File size: 57,814 Bytes
697d349 a8124a8 697d349 a8124a8 697d349 60a2fc8 697d349 a8124a8 697d349 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 | #!/usr/bin/env python3
"""
Single-episode diagnostics for the Xperience-10M task suite artifacts.
This script is intentionally artifact-driven. It consumes the already exported
one-episode shared feature table and prediction files, validates their shape and
hashes, and writes diagnostics that can be manually traced back to those inputs.
It does not invent labels or claim multi-episode generalization.
"""
from __future__ import annotations
import argparse
import csv
import hashlib
import html
import json
import math
import sys
from collections import OrderedDict
from pathlib import Path
from typing import Iterable
import numpy as np
TASKS = [
"timeline_action",
"timeline_subtask",
"transition_detection",
"next_action",
"hand_trajectory_forecast",
"contact_prediction",
"object_relevance",
"caption_grounding",
"cross_modal_retrieval",
"modality_reconstruction",
"temporal_order",
"misalignment_detection",
]
TASK_DISPLAY = {
"timeline_action": "Current Action Recognition",
"timeline_subtask": "Current Subtask Recognition",
"transition_detection": "Action Transition Detection",
"next_action": "Next-Action Prediction",
"hand_trajectory_forecast": "Future Hand Motion Forecasting",
"contact_prediction": "Contact State Prediction",
"object_relevance": "Relevant Object Prediction",
"caption_grounding": "Language-to-Time Grounding",
"cross_modal_retrieval": "Cross-Modal Window Retrieval",
"modality_reconstruction": "Sensor-to-Visual Reconstruction",
"temporal_order": "Temporal Order Verification",
"misalignment_detection": "Cross-Modal Misalignment Detection",
}
GROUP_DISPLAY = {
"all_features": "All Features",
"video": "Video",
"depth": "Depth",
"pose_slam": "Pose + SLAM",
"motion_capture": "Motion Capture",
"inertial": "Inertial",
"audio": "Audio",
"language": "Language",
"no_language": "All Except Language",
"motion_pose_inertial": "Motion + Pose + IMU",
}
def parse_args() -> argparse.Namespace:
workspace_default = Path(__file__).resolve().parents[1]
parser = argparse.ArgumentParser(description="Run single-episode diagnostics on real exported artifacts.")
parser.add_argument("--workspace", type=Path, default=workspace_default)
parser.add_argument(
"--suite-dir",
type=Path,
default=workspace_default / "results/episode_task_suite",
help="Existing single-episode task-suite artifact directory.",
)
parser.add_argument(
"--output-dir",
type=Path,
default=workspace_default / "results/single_episode_diagnostics",
help="Where to write new diagnostics. Existing task-suite outputs are not overwritten.",
)
parser.add_argument("--test-fraction", type=float, default=0.30)
parser.add_argument("--future-offset-windows", type=int, default=4)
parser.add_argument("--misalignment-shift-windows", type=int, default=8)
parser.add_argument("--ridge-l2", type=float, default=10.0)
parser.add_argument(
"--annotation",
type=Path,
default=None,
help="Optional raw annotation.hdf5. When provided, object relevance labels are exported from caption_frame_info_map.",
)
parser.add_argument(
"--homie-toolkit",
type=Path,
default=None,
help="Optional HOMIE-toolkit path. If omitted, inferred from --annotation when possible.",
)
return parser.parse_args()
def write_json(path: Path, data: dict | list) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(data, indent=2, ensure_ascii=False), encoding="utf-8")
def write_text(path: Path, text: str) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(text, encoding="utf-8")
def write_csv(path: Path, rows: list[dict], fieldnames: list[str] | None = None) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
if fieldnames is None:
keys: OrderedDict[str, None] = OrderedDict()
for row in rows:
for key in row:
keys.setdefault(key, None)
fieldnames = list(keys.keys())
with path.open("w", newline="", encoding="utf-8") as fp:
writer = csv.DictWriter(fp, fieldnames=fieldnames, lineterminator="\n")
writer.writeheader()
for row in rows:
writer.writerow({k: row.get(k, "") for k in fieldnames})
def read_csv(path: Path) -> list[dict]:
with path.open(newline="", encoding="utf-8") as fp:
return list(csv.DictReader(fp))
def sha256(path: Path) -> str:
h = hashlib.sha256()
with path.open("rb") as fp:
for chunk in iter(lambda: fp.read(1024 * 1024), b""):
h.update(chunk)
return h.hexdigest()
def public_artifact_path(path: Path, repo_root: Path) -> str:
path = path.resolve()
try:
return str(path.relative_to(repo_root))
except ValueError:
pass
if path.name == "annotation.hdf5":
return "external_raw_sample/annotation.hdf5"
return path.name
def public_source_reference(value: object) -> object:
if value in (None, ""):
return value
text = str(value)
if text.startswith("/") or "/" + "Users/" in text or "/" + "private/" in text:
path = Path(text)
if path.name == "annotation.hdf5":
return "external_raw_sample/annotation.hdf5"
return path.name
return text
def load_inputs(suite_dir: Path) -> tuple[np.ndarray, np.ndarray, np.ndarray, list[dict], list[dict], dict]:
npz_path = suite_dir / "shared_windows.npz"
windows_path = suite_dir / "windows.csv"
manifest_path = suite_dir / "feature_manifest.json"
summary_path = suite_dir / "summary_report.json"
required = [npz_path, windows_path, manifest_path, summary_path]
missing = [str(p) for p in required if not p.exists()]
if missing:
raise FileNotFoundError(f"Missing required input artifacts: {missing}")
npz = np.load(npz_path)
X = np.asarray(npz["X"], dtype=np.float32)
starts = np.asarray(npz["starts"], dtype=np.int64)
ends = np.asarray(npz["ends"], dtype=np.int64)
windows = read_csv(windows_path)
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
summary = json.loads(summary_path.read_text(encoding="utf-8"))
if X.ndim != 2:
raise ValueError(f"Expected X to be 2-D, got shape {X.shape}")
if len(windows) != X.shape[0]:
raise ValueError(f"windows.csv rows ({len(windows)}) do not match X rows ({X.shape[0]})")
if len(starts) != X.shape[0] or len(ends) != X.shape[0]:
raise ValueError("starts/ends arrays do not match X rows")
for i, row in enumerate(windows):
if int(row["start_frame"]) != int(starts[i]) or int(row["end_frame"]) != int(ends[i]):
raise ValueError(f"Window start/end mismatch at row {i}")
cursor = 0
for block in manifest:
start, end, dim = int(block["start"]), int(block["end"]), int(block["dim"])
if start != cursor or end <= start or end - start != dim:
raise ValueError(f"Feature manifest has a gap, overlap, or bad dim at block {block}")
cursor = end
if cursor != X.shape[1]:
raise ValueError(f"Feature manifest ends at {cursor}, but X has {X.shape[1]} columns")
return X, starts, ends, windows, manifest, summary
def chronological_split(n: int, test_fraction: float) -> tuple[np.ndarray, np.ndarray]:
if n < 2:
raise ValueError("Need at least two samples for a chronological split.")
split = int(round(n * (1.0 - test_fraction)))
split = max(1, min(split, n - 1))
return np.arange(split, dtype=np.int64), np.arange(split, n, dtype=np.int64)
def block_indices(manifest: list[dict], include: Iterable[str] | None = None, exclude: Iterable[str] | None = None) -> np.ndarray:
include = list(include or [])
exclude = list(exclude or [])
idxs: list[int] = []
for block in manifest:
name = str(block["name"])
if include and not any(name == p or name.startswith(p) for p in include):
continue
if exclude and any(name == p or name.startswith(p) for p in exclude):
continue
idxs.extend(range(int(block["start"]), int(block["end"])))
return np.asarray(idxs, dtype=np.int64)
def modality_groups(manifest: list[dict]) -> dict[str, np.ndarray]:
all_idx = block_indices(manifest)
language = block_indices(manifest, ["caption_objects_interaction_text"])
groups = {
"all_features": all_idx,
"video": block_indices(manifest, ["video_"]),
"depth": block_indices(manifest, ["depth_confidence"]),
"pose_slam": block_indices(manifest, ["camera_translation", "camera_rotation_matrix", "slam_point_cloud", "calibration"]),
"motion_capture": block_indices(manifest, ["hand_left_joints", "hand_right_joints", "body_joints", "body_contacts"]),
"inertial": block_indices(manifest, ["imu_accel_gyro"]),
"audio": block_indices(manifest, ["audio_"]),
"language": language,
"no_language": np.setdiff1d(all_idx, language),
}
return {name: idx for name, idx in groups.items() if len(idx) > 0}
def encode_labels(labels: Iterable[str]) -> tuple[np.ndarray, list[str]]:
seen: OrderedDict[str, int] = OrderedDict()
encoded = []
for label in labels:
label = str(label)
if label not in seen:
seen[label] = len(seen)
encoded.append(seen[label])
return np.asarray(encoded, dtype=np.int64), list(seen.keys())
def extract_objects(info: dict) -> list[str]:
objects = info.get("objects")
if isinstance(objects, list):
return [str(x).strip() for x in objects if str(x).strip()]
if objects:
return [str(objects).strip()]
return []
def infer_homie_toolkit(annotation: Path, explicit: Path | None) -> Path | None:
if explicit is not None:
return explicit
annotation = annotation.resolve()
candidates = []
for parent in annotation.parents:
candidates.append(parent / "HOMIE-toolkit")
for candidate in candidates:
if candidate.exists():
return candidate
return None
def load_object_targets_from_annotation(annotation: Path, windows: list[dict], toolkit: Path | None) -> dict:
annotation = annotation.resolve()
if not annotation.exists():
raise FileNotFoundError(annotation)
toolkit = infer_homie_toolkit(annotation, toolkit)
if toolkit is None or not toolkit.exists():
raise FileNotFoundError(f"HOMIE-toolkit not found for annotation {annotation}")
sys.path.insert(0, str(toolkit))
from data_loader import load_from_annotation_hdf5
ann = load_from_annotation_hdf5(annotation, 0, None, load_slam_point_cloud=False)
frame_info = ann["caption_frame_info_map"]
vocab: OrderedDict[str, int] = OrderedDict()
labels: list[list[str]] = []
rows_out: list[dict] = []
for row in windows:
counts: OrderedDict[str, int] = OrderedDict()
for frame in range(int(row["start_frame"]), int(row["end_frame"]) + 1):
for obj in extract_objects(frame_info.get(frame, {})):
counts[obj] = counts.get(obj, 0) + 1
objects = list(counts.keys())
for obj in objects:
if obj not in vocab:
vocab[obj] = len(vocab)
labels.append(objects)
rows_out.append({
"window_index": int(row["window_index"]),
"start_frame": int(row["start_frame"]),
"end_frame": int(row["end_frame"]),
"center_frame": int(row["center_frame"]),
"objects": "|".join(objects),
"object_count": int(len(objects)),
})
if not vocab:
raise ValueError("No object labels found in annotation caption_frame_info_map.")
Y = np.zeros((len(windows), len(vocab)), dtype=np.float32)
for i, objects in enumerate(labels):
for obj in objects:
Y[i, vocab[obj]] = 1.0
return {
"Y": Y,
"labels": labels,
"vocab": list(vocab.keys()),
"rows": rows_out,
"annotation": "external_raw_sample/annotation.hdf5",
"toolkit": "HOMIE-toolkit",
"source_note": (
"Object labels were exported from a raw Xperience-10M sample annotation. "
"The public artifact stores source type and hash instead of machine-specific file paths."
),
}
def standardize(train: np.ndarray, test: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
mean = train.mean(axis=0, keepdims=True)
std = train.std(axis=0, keepdims=True)
std[std < 1e-6] = 1.0
return (train - mean) / std, (test - mean) / std
def standardize_train_apply(train: np.ndarray, *arrays: np.ndarray) -> list[np.ndarray]:
mean = train.mean(axis=0, keepdims=True)
std = train.std(axis=0, keepdims=True)
std[std < 1e-6] = 1.0
return [(arr - mean) / std for arr in arrays]
def ridge_predict(X_train: np.ndarray, Y_train: np.ndarray, X_test: np.ndarray, l2: float) -> np.ndarray:
X_train = np.asarray(X_train, dtype=np.float32)
X_test = np.asarray(X_test, dtype=np.float32)
Y_train = np.asarray(Y_train, dtype=np.float32)
Xb = np.concatenate([X_train, np.ones((X_train.shape[0], 1), dtype=np.float32)], axis=1)
Xtb = np.concatenate([X_test, np.ones((X_test.shape[0], 1), dtype=np.float32)], axis=1)
if Xb.shape[0] <= Xb.shape[1]:
K = Xb @ Xb.T
K.flat[:: K.shape[0] + 1] += l2
alpha = np.linalg.solve(K, Y_train)
return Xtb @ Xb.T @ alpha
A = Xb.T @ Xb
A.flat[:: A.shape[0] + 1] += l2
W = np.linalg.solve(A, Xb.T @ Y_train)
return Xtb @ W
def classification_metrics(y_true: np.ndarray, y_pred: np.ndarray) -> dict:
classes = np.unique(np.concatenate([y_true, y_pred]))
f1s = []
recalls = []
for cls in classes:
tp = int(((y_true == cls) & (y_pred == cls)).sum())
fp = int(((y_true != cls) & (y_pred == cls)).sum())
fn = int(((y_true == cls) & (y_pred != cls)).sum())
precision = tp / (tp + fp) if tp + fp else 0.0
recall = tp / (tp + fn) if tp + fn else 0.0
f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0.0
f1s.append(f1)
recalls.append(recall)
return {
"accuracy": float((y_true == y_pred).mean()) if len(y_true) else 0.0,
"macro_f1": float(np.mean(f1s)) if f1s else 0.0,
"balanced_accuracy": float(np.mean(recalls)) if recalls else 0.0,
}
def multilabel_metrics(Y_true: np.ndarray, Y_pred: np.ndarray) -> dict:
Y_true = Y_true.astype(np.int64)
Y_pred = Y_pred.astype(np.int64)
tp = int(((Y_true == 1) & (Y_pred == 1)).sum())
fp = int(((Y_true == 0) & (Y_pred == 1)).sum())
fn = int(((Y_true == 1) & (Y_pred == 0)).sum())
precision = tp / (tp + fp) if tp + fp else 0.0
recall = tp / (tp + fn) if tp + fn else 0.0
micro_f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0.0
per_f1 = []
for j in range(Y_true.shape[1]):
tpj = int(((Y_true[:, j] == 1) & (Y_pred[:, j] == 1)).sum())
fpj = int(((Y_true[:, j] == 0) & (Y_pred[:, j] == 1)).sum())
fnj = int(((Y_true[:, j] == 1) & (Y_pred[:, j] == 0)).sum())
pj = tpj / (tpj + fpj) if tpj + fpj else 0.0
rj = tpj / (tpj + fnj) if tpj + fnj else 0.0
per_f1.append(2 * pj * rj / (pj + rj) if pj + rj else 0.0)
return {
"micro_f1": float(micro_f1),
"macro_f1": float(np.mean(per_f1)) if per_f1 else 0.0,
"exact_match": float(np.mean(np.all(Y_true == Y_pred, axis=1))) if len(Y_true) else 0.0,
"precision": float(precision),
"recall": float(recall),
}
def regression_metrics(y_true: np.ndarray, y_pred: np.ndarray) -> dict:
err = y_pred - y_true
mse = float(np.mean(err ** 2))
mae = float(np.mean(np.abs(err)))
denom = float(np.sum((y_true - y_true.mean(axis=0, keepdims=True)) ** 2))
r2 = 1.0 - float(np.sum(err ** 2)) / denom if denom > 1e-12 else 0.0
return {"mse": mse, "mae": mae, "r2": r2}
def retrieval_metrics(pred_query: np.ndarray, target: np.ndarray) -> dict:
pred_query = pred_query.astype(np.float32)
target = target.astype(np.float32)
q_norm = pred_query / np.maximum(np.linalg.norm(pred_query, axis=1, keepdims=True), 1e-8)
t_norm = target / np.maximum(np.linalg.norm(target, axis=1, keepdims=True), 1e-8)
sims = q_norm @ t_norm.T
ranks = []
for i in range(sims.shape[0]):
order = np.argsort(-sims[i])
rank = int(np.where(order == i)[0][0]) + 1
ranks.append(rank)
ranks_arr = np.asarray(ranks, dtype=np.float32)
return {
"mrr": float(np.mean(1.0 / ranks_arr)) if len(ranks_arr) else 0.0,
"top1_accuracy": float(np.mean(ranks_arr <= 1)) if len(ranks_arr) else 0.0,
"top5_accuracy": float(np.mean(ranks_arr <= 5)) if len(ranks_arr) else 0.0,
"top10_accuracy": float(np.mean(ranks_arr <= 10)) if len(ranks_arr) else 0.0,
"median_rank": float(np.median(ranks_arr)) if len(ranks_arr) else 0.0,
"mean_rank": float(np.mean(ranks_arr)) if len(ranks_arr) else 0.0,
"num_queries": int(len(ranks_arr)),
}
def onehot(y: np.ndarray, n_classes: int) -> np.ndarray:
out = np.zeros((len(y), n_classes), dtype=np.float32)
out[np.arange(len(y)), y] = 1.0
return out
def fit_classification(
X: np.ndarray,
labels: np.ndarray,
train_idx: np.ndarray,
test_idx: np.ndarray,
l2: float,
) -> tuple[dict, np.ndarray]:
y, class_names = encode_labels(labels)
train_classes = set(int(x) for x in y[train_idx])
test_classes = set(int(x) for x in y[test_idx])
unseen = [class_names[i] for i in sorted(test_classes - train_classes)]
X_train, X_test = standardize(X[train_idx], X[test_idx])
scores = ridge_predict(X_train, onehot(y[train_idx], len(class_names)), X_test, l2)
pred = scores.argmax(axis=1)
metrics = classification_metrics(y[test_idx], pred)
metrics.update({
"num_classes": len(class_names),
"num_train": int(len(train_idx)),
"num_test": int(len(test_idx)),
"unseen_test_classes": "|".join(unseen),
"unseen_test_class_count": int(len(unseen)),
})
return metrics, pred
def fit_multilabel(
X: np.ndarray,
Y: np.ndarray,
train_idx: np.ndarray,
test_idx: np.ndarray,
l2: float,