generate_histogram.py

import os
import torch
import torch.nn.functional as F
import pickle
from tqdm import tqdm
import random
import numpy as np
import matplotlib.pyplot as plt

# ===============================================================
# --- CONFIGURATION: SET YOUR ANALYSIS PARAMETERS HERE ---
# ===============================================================

# --- Task: Analyze TEXT Embeddings ---
# Set POOL_MODE to True because text embeddings are variable-length.
INPUT_FILE_PATH    = "/media/RTCIN7TBDriveA/Interns/RDT2/gte3kor/complex_image_search-main/results/coreset/embeddings/text_embeddings.pkl"
OUTPUT_PLOT_PATH   = "text_distance_histogram.png"
POOL_MODE          = True  # <<< IMPORTANT: True for Text
PLOT_TITLE         = "Distance Distribution of Text Embeddings"

# --- Task: Analyze IMAGE Embeddings ---
# Uncomment the block below and comment out the Text block above.
# Set POOL_MODE to False because image embeddings are already fixed-size.
# INPUT_FILE_PATH    = "/media/RTCIN7TBDriveA/Interns/RDT2/gte3kor/complex_image_search-main/results/coreset/embeddings/image_embeddings.pkl"
# OUTPUT_PLOT_PATH   = "image_distance_histogram.png"
# POOL_MODE          = False # <<< IMPORTANT: False for Images
# PLOT_TITLE         = "Distance Distribution of Image Embeddings"


# --- General Settings ---
SAMPLE_SIZE = 2000      # Number of embeddings to sample for the analysis.
DEVICE      = 'cuda:2'    # Device to use for computation ('cuda' or 'cpu').

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


def load_embeddings(file_path):
    """Loads a pickle file and returns its content."""
    print(f"Loading embeddings from '{file_path}'...")
    if not os.path.exists(file_path):
        print(f"Error: File not found at '{file_path}'")
        return None
    with open(file_path, 'rb') as f:
        return pickle.load(f)

def main():
    """Main function to perform embedding analysis."""
    device = torch.device(DEVICE if torch.cuda.is_available() and DEVICE == 'cuda' else 'cpu')
    
    print("=" * 60)
    print("--- Running Embedding Distance Diagnosis ---")
    print(f"  Input File:     {INPUT_FILE_PATH}")
    print(f"  Output Plot:    {OUTPUT_PLOT_PATH}")
    print(f"  Pooling Mode:   {'Enabled' if POOL_MODE else 'Disabled'}")
    print(f"  Sample Size:    {SAMPLE_SIZE}")
    print(f"  Device:         {device}")
    print("=" * 60)

    # 1. Load the embeddings
    original_embeddings_dict = load_embeddings(INPUT_FILE_PATH)
    if not original_embeddings_dict:
        print("No embeddings loaded. Exiting.")
        return

    # 2. Pre-process embeddings based on the POOL_MODE
    print("Pre-processing embeddings...")
    processed_embeddings = {}
    for key, emb in tqdm(original_embeddings_dict.items(), desc="Processing Embeddings"):
        if isinstance(emb, np.ndarray):
            emb_tensor = torch.from_numpy(emb.flatten()).float()
        else:
            emb_tensor = emb.float()

        # Conditional Pooling
        if POOL_MODE and emb_tensor.dim() > 1:
            emb_tensor = emb_tensor.mean(dim=0)
        
        # Always L2-normalize for fair comparison
        processed_embeddings[key] = F.normalize(emb_tensor, p=2, dim=0)

    # 3. Take a random sample
    if len(processed_embeddings) < SAMPLE_SIZE:
        print(f"Warning: Dataset size ({len(processed_embeddings)}) is smaller than sample size. Using all data.")
        keys_to_sample = list(processed_embeddings.keys())
    else:
        keys_to_sample = random.sample(list(processed_embeddings.keys()), SAMPLE_SIZE)
    
    sample_tensor = torch.stack([processed_embeddings[key] for key in keys_to_sample]).to(device)

    # 4. Calculate pairwise distances
    print(f"\nCalculating pairwise distances for {len(keys_to_sample)} samples...")
    distances = torch.cdist(sample_tensor, sample_tensor)
    distances_flat = distances[torch.triu(torch.ones_like(distances), diagonal=1).bool()].cpu().numpy()

    # 5. Analyze and print statistics
    min_dist = np.min(distances_flat)
    max_dist = np.max(distances_flat)
    mean_dist = np.mean(distances_flat)
    median_dist = np.median(distances_flat)

    print("\n--- DISTANCE ANALYSIS (from sample) ---")
    print(f"Minimum Distance: {min_dist:.4f}")
    print(f"Maximum Distance: {max_dist:.4f}")
    print(f"Mean Distance:    {mean_dist:.4f}")
    print(f"Median Distance:  {median_dist:.4f}")
    print("---------------------------------------")

    # 6. Plot and save the histogram
    print("Generating distance histogram...")
    plt.figure(figsize=(10, 6))
    plt.hist(distances_flat, bins=100, color='skyblue', edgecolor='black')
    plt.title(PLOT_TITLE)
    plt.xlabel('L2 Distance (after normalization)')
    plt.ylabel('Frequency')
    plt.axvline(mean_dist, color='r', linestyle='dashed', linewidth=2, label=f'Mean: {mean_dist:.2f}')
    plt.axvline(median_dist, color='g', linestyle='dashed', linewidth=2, label=f'Median: {median_dist:.2f}')
    plt.legend()
    
    output_dir = os.path.dirname(OUTPUT_PLOT_PATH)
    if output_dir:
        os.makedirs(output_dir, exist_ok=True)
        
    plt.savefig(OUTPUT_PLOT_PATH)
    print(f"✅ Histogram saved to '{OUTPUT_PLOT_PATH}'")


if __name__ == '__main__':
    main()