first element is a dictionary containing masks
        masks = None
        if isinstance(ann, list) and len(ann) > 0 and isinstance(ann[0], dict) and 'masks' in ann[0]:
            mask_tensor = ann[0]['masks']
            if mask_tensor is not None and mask_tensor.numel() > 0: # Check if tensor is not None and not empty
                 masks = mask_tensor.cpu().numpy()
                 print(f"Found {len(masks)} masks with shape: {masks.shape}")
            else:
                 print("Annotation 'masks' tensor is None or empty.")
        else:
            print(f"No masks found or annotation format unexpected. ann type: {type(ann)}")
            if isinstance(ann, list) and len(ann) > 0:
                print(f"First element of ann: {ann[0]}")


        # Prepare output image (start with original)
        output_image = image_pil.copy()

        # Draw masks if found
        if masks is not None and len(masks) > 0:
            # Ensure output_image is RGBA for compositing
            overlay = Image.new('RGBA', output_image.size, (0, 0, 0, 0))
            draw = ImageDraw.Draw(overlay)

            for i, mask in enumerate(masks):
                # Ensure mask is boolean/binary before converting
                binary_mask = (mask > 0) # Use threshold 0 for binary mask from FastSAM output
                mask_uint8 = binary_mask.astype(np.uint8) * 255
                if mask_uint8.max() == 0: # Skip empty masks
                    # print(f"Skipping empty mask {i}")
                    continue

                color = (random.randint(50, 255), random.randint(50, 255), random.randint(50, 255), 180) # RGBA color
                try:
                    mask_image = Image.fromarray(mask_uint8, mode='L') # Grayscale mask
                    # Draw the mask onto the overlay
                    draw.bitmap((0, 0), mask_image, fill=color)
                except Exception as draw_err:
                    print(f"Error drawing mask {i}: {draw_err}")
                    traceback.print_exc()
                    continue # Skip this mask

            # Composite the overlay onto the image
            try:
                 output_image_rgba = output_image.convert('RGBA')
                 output_image_composited = Image.alpha_composite(output_image_rgba, overlay)
                 output_image = output_image_composited.convert('RGB') # Convert back to RGB for Gradio
                 status_message = f"Segmentation complete. Found {len(masks)} masks."
                 print("Mask drawing and compositing finished.")
            except Exception as comp_err:
                 print(f"Error during alpha compositing: {comp_err}")
                 traceback.print_exc()
                 output_image = image_pil # Fallback to original image
                 status_message = "Error during mask visualization."

        else:
            print("No masks detected or processed for 'segment everything' mode.")
            status_message = "No segments found or processed."
            output_image = image_pil # Return original image if no masks

        # Save for debugging before returning
        if output_image:
             try:
                 debug_path = "debug_fastsam_everything_output.png"
                 output_image.save(debug_path)
                 print(f"Saved debug output to {debug_path}")
             except Exception as save_err:
                 print(f"Failed to save debug image: {save_err}")

        return output_image, status_message # Return image and status message

    except Exception as e:
        print(f"Error during FastSAM 'everything' processing: {e}")
        traceback.print_exc()
        # Return original image and error message in case of failure
        return image_pil, f"Error during processing: {e}"


def run_text_prompted_segmentation(image_pil: Image.Image, text_prompts: str, conf_threshold: float = 0.4, iou_threshold: float = 0.9):
     # Add input type check
    if not isinstance(image_pil, Image.Image):
         print(f"FastSAM Text input is not a PIL Image, type: {type(image_pil)}")
         if isinstance(image_pil, np.ndarray):
             try:
                 image_pil = Image.fromarray(image_pil)
                 print("Converted numpy input to PIL Image for FastSAM Text.")
             except Exception as e:
                 print(f"Failed to convert numpy array to PIL Image: {e}")
                 return None, "Error: Invalid image input format."
         else:
             return None, "Error: Please provide a valid image."

