Buckets:
| """Boundary audit: how far out on the real line does our construction reach? | |
| The paper's Theorem 1 domain is [-M_sigma, M_sigma] with M_sigma = 2^(emax-2) = 2^125 | |
| (float32). Our independent construction has its own overflow budget; we measure it.""" | |
| import json, os, sys | |
| import numpy as np | |
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) | |
| from construct import Construction, ACT_NAMES | |
| OUT = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "outputs") | |
| SEED = 20260725 | |
| rows = [] | |
| for name in ACT_NAMES: | |
| best = None | |
| for hp in range(-23, 126, 8): | |
| Z = (2 * np.arange(-4, 5)).reshape(-1, 1) | |
| try: | |
| C = Construction(name, np.float32, hp=hp, Z=Z, d=1, mode="thm31", n_extra=4) | |
| C.draw_targets(seed=SEED + 21) | |
| net = C.build() | |
| bad = 0 | |
| for i in range(C.m): | |
| v, g = net.value_and_grad(C.X[i], C.hstar[i]) | |
| bad += int(v[0] != C.fstar[i]) + int(g[0] != C.gstar[i, 0]) | |
| if bad == 0: | |
| best = dict(hp=hp, x_absmax=float(np.max(np.abs(C.X))), | |
| log2_x_absmax=float(np.log2(float(np.max(np.abs(C.X)))))) | |
| else: | |
| rows.append(dict(activation=name, hp=hp, status=f"{bad} mismatches")) | |
| break | |
| except Exception as e: | |
| rows.append(dict(activation=name, hp=hp, | |
| status=f"{type(e).__name__}: {str(e)[:60]}")) | |
| break | |
| rows.append(dict(activation=name, largest_verified=best, | |
| paper_domain_log2=125, | |
| note="points x = k*2^hp, |k| <= 8, 9 points, L=9, float32")) | |
| print(name, best, rows[-2].get("status")) | |
| with open(os.path.join(OUT, "claim1_wide_domain.json"), "w") as f: | |
| json.dump(rows, f, indent=1, default=float) | |
| print("wrote claim1_wide_domain.json") | |
Xet Storage Details
- Size:
- 1.87 kB
- Xet hash:
- cc41b3093acdcb480c99c9abfc5cb7285ad670b4fe620f5507fed6ae7bb0649e
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.