File size: 5,144 Bytes
9b789e2 | 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 | """926/C — summary.md from C/results.json, C/seeds.csv, B (released diffusion weights) and T7 (sampling seeds)."""
import csv
import json
from pathlib import Path
import numpy as np
R = Path("/home/hubin/926/results")
T7 = Path("/home/hubin/JAMIETSENG/9.24torun/results/T7_best_of_five/results.json")
def ci(d):
return f"{d['delta']:+.4f} [{d['ci95_subject'][0]:+.4f}, {d['ci95_subject'][1]:+.4f}]"
def main():
c = json.load(open(R / "C/results.json"))
b = json.load(open(R / "B/results.json"))
t7 = json.load(open(T7))
L = ["# C — more than one decoder\n",
"All tests here are secondary (main endpoint = A). Unit = subject (12), 95% CI = subject bootstrap. "
"Wrong category = (c+1) mod 72. Target / wrong means use the 16 train trials (objects 00–07) of that category.\n",
"## Released Neuro-3D point-diffusion weights (200k), same split\n",
f"From B (frozen decoder, 200 steps, identical sampling seed per input; {b['n_subjects']} subjects):\n",
"| input | Chamfer-L2 | Δ vs paired [95% CI] |", "|---|---|---|"]
t = b["per_condition"]
L.append(f"| paired (4-trial avg) | {t['paired_4avg']['chamfer_l2']['mean']:.4f} | — |")
for k, lab in (("same_category_swap", "same-category swap"), ("target_within", "target-class mean"),
("wrong_within", "wrong-class mean")):
L.append(f"| {lab} | {t[k]['chamfer_l2']['mean']:.4f} | {ci(b['delta_vs_paired_4avg'][k]['chamfer_l2'])} |")
L += ["\nSame pattern as on the modified Gaussian head: swap ≈ paired, target mean much better, wrong mean worse.\n",
"## Three small baselines (seed mean over 5 encoder seeds)\n",
"| decoder | input | category hit | Chamfer-L2 | Δ Chamfer vs paired [95% CI] | F@0.1 | EMD |", "|---|---|---|---|---|---|---|"]
names = {"classify_medoid": "classify → train medoid", "latent_regression": "EEG → CLIP point latent (ridge) → nearest train object",
"retrieval": "nearest train EEG → its object"}
for dec, conds in c["fingerprint_seed_mean"].items():
for cond, v in conds.items():
m = v["mean"]
d = ci(v["delta_chamfer_l2"]) if "delta_chamfer_l2" in v else "—"
L.append(f"| {names[dec]} | {cond} | {m['category_hit']:.3f} | {m['chamfer_l2']:.4f} | {d} | {m['f01']:.4f} | {m['emd_256']:.4f} |")
L += ["\nOn all three baselines the target-class mean beats paired by 0.057–0.069 with 12/12 subjects, and swap moves "
"Chamfer by ≤ 0.0023. The wrong-class mean is only reliably worse for latent regression (+0.0092); for "
"classify→medoid and retrieval it is indistinguishable from paired, because paired single-subject EEG already "
"decodes near chance (category hit ≈ 2%, chance 1.4%), so there is little category signal left to remove.\n",
"## Category readout, 5 training seeds (validation-safe encoders; test objects 08–09, 4-trial average)\n",
"| seed | pool | top-1 | top-5 | top-10 | MRR | NLL |", "|---|---|---|---|---|---|---|"]
rows = list(csv.DictReader(open(R / "C/seeds.csv")))
per = {}
for r in rows:
key = "pooled" if r["subject"] == "pooled12" else "single"
per.setdefault((key, r["seed"]), []).append([float(r[k]) for k in ("top1", "top5", "top10", "mrr", "nll")])
agg = {"single": [], "pooled": []}
for (key, seed), v in sorted(per.items(), key=lambda x: (x[0][0] != "single", int(x[0][1]))):
m = np.mean(v, 0); agg[key].append(m)
lab = "single-subject (mean of 12)" if key == "single" else "pooled 12 (mean posterior)"
L.append(f"| {seed} | {lab} | " + " | ".join(f"{x:.4f}" for x in m) + " |")
for key in ("single", "pooled"):
a = np.array(agg[key])
L.append(f"| **mean ± std** | {key} | " + " | ".join(f"{x:.4f} ± {s:.4f}" for x, s in zip(a.mean(0), a.std(0, ddof=1))) + " |")
L.append("\nChance: top-1 0.0139, top-5 0.069, top-10 0.139, NLL ln 72 = 4.277.\n")
L += ["## Sampling seed (released diffusion, separate from training seeds)\n",
"From 9.24torun/T7: same 200k weights, 12 subjects × 144 objects, 5 sampling seeds, 1000 steps, single sample per seed.\n",
"| sampling seed | Chamfer-L1 | Chamfer-L2 | F@0.1 | EMD |", "|---|---|---|---|---|"]
ps = t7["per_seed"]
for s, v in ps.items():
L.append(f"| {s} | {v['chamfer_l1']:.4f} | {v['chamfer_l2']:.4f} | {v['f_at_01']:.4f} | {v['emd_256']:.4f} |")
a = np.array([[v["chamfer_l1"], v["chamfer_l2"], v["f_at_01"], v["emd_256"]] for v in ps.values()])
L.append("| **mean ± std** | " + " | ".join(f"{x:.4f} ± {s:.4f}" for x, s in zip(a.mean(0), a.std(0, ddof=1))) + " |")
L += ["\n## Second encoder / splat checkpoints\n",
"None in the repository (EEG-to-3DGS, BrainGS, BrainSSD); recorded in `results/MISSING_CKPT.md`, no training started.\n",
"Files: `results.json`, `seeds.csv`, `per_object.csv` (object-level, labelled as such)."]
(R / "C/summary.md").write_text("\n".join(L) + "\n")
print("\n".join(L))
if __name__ == "__main__":
main()
|