    # Ensure model is loaded
    if not load_fastsam_model() or not fastsam_lib_imported or FastSAMPrompt is None:
        return image_pil, "Error: FastSAM Model not loaded or library unavailable." # Return original image on load fail
    if not text_prompts:
        return image_pil, "Please enter text prompts (e.g., 'person, dog')."

    prompts = [p.strip() for p in text_prompts.split(',') if p.strip()]
    if not prompts:
        return image_pil, "No valid text prompts entered."

    print(f"Running FastSAM text-prompted segmentation for: {prompts} with conf={conf_threshold}, iou={iou_threshold}")
    output_image = None
    status_message = "Processing..."

    try:
        # Ensure image is RGB
        if image_pil.mode != "RGB":
            print(f"Converting image from {image_pil.mode} to RGB for FastSAM.")
            image_pil_rgb = image_pil.convert("RGB")
        else:
            image_pil_rgb = image_pil

        image_np_rgb = np.array(image_pil_rgb)
        print(f"Input image shape for FastSAM Text: {image_np_rgb.shape}")

        # Run FastSAM once to get all potential segments
        everything_results = fastsam_model(
            image_np_rgb, device=DEVICE, retina_masks=True, imgsz=640, # Use consistent args
            conf=conf_threshold, iou=iou_threshold, verbose=True
        )

        # Check results
        if everything_results is None or not isinstance(everything_results, list) or len(everything_results) == 0:
            print("FastSAM model returned None or empty results for text prompt base.")
            return image_pil, "FastSAM did not return base results."

        # Initialize FastSAMPrompt
        if FastSAMPrompt is None:
             print("FastSAMPrompt class is not available.")
             return image_pil, "Error: FastSAMPrompt class not loaded."
        prompt_process = FastSAMPrompt(image_np_rgb, everything_results, device=DEVICE)

        all_matching_masks = []
        found_prompts_details = [] # Store details like 'prompt (count)'

        # Process each text prompt
        for text in prompts:
            print(f"  Processing prompt: '{text}'")
            # Get annotation for the specific text prompt
            ann = prompt_process.text_prompt(text=text)

            # Check annotation format and extract masks
            current_masks = None
            num_found = 0
            # Adjust check based on actual structure of 'ann' for text_prompt
            if isinstance(ann, list) and len(ann) > 0 and isinstance(ann[0], dict) and 'masks' in ann[0]:
                 mask_tensor = ann[0]['masks']
                 if mask_tensor is not None and mask_tensor.numel() > 0:
                    current_masks = mask_tensor.cpu().numpy()
                    num_found = len(current_masks)
                    print(f"    Found {num_found} mask(s) for '{text}'. Shape: {current_masks.shape}")
                    all_matching_masks.extend(current_masks) # Add found masks to the list
                 else:
                    print(f"    Annotation 'masks' tensor is None or empty for '{text}'.")
            else:
                 print(f"    No masks found or annotation format unexpected for '{text}'. ann type: {type(ann)}")
                 if isinstance(ann, list) and len(ann) > 0:
                     print(f"    First element of ann for '{text}': {ann[0]}")

            found_prompts_details.append(f"{text} ({num_found})") # Record count for status

        # Prepare output image
        output_image = image_pil.copy()
        status_message = f"Results: {', '.join(found_prompts_details)}" if found_prompts_details else "No matches found for any prompt."

        # Draw all collected masks if any were found
        if all_matching_masks:
            print(f"Total masks collected across all prompts: {len(all_matching_masks)}")
            # Stack masks if needed (optional, can draw one by one)
            # masks_np = np.stack(all_matching_masks, axis=0)
            # print(f"Total masks stacked shape: {masks_np.shape}")

            overlay = Image.new('RGBA', output_image.size, (0, 0, 0, 0))
            draw = ImageDraw.Draw(overlay)

            for i, mask in enumerate(all_matching_masks): # Iterate through collected masks
                binary_mask = (mask > 0)
                mask_uint8 = binary_mask.astype(np.uint8) * 255
                if mask_uint8.max() == 0:
                    continue # Skip empty masks

                # Assign a unique color per mask or per prompt (using random here)
                color = (random.randint(50, 255), random.randint(50, 255), random.randint(50, 255), 180)
                try:
                    mask_image = Image.fromarray(mask_uint8, mode='L')
                    draw.bitmap((0, 0), mask_image, fill=color)
                except Exception as draw_err:
                    print(f"Error drawing collected mask {i}: {draw_err}")
                    traceback.print_exc()
                    continue

