| """Re-render and re-score submitted scenes from the episode tables: an independent check of the recorded scores. |
| |
| uv run python scripts/rescore.py --setting medium_effort |
| uv run python scripts/rescore.py --episode gpt-6-astra__nerf_lego__sculpt__eval |
| |
| The stored Gaussians are rounded to 3 decimals, as the environment logs them, so a re-score can differ from the |
| recorded score by a little (medium effort: mean |difference| 0.02 points, largest 0.22). |
| """ |
| import argparse |
| import json |
| from pathlib import Path |
|
|
| import numpy as np |
| import pyarrow.parquet as pq |
| import torch |
|
|
| from splatbench.render import empty_gaussians |
| from splatbench.scene import Scene |
| from splatbench.score import evaluate, splat_score |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("--data", default="data") |
| ap.add_argument("--setting", default="medium_effort", |
| choices=["medium_effort", "high_effort", "max_effort", "early_submit_refused", "one_shot"]) |
| ap.add_argument("--episode", default=None) |
| ap.add_argument("--limit", type=int, default=None) |
| ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu") |
| a = ap.parse_args() |
|
|
| rows = pq.read_table(Path(a.data) / "episodes" / f"{a.setting}.parquet", |
| columns=["episode_id", "task_id", "score", "final_gaussians"]).to_pylist() |
| if a.episode: |
| rows = [r for r in rows if r["episode_id"] == a.episode] |
| rows = rows[:a.limit] if a.limit else rows |
| scenes, diffs = {}, [] |
| for r in rows: |
| scene = r["task_id"].split("__")[0] |
| if scene not in scenes: |
| sd = Path(a.data) / "bench" / "scenes" / scene |
| scenes[scene] = (Scene.load(sd), json.loads((sd / "norm.json").read_text())) |
| sc, norm = scenes[scene] |
| G = r["final_gaussians"] |
| f = lambda k: torch.tensor([g[k] for g in G], dtype=torch.float32, device=a.device) |
| gs = dict(means=f("pos"), scales=f("scale"), quats=f("quat"), colors=f("color"), opacities=f("opacity")) if G \ |
| else empty_gaussians(a.device) |
| s = splat_score(evaluate(gs, sc, "test", a.device, with_lpips=True)["mean"], norm["test"])["score"] |
| diffs.append(s - r["score"]) |
| print(f"{r['episode_id']:58s} recorded {r['score']:7.2f} re-scored {s:7.2f} difference {s - r['score']:+.3f}", flush=True) |
| d = np.abs(np.array(diffs)) |
| print(f"{len(d)} episodes: mean |difference| {d.mean():.3f}, largest {d.max():.3f}") |
|
|
