File size: 13,399 Bytes
45e45cb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 | #!/usr/bin/env python3
"""
Render summary figures + stats from episode_results/ (no GPU needed).
Reads: <results-root>/episode_results/<chunk>_<episode>/<mode>.json
Writes: <results-root>/summary/
fig1_absolute_scores.png 4 checkpoints x episodes x 5 modes
fig2_metrics.png Anchor Range / Anchor Std / Reference Error
fig3_by_length.png range vs video length
summary.md mean / median / p90, %>threshold
top10/rankNN_<episode>.png full 5-curve overlays, worst episodes
The 5 modes are the 3 GRM BEFORE-anchoring modes (incremental / forward /
backward) plus 2 sampling-density variants (interval_half / interval_double).
Because the density variants sample a different number of frames, checkpoints
are aligned by physical AFTER-frame index: the incremental (baseline) sequence
defines the 4 target frames (1/4, 2/4, 3/4, end); every mode contributes the
score of its AFTER frame closest to each target.
Metric (same as Robometer):
Anchor Range = max - min of the 5 mode scores at the same physical frame
Anchor Std = std of the 5 mode scores
Reference Error= mode score - incremental score (signed)
threshold = 20 pts
Run with any python that has numpy + matplotlib:
python render_figures.py [--results-root PATH] [--top-n 10]
"""
from __future__ import annotations
import argparse
import json
from collections import defaultdict
from pathlib import Path
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
MODES = ["incremental", "forward", "backward", "interval_half", "interval_double"]
BASELINE = "incremental"
MODE_COLORS = {
"incremental": "tab:blue", "forward": "tab:orange", "backward": "tab:green",
"interval_half": "tab:red", "interval_double": "tab:purple",
}
FRACS = ["1/4", "2/4", "3/4", "end"]
THRESHOLD = 20.0
def parse_args():
p = argparse.ArgumentParser()
base = Path(__file__).resolve().parent.parent
p.add_argument("--results-root", default=str(base / "results_full"))
p.add_argument("--top-n", type=int, default=10)
return p.parse_args()
def load_episodes(results_root: Path):
"""-> {ep_key: {mode: payload}} for episodes with all 5 modes present."""
data = {}
ep_root = results_root / "episode_results"
if not ep_root.exists():
return data
for ep_dir in sorted(ep_root.iterdir()):
if not ep_dir.is_dir():
continue
modes = {}
for m in MODES:
f = ep_dir / f"{m}.json"
if f.exists():
modes[m] = json.loads(f.read_text())
if len(modes) == len(MODES):
data[ep_dir.name] = modes
return data
def target_frames(baseline_after: list[int]) -> list[int]:
"""4 physical AFTER-frame indices at 1/4, 2/4, 3/4, end of the baseline seq."""
L = len(baseline_after)
ps = [int(round((L - 1) * k / 4)) for k in (1, 2, 3, 4)]
return [baseline_after[p] for p in ps]
def score_at(payload: dict, target_af: int) -> float:
"""Score of the AFTER frame closest to target_af (physical alignment)."""
afs = payload["after_frames"]
j = min(range(len(afs)), key=lambda i: abs(afs[i] - target_af))
return payload["scores_100"][j]
def video_length_s(payload: dict) -> float:
return payload["total_raw_frames"] / max(payload.get("native_fps", 0.0), 1e-6)
def main():
args = parse_args()
root = Path(args.results_root)
out = root / "summary"
(out / "top10").mkdir(parents=True, exist_ok=True)
data = load_episodes(root)
eps = sorted(data)
print(f"episodes with all {len(MODES)} modes: {len(eps)}")
if not eps:
return
# ββ extract checkpoint scores + metrics βββββββββββββββββββββββββββββββ
score = defaultdict(dict)
rng_, std_ = {}, {}
ref_err = defaultdict(list) # mode -> [score - incremental score]
fb_diff = [] # forward - backward per (ep, frac)
for ep in eps:
base_after = data[ep][BASELINE]["after_frames"]
tgts = target_frames(base_after)
