import json import glob import os import pandas as pd import numpy as np # 1. 基础配置 methods = [ "vanilla_3dgs", "analyticsplatting", "erankgs", "ges", "lightgaussian", "minisplatting", "opti3dgs", "pgsr", "steepgs", "3dgsmcmc" ] scene_mapping = {"bonsai": "bonsai", "lego": "Lego"} seeds = {"seed0": "", "seed1": "_seed1", "seed2": "_seed2"} raw_records = [] # 2. 扫描数据 print("========================================================") print("开始扫描 Seed 目录...") for method in methods: for norm_scene, dir_scene in scene_mapping.items(): for seed_id, seed_suffix in seeds.items(): dir_name = f"{method}{seed_suffix}_{dir_scene}" search_dir = f"outputs/{dir_name}" # vanilla_3dgs 特殊 fallback 逻辑 if method == "vanilla_3dgs" and seed_id == "seed0": if not os.path.exists(search_dir) or not glob.glob(f"{search_dir}/metrics_test_iter*.json"): search_dir = f"outputs/{method}_{dir_scene}_bak" json_files = sorted(glob.glob(f"{search_dir}/metrics_test_iter*.json")) psnr, ssim, lpips = np.nan, np.nan, np.nan source_file = "MISSING" status = "MISSING" if json_files: target_file = json_files[-1] # 默认取排序后的最后一个 for jf in json_files: if "30000" in jf: target_file = jf break source_file = target_file try: with open(target_file, 'r') as f: data = json.load(f) # 兼容不同 JSON 层级 photo = data.get("photometric", data) psnr = photo.get("PSNR", photo.get("psnr", np.nan)) ssim = photo.get("SSIM", photo.get("ssim", np.nan)) lpips = photo.get("LPIPS", photo.get("lpips", np.nan)) status = "OK" except Exception: pass # 基于业务逻辑的 Status 调整 if method == "erankgs" and norm_scene == "lego" and psnr < 20: # ~17dB status = "METHOD_FAILURE" raw_records.append({ "SceneNormalized": norm_scene, "SceneDirName": dir_scene, "Method": method, "Seed": seed_id, "PSNR": psnr, "SSIM": ssim, "LPIPS": lpips, "Status": status, "SourceFile": source_file }) # 输出 Raw Table os.makedirs("outputs/phase4", exist_ok=True) df_raw = pd.DataFrame(raw_records) raw_path = "outputs/phase4/task_4_1_seed_variance_raw_reconstructed.csv" df_raw.to_csv(raw_path, index=False) # 3. 生成 Summary Table summary_records = [] for (norm_scene, method), group in df_raw.groupby(["SceneNormalized", "Method"]): psnrs = group["PSNR"].dropna() n_seeds = len(psnrs) psnr_mean = psnrs.mean() if n_seeds > 0 else np.nan psnr_std = psnrs.std(ddof=1) if n_seeds > 1 else np.nan psnr_min = psnrs.min() if n_seeds > 0 else np.nan psnr_max = psnrs.max() if n_seeds > 0 else np.nan psnr_range = (psnr_max - psnr_min) if n_seeds > 0 else np.nan statuses = group["Status"].tolist() if "METHOD_FAILURE" in statuses: final_status = "METHOD_FAILURE" elif method == "analyticsplatting" and norm_scene == "lego" and psnr_std > 1.0: final_status = "OUTLIER" elif "MISSING" in statuses and n_seeds < 3: final_status = "MISSING_SEEDS" else: final_status = "OK" summary_records.append({ "Scene": norm_scene, "Method": method, "NSeeds": n_seeds, "PSNR_mean": psnr_mean, "PSNR_std": psnr_std, "PSNR_min": psnr_min, "PSNR_max": psnr_max, "PSNR_range": psnr_range, "Status": final_status }) df_summary = pd.DataFrame(summary_records) sum_path = "outputs/phase4/task_4_1_seed_variance_summary_reconstructed.csv" df_summary.to_csv(sum_path, index=False) # 4. 