Gaze-LIPE / scripts /validate_kd_clean_cache.py
thanhhuyvan's picture
Publish KD reproducibility investigation
178f61f
Raw
History Blame Contribute Delete
9.03 kB
"""Read-only structural and provenance validation for a clean KD cache."""
from __future__ import annotations
import argparse
import hashlib
import json
from pathlib import Path
import sys
ROOT = Path(r"E:\Gaze_estimation")
sys.path.insert(0, str(ROOT / ".codex_deps"))
import h5py
import numpy as np
MANIFEST_ROOT = ROOT / "data" / "processed_kd_clean_v1" / "manifests"
CACHE_ROOT = ROOT / "data" / "processed_kd_clean_v1" / "cache"
def file_sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as stream:
for block in iter(lambda: stream.read(1024 * 1024), b""):
digest.update(block)
return digest.hexdigest().upper()
def decode(values: np.ndarray) -> list[str]:
return [value.decode("utf-8") if isinstance(value, bytes) else str(value) for value in values]
def softmax(values: np.ndarray) -> np.ndarray:
shifted = values.astype(np.float64) - values.max(axis=1, keepdims=True)
exp = np.exp(shifted)
return exp / exp.sum(axis=1, keepdims=True)
def vector_from_angles(pitch_deg: np.ndarray, yaw_deg: np.ndarray) -> np.ndarray:
pitch, yaw = np.deg2rad(pitch_deg), np.deg2rad(yaw_deg)
return np.column_stack(
(-np.cos(pitch) * np.sin(yaw), -np.sin(pitch), -np.cos(pitch) * np.cos(yaw))
)
def angular_error(first: np.ndarray, second: np.ndarray) -> np.ndarray:
first = first / np.linalg.norm(first, axis=1, keepdims=True)
second = second / np.linalg.norm(second, axis=1, keepdims=True)
return np.rad2deg(np.arccos(np.clip(np.sum(first * second, axis=1), -1.0, 1.0)))
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--participant", required=True)
parser.add_argument("--tag", required=True)
args = parser.parse_args()
cache_path = CACHE_ROOT / f"{args.participant}.{args.tag}.h5"
processing_path = CACHE_ROOT / f"{args.participant}.{args.tag}.processing.jsonl"
summary_path = CACHE_ROOT / f"{args.participant}.{args.tag}.summary.json"
source_path = MANIFEST_ROOT / f"{args.participant}.source.jsonl"
output_path = CACHE_ROOT / f"{args.participant}.{args.tag}.validation.json"
if output_path.exists():
raise FileExistsError(f"refusing to overwrite validation artifact: {output_path}")
summary = json.loads(summary_path.read_text(encoding="utf-8"))
source_rows = {
row["sample_id"]: row
for row in (json.loads(line) for line in source_path.read_text(encoding="utf-8").splitlines())
}
decisions = [json.loads(line) for line in processing_path.read_text(encoding="utf-8").splitlines()]
errors: list[str] = []
if file_sha256(cache_path) != summary["cache_sha256"]:
errors.append("cache SHA-256 differs from summary")
if file_sha256(processing_path) != summary["processing_manifest_sha256"]:
errors.append("processing manifest SHA-256 differs from summary")
if len(decisions) != summary["source_rows_examined"]:
errors.append("processing-decision count differs from examined-row summary")
if sum(bool(row["accepted"]) for row in decisions) != summary["accepted_rows"]:
errors.append("accepted decision count differs from summary")
required = {
"sample_id", "relative_frame_path", "participant", "day", "frame_id",
"raw_image_sha256", "source_index", "annotation_row", "left_patches",
"right_patches", "landmarks", "left_gaze", "right_gaze",
"left_affine_matrix", "right_affine_matrix", "left_roll_deg", "right_roll_deg",
"teacher_pitch_logits_raw", "teacher_yaw_logits_raw", "teacher_pitch_logits",
"teacher_yaw_logits", "teacher_aligned_vector", "teacher_pitch_deg",
"teacher_yaw_deg", "teacher_error_deg",
}
metrics: dict[str, float | int | bool] = {}
with h5py.File(cache_path, "r") as handle:
missing = sorted(required - set(handle.keys()))
if missing:
errors.append(f"missing datasets: {missing}")
row_counts = {name: int(handle[name].shape[0]) for name in required if name in handle}
if len(set(row_counts.values())) != 1:
errors.append(f"dataset row counts differ: {row_counts}")
count = min(row_counts.values()) if row_counts else 0