            # Composite the overlay
            try:
                output_image_rgba = output_image.convert('RGBA')
                output_image_composited = Image.alpha_composite(output_image_rgba, overlay)
                output_image = output_image_composited.convert('RGB')
                print("Text prompt mask drawing and compositing finished.")
            except Exception as comp_err:
                print(f"Error during alpha compositing for text prompts: {comp_err}")
                traceback.print_exc()
                output_image = image_pil # Fallback
                status_message += " (Error during visualization)"
        else:
            print("No matching masks found for any text prompt.")
            # status_message is already set

        # Save for debugging
        if output_image:
             try:
                 debug_path = "debug_fastsam_text_output.png"
                 output_image.save(debug_path)
                 print(f"Saved debug output to {debug_path}")
             except Exception as save_err:
                 print(f"Failed to save debug image: {save_err}")

        return output_image, status_message

    except Exception as e:
        print(f"Error during FastSAM text-prompted processing: {e}")
        traceback.print_exc()
        # Return original image and error message
        return image_pil, f"Error during processing: {e}"

# --- Gradio Interface ---

print("Attempting to preload models...")
load_clip_model() # Preload CLIP
load_fastsam_model()  # Preload FastSAM
print("Preloading finished (check logs above for errors).")


# --- Gradio Interface Definition ---
# (Your Gradio Blocks code remains largely the same, but ensure the outputs match the function returns)

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# CLIP & FastSAM Demo")
    gr.Markdown("Explore Zero-Shot Classification, 'Segment Everything', and Text-Prompted Segmentation.")

    with gr.Tabs():
        # --- CLIP Tab ---
        with gr.TabItem("CLIP Zero-Shot Classification"):
            # ... (CLIP UI definition - seems ok) ...
            clip_button.click(
                run_clip_zero_shot,
                inputs=[clip_input_image, clip_text_labels],
                # Output matches: Label (dict/str), Image (PIL/None)
                outputs=[clip_output_label, clip_output_image_display]
            )
            # ... (CLIP Examples - seems ok) ...


        # --- FastSAM Everything Tab ---
        with gr.TabItem("FastSAM Segment Everything"):
            gr.Markdown("Upload an image to segment all objects/regions.")
            with gr.Row():
                with gr.Column(scale=1):
                    fastsam_input_image_all = gr.Image(type="pil", label="Input Image")
                    with gr.Row():
                        fastsam_conf_all = gr.Slider(minimum=0.1, maximum=1.0, value=0.4, step=0.05, label="Confidence Threshold")
                        fastsam_iou_all = gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="IoU Threshold")
                    fastsam_button_all = gr.Button("Run FastSAM Segmentation", variant="primary")
                with gr.Column(scale=1):
                    # Output for the image
                    fastsam_output_image_all = gr.Image(type="pil", label="Segmented Image")
                    # Add a Textbox for status messages/errors
                    fastsam_status_all = gr.Textbox(label="Status", interactive=False)

            fastsam_button_all.click(
                run_fastsam_segmentation,
                inputs=[fastsam_input_image_all, fastsam_conf_all, fastsam_iou_all],
                # Outputs: Image (PIL/None), Status (str)
                outputs=[fastsam_output_image_all, fastsam_status_all] # Updated outputs
            )
            # Update examples if needed to match new output structure (add None/str for status)
            # Note: Examples might need adjustment if they expect only image output
            gr.Examples(
                 examples=[
                     ["examples/dogs.jpg", 0.4, 0.9],
                     ["examples/fruits.jpg", 0.5, 0.8],
                     ["examples/lion.jpg", 0.45, 0.9],
                 ],
                 inputs=[fastsam_input_image_all, fastsam_conf_all, fastsam_iou_all],
                 # Need to adjust outputs for examples if function signature changed
                 # This might require a wrapper if examples expect single output
                 # For now, comment out example outputs or adjust function signature for examples
                 outputs=[fastsam_output_image_all, fastsam_status_all],
                 fn=run_fastsam_segmentation,
                 cache_examples=False, # Keep False for debugging
             )