for frac, taf in zip(FRACS, tgts):
for m in MODES:
score[(ep, frac)][m] = score_at(data[ep][m], taf)
arr = np.array([score[(ep, frac)][m] for m in MODES])
rng_[(ep, frac)] = float(arr.max() - arr.min())
std_[(ep, frac)] = float(arr.std(ddof=0))
for m in MODES:
if m == BASELINE:
continue
ref_err[m].append(score[(ep, frac)][m] - score[(ep, frac)][BASELINE])
fb_diff.append(score[(ep, frac)]["forward"] - score[(ep, frac)]["backward"])
ep_mean_rng = {e: np.mean([rng_[(e, f)] for f in FRACS]) for e in eps}
ep_max_rng = {e: max(rng_[(e, f)] for f in FRACS) for e in eps}
order = sorted(eps, key=lambda e: -ep_mean_rng[e])
x = np.arange(len(order))
all_rng = list(rng_.values())
pct_all = 100.0 * np.mean(np.array(all_rng) > THRESHOLD)
non_base = [m for m in MODES if m != BASELINE]
# ββ fig1: absolute scores βββββββββββββββββββββββββββββββββββββββββββββ
fig, axes = plt.subplots(4, 1, figsize=(16, 14), sharex=True)
for ax, frac in zip(axes, FRACS):
for i, e in enumerate(order):
vals = [score[(e, frac)][m] for m in MODES]
ax.plot([i, i], [min(vals), max(vals)], color="0.85", lw=1, zorder=1)
for m in MODES:
ax.scatter(x, [score[(e, frac)][m] for e in order],
s=8, color=MODE_COLORS[m], label=m, zorder=2)
ax.set_ylabel("score (0-100)")
ax.set_title(f"checkpoint {frac}", loc="left", fontsize=11)
ax.set_ylim(0, 100)
ax.grid(alpha=0.2)
axes[0].legend(ncol=5, fontsize=9, loc="upper right")
axes[-1].set_xlabel("episode (sorted by mean Anchor Range, desc)")
fig.suptitle("Summary of absolute progress scores - 5 anchoring modes per episode\n"
"(gray bar = min-max spread at the same physical frame)", y=0.995)
fig.tight_layout()
fig.savefig(out / "fig1_absolute_scores.png", dpi=150)
plt.close(fig)
# ββ fig2: metrics βββββββββββββββββββββββββββββββββββββββββββββββββββββ
fig, axes = plt.subplots(2, 2, figsize=(15, 11))
ax = axes[0][0]
for frac in FRACS:
vals = sorted((rng_[(e, frac)] for e in eps), reverse=True)
pct = 100.0 * np.mean(np.array(vals) > THRESHOLD)
ax.plot(vals, label=f"{frac} ({pct:.0f}% > {THRESHOLD:.0f} pts)")
ax.axhline(THRESHOLD, color="red", ls="--", lw=1)
ax.set_xlabel("episode rank (desc)")
ax.set_ylabel("Anchor Range (pts)")
ax.set_title("(A) Anchor Range per checkpoint, sorted")
ax.legend(fontsize=9)
ax.grid(alpha=0.2)
ax = axes[0][1]
ax.hist(all_rng, bins=30, color="tab:red", alpha=0.75)
ax.axvline(THRESHOLD, color="black", ls="--", lw=1.5,
label=f"{THRESHOLD:.0f}-pt threshold")
ax.set_xlabel("Anchor Range (pts)")
ax.set_ylabel("count (episode x checkpoint)")
ax.set_title(f"(B) Anchor Range distribution - {pct_all:.0f}% above threshold")
ax.legend(fontsize=9)
ax.grid(alpha=0.2)
ax = axes[1][0]
ax.hist(list(std_.values()), bins=30, color="tab:blue", alpha=0.75)
ax.set_xlabel("Anchor Std (pts)")
ax.set_ylabel("count (episode x checkpoint)")
ax.set_title("(C) Anchor Std distribution")
ax.grid(alpha=0.2)
ax = axes[1][1]
ax.boxplot([ref_err[m] for m in non_base],
labels=[m.replace("_", "\n") for m in non_base], showmeans=True)
ax.axhline(0, color="black", lw=1)
ax.set_ylabel("score - incremental score (pts)")
ax.set_title("(D) Reference Error vs incremental (signed)")
ax.grid(alpha=0.2)
fig.suptitle("Prefix-robustness metrics (Robo-Dopamine, dense curves)", y=0.995)
fig.tight_layout()
fig.savefig(out / "fig2_metrics.png", dpi=150)
plt.close(fig)
# ββ fig3: range vs video length βββββββββββββββββββββββββββββββββββββββ
fig, ax = plt.subplots(figsize=(10, 6))
dur = np.array([video_length_s(data[e][BASELINE]) for e in eps])
mrng = np.array([ep_max_rng[e] for e in eps])
qs = np.quantile(dur, [0, 0.25, 0.5, 0.75, 1.0])
groups, labels = [], []
for lo, hi in zip(qs[:-1], qs[1:]):
m = (dur >= lo) & (dur <= hi)
groups.append(mrng[m])
labels.append(f"{lo:.0f}-{hi:.0f}s\n(n={int(m.sum())})")
