txus's picture
download
raw
2.39 kB
"""Pull the headline numbers from the job artifacts into one compact summary,
so the logbook can cite exact values."""
import glob
import json
import os
import sys
OUT = sys.argv[1] if len(sys.argv) > 1 else "outputs"
def g(d, *ks, default=None):
for k in ks:
d = d.get(k, {}) if isinstance(d, dict) else {}
return d if d != {} else default
ev = json.load(open(f"{OUT}/eval.json")) if os.path.exists(f"{OUT}/eval.json") else {}
print("=" * 60)
if "newton_alignment" in ev:
na = ev["newton_alignment"]
print("NEWTON ALIGNMENT (cosine to Newton dir, quadratics N=%s)" % na.get("N"))
for k in ("lsr1_proj", "lsr1_noproj", "lbfgs"):
v = na[k]
print(f" {k:12s}: iter1={v[0]:+.3f} mid={v[len(v)//2]:+.3f} "
f"final={v[-1]:+.3f} mean={sum(v)/len(v):+.3f}")
if "performance" in ev:
pf = ev["performance"]
print("\nPERFORMANCE PROFILES (AUC over log-tau, higher=better)")
for s, a in sorted(pf["aucs"].items(), key=lambda x: -x[1]):
rho1 = pf["profiles"][s][0]
print(f" {s:12s}: AUC={a:.3f} rho(tau=1)={rho1:.3f}")
print("\n per-type median iters-to-target (L-SR1 vs best baseline):")
for t in pf["per_type"]:
med = t["median"]; solved = t["solved"]
base = {k: v for k, v in med.items() if k != "L-SR1" and v is not None}
bb = min(base, key=base.get) if base else None
print(f" {t['family']:11s} N={str(t['N']):>4s} cond={str(t['cond']):>7s}"
f" | L-SR1 med={med['L-SR1']} solved={solved['L-SR1']}"
f" | best base={bb}={med.get(bb) if bb else None}")
# wall-clock from histories
print("\nWALL-CLOCK (meta-training)")
for h in sorted(glob.glob(f"{OUT}/*_hist.json")):
d = json.load(open(h))
name = os.path.basename(h).replace("_hist.json", "")
it = d["history"][-1]["iter"] if d["history"] else 0
print(f" {name:20s}: {it} iters in {d.get('wall_time_s', 0)/60:.1f} min")
if os.path.exists(f"{OUT}/efficiency.json"):
e = json.load(open(f"{OUT}/efficiency.json"))
print("\nEFFICIENCY (synthetic proxy -- NOT a Claim-6 reproduction)")
print(f" L-SR1 update params={e['lsr1_update_params']:,} "
f"lgd-style params={e['lgd_update_params']:,}")
print(f" lsr1={e['lsr1_ms']:.1f}ms lgd={e['lgd_ms']:.1f}ms "
f"mem lsr1={e['lsr1_mem_gib']} lgd={e['lgd_mem_gib']}")
print("=" * 60)

Xet Storage Details

Size:
2.39 kB
·
Xet hash:
219b971dd368dc36aef123d619068f218394870648d55fdd04f4e09e07337358

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.