生成 Kendall Tau Checked CSV tau_file = "outputs/phase4/task_4_2_kendall_tau.json" tau_path = "outputs/phase4/task_4_2_kendall_tau_checked.csv" if os.path.exists(tau_file): with open(tau_file, 'r') as f: tau_data = json.load(f) tau_records = [] for scene in ["bonsai", "Lego"]: if scene in tau_data: sc_data = tau_data[scene] tau_records.append({ "Scene": "lego" if scene == "Lego" else "bonsai", "default_vs_seed1_tau": sc_data.get("default_vs_seed1", {}).get("tau"), "default_vs_seed1_pval": sc_data.get("default_vs_seed1", {}).get("pval"), "default_vs_seed2_tau": sc_data.get("default_vs_seed2", {}).get("tau"), "default_vs_seed2_pval": sc_data.get("default_vs_seed2", {}).get("pval"), "seed1_vs_seed2_tau": sc_data.get("seed1_vs_seed2", {}).get("tau"), "seed1_vs_seed2_pval": sc_data.get("seed1_vs_seed2", {}).get("pval"), "median_tau": sc_data.get("median_tau"), "min_tau": sc_data.get("min_tau"), "valid_methods_count": len(sc_data.get("valid_methods", [])) }) df_tau = pd.DataFrame(tau_records) df_tau.to_csv(tau_path, index=False) else: print(f"⚠️ 警告: Kendall Tau 源文件缺失 {tau_file}") # 5. Sanity Checks 输出 print("\n========================================================") print("SANITY CHECKS") print("========================================================") # 基础计数 print(f"每个 scene 找到的 method 数量: bonsai = {len(df_summary[df_summary['Scene'] == 'bonsai'])}, lego = {len(df_summary[df_summary['Scene'] == 'lego'])}") print(f"每个 scene 找到的 raw seed rows: bonsai = {len(df_raw[df_raw['SceneNormalized'] == 'bonsai'])}, lego = {len(df_raw[df_raw['SceneNormalized'] == 'lego'])}") # 特殊 Cell 验证 try: ana_lego = df_summary[(df_summary['Method'] == 'analyticsplatting') & (df_summary['Scene'] == 'lego')].iloc[0] print(f"\n[验证] analyticsplatting x Lego") print(f" - PSNR_mean: {ana_lego['PSNR_mean']:.3f}") print(f" - PSNR_std: {ana_lego['PSNR_std']:.3f} (预期 ~1.33)") print(f" - PSNR_range: {ana_lego['PSNR_range']:.3f} (预期 ~2.34)") print(f" - Status: {ana_lego['Status']} (预期 OUTLIER)") except Exception as e: print("\n[错误] 未能定位 analyticsplatting x Lego 的聚合数据") try: era_lego = df_summary[(df_summary['Method'] == 'erankgs') & (df_summary['Scene'] == 'lego')].iloc[0] print(f"\n[验证] erankgs x Lego") print(f" - PSNR_mean: {era_lego['PSNR_mean']:.3f} (预期 ~17)") print(f" - PSNR_std: {era_lego['PSNR_std']:.3f}") print(f" - PSNR_range: {era_lego['PSNR_range']:.3f}") print(f" - Status: {era_lego['Status']} (预期 METHOD_FAILURE)") except Exception as e: print("\n[错误] 未能定位 erankgs x Lego 的聚合数据") # 队列级中位数与 P95 (排除 METHOD_FAILURE) valid_stds_incl = df_summary[df_summary["Status"].isin(["OK", "OUTLIER"])]["PSNR_std"].dropna() valid_stds_excl = df_summary[df_summary["Status"] == "OK"]["PSNR_std"].dropna() med_incl = valid_stds_incl.median() p95_incl = valid_stds_incl.quantile(0.95) med_excl = valid_stds_excl.median() p95_excl = valid_stds_excl.quantile(0.95) print("\n[验证] Cohort PSNR Std 统计") print(f" - Median (including outlier): {med_incl:.4f}") print(f" - P95 (including outlier): {p95_incl:.4f}") print(f" - Median (excluding outlier): {med_excl:.4f}") print(f" - P95 (excluding outlier): {p95_excl:.4f}") if abs(p95_incl - 0.384) > 0.005 and abs(p95_excl - 0.384) > 0.005: print(f"\n⚠️ 论文一致性警告:") print(f"不一致：论文中声明 P95=0.384 dB，但重算结果为包含Outlier: {p95_incl:.3f}, 排除Outlier: {p95_excl:.3f}") else: print(f"\n✅ P95=0.384 dB 与论文声明一致") # Kendall Tau if 'df_tau' in locals(): print("\n[验证] Kendall Tau 值") b_tau = df_tau[df_tau["Scene"] == "bonsai"]["median_tau"].iloc[0] l_tau = df_tau[df_tau["Scene"] == "lego"]["median_tau"].iloc[0] print(f" - bonsai median_tau: {b_tau:.4f} (预期: 0.7778)") print(f" - Lego median_tau: {l_tau:.4f} (预期: 0.8571)") print("\n========================================================") print("生成的文件清单:") print(f"1. {raw_path}") print(f"2. {sum_path}") print(f"3. {tau_path}") print("========================================================")