ax.boxplot(groups, labels=labels, showmeans=True)
ax.axhline(THRESHOLD, color="red", ls="--", lw=1)
ax.set_xlabel("video length (quartile bins)")
ax.set_ylabel("max Anchor Range over 4 checkpoints (pts)")
ax.set_title("Anchor Range vs video length")
ax.grid(alpha=0.2)
fig.tight_layout()
fig.savefig(out / "fig3_by_length.png", dpi=150)
plt.close(fig)
# ββ summary.md ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def stats(vals):
a = np.array(vals)
return (f"mean {a.mean():.2f} | median {np.median(a):.2f} | "
f"p90 {np.quantile(a, 0.9):.2f} | max {a.max():.2f}")
cam = data[eps[0]][BASELINE]["camera"]
lines = ["# Robo-Dopamine prefix-robustness - full batch summary (dense curves)", ""]
lines += [f"Episodes: **{len(eps)}** | camera: {cam} | "
f"modes: {', '.join(MODES)} | threshold: {THRESHOLD:.0f} pts", ""]
lines += ["Score axis = the GRM pipeline `progress` field (0-100), i.e. the "
"per-mode task-completion estimate. The 3 anchor modes share "
f"frame_interval={data[eps[0]]['incremental']['frame_interval']}; "
"interval_half / interval_double halve / double it and are aligned "
"to each checkpoint by the closest physical AFTER frame.", ""]
lines += ["## Anchor Range (max - min of 5 mode scores, same frame)", "",
"| checkpoint | stats | % > threshold |", "|---|---|---|"]
for frac in FRACS:
vals = [rng_[(e, frac)] for e in eps]
pct = 100.0 * np.mean(np.array(vals) > THRESHOLD)
lines.append(f"| {frac} | {stats(vals)} | **{pct:.1f}%** |")
lines.append(f"| all | {stats(all_rng)} | **{pct_all:.1f}%** |")
ep_any = 100.0 * np.mean([ep_max_rng[e] > THRESHOLD for e in eps])
lines += ["", f"Episodes with >= 1 checkpoint above threshold: **{ep_any:.1f}%**", ""]
lines += ["## Anchor Std", "", f"All cells: {stats(list(std_.values()))}", ""]
lines += ["## Reference Error vs incremental (signed, pts)", "",
"| mode | mean | median | std |", "|---|---|---|---|"]
for m in non_base:
a = np.array(ref_err[m])
lines.append(f"| {m} | {a.mean():+.2f} | {np.median(a):+.2f} | {a.std():.2f} |")
lines.append("")
fb = np.array(fb_diff)
lines += ["## Anchor bias: forward vs backward (systematic)", "",
f"forward - backward (same frame): mean **{fb.mean():+.2f}** | "
f"median {np.median(fb):+.2f} | std {fb.std():.2f} pts", "",
"A large non-zero mean is direct evidence the model's completion "
"estimate depends on whether it is anchored to the start or the goal.", ""]
# ββ top10: full-curve overlays ββββββββββββββββββββββββββββββββββββββββ
worst = sorted(eps, key=lambda e: -ep_max_rng[e])[:args.top_n]
lines += [f"## Top {args.top_n} least-robust episodes (by max Anchor Range)", "",
"| rank | episode | max range | mean range | figure |",
"|---|---|---|---|---|"]
for rank, e in enumerate(worst, 1):
base_after = data[e][BASELINE]["after_frames"]
tgts = target_frames(base_after)
fname = f"rank{rank:02d}_{e}.png"
lines.append(f"| {rank} | {e} | {ep_max_rng[e]:.1f} | "
f"{ep_mean_rng[e]:.1f} | top10/{fname} |")
fig, ax = plt.subplots(figsize=(13, 6))
for m in MODES:
pay = data[e][m]
xs = np.array(pay["after_frames"], dtype=float)
ax.plot(xs, pay["scores_100"], color=MODE_COLORS[m], lw=1.5,
marker=".", ms=3, label=m)
for frac, taf in zip(FRACS, tgts):
ax.axvline(taf, color="0.6", ls=":", lw=1)
ax.text(taf, 97, frac, ha="center", fontsize=8, color="0.4")
ax.set_xlabel("physical AFTER-frame index (original video frames)")
ax.set_ylabel("progress score (0-100)")
ax.set_ylim(0, 100)
ax.grid(alpha=0.2)
ax.legend(fontsize=9)
task = data[e][BASELINE]["task"]
ax.set_title(f"#{rank} {e} max range {ep_max_rng[e]:.1f}\n{task[:110]}",
fontsize=10)
fig.tight_layout()
fig.savefig(out / "top10" / fname, dpi=140)
plt.close(fig)
(out / "summary.md").write_text("\n".join(lines))
print("written:", out)
if __name__ == "__main__":
main()
|