        # --- Text-Prompted Segmentation Tab ---
        with gr.TabItem("Text-Prompted Segmentation"):
            gr.Markdown("Upload an image and provide comma-separated prompts (e.g., 'person, dog').")
            with gr.Row():
                with gr.Column(scale=1):
                    prompt_input_image = gr.Image(type="pil", label="Input Image")
                    prompt_text_input = gr.Textbox(label="Comma-Separated Text Prompts", placeholder="e.g., glasses, watch")
                    with gr.Row():
                        prompt_conf = gr.Slider(minimum=0.1, maximum=1.0, value=0.4, step=0.05, label="Confidence Threshold")
                        prompt_iou = gr.Slider(minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="IoU Threshold")
                    prompt_button = gr.Button("Segment by Text", variant="primary")
                with gr.Column(scale=1):
                    # Output Image
                    prompt_output_image = gr.Image(type="pil", label="Text-Prompted Segmentation")
                    # Status Textbox (already exists, correctly)
                    prompt_status_message = gr.Textbox(label="Status", interactive=False)

            prompt_button.click(
                run_text_prompted_segmentation,
                inputs=[prompt_input_image, prompt_text_input, prompt_conf, prompt_iou],
                # Outputs: Image (PIL/None), Status (str) - Matches function
                outputs=[prompt_output_image, prompt_status_message]
            )
            # Update examples similarly if needed
            gr.Examples(
                 examples=[
                     ["examples/dog_bike.jpg", "person, bicycle", 0.4, 0.9],
                     ["examples/astronaut.jpg", "person, helmet", 0.35, 0.9],
                     ["examples/dogs.jpg", "dog", 0.4, 0.9],
                     ["examples/fruits.jpg", "banana, apple", 0.5, 0.8],
                     ["examples/teacher.jpg", "person, glasses", 0.4, 0.9],
                 ],
                 inputs=[prompt_input_image, prompt_text_input, prompt_conf, prompt_iou],
                 outputs=[prompt_output_image, prompt_status_message],
                 fn=run_text_prompted_segmentation,
                 cache_examples=False, # Keep False for debugging
             )


    # --- Example File Download ---
    # (Download logic seems okay, ensure 'wget' is installed: pip install wget)
    if not os.path.exists("examples"):
        os.makedirs("examples")
        print("Created 'examples' directory.")
    example_files = {
        "astronaut.jpg": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/d1/Astronaut_-_St._Jean_Bay.jpg/640px-Astronaut_-_St._Jean_Bay.jpg",
        "dog_bike.jpg": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/gradio/outputs_multimodal.jpg",
        "clip_logo.png": "https://raw.githubusercontent.com/openai/CLIP/main/CLIP.png",
        "dogs.jpg": "https://raw.githubusercontent.com/ultralytics/assets/main/im/image8.jpg",
        "fruits.jpg": "https://raw.githubusercontent.com/ultralytics/assets/main/im/image9.jpg",
        "lion.jpg": "https://huggingface.co/spaces/gradio/image-segmentation/resolve/main/images/lion.jpg",
        "teacher.jpg": "https://images.pexels.com/photos/848117/pexels-photo-848117.jpeg?auto=compress&cs=tinysrgb&w=600"
    }
    def download_example_file(filename, url, retries=3):
        filepath = os.path.join("examples", filename)
        if not os.path.exists(filepath):
            print(f"Attempting to download {filename}...")
            for attempt in range(retries):
                try:
                    wget.download(url, filepath)
                    print(f"Downloaded {filename} successfully.")
                    return # Exit function on success
                except Exception as e:
                    print(f"Download attempt {attempt + 1}/{retries} for {filename} failed: {e}")
                    if os.path.exists(filepath): # Clean up partial download
                        try: os.remove(filepath)
                        except OSError: pass
                    if attempt + 1 == retries:
                        print(f"Failed to download {filename} after {retries} attempts.")
        else:
             print(f"Example file {filename} already exists.")

    # Trigger downloads
    for filename, url in example_files.items():
        download_example_file(filename, url)
    print("Example file check/download complete.")


# --- Launch App ---
if __name__ == "__main__":
    print("Launching Gradio Demo...")
    demo.launch(debug=True) # Keep debug=True