File size: 16,433 Bytes
10979b5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 | #!/usr/bin/env python3
"""Validate and merge the 18 formal E1 trained-feature audit shards."""
from __future__ import annotations
import argparse
import csv
import json
import math
import sys
from collections import defaultdict
from datetime import UTC, datetime
from pathlib import Path
from typing import Any
PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(PROJECT_ROOT) not in sys.path:
sys.path.insert(0, str(PROJECT_ROOT))
from run_e1_trained_features import LAYER_ORDER, _aggregate, _format, _write_csv
Row = dict[str, Any]
EXPECTED_GATES = (
"relu6_self",
"relu_self",
"gelu_self",
"smooth_clipped_self",
"identity",
"no_gate",
)
EXPECTED_PAIRS = {(gate, seed) for gate in EXPECTED_GATES for seed in range(3)}
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--shard-root", type=Path, required=True)
parser.add_argument("--output-dir", type=Path, required=True)
return parser.parse_args()
def _convert(value: str) -> Any:
if value == "":
return None
if value == "True":
return True
if value == "False":
return False
try:
if value.lstrip("-").isdigit():
return int(value)
return float(value)
except ValueError:
return value
def _read_csv(path: Path) -> list[Row]:
with path.open(newline="", encoding="utf-8") as handle:
return [{key: _convert(value) for key, value in row.items()} for row in csv.DictReader(handle)]
def _weighted_gate_seed_rows(gate_rows: list[Row]) -> list[Row]:
groups: dict[tuple[str, int, str], list[Row]] = defaultdict(list)
for row in gate_rows:
if float(row["cutoff"]) == 1.0:
groups[(row["gate"], int(row["seed"]), row["stage"])].append(row)
metrics = (
"negative_fraction",
"active_0_to_6_fraction",
"above_reference_6_fraction",
"actual_clip_crossing_fraction",
"clip_value_mean",
)
result: list[Row] = []
for (gate, seed, stage), rows in sorted(groups.items()):
total = sum(int(row["element_count"]) for row in rows)
output: Row = {
"gate": gate,
"seed": seed,
"stage": stage,
"element_count": total,
}
for metric in metrics:
valid = [row for row in rows if row[metric] is not None]
output[metric] = (
sum(float(row[metric]) * int(row["element_count"]) for row in valid)
/ sum(int(row["element_count"]) for row in valid)
if valid
else None
)
result.append(output)
return result
def _nonfinite_diagnostics(tables: dict[str, list[Row]]) -> tuple[list[Row], list[Row]]:
"""Return explicit nonfinite cells and an observation-level summary."""
cells: list[Row] = []
observations: dict[tuple[Any, ...], dict[str, Any]] = {}
for table, rows in tables.items():
for row_index, row in enumerate(rows, start=2):
for metric, value in row.items():
if not isinstance(value, float) or math.isfinite(value):
continue
cell = {
"table": table,
"row": row_index,
"gate": row.get("gate"),
"seed": row.get("seed"),
"cutoff": row.get("cutoff"),
"layer": row.get("layer"),
"metric": metric,
"nonfinite_value": "nan" if math.isnan(value) else "inf" if value > 0 else "-inf",
}
cells.append(cell)
key = (
table,
row.get("gate"),
row.get("seed"),
row.get("cutoff"),
row.get("layer"),
)
observation = observations.setdefault(
key,
{
"table": table,
"gate": row.get("gate"),
"seed": row.get("seed"),
"cutoff": row.get("cutoff"),
"layer": row.get("layer"),
