app.py

import gradio as gr
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
import torch
import pickle
from pathlib import Path
import os
import spaces

# Load model/processor
model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
model.eval()

DATASET_DIR = Path("dataset")
CACHE_FILE = "cache.pkl"

# Define supported image formats
IMAGE_EXTENSIONS = ["*.jpg", "*.jpeg", "*.png", "*.bmp", "*.gif", "*.webp", "*.tiff", "*.tif"]

def get_all_image_files():
    """Get all image files from dataset directory"""
    image_files = []
    for ext in IMAGE_EXTENSIONS:
        image_files.extend(DATASET_DIR.glob(ext))
        image_files.extend(DATASET_DIR.glob(ext.upper()))  # Also check uppercase
    return image_files

def get_embedding(image: Image.Image, device="cpu"):
    # Use CLIP's built-in preprocessing
    inputs = processor(images=image, return_tensors="pt").to(device)
    model_device = model.to(device)
    with torch.no_grad():
        emb = model_device.get_image_features(**inputs)
    # L2 normalize the embeddings
    emb = emb / emb.norm(p=2, dim=-1, keepdim=True)
    return emb

@spaces.GPU
def get_reference_embeddings():
    # Get all current image files
    current_image_files = get_all_image_files()
    current_images = set(img_path.name for img_path in current_image_files)
    
    # Load existing cache if it exists
    cached_embeddings = {}
    if os.path.exists(CACHE_FILE):
        with open(CACHE_FILE, "rb") as f:
            cached_embeddings = pickle.load(f)
    
    # Check if cache is up to date
    cached_images = set(cached_embeddings.keys())
    
    # If cache is missing images or has extra images, rebuild
    if current_images != cached_images:
        print(f"Cache outdated. Current: {len(current_images)}, Cached: {len(cached_images)}")
        embeddings = {}
        device = "cuda" if torch.cuda.is_available() else "cpu"
        
        for img_path in current_image_files:
            print(f"Processing {img_path.name}...")
            try:
                img = Image.open(img_path).convert("RGB")
                emb = get_embedding(img, device=device)
                embeddings[img_path.name] = emb.cpu()
            except Exception as e:
                print(f"Error processing {img_path.name}: {e}")
                continue
        
        # Save updated cache
        with open(CACHE_FILE, "wb") as f:
            pickle.dump(embeddings, f)
        print(f"Cache updated with {len(embeddings)} images")
        return embeddings
    else:
        print(f"Using cached embeddings for {len(cached_embeddings)} images")
        return cached_embeddings

reference_embeddings = get_reference_embeddings()

@spaces.GPU
def search_similar(query_img):
    # Refresh embeddings to catch any new images
    global reference_embeddings
    reference_embeddings = get_reference_embeddings()
    
    query_emb = get_embedding(query_img, device="cuda")
    results = []
    
    for name, ref_emb in reference_embeddings.items():
        # Move reference embedding to same device as query
        ref_emb_gpu = ref_emb.to("cuda")
        # Compute cosine similarity
        sim = torch.nn.functional.cosine_similarity(query_emb, ref_emb_gpu, dim=1).item()
        results.append((name, sim))
    
    results.sort(key=lambda x: x[1], reverse=True)
    
    # Filter out low similarity results (adjust threshold as needed)
    SIMILARITY_THRESHOLD = 0.2  # Only show results above 20% similarity
    filtered_results = [(name, score) for name, score in results if score > SIMILARITY_THRESHOLD]
    
    if not filtered_results:
        return [("No similar images found", "No matches above similarity threshold")]
    
    # Return top 5 results
    return [(f"dataset/{name}", f"Score: {score:.4f}") for name, score in filtered_results[:5]]

@spaces.GPU
def add_image(name: str, image):
    if not name.strip():
        return "Please provide a valid image name."
    
    # Save as PNG to preserve quality for all input formats
    path = DATASET_DIR / f"{name}.png"
    image.save(path, "PNG")
    
    # Use GPU for consistency if available
    device = "cuda" if torch.cuda.is_available() else "cpu"
    emb = get_embedding(image, device=device)
    
    # Add to current embeddings and save cache
    reference_embeddings[f"{name}.png"] = emb.cpu()
    
    with open(CACHE_FILE, "wb") as f:
        pickle.dump(reference_embeddings, f)
    
    return f"Image '{name}' added to dataset. Total images: {len(reference_embeddings)}"

search_interface = gr.Interface(fn=search_similar,
                                inputs=gr.Image(type="pil", label="Query Image"),
                                outputs=gr.Gallery(label="Top Matches", columns=5),
                                allow_flagging="never")

add_interface = gr.Interface(fn=add_image,
                             inputs=[gr.Text(label="Image Name"), gr.Image(type="pil", label="Product Image")],
                             outputs="text",
                             allow_flagging="never")

demo = gr.TabbedInterface([search_interface, add_interface], tab_names=["Search", "Add Product"])
demo.launch(mcp_server=True)