File size: 3,406 Bytes
59839e3 | 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 | #!/usr/bin/env python3
"""Select v802 component weights + decoder on dev seeds; confirm on check seeds.
Components are OOF files saved by the v802 experiments as
artifacts/models/v802_trials/oof/<prefix>_seed<seed>.npz (key "probs"),
plus the special component "v801" (re-run v801 legacy+kNN mix).
"""
from __future__ import annotations
import argparse
import itertools
import json
from pathlib import Path
import numpy as np
from v802_common import (
CHECK_SEEDS, DEV_SEEDS, ROOT, dump, evaluate, load_data, make_folds, summarize, v801_oof,
)
TRIALS = ROOT / "artifacts" / "models" / "v802_trials"
def component(data, prefix: str, seed: int) -> np.ndarray:
if prefix == "v801":
base = v801_oof(data, seed, TRIALS / "cache")
return 0.8 * base["legacy"] + 0.2 * base["knn"]
stored = np.load(TRIALS / "oof" / f"{prefix}_seed{seed}.npz")
assert (stored["names"] == data.names[data.target_idx]).all()
return stored["probs"]
def weight_grid(count: int, step: float) -> list[tuple[float, ...]]:
ticks = np.round(np.arange(0, 1 + 1e-9, step), 3)
return [w for w in itertools.product(ticks, repeat=count) if abs(sum(w) - 1) < 1e-6]
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--components", nargs="+", required=True)
parser.add_argument("--step", type=float, default=0.1)
parser.add_argument("--priors", type=float, nargs="+", default=[0.0, 0.25, 0.5, 1.0])
parser.add_argument("--cards", type=float, nargs="+", default=[0.0, 0.25, 0.5, 1.0])
parser.add_argument("--top", type=int, default=5)
parser.add_argument("--name", required=True)
parser.add_argument("--feature-meta", type=Path,
default=ROOT / "artifacts/features/v802/dinov3-raw-640x352/meta.json")
args = parser.parse_args()
data = load_data(json.loads(args.feature_meta.read_text())["file_names"])
def run(seeds, weights, prior, card):
rows = []
for seed in seeds:
folds = make_folds(data, seed)
probs = sum(w * cache[(c, seed)] for w, c in zip(weights, args.components) if w)
rows.append(evaluate(data, probs, folds, prior, card))
return summarize(rows)
cache = {(c, s): component(data, c, s) for c in args.components for s in DEV_SEEDS}
results = []
for weights in weight_grid(len(args.components), args.step):
for prior in args.priors:
for card in args.cards:
summary = run(DEV_SEEDS, weights, prior, card)
results.append({"weights": weights, "prior": prior, "card": card, "dev": summary})
results.sort(key=lambda r: (r["dev"]["mean_exact"], r["dev"]["min_exact"]), reverse=True)
top = results[: args.top]
have_check = all((TRIALS / "oof" / f"{c}_seed{s}.npz").exists() or c == "v801"
for c in args.components for s in CHECK_SEEDS)
if have_check:
cache.update({(c, s): component(data, c, s) for c in args.components for s in CHECK_SEEDS})
for row in top:
row["check"] = run(CHECK_SEEDS, row["weights"], row["prior"], row["card"])
for row in top:
print(json.dumps(row))
dump(TRIALS / f"{args.name}.json", {"components": args.components, "top": top,
"evaluated": len(results), "check_available": have_check})
if __name__ == "__main__":
main()
|