"nonfinite_field_count": 0,
"nonfinite_metrics": set(),
},
)
observation["nonfinite_field_count"] += 1
observation["nonfinite_metrics"].add(metric)
summary: list[Row] = []
for observation in observations.values():
summary.append(
{
**observation,
"nonfinite_metrics": ";".join(sorted(observation["nonfinite_metrics"])),
}
)
summary.sort(
key=lambda row: (
str(row["table"]),
str(row["gate"]),
int(row["seed"]) if row["seed"] is not None else -1,
float(row["cutoff"]) if row["cutoff"] is not None else -1.0,
str(row["layer"]),
)
)
return cells, summary
def _report(
sample_hash: str,
accuracy_summary: list[Row],
auc_summary: list[Row],
feature_summary: list[Row],
transfer_summary: list[Row],
gate_summary: list[Row],
nonfinite_cells: list[Row],
nonfinite_summary: list[Row],
) -> str:
lines = [
"# E1 Trained CIFAR-100 Feature Audit",
"",
"Status: complete formal 6-gate x 3-seed matrix.",
"",
"## Frozen Protocol",
"",
"- Exactly 18 `checkpoint_last.pt` files at fixed epoch 100.",
"- `checkpoint_best.pt` is excluded to avoid CIFAR-100 test-selection leakage.",
f"- Every run uses the same 10,000 ordered test samples; index hash `{sample_hash}`.",
"- Butterworth order 4, cutoffs 0, 0.125, 0.25, 0.5, 0.75, and identity endpoint 1.",
"",
"## Frequency-Accuracy AUC",
"",
"| Gate | Seeds | Identity Top-1 | Top-1 AUC |",
"|---|---:|---:|---:|",
]
for row in auc_summary:
lines.append(
f"| {row['gate']} | {row['count']} | {_format(row['identity_top1_mean'])} +/- {_format(row['identity_top1_std'])} | "
f"{_format(row['frequency_accuracy_auc_top1_mean'])} +/- {_format(row['frequency_accuracy_auc_top1_std'])} |"
)
lines.extend(
[
"",
"## Accuracy Curve",
"",
"| Gate | Cutoff | Top-1 | Top-5 |",
"|---|---:|---:|---:|",
]
)
for row in accuracy_summary:
lines.append(
f"| {row['gate']} | {row['cutoff']} | {_format(row['top1_mean'])} +/- {_format(row['top1_std'])} | "
f"{_format(row['top5_mean'])} +/- {_format(row['top5_std'])} |"
)
lines.extend(
[
"",
"## Identity-Input Feature PSD",
"",
"| Gate | Layer | Size | Centroid | High/low | Entropy |",
"|---|---|---|---:|---:|---:|",
]
)
for row in feature_summary:
lines.append(
f"| {row['gate']} | {row['layer']} | {row['height']}x{row['width']} | "
f"{_format(row['spectral_centroid_mean'])} | {_format(row['high_low_ratio_mean'])} | "
f"{_format(row['spectral_entropy_mean'])} |"
)
lines.extend(
[
"",
"## Cutoff 0.25 Feature Transfer",
"",
"| Gate | Layer | Centroid delta | Entropy delta |",
"|---|---|---:|---:|",
]
)
for row in transfer_summary:
lines.append(
f"| {row['gate']} | {row['layer']} | {_format(row['centroid_delta_vs_identity_mean'])} | "
f"{_format(row['entropy_delta_vs_identity_mean'])} |"
)
lines.extend(
[
"",
"## Identity-Input Gate Regions",
"",
"| Gate | Stage | Negative | Active [0,6) | Above 6 | Clip mean | Actual clip crossing |",
"|---|---|---:|---:|---:|---:|---:|",
]
)
for row in gate_summary:
lines.append(
f"| {row['gate']} | {row['stage']} | {_format(row['negative_fraction_mean'])} | "
f"{_format(row['active_0_to_6_fraction_mean'])} | {_format(row['above_reference_6_fraction_mean'])} | "
f"{_format(row['clip_value_mean_mean'])} | {_format(row['actual_clip_crossing_fraction_mean'])} |"
)
lines.extend(
[
"",
"## Nonfinite Diagnostics",
"",
f"- Nonfinite numeric cells: {len(nonfinite_cells)}.",
