File size: 5,717 Bytes
d04bc2d | 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 | """926/B — aggregate per-subject diffusion runs into results.json / summary.md."""
import json
from pathlib import Path
import numpy as np
B = Path("/home/hubin/926/results/B")
METRICS = ["chamfer_l2", "chamfer_l1", "fscore_0.1", "fscore_0.05", "emd_256"]
LOWER_BETTER = {"chamfer_l2": True, "chamfer_l1": True, "fscore_0.1": False, "fscore_0.05": False, "emd_256": True}
REF = "paired_4avg"
def boot(d, n=10000, seed=0):
rng = np.random.default_rng(seed)
m = d[rng.integers(0, len(d), (n, len(d)))].mean(1)
return [float(np.quantile(m, 0.025)), float(np.quantile(m, 0.975))]
def main():
subs = {}
for p in sorted((B / "per_subject").glob("sub*.json")):
d = json.load(open(p))
by = {}
for r in d["rows"]:
by.setdefault(r["condition"], []).append(r)
subs[d["subject"]] = {c: {k: float(np.mean([r[k] for r in rs])) for k in METRICS} | {"n": len(rs)}
for c, rs in by.items()}
conds = [c for c in json.load(open(next((B / "per_subject").glob("sub*.json"))))["conditions"]] \
if "conditions" in json.load(open(next((B / "per_subject").glob("sub*.json")))) else list(next(iter(subs.values())))
complete = [s for s in subs if all(c in subs[s] and subs[s][c]["n"] == 144 for c in conds)]
S = sorted(complete)
def delta(c, ref, k):
d = np.array([subs[s][c][k] - subs[s][ref][k] for s in S])
return {"delta": float(d.mean()), "ci95_subject": boot(d), "per_subject": dict(zip(S, d.tolist())),
"n_worse": int((d > 0).sum() if LOWER_BETTER[k] else (d < 0).sum())}
table = {c: {k: {"mean": float(np.mean([subs[s][c][k] for s in S])),
"sd_subject": float(np.std([subs[s][c][k] for s in S], ddof=1)) if len(S) > 1 else None}
for k in METRICS} for c in conds}
deltas = {c: {k: delta(c, REF, k) for k in METRICS} for c in conds if c != REF}
deltas_single = {c: delta(c, "paired_single", "chamfer_l2") for c in conds if c != "paired_single"}
checks = []
def chk(name, c, ref, want):
d = delta(c, ref, "chamfer_l2")
lo, hi = d["ci95_subject"]
ok = hi < 0 if want == "better" else lo > 0
checks.append({"check": name, "input": c, "reference": ref, "want": f"{want} than reference (Chamfer-L2)",
"delta": d["delta"], "ci95_subject": d["ci95_subject"], "pass": bool(ok)})
chk("trial-matched target mean", "target_trialmatched4", REF, "better")
chk("trial-matched wrong mean", "wrong_trialmatched4", REF, "worse")
chk("within-subject target mean", "target_within", REF, "better")
chk("within-subject wrong mean", "wrong_within", REF, "worse")
for m in ("varnorm", "spectrum", "cov"):
chk(f"target mean vs moment-matched paired ({m})", "target_within", f"paired_match_{m}", "better")
chk(f"wrong mean vs moment-matched paired ({m})", "wrong_within", f"paired_match_{m}", "worse")
passed = len(S) > 1 and all(c["pass"] for c in checks)
meta = json.load(open(next((B / "per_subject").glob("sub*.json"))))
res = {"experiment": "926_B_mean_input_controls",
"decoder": "released Neuro-3D point diffusion, 200k steps, frozen; 200 sampling steps; batch 48; "
"seed = 20260722 + batch start (identical across inputs)",
"checkpoint_step": meta.get("checkpoint_step"), "subjects_complete": S, "n_subjects": len(S),
"objects_per_subject": 144, "reference": REF,
"note": "diffusion-to-target distances (~0.4) live in their own column and are NOT the main endpoint (A)",
"wrong_category": "(c + 1) mod 72", "per_condition": table, "delta_vs_paired_4avg": deltas,
"delta_vs_paired_single_chamfer_l2": deltas_single, "pass_checks": checks,
"decision": "keep" if passed else "drop", "test_status": "secondary (main endpoint is A)",
"per_subject": subs}
json.dump(res, open(B / "results.json", "w"), indent=1)
L = [f"# B — mean inputs to the frozen diffusion decoder\n",
f"**decision: {res['decision']}** (secondary test; {len(S)} subjects complete: {', '.join(S)})\n",
"Released Neuro-3D point diffusion (200k), frozen; same targets, weights and sampling seed for every input. "
"Wrong category = (c+1) mod 72. Statistical unit = subject; 95% CI = bootstrap over subjects. "
"These diffusion-to-target distances (~0.4) are a separate column and are not the main endpoint.\n",
"## Per input (subject means)\n",
"| input | Chamfer-L2 | Δ vs paired 4-avg [95% CI] | F@0.1 | Δ F@0.1 | EMD | Δ EMD |", "|---|---|---|---|---|---|---|"]
for c in conds:
t = table[c]
if c == REF:
L.append(f"| {c} | {t['chamfer_l2']['mean']:.4f} | — | {t['fscore_0.1']['mean']:.4f} | — | {t['emd_256']['mean']:.4f} | — |")
continue
d = deltas[c]
f = lambda k: f"{d[k]['delta']:+.4f} [{d[k]['ci95_subject'][0]:+.4f}, {d[k]['ci95_subject'][1]:+.4f}]"
L.append(f"| {c} | {t['chamfer_l2']['mean']:.4f} | {f('chamfer_l2')} | {t['fscore_0.1']['mean']:.4f} | "
f"{f('fscore_0.1')} | {t['emd_256']['mean']:.4f} | {f('emd_256')} |")
L += ["\n## Pass checks (Chamfer-L2, CI must exclude 0)\n", "| check | Δ [95% CI] | pass |", "|---|---|---|"]
for c in checks:
L.append(f"| {c['check']} (`{c['input']}` − `{c['reference']}`) | {c['delta']:+.4f} "
f"[{c['ci95_subject'][0]:+.4f}, {c['ci95_subject'][1]:+.4f}] | {'yes' if c['pass'] else 'no'} |")
(B / "summary.md").write_text("\n".join(L) + "\n")
print("\n".join(L))
if __name__ == "__main__":
main()
|