if count != summary["accepted_rows"]:
errors.append("H5 row count differs from accepted-row summary")
sample_ids = decode(handle["sample_id"][:])
paths = decode(handle["relative_frame_path"][:])
hashes = decode(handle["raw_image_sha256"][:])
source_indices = handle["source_index"][:]
if len(sample_ids) != len(set(sample_ids)):
errors.append("H5 sample IDs are not unique")
for output_index, sample_id in enumerate(sample_ids):
source = source_rows.get(sample_id)
if source is None:
errors.append(f"H5 sample_id not found in source manifest: {sample_id}")
continue
if source["relative_frame_path"] != paths[output_index]:
errors.append(f"path mismatch at H5 row {output_index}")
if source["raw_image_sha256"] != hashes[output_index]:
errors.append(f"image hash mismatch at H5 row {output_index}")
if source["source_index"] != int(source_indices[output_index]):
errors.append(f"source index mismatch at H5 row {output_index}")
numeric_names = [name for name in required if name in handle and handle[name].dtype.kind not in "OSU"]
if any(not np.isfinite(handle[name][:]).all() for name in numeric_names):
errors.append("one or more numeric datasets contain non-finite values")
if handle["teacher_pitch_logits"].shape[1:] != (90,) or handle["teacher_yaw_logits"].shape[1:] != (90,):
errors.append("aligned teacher logits do not have 90 bins")
if handle["left_patches"].shape[1:] != (4, 16, 16):
errors.append("left patches do not use the declared V16 shape")
if handle["landmarks"].shape[1:] != (478, 2):
errors.append("landmarks do not have shape [N,478,2]")
aligned_pitch_probability = softmax(handle["teacher_pitch_logits"][:])
aligned_yaw_probability = softmax(handle["teacher_yaw_logits"][:])
metrics["max_aligned_pitch_probability_sum_error"] = float(
np.max(np.abs(aligned_pitch_probability.sum(axis=1) - 1.0))
)
metrics["max_aligned_yaw_probability_sum_error"] = float(
np.max(np.abs(aligned_yaw_probability.sum(axis=1) - 1.0))
)
target_deg = np.rad2deg(handle["left_gaze"][:].astype(np.float64))
target_vector = vector_from_angles(target_deg[:, 0], target_deg[:, 1])
aligned_vector = handle["teacher_aligned_vector"][:].astype(np.float64)
recomputed_error = angular_error(aligned_vector, target_vector)
stored_error = handle["teacher_error_deg"][:].astype(np.float64)
metrics["max_teacher_error_recompute_difference_deg"] = float(
np.max(np.abs(recomputed_error - stored_error))
)
metrics["teacher_aligned_error_mean_deg"] = float(recomputed_error.mean())
bin_definition = str(handle.attrs["teacher_bin_centers_deg"])
if bin_definition == "index * 4 - 180":
binwidth, offset = 4.0, 180.0
elif bin_definition == "index * 2 - 90":
binwidth, offset = 2.0, 90.0
else:
errors.append(f"unknown teacher bin definition: {bin_definition}")
binwidth, offset = np.nan, np.nan
raw_pitch = softmax(handle["teacher_pitch_logits_raw"][:]) @ np.arange(90) * binwidth - offset
raw_yaw = softmax(handle["teacher_yaw_logits_raw"][:]) @ np.arange(90) * binwidth - offset
raw_vector = vector_from_angles(raw_pitch, raw_yaw)
metrics["teacher_raw_vs_roll_corrected_target_error_mean_deg"] = float(
angular_error(raw_vector, target_vector).mean()
)
metrics["teacher_alignment_error_change_deg"] = float(
metrics["teacher_aligned_error_mean_deg"]
- metrics["teacher_raw_vs_roll_corrected_target_error_mean_deg"]
)
teacher_audit = json.loads(handle.attrs["teacher_loader_audit"])
if teacher_audit["mapped_tensor_count"] != teacher_audit["checkpoint_tensor_count"]:
errors.append("embedded strict-loader audit does not map every checkpoint tensor")
report = {
"schema": "mpiigaze-kd-clean-cache-validation-v1",
"participant": args.participant,
"tag": args.tag,
"cache_path": str(cache_path.resolve()),
"cache_sha256": file_sha256(cache_path),
"rows": summary["accepted_rows"],
"metrics": metrics,
"errors": errors,
"pass": not errors,
}
output_path.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8", newline="\n")
print(json.dumps(report, indent=2))
if errors:
raise SystemExit(1)
if __name__ == "__main__":
main()