File size: 10,106 Bytes
178f61f | 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 | """Batched equivalent of generate_kd_clean_cache.py for CPU throughput.
Landmark detection and every preprocessing operation remain row-serial and use the
same functions as the audited generator. Only strict-teacher forward calls are
grouped into batches; output rows retain manifest order. Existing files are never
overwritten.
"""
from __future__ import annotations
import argparse
from datetime import datetime, timezone
import json
from pathlib import Path
import sys
ROOT = Path(r"E:\Gaze_estimation")
sys.path.insert(0, str(ROOT / ".codex_deps"))
sys.path.insert(0, str(ROOT))
import cv2
import numpy as np
import torch
import scripts.generate_kd_clean_cache as base
from src.models.teacher_strict import audit_to_dict, load_teacher_model_strict
from src.utils.preprocess import GazePreprocessor
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--participant", required=True, choices=[f"p{i:02d}" for i in range(15)])
parser.add_argument("--tag", default="official448_full")
parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
parser.add_argument("--teacher-batch-size", type=int, default=16)
parser.add_argument("--max-source-rows", type=int)
parser.add_argument("--max-accepted", type=int)
args = parser.parse_args()
if args.teacher_batch_size < 1:
raise ValueError("teacher batch size must be positive")
manifest_path = base.MANIFEST_ROOT / f"{args.participant}.source.jsonl"
summary_path = base.MANIFEST_ROOT / f"{args.participant}.source.summary.json"
validation_path = base.MANIFEST_ROOT / f"{args.participant}.source.validation.json"
summary = json.loads(summary_path.read_text(encoding="utf-8"))
validation = json.loads(validation_path.read_text(encoding="utf-8"))
if not validation.get("pass") or base.file_sha256(manifest_path) != summary["manifest_sha256"]:
raise RuntimeError("source manifest is not validated or its hash changed")
base.OUTPUT_ROOT.mkdir(parents=True, exist_ok=True)
output_path = base.OUTPUT_ROOT / f"{args.participant}.{args.tag}.h5"
decision_path = base.OUTPUT_ROOT / f"{args.participant}.{args.tag}.processing.jsonl"
summary_output = base.OUTPUT_ROOT / f"{args.participant}.{args.tag}.summary.json"
for path in (output_path, decision_path, summary_output):
if path.exists():
raise FileExistsError(f"refusing to overwrite clean artifact: {path}")
teacher, teacher_audit = load_teacher_model_strict(base.CHECKPOINT, device=args.device)
preprocessor = GazePreprocessor(model_path=str(base.LANDMARK_MODEL))
accepted, decisions, pending = [], [], []
def flush_teacher():
if not pending:
return
tensor = torch.from_numpy(np.stack([item[1] for item in pending])).to(args.device)
with torch.inference_mode():
raw_pitch, raw_yaw = teacher(tensor)
raw_pitch = raw_pitch.detach().cpu().numpy().astype(np.float32)
raw_yaw = raw_yaw.detach().cpu().numpy().astype(np.float32)
for batch_index, (output_index, _) in enumerate(pending):
record = accepted[output_index]
pitch_logits, yaw_logits = raw_pitch[batch_index], raw_yaw[batch_index]
aligned_pitch, aligned_yaw, teacher_vector = base.roll_align_teacher_distribution(
pitch_logits, yaw_logits, float(record["left_roll_deg"])
)
teacher_pitch, teacher_yaw = base.angles_from_vectors(teacher_vector[None, :])
record.update({
"teacher_pitch_logits_raw": pitch_logits,
"teacher_yaw_logits_raw": yaw_logits,
"teacher_pitch_logits": aligned_pitch,
"teacher_yaw_logits": aligned_yaw,
"teacher_aligned_vector": teacher_vector.astype(np.float32),
"teacher_pitch_deg": np.float32(teacher_pitch[0]),
"teacher_yaw_deg": np.float32(teacher_yaw[0]),
"teacher_error_deg": np.float32(base.angular_error_deg(teacher_vector, record.pop("_left_aligned_vector"))),
})
pending.clear()
examined = 0
for encoded in manifest_path.read_text(encoding="utf-8").splitlines():
if args.max_source_rows is not None and examined >= args.max_source_rows:
break
if args.max_accepted is not None and len(accepted) >= args.max_accepted:
break
row = json.loads(encoded); examined += 1
image_path = base.ORIGINAL_ROOT / Path(row["relative_frame_path"])
if base.file_sha256(image_path) != row["raw_image_sha256"]:
base.append_or_reject(decisions, row, False, "raw_image_hash_mismatch", None); continue
frame = cv2.imread(str(image_path))
if frame is None:
base.append_or_reject(decisions, row, False, "opencv_decode_failed", None); continue
landmarks = preprocessor.get_landmarks(frame)
if landmarks is None:
base.append_or_reject(decisions, row, False, "face_landmarks_not_found", None); continue
