| """Read eval.json across budgets and produce the success-vs-demos figure + table.""" |
| import json, os, argparse |
| import numpy as np |
| import matplotlib; matplotlib.use("Agg") |
| import matplotlib.pyplot as plt |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("--budgets", nargs="+", type=int, default=[10, 50, 100, 1000]) |
| ap.add_argument("--seed", type=int, default=0) |
| ap.add_argument("--tag", default="state") |
| ap.add_argument("--outdir", default="results") |
| a = ap.parse_args() |
| os.makedirs(a.outdir, exist_ok=True) |
|
|
| xs, ys, los, his, rows = [], [], [], [], [] |
| for b in a.budgets: |
| p = f"runs/gcbc_{a.tag}_b{b:04d}_seed{a.seed}/eval.json" |
| if not os.path.exists(p): |
| print("missing", p); continue |
| d = json.load(open(p)); o = d["overall"] |
| xs.append(b); ys.append(o["rate"]); los.append(o["ci"][0]); his.append(o["ci"][1]) |
| rows.append((b, o["k"], o["n"], o["rate"], o["ci"][0], o["ci"][1])) |
|
|
| ys = np.array(ys); los = np.array(los); his = np.array(his) |
| plt.figure(figsize=(7, 4.5)) |
| plt.errorbar(xs, ys, yerr=[ys - los, his - ys], fmt="o-", capsize=4, lw=2, color="#2563eb") |
| plt.xscale("log") |
| plt.xlabel("number of demonstrations (log scale)") |
| plt.ylabel("online success rate") |
| plt.title("Cube-single GCBC (state) — success vs. demonstrations\n95% Wilson CI, 50 rollouts x 5 tasks") |
| plt.grid(True, which="both", alpha=0.3); plt.ylim(0, max(0.5, his.max() * 1.15)) |
| plt.tight_layout() |
| fig = os.path.join(a.outdir, f"success_vs_demos_{a.tag}.png") |
| plt.savefig(fig, dpi=130); print("figure ->", fig) |
|
|
| csv = os.path.join(a.outdir, f"scaling_{a.tag}.csv") |
| with open(csv, "w") as f: |
| f.write("demos,successes,rollouts,success_rate,ci_low,ci_high\n") |
| for r in rows: f.write("%d,%d,%d,%.4f,%.4f,%.4f\n" % r) |
| print("table ->", csv) |
| print("\nDemos Success 95% Wilson CI") |
| for b, k, n, p, lo, hi in rows: |
| print(f"{b:5d} {p:6.1%} [{lo:.1%}, {hi:.1%}] ({k}/{n})") |
|
|
