Buckets:
| #!/usr/bin/env python | |
| """Plot the reward / progress columns of a SPIRAL-labelled dataset. | |
| Reads what is actually written to disk - no model inference - so it shows exactly | |
| what SPIRAL will consume. | |
| python tools/plot_spiral_labels.py --repo-id <ns>/<name> --episodes 5 | |
| """ | |
| from __future__ import annotations | |
| import argparse | |
| import json | |
| from pathlib import Path | |
| import matplotlib | |
| matplotlib.use("Agg") | |
| import matplotlib.pyplot as plt # noqa: E402 | |
| import numpy as np # noqa: E402 | |
| import pyarrow.parquet as pq # noqa: E402 | |
| def main() -> None: | |
| ap = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter) | |
| ap.add_argument("--repo-id", required=True) | |
| ap.add_argument("--root", default=None) | |
| ap.add_argument("--episodes", type=int, default=5) | |
| ap.add_argument("--episode-ids", default=None) | |
| ap.add_argument("--seed", type=int, default=0) | |
| ap.add_argument("--out", default="spiral_labels.png") | |
| args = ap.parse_args() | |
| root = Path(args.root) if args.root else Path.home()/".cache/huggingface/lerobot"/args.repo_id | |
| info = json.loads((root/"meta/info.json").read_text()) | |
| fps = float(info["fps"]) | |
| eps = [json.loads(l) for l in (root/"meta/episodes.jsonl").read_text().splitlines() if l.strip()] | |
| ids = ([int(x) for x in args.episode_ids.split(",")] if args.episode_ids else | |
| sorted(np.random.default_rng(args.seed).choice( | |
| [e["episode_index"] for e in eps], size=min(args.episodes, len(eps)), replace=False).tolist())) | |
| rows = [] | |
| for ep in ids: | |
| t = pq.read_table(root/info["data_path"].format( | |
| episode_chunk=ep//int(info.get("chunks_size", 1000)), episode_index=ep)) | |
| r = np.array([x[0] for x in t.column("reward").to_pylist()], dtype=np.float64) | |
| p = np.array([x[0] for x in t.column("progress").to_pylist()], dtype=np.float64) | |
| rows.append((ep, np.arange(len(r))/fps, r, p)) | |
| fig, axes = plt.subplots(len(rows), 2, figsize=(11, 2.3*len(rows))) | |
| axes = np.atleast_2d(axes) | |
| for i, (ep, tt, r, p) in enumerate(rows): | |
| a0, a1 = axes[i] | |
| a0.plot(tt, p, color="#20c4c0", lw=2) | |
| a0.set_ylabel(f"ep {ep}\nprogress", fontsize=9, fontweight="bold") | |
| a0.grid(alpha=.3, color="white", lw=1.2); a0.set_facecolor("#eef0f4") | |
| a1.plot(tt, r, color="#f0a832", lw=1.4) | |
| a1.axhline(0, color="#999", lw=0.8, ls="--") | |
| a1.set_ylabel("reward", fontsize=9, fontweight="bold") | |
| a1.grid(alpha=.3, color="white", lw=1.2); a1.set_facecolor("#eef0f4") | |
| for a in (a0, a1): | |
| for s in a.spines.values(): s.set_visible(False) | |
| if i == 0: | |
| a0.set_title("progress column (SARM2 value, paper's r*)", fontsize=10) | |
| a1.set_title("reward column (per-frame delta -> SPIRAL TD)", fontsize=10) | |
| if i == len(rows)-1: | |
| a0.set_xlabel("Time (s)"); a1.set_xlabel("Time (s)") | |
| fig.tight_layout(); fig.savefig(args.out, dpi=140) | |
| allr = np.concatenate([r for _, _, r, _ in rows]) | |
| allp = np.concatenate([p for _, _, _, p in rows]) | |
| print(f"[out] {Path(args.out).resolve()}") | |
| print(f" reward : mean {allr.mean():.5f} max {allr.max():.5f} negative {100*(allr<0).mean():.1f}%") | |
| print(f" progress: range [{allp.min():.3f}, {allp.max():.3f}]") | |
| for ep, _, r, p in rows: | |
| print(f" ep {ep:>3}: sum(reward) {r.sum():+.4f} | progress {p[0]:+.3f} -> {p[-1]:+.3f}") | |
| if __name__ == "__main__": | |
| main() | |
Xet Storage Details
- Size:
- 3.47 kB
- Xet hash:
- 586efddd312856f66cf983dda5d0ba46aa7fe81815e3f6e68bc4807f1f640fa1
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.