crop = base.face_crop(frame, landmarks)
if crop is None:
base.append_or_reject(decisions, row, False, "teacher_face_crop_empty", None); continue
left_eye, left_angle, left_matrix = base.normalize_eye_with_matrix(frame, landmarks, preprocessor.LEFT_CORNERS, preprocessor)
right_eye, right_angle, right_matrix = base.normalize_eye_with_matrix(frame, landmarks, preprocessor.RIGHT_CORNERS, preprocessor)
landmark_array = np.asarray([[point.x, point.y] for point in landmarks], dtype=np.float32)
left_center = landmark_array[preprocessor.LEFT_CORNERS].mean(axis=0)
right_center = landmark_array[preprocessor.RIGHT_CORNERS].mean(axis=0)
target = np.asarray(row["target_ccs"], dtype=np.float64)
left_vector = target - np.asarray(row["left_eye_ccs"], dtype=np.float64)
right_vector = target - np.asarray(row["right_eye_ccs"], dtype=np.float64)
left_vector /= np.linalg.norm(left_vector); right_vector /= np.linalg.norm(right_vector)
left_aligned = base.rotation_matrix_z(left_angle) @ left_vector
right_aligned = base.rotation_matrix_z(right_angle) @ right_vector
record = dict(row)
record.update({
"left_patches": preprocessor.extract_patches(left_eye, patch_size=16),
"right_patches": preprocessor.extract_patches(right_eye, patch_size=16),
"landmarks": landmark_array - (left_center + right_center) / 2.0,
"left_gaze": np.asarray(preprocessor.gaze_3d_to_mag(left_aligned), dtype=np.float32),
"right_gaze": np.asarray(preprocessor.gaze_3d_to_mag(right_aligned), dtype=np.float32),
"left_affine_matrix": left_matrix, "right_affine_matrix": right_matrix,
"left_roll_deg": np.float32(left_angle), "right_roll_deg": np.float32(right_angle),
"_left_aligned_vector": left_aligned,
})
output_index = len(accepted); accepted.append(record)
pending.append((output_index, crop))
base.append_or_reject(decisions, row, True, "", output_index)
if len(pending) >= args.teacher_batch_size:
flush_teacher()
flush_teacher()
if not accepted:
raise RuntimeError("no rows were accepted; no cache written")
attributes = {
"schema": "mpiigaze-kd-clean-cache-v2-official448",
"created_utc": datetime.now(timezone.utc).isoformat(), "git_commit": base.git_commit(),
"participant": args.participant, "partial": args.max_source_rows is not None or args.max_accepted is not None,
"source_manifest_sha256": summary["manifest_sha256"],
"teacher_loader_audit": audit_to_dict(teacher_audit),
"teacher_checkpoint_sha256": teacher_audit.checkpoint_sha256,
"generator_script_sha256": base.file_sha256(Path(__file__)),
"serial_reference_script_sha256": base.file_sha256(Path(base.__file__)),
"teacher_strict_script_sha256": base.file_sha256(ROOT / "src" / "models" / "teacher_strict.py"),
"preprocess_script_sha256": base.file_sha256(ROOT / "src" / "utils" / "preprocess.py"),
"landmark_model_sha256": base.file_sha256(base.LANDMARK_MODEL),
"patch_size": 16, "teacher_input_size": 448, "teacher_batch_size": args.teacher_batch_size,
"teacher_bin_centers_deg": "index * 4 - 180",
"teacher_protocol_basis": "recovered checkpoint key layout plus documented Gaze360 ResNet configuration",
"training_target": "left-eye gaze rotated by left eye affine roll angle",
"teacher_alignment": "independent pitch/yaw joint distribution rotated by same left-eye Z roll; marginals rebinned to 90 bins",
}
base.create_h5(output_path, accepted, attributes)
with decision_path.open("x", encoding="utf-8", newline="\n") as stream:
for decision in decisions:
stream.write(json.dumps(decision, sort_keys=True, separators=(",", ":")) + "\n")
errors = np.asarray([row["teacher_error_deg"] for row in accepted])
run_summary = {
"schema": "mpiigaze-kd-clean-cache-summary-v1", "participant": args.participant,
"source_rows_examined": examined, "accepted_rows": len(accepted), "rejected_rows": examined - len(accepted),
"cache_path": str(output_path.resolve()), "cache_sha256": base.file_sha256(output_path),
"processing_manifest_path": str(decision_path.resolve()), "processing_manifest_sha256": base.file_sha256(decision_path),
"source_manifest_sha256": summary["manifest_sha256"], "teacher_checkpoint_sha256": teacher_audit.checkpoint_sha256,
"strict_loader_inference_sha256": teacher_audit.inference_sha256,
"teacher_error_mean_deg": float(errors.mean()), "teacher_error_median_deg": float(np.median(errors)),
"teacher_batch_size": args.teacher_batch_size,
}
summary_output.write_text(json.dumps(run_summary, indent=2) + "\n", encoding="utf-8", newline="\n")
print(json.dumps(run_summary, indent=2))
if __name__ == "__main__":
main()
|