smlr.py

import torch
import numpy as np
import shutil
from PIL import Image
from pathlib import Path
from collections import defaultdict
from transformers import CLIPProcessor, CLIPModel
import argparse
from tqdm.auto import tqdm
from annoy import AnnoyIndex
from scipy.cluster.hierarchy import linkage, fcluster
from scipy.spatial.distance import squareform

# Main function to process images and cluster them based on conceptual similarity using CLIP embeddings.
def process_images(image_directory, clip_model, threshold, batch_size):
    image_directory = Path(image_directory)
    device = "cuda" if torch.cuda.is_available() else "cpu"

    embeddings_file = image_directory / 'embeddings.npy'
    regenerate_embeddings = check_and_load_embeddings(embeddings_file)

    # Load CLIP model and processor
    model = CLIPModel.from_pretrained(clip_model).to(device)
    processor = CLIPProcessor.from_pretrained(clip_model)
    allowed_extensions = {".jpeg", ".jpg", ".png", ".webp"}

    images_to_paths, all_image_ids = get_images_to_paths(image_directory, allowed_extensions)
    damaged_image_ids, all_embeddings = generate_embeddings(all_image_ids, images_to_paths, model, processor, device, batch_size, regenerate_embeddings, embeddings_file)

    if regenerate_embeddings:
        np.save(embeddings_file, all_embeddings)

    print("Building Annoy index...")
    annoy_index = build_annoy_index(all_embeddings)

    print("Computing distance matrix...")
    distances = compute_distance_matrix(all_embeddings, annoy_index)

    print("Applying hierarchical clustering...")
    labels = apply_clustering(distances, threshold)

    image_id_clusters = build_image_clusters(all_image_ids, labels)
    organize_images(images_to_paths, image_directory, image_id_clusters, damaged_image_ids)

# Check for existing embeddings file and load it if found, otherwise generate new embeddings
def check_and_load_embeddings(embeddings_file):
    if embeddings_file.exists():
        use_existing_embeddings = input("Embeddings file found. Do you want to use existing embeddings? (Y/N) ").strip().lower()
        if use_existing_embeddings in ('', 'y', 'yes'):
            print("Loading embeddings from file...")
            all_embeddings = np.load(embeddings_file)
            return False
    return True

# Get the paths of all images in the given directory and return the image ids and their paths
def get_images_to_paths(image_directory, allowed_extensions):
    images_to_paths = {
        image_path.stem: image_path
        for image_path in image_directory.iterdir()
        if image_path.suffix.lower() in allowed_extensions
    }
    return images_to_paths, list(images_to_paths.keys())

# Generate CLIP embeddings for all images, handling damaged images if any
def generate_embeddings(all_image_ids, images_to_paths, model, processor, device, batch_size, regenerate_embeddings, embeddings_file):
    if not regenerate_embeddings:
        return set(), np.load(embeddings_file)

    damaged_image_ids, all_embeddings = set(), []
    progress_bar = tqdm(total=len(all_image_ids), desc="Generating CLIP embeddings")

    for i in range(0, len(all_image_ids), batch_size):
        batch_image_ids, batch_images = process_image_batch(all_image_ids, i, batch_size, images_to_paths, damaged_image_ids)
        inputs = processor(images=batch_images, return_tensors="pt", padding=True).to(device)

        with torch.no_grad():
            outputs = model.get_image_features(**inputs)

        all_embeddings.extend(outputs.cpu().numpy())
        progress_bar.update(len(batch_image_ids))

    progress_bar.close()
    return damaged_image_ids, all_embeddings

# Process a batch of images, returning their ids and loaded images, while identifying damaged images
def process_image_batch(all_image_ids, start_idx, batch_size, images_to_paths, damaged_image_ids):
    batch_image_ids = all_image_ids[start_idx: start_idx + batch_size]
    batch_images = []

    for image_id in batch_image_ids:
        try:
            image = Image.open(images_to_paths[image_id])
            image.load()
            batch_images.append(image)
        except OSError:
            print(f"\nError processing image {images_to_paths[image_id]}, marking as corrupted.")
            damaged_image_ids.add(image_id)

    return batch_image_ids, batch_images

# Build an Annoy index using the generated CLIP embeddings
def build_annoy_index(all_embeddings):
    embeddings = np.array(all_embeddings)
    n_dimensions = embeddings.shape[1]

    annoy_index = AnnoyIndex(n_dimensions, "angular")
    for i, embedding in enumerate(embeddings):
        annoy_index.add_item(i, embedding)

    annoy_index.build(100)
    return annoy_index

# Compute the distance matrix of the embeddings using the Annoy index
def compute_distance_matrix(all_embeddings, annoy_index):
    n = len(all_embeddings)
    distances = []

    for i in range(n):
        for j in range(i + 1, n):
            distance = annoy_index.get_distance(i, j)
            distances.append(distance)

    return distances

# Apply hierarchical clustering on the computed distance matrix with the given threshold
def apply_clustering(distances, threshold):
    condensed_distances = np.array(distances)
    Z = linkage(condensed_distances, method='average', optimal_ordering=True)
    return fcluster(Z, t=threshold, criterion='distance')

# Build clusters of image ids based on the clustering labels
def build_image_clusters(all_image_ids, labels):
    image_id_clusters = defaultdict(set)

    for image_id, cluster_label in zip(all_image_ids, labels):
        image_id_clusters[cluster_label].add(image_id)

    return image_id_clusters

# Organize images into separate folders for clusters, unique images, and corrupted images
def organize_images(images_to_paths, image_directory, image_id_clusters, damaged_image_ids):
    for idx, image_id_cluster in enumerate(image_id_clusters.values()):
        if len(image_id_cluster) < 2:
            continue

        move_images_to_directory(image_directory, f"cluster_{idx}", image_id_cluster, images_to_paths)

    unique_image_ids = set(images_to_paths.keys()) - set(damaged_image_ids) - {image_id for cluster in image_id_clusters.values() for image_id in cluster if len(cluster) >= 2}
    move_images_to_directory(image_directory, "unique", unique_image_ids, images_to_paths)

    if damaged_image_ids:
        move_images_to_directory(image_directory, "corrupted", damaged_image_ids, images_to_paths)

# Move images to the specified folder within the image_directory
def move_images_to_directory(image_directory, folder_name, image_ids, images_to_paths):
    output_directory = image_directory / folder_name
    output_directory.mkdir(parents=True, exist_ok=True)

    for image_id in image_ids:
        source = images_to_paths[image_id]
        destination = output_directory / source.name
        shutil.move(source, destination)

def main():
    parser = argparse.ArgumentParser(description="Finding conceptually similar images using CLIP and Annoy-based hierarchical clustering.")
    parser.add_argument("--image_directory", type=str, required=True, help="Path to the directory containing the images to cluster.")
    parser.add_argument("--clip_model", type=str, default="openai/clip-vit-large-patch14-336", help="CLIP model to use for generating embeddings. Options listed from least to most demanding: openai/clip-vit-base-patch16, openai/clip-vit-base-patch32, openai/clip-vit-large-patch14, openai/clip-vit-large-patch14-336 (default)")
    parser.add_argument("--threshold", type=float, default=0.22, help="Threshold for hierarchical clustering. Lower values will reduce false positives but may miss more. (default: 0.22)")
    parser.add_argument("--batch_size", type=int, default=192, help="Batch size for generating CLIP embeddings. Higher values will require more VRAM. (default: 192)")
    args = parser.parse_args()

    process_images(args.image_directory, args.clip_model, args.threshold, args.batch_size)

if __name__ == "__main__":
    main()