f"- Affected table observations: {len(nonfinite_summary)}.",
"- Aggregates exclude nonfinite values. If no finite observation remains, the statistic is JSON null and report NA; values are never replaced by zero.",
"",
"| Table | Gate | Seed | Cutoff | Layer | Nonfinite fields |",
"|---|---|---:|---:|---|---|",
]
)
for row in nonfinite_summary:
lines.append(
f"| {row['table']} | {row['gate']} | {row['seed']} | {row['cutoff']} | "
f"{row['layer']} | {row['nonfinite_metrics']} |"
)
lines.extend(
[
"",
"## Interpretation Limits",
"",
"- Stage3 is 2x2 and has no non-DC FFT cell below normalized radius 0.5; its high/low ratio is NA.",
"- Stage4 and pre_classifier are 1x1, so their spatial PSD values are NA.",
"- Frequency-Accuracy AUC measures causal sensitivity to this input filter protocol, not model function frequency.",
"",
]
)
return "\n".join(lines)
def main() -> None:
args = parse_args()
shard_root = args.shard_root.resolve()
output_dir = args.output_dir.resolve()
allowed = Path("/tmp/gmnet_runs/e1_trained_features/full").resolve()
if output_dir != allowed:
raise ValueError(f"formal merged output must be {allowed}")
result_paths = sorted(shard_root.glob("*/results.json"))
if len(result_paths) != 18:
raise ValueError(f"expected 18 shard results, found {len(result_paths)}")
manifests: list[Row] = []
accuracy_rows: list[Row] = []
auc_rows: list[Row] = []
feature_rows: list[Row] = []
gate_rows: list[Row] = []
sample_hashes: set[str] = set()
pairs: set[tuple[str, int]] = set()
for result_path in result_paths:
result = json.loads(result_path.read_text(encoding="utf-8"))
protocol = result["protocol"]
if result["status"] != "full":
raise ValueError(f"non-full shard: {result_path}")
if protocol["checkpoint_name"] != "checkpoint_last.pt":
raise ValueError(f"wrong checkpoint policy: {result_path}")
if protocol["require_epochs_completed"] != 100 or protocol["sample_count"] != 10_000:
raise ValueError(f"wrong frozen protocol: {result_path}")
if protocol["cutoffs"] != [0.0, 0.125, 0.25, 0.5, 0.75, 1.0]:
raise ValueError(f"wrong cutoffs: {result_path}")
sample_hashes.add(protocol["sample_indices_sha256"])
shard = result_path.parent
shard_manifest = _read_csv(shard / "checkpoint_manifest.csv")
if len(shard_manifest) != 1:
raise ValueError(f"expected one checkpoint per shard: {shard}")
manifest = shard_manifest[0]
if manifest["checkpoint_name"] != "checkpoint_last.pt" or manifest["checkpoint_epochs_completed"] != 100:
raise ValueError(f"checkpoint is not fixed epoch100 last: {shard}")
pair = (manifest["gate"], int(manifest["seed"]))
if pair in pairs:
raise ValueError(f"duplicate shard {pair}")
pairs.add(pair)
manifests.append(manifest)
accuracy_rows.extend(_read_csv(shard / "accuracy_curve.csv"))
auc_rows.extend(_read_csv(shard / "accuracy_auc.csv"))
feature_rows.extend(_read_csv(shard / "feature_metrics.csv"))
gate_rows.extend(_read_csv(shard / "gate_regions.csv"))
if pairs != EXPECTED_PAIRS:
raise ValueError(f"checkpoint matrix mismatch; missing={sorted(EXPECTED_PAIRS - pairs)}")
if len(sample_hashes) != 1:
raise ValueError(f"sample subsets differ across shards: {sample_hashes}")
sample_hash = next(iter(sample_hashes))
nonfinite_cells, nonfinite_summary = _nonfinite_diagnostics(
{
"checkpoint_manifest": manifests,
