File size: 2,563 Bytes
4093113 | 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 | #!/usr/bin/env python3
"""Frozen epoch-zero control for the million-epoch Deep-UFM run."""
from __future__ import annotations
import argparse
import io
import json
import zipfile
from pathlib import Path
import numpy as np
import torch
from verify_native_relu_ufm import D, K, N, N_PER_CLASS, analyse
INITIAL_STD = 0.1
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--seed", type=int, default=71)
args = parser.parse_args()
args.output.mkdir(parents=True, exist_ok=True)
torch.manual_seed(args.seed)
arrays = {
"Y": torch.eye(K, dtype=torch.float32)
.repeat_interleave(N_PER_CLASS, dim=1)
.numpy(),
"H1": (torch.randn(D, N, dtype=torch.float32) * INITIAL_STD).numpy(),
}
for index in range(1, 5):
arrays[f"W{index}"] = (
torch.randn(D, D, dtype=torch.float32) * INITIAL_STD
).numpy()
arrays["W5"] = (
torch.randn(K, D, dtype=torch.float32) * INITIAL_STD
).numpy()
state = args.output / "epoch_zero_state.npz"
with zipfile.ZipFile(state, "w", compression=zipfile.ZIP_STORED) as archive:
for name, array in arrays.items():
payload = io.BytesIO()
np.lib.format.write_array(
payload, np.asanyarray(array), allow_pickle=False
)
info = zipfile.ZipInfo(f"{name}.npy", (1980, 1, 1, 0, 0, 0))
info.compress_type = zipfile.ZIP_STORED
info.external_attr = 0o600 << 16
archive.writestr(info, payload.getvalue())
oracle = analyse(state)
control = {
"control": (
"same source-scale architecture, target, seed and initialization "
"before any of the registered gradient-descent epochs"
),
"oracle": oracle,
"training_accuracy_below_one": oracle["fit"]["accuracy"] < 1.0,
"nine_outlier_claim_absent_at_initialization": not oracle["hessian"][
"nine_outlier_gate"
],
}
control["destructive_control_pass"] = bool(
control["training_accuracy_below_one"]
and control["nine_outlier_claim_absent_at_initialization"]
)
(args.output / "epoch_zero_control.json").write_text(
json.dumps(control, indent=2, sort_keys=True) + "\n",
encoding="utf-8",
)
print(json.dumps(control, indent=2, sort_keys=True))
if not control["destructive_control_pass"]:
raise SystemExit(2)
if __name__ == "__main__":
main()
|