import os
import json
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt

def main():
    data = []
    out_dir = "outputs"
    
    if not os.path.exists(out_dir):
        return

    for d in os.listdir(out_dir):
        json_path = os.path.join(out_dir, d, "metrics.json")
        if os.path.exists(json_path):
            try:
                with open(json_path, 'r') as f:
                    j = json.load(f)
                    
                flags = j.get("pathology_flags", [])
                
                row = {
                    "method": j.get("method", "unknown"),
                    "scene": j.get("scene", "unknown"),
                    "PSNR": j.get("photometric", {}).get("PSNR", 0.0),
                    "LPIPS": j.get("photometric", {}).get("LPIPS", 0.0),
                    "N_gaussians": j.get("geometric", {}).get("N_gaussians", 0),
                    "gamma_median": j.get("geometric", {}).get("gamma_median", 0.0),
                    "alpha_mean": j.get("geometric", {}).get("alpha_mean", 0.0),
                    "is_collapsed": len(flags) > 0,
                    "flags": ",".join(flags)
                }
                data.append(row)
            except Exception:
                pass

    if not data:
        return

    df = pd.DataFrame(data)
    df.to_csv("SplatAtlas_Full_Report.csv", index=False)
    
    plt.figure(figsize=(10, 6))
    sns.scatterplot(
        data=df, 
        x="PSNR", 
        y="gamma_median", 
        hue="method", 
        style="is_collapsed", 
        s=150, 
        alpha=0.8
    )
    plt.axhline(18.0, color='red', linestyle='--', label='Pathology Threshold (\u03b3=18)')
    plt.yscale('log')
    plt.title("Photometric-to-Geometric Compensation (Cat 1-5)")
    plt.xlabel("PSNR (Higher is better)")
    plt.ylabel("Anisotropy \u03b3 (Lower is better, log scale)")
    plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left')
    plt.tight_layout()
    plt.savefig("compensation_dynamics.png", dpi=300)

if __name__ == "__main__":
    main()