File size: 8,789 Bytes
02443ff | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 | from __future__ import annotations
import argparse
from collections import defaultdict
import hashlib
import json
from pathlib import Path
import sys
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "src"))
from anima_style_probe.factor_interventions import ( # noqa: E402
FACTORS,
INTERVENTION_VERSION,
LEVELS,
TRAIN_FAMILIES,
VALIDATION_FAMILIES,
)
def stable_int(*values: object) -> int:
text = "|".join(map(str, values))
return int.from_bytes(hashlib.blake2b(text.encode(), digest_size=8).digest(), "little")
def read_records(paths: list[Path]) -> list[dict]:
records: list[dict] = []
seen: set[str] = set()
for path in paths:
payload = json.loads(path.read_text(encoding="utf-8"))
for row in payload["records"]:
record_id = str(row["record_id"])
if record_id in seen:
raise RuntimeError(f"duplicate packed record: {record_id}")
seen.add(record_id)
records.append(row)
return records
def intensity_for(key: object) -> tuple[str, int]:
value = stable_int("intensity", key) % 10
level = "weak" if value < 4 else "medium" if value < 8 else "strong"
sign = 1 if stable_int("sign", key) % 2 else -1
return level, sign
def intervention_row(
source: dict,
factor: str,
family: str,
level: str,
sign: int,
*,
anchor_kind: str,
repeat_of: str | None = None,
validation: bool = False,
) -> dict:
suffix = f"{factor}-{family}-{level}-{'p' if sign > 0 else 'n'}"
if repeat_of is not None:
suffix += "-repeat"
intervention_id = f"{source['record_id']}__{suffix}"
return {
"record_id": intervention_id,
"source_record_id": source["record_id"],
"style_id": source["style_id"],
"source": source["source"],
"split": "validation" if validation else "train",
"shard": source["shard"],
"factor": factor,
"factor_index": FACTORS.index(factor),
"family": family,
"level": level,
"sign": sign,
"signed_intensity": sign * LEVELS[level],
"operation_seed": stable_int("operation", intervention_id) & ((1 << 63) - 1),
"transform_version": INTERVENTION_VERSION,
"anchor_kind": anchor_kind,
"repeat_of": repeat_of,
"panel": False,
"anima_pilot": False,
}
def build_training(records: list[dict]) -> list[dict]:
by_style: dict[str, list[dict]] = defaultdict(list)
for row in records:
by_style[str(row["style_id"])].append(row)
if len(by_style) != 8_000:
raise RuntimeError(f"expected 8,000 train identities, found {len(by_style)}")
output: list[dict] = []
for style_id, rows in sorted(by_style.items()):
if len(rows) != 40:
raise RuntimeError(f"{style_id} has {len(rows)} optimization records, expected 40")
ordered = sorted(rows, key=lambda row: stable_int("anchor", row["record_id"]))
shared = ordered[0]
specific = iter(ordered[1:9])
factor_rows: dict[str, list[dict]] = defaultdict(list)
for factor in FACTORS:
families = list(TRAIN_FAMILIES[factor])
rotation = stable_int("family", style_id, factor) % len(families)
families = families[rotation:] + families[:rotation]
anchors = [shared, next(specific), next(specific)]
for index, (source, family) in enumerate(zip(anchors, families, strict=True)):
level, sign = intensity_for((style_id, factor, family))
row = intervention_row(
source,
factor,
family,
level,
sign,
anchor_kind="shared" if index == 0 else "factor_specific",
)
output.append(row)
factor_rows[factor].append(row)
repeat_factor = FACTORS[stable_int("repeat-factor", style_id) % len(FACTORS)]
base = factor_rows[repeat_factor][stable_int("repeat-row", style_id) % 3]
second_level = {"weak": "medium", "medium": "strong", "strong": "medium"}[base["level"]]
source = next(row for row in rows if row["record_id"] == base["source_record_id"])
output.append(
intervention_row(
source,
base["factor"],
base["family"],
second_level,
base["sign"],
anchor_kind="intensity_repeat",
repeat_of=base["record_id"],
)
)
if len(output) != 104_000:
raise RuntimeError(f"expected 104,000 train variants, found {len(output)}")
by_stratum: dict[tuple[str, str], list[dict]] = defaultdict(list)
for row in output:
if row["anchor_kind"] != "intensity_repeat":
by_stratum[(row["source"], row["factor"])].append(row)
for rows in by_stratum.values():
for row in sorted(rows, key=lambda item: stable_int("panel", item["record_id"]))[:128]:
row["panel"] = True
if sum(row["panel"] for row in output) != 1_024:
raise RuntimeError("failed to build balanced 1,024-record panel")
return output
def build_validation(records: list[dict]) -> list[dict]:
by_source_style: dict[str, dict[str, list[dict]]] = defaultdict(lambda: defaultdict(list))
for row in records:
if row.get("split") == "validation":
by_source_style[row["source"]][row["style_id"]].append(row)
output: list[dict] = []
for source in ("synthetic", "human"):
styles = sorted(
by_source_style[source], key=lambda style: stable_int("validation-style", style)
)[:256]
if len(styles) != 256:
raise RuntimeError(f"{source} has only {len(styles)} unseen validation identities")
for style_id in styles:
base = min(
by_source_style[source][style_id],
key=lambda row: stable_int("validation-record", row["record_id"]),
)
for factor in FACTORS:
family = VALIDATION_FAMILIES[factor]
sign = 1 if stable_int("validation-sign", style_id, factor) % 2 else -1
for level in ("weak", "strong"):
output.append(
intervention_row(
base,
factor,
family,
level,
sign,
anchor_kind="validation_intensity",
validation=True,
)
)
if len(output) != 4_096:
raise RuntimeError(f"expected 4,096 validation variants, found {len(output)}")
for row in output:
row["anima_pilot"] = True
return output
def main() -> int:
parser = argparse.ArgumentParser(description="Build the factor-intervention subset manifest.")
parser.add_argument("--packed-root", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
train_paths = sorted(args.packed_root.glob("train-rank*/features-*.json"))
validation_paths = sorted(args.packed_root.glob("validation-*/features-*.json"))
if not train_paths or not validation_paths:
raise FileNotFoundError("packed train or validation metadata is missing")
train = build_training(read_records(train_paths))
validation = build_validation(read_records(validation_paths))
rows = train + validation
if len({row["record_id"] for row in rows}) != len(rows):
raise RuntimeError("duplicate intervention record IDs")
args.output.parent.mkdir(parents=True, exist_ok=True)
temporary = args.output.with_suffix(args.output.suffix + ".tmp")
with temporary.open("w", encoding="utf-8", newline="\n") as handle:
for row in rows:
handle.write(json.dumps(row, ensure_ascii=False, separators=(",", ":")) + "\n")
temporary.replace(args.output)
summary = {
"status": "complete",
"train_variants": len(train),
"validation_variants": len(validation),
"unique_train_anchors": len({row["source_record_id"] for row in train}),
"panel": sum(row["panel"] for row in train),
"anima_pilot": sum(row["anima_pilot"] for row in validation),
"transform_version": INTERVENTION_VERSION,
"output": str(args.output),
}
args.output.with_suffix(".summary.json").write_text(
json.dumps(summary, indent=2) + "\n", encoding="utf-8"
)
print(json.dumps(summary, indent=2))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|