Ethanfel
cb92bd4f6e
Correct conflict-sparsification skips and scoring, preserve experimental precision, add native-factor NP/CT exports, and harden mixed H3/CLIP detection and external evaluation. Include regression coverage and reproducible local H3 benchmark, review, and waveform diagnostics with evidence records. Keep experimental preference ranking and package version unchanged.
69 lines
2.9 KiB
Python
69 lines
2.9 KiB
Python
"""Audit completed frozen H3 jobs once, without queue polling or quality labels."""
|
|
import argparse
|
|
import json
|
|
from pathlib import Path
|
|
import subprocess
|
|
import sys
|
|
|
|
try:
|
|
from .h3_av_review import output_video
|
|
from .h3_benchmark import digest, save_new
|
|
from .h3_study_record import collect_render_run
|
|
except ImportError:
|
|
from h3_av_review import output_video
|
|
from h3_benchmark import digest, save_new
|
|
from h3_study_record import collect_render_run
|
|
|
|
|
|
def audit(plan_path, stage, seed, out=None):
|
|
plan = json.loads(plan_path.read_text())
|
|
if out is not None and out.exists():
|
|
raise FileExistsError(out)
|
|
records, missing = [], []
|
|
for job in plan["jobs"]:
|
|
if job["stage"] != stage or job["seed"] != seed:
|
|
continue
|
|
run = Path(plan["root"]) / job["id"]
|
|
history_path = run / "history.json"
|
|
if not history_path.exists():
|
|
missing.append(job["id"])
|
|
continue
|
|
history = json.loads(history_path.read_text())
|
|
video = output_video(history)
|
|
audit_dir = run / "media-audit"
|
|
if not audit_dir.exists():
|
|
subprocess.run([sys.executable, str(Path(__file__).with_name("h3_video_metrics.py")),
|
|
"--video", str(video), "--out", str(audit_dir)], check=True,
|
|
stdout=subprocess.DEVNULL)
|
|
# Existing incomplete/changed audits are rejected, not overwritten.
|
|
record = collect_render_run(run)
|
|
recorded_graph = history["prompt"][2]
|
|
prepared_graph = json.loads((run / "prompt_api.json").read_text())
|
|
if recorded_graph != prepared_graph:
|
|
raise ValueError(f"Executed graph differs from prepared graph: {run}")
|
|
records.append(record)
|
|
print(json.dumps(dict(audited=job["id"], frames=record["metrics"]["decoded_frames"],
|
|
audio_clipped_fraction=record["metrics"]["audio"]["clipped_fraction"])), flush=True)
|
|
if not records and not missing:
|
|
raise ValueError("No matching frozen cases")
|
|
record = dict(plan_sha256=digest(plan_path), stage=stage, seed=seed, renders=records,
|
|
incomplete_jobs=missing, quality_labels=[],
|
|
note="Whole-clip technical diagnostics only; no full-motion or listening judgment.")
|
|
if out is not None:
|
|
save_new(out, record)
|
|
return dict(audited=len(records), incomplete=len(missing), record=str(out) if out else None)
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
parser.add_argument("--plan", type=Path, required=True)
|
|
parser.add_argument("--stage", choices=("calibration", "heldout"), default="calibration")
|
|
parser.add_argument("--seed", type=int, default=2026090803)
|
|
parser.add_argument("--out", type=Path)
|
|
args = parser.parse_args()
|
|
print(json.dumps(audit(args.plan, args.stage, args.seed, args.out)))
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|