"accuracy_curve": accuracy_rows,
"accuracy_auc": auc_rows,
"feature_metrics": feature_rows,
"gate_regions": gate_rows,
}
)
accuracy_summary = _aggregate(
accuracy_rows, ("gate", "cutoff"), ("top1", "top5")
)
auc_summary = _aggregate(
auc_rows,
("gate",),
("identity_top1", "identity_top5", "frequency_accuracy_auc_top1", "frequency_accuracy_auc_top5"),
)
identity_features = [row for row in feature_rows if row["cutoff"] == 1.0]
feature_summary = _aggregate(
identity_features,
("gate", "layer", "height", "width"),
("spectral_centroid", "high_low_ratio", "spectral_entropy"),
)
feature_summary.sort(key=lambda row: (row["gate"], LAYER_ORDER.index(row["layer"])))
transfer_rows = [row for row in feature_rows if row["cutoff"] == 0.25]
transfer_summary = _aggregate(
transfer_rows,
("gate", "layer"),
("centroid_delta_vs_identity", "entropy_delta_vs_identity"),
)
transfer_summary.sort(key=lambda row: (row["gate"], LAYER_ORDER.index(row["layer"])))
gate_seed_rows = _weighted_gate_seed_rows(gate_rows)
gate_summary = _aggregate(
gate_seed_rows,
("gate", "stage"),
(
"negative_fraction",
"active_0_to_6_fraction",
"above_reference_6_fraction",
"clip_value_mean",
"actual_clip_crossing_fraction",
),
)
output_dir.mkdir(parents=True, exist_ok=True)
_write_csv(output_dir / "checkpoint_manifest.csv", manifests)
_write_csv(output_dir / "accuracy_curve.csv", accuracy_rows)
_write_csv(output_dir / "accuracy_auc.csv", auc_rows)
_write_csv(output_dir / "feature_metrics.csv", feature_rows)
_write_csv(output_dir / "gate_regions.csv", gate_rows)
_write_csv(output_dir / "accuracy_summary.csv", accuracy_summary)
_write_csv(output_dir / "accuracy_auc_summary.csv", auc_summary)
_write_csv(output_dir / "feature_summary.csv", feature_summary)
_write_csv(output_dir / "feature_transfer_summary.csv", transfer_summary)
_write_csv(output_dir / "gate_seed_summary.csv", gate_seed_rows)
_write_csv(output_dir / "gate_summary.csv", gate_summary)
if nonfinite_cells:
_write_csv(output_dir / "nonfinite_cells.csv", nonfinite_cells)
_write_csv(output_dir / "nonfinite_summary.csv", nonfinite_summary)
merged = {
"schema_version": 1,
"experiment_id": "E1-trained-cifar100-feature-audit-formal",
"timestamp_utc": datetime.now(UTC).isoformat(),
"status": "complete",
"checkpoint_policy": "checkpoint_last.pt at fixed epoch 100",
"checkpoint_count": len(manifests),
"sample_count_per_run": 10_000,
"sample_indices_sha256": sample_hash,
"accuracy_auc_summary": auc_summary,
"feature_identity_summary": feature_summary,
"feature_transfer_cutoff_0_25_summary": transfer_summary,
"gate_identity_summary": gate_summary,
"nonfinite_diagnostics": {
"numeric_cell_count": len(nonfinite_cells),
"affected_observation_count": len(nonfinite_summary),
"handling": "Excluded from aggregate statistics; all-nonfinite aggregates are null, never zero.",
"observations": nonfinite_summary,
},
}
(output_dir / "results.json").write_text(
json.dumps(merged, indent=2, allow_nan=False) + "\n", encoding="utf-8"
)
(output_dir / "REPORT.md").write_text(
_report(
sample_hash,
accuracy_summary,
auc_summary,
feature_summary,
transfer_summary,
gate_summary,
nonfinite_cells,
nonfinite_summary,
),
encoding="utf-8",
)
print(f"Merged 18 formal shards: {output_dir}")
if __name__ == "__main__":
main()
|