Ethanfel
011c915fa7
Publish the existing H3 merge/export and autotuner correctness fixes, offline character and AV calibration tools, provenance audits, and regression coverage. Add a dictionary-preserving storage hook for offline mapped TIES exports. Keep scoring weights and defaults unchanged, retain NP/CT as opt-in, and exclude the rejected F08 search-admission prototype. Validate the exact staged tree: 1052 Python tests, 69 subtests, 3 JS tests, and real ComfyUI CPU export round trips pass; 3 CUDA-only tests skipped.
100 lines
5.1 KiB
Python
100 lines
5.1 KiB
Python
"""Reconstruct Synchformer model inputs from original clips without GPU inference."""
|
|
from __future__ import annotations
|
|
|
|
import argparse
|
|
import hashlib
|
|
from pathlib import Path
|
|
|
|
try:
|
|
from .h3_synchformer import (CFG, CKPT, POLICY, REPO, ROOT, audio_control, control_gates,
|
|
decode_audio, decode_video, digest, isolate, now, read_json, save_new, tensor_record)
|
|
except ImportError:
|
|
from h3_synchformer import (CFG, CKPT, POLICY, REPO, ROOT, audio_control, control_gates,
|
|
decode_audio, decode_video, digest, isolate, now, read_json, save_new, tensor_record)
|
|
|
|
|
|
def audit(name):
|
|
import numpy as np
|
|
import torch
|
|
import torchaudio
|
|
if not name or Path(name).name != name or name in (".", ".."):
|
|
raise ValueError("Simple run name required")
|
|
isolate()
|
|
torch.set_num_threads(4)
|
|
run = ROOT / "runs" / name
|
|
start, complete = read_json(run / "start.json"), read_json(run / "complete.json")
|
|
policy = read_json(POLICY)
|
|
manifest_path = REPO / policy["manifest"]
|
|
manifest = read_json(manifest_path)
|
|
pins = {str(POLICY): start["policy_sha256"], str(manifest_path): start["manifest_sha256"],
|
|
str(ROOT / "setup-receipt.json"): start["receipt_sha256"],
|
|
str(REPO / "scripts/h3_synchformer.py"): start["runner_sha256"],
|
|
str(REPO / "scripts/h3_local_av_evaluator.py"): start["audio_helper_sha256"]}
|
|
for path, sha in pins.items():
|
|
if digest(path) != sha:
|
|
raise ValueError(f"Changed pinned artifact: {path}")
|
|
receipt = read_json(ROOT / "setup-receipt.json")
|
|
pins.update({str(CFG): receipt["config_sha256"], str(CKPT): receipt["checkpoint_sha256"]})
|
|
for path in (CFG, CKPT):
|
|
if digest(path) != pins[str(path)]:
|
|
raise ValueError("Changed model asset")
|
|
mel = torchaudio.transforms.MelSpectrogram(sample_rate=16000, win_length=400,
|
|
hop_length=160, n_fft=1024, n_mels=128)
|
|
by_case = {c["case_id"]: c for c in manifest["cases"]}
|
|
reconstructed = {}
|
|
rows = []
|
|
for case_id, condition, repeat in start["jobs"]:
|
|
case = by_case[case_id]
|
|
if case_id not in reconstructed:
|
|
if digest(case["path"]) != case["sha256"]:
|
|
raise ValueError("Original media changed")
|
|
pins[case["path"]] = case["sha256"]
|
|
video, vi = decode_video(case["path"])
|
|
if vi["source_frame_indices"] != [round(i * 24 / 25) for i in range(125)]:
|
|
raise ValueError("PTS mapping does not match independently calculated 24-to-25-fps mapping")
|
|
audio, ai = decode_audio(case["path"])
|
|
reconstructed[case_id] = video, vi, audio, ai
|
|
video, vi, audio, ai = reconstructed[case_id]
|
|
# Reconstruct from declared ranges, without invoking Synchformer's Compose.
|
|
controlled = audio_control(audio, condition)
|
|
v = torch.stack([video[2 + i * 8:18 + i * 8, :, 16:240, 101:325] for i in range(14)])
|
|
v = ((v.half().div(255.) - .5) / .5).unsqueeze(0)
|
|
a = torch.from_numpy(np.stack([controlled[1280 + i * 5120:11520 + i * 5120] for i in range(14)]))
|
|
a = torch.log(mel(a) + 1e-6)
|
|
if a.shape != (14, 128, 65):
|
|
raise ValueError("Unexpected raw spectrogram shape")
|
|
a = torch.nn.functional.pad(a, (0, 1), value=0.)
|
|
a = ((a + 4.2677393) / (2 * 4.5689974)).unsqueeze(0).unsqueeze(2)
|
|
stem = f"{case_id}-{condition}-r{repeat}"
|
|
path, ipath = run / (stem + ".json"), run / (stem + ".input.json")
|
|
row, before = read_json(path), read_json(ipath)
|
|
if any(row[k] != value for k, value in before.items()):
|
|
raise ValueError("Input record altered after inference")
|
|
info = row["input"]
|
|
if (info["video_tensor"] != tensor_record(v) or info["audio_tensor"] != tensor_record(a)
|
|
or info["waveform_sha256"] != hashlib.sha256(controlled.tobytes()).hexdigest()
|
|
or info["audio_peak"] != float(np.max(np.abs(controlled)))
|
|
or info["video"] != vi or info["audio"] != ai
|
|
or info["segment_video_ranges"] != [[2 + i * 8, 18 + i * 8] for i in range(14)]
|
|
or info["segment_audio_ranges"] != [[1280 + i * 5120, 11520 + i * 5120] for i in range(14)]):
|
|
raise ValueError(f"Reconstructed model input mismatch: {stem}")
|
|
pins[str(path)] = digest(path)
|
|
pins[str(ipath)] = digest(ipath)
|
|
rows.append(row)
|
|
gates = control_gates(rows)
|
|
if gates != complete["gates"] or len(rows) != complete["cases"]:
|
|
raise ValueError("Completion summary does not match raw evidence")
|
|
for path in (run / "start.json", run / "complete.json", Path(__file__).resolve()):
|
|
pins[str(path)] = digest(path)
|
|
report = {"completed_at_utc": now(), "cases": len(rows), "gates": gates,
|
|
"independently_reconstructed_model_inputs": True, "gpu_inference_rerun": False,
|
|
"original_media_unchanged": True, "artifact_sha256": pins}
|
|
save_new(run / "control-audit.json", report)
|
|
print({k: v for k, v in report.items() if k != "artifact_sha256"})
|
|
|
|
|
|
if __name__ == "__main__":
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
parser.add_argument("run_name")
|
|
audit(parser.parse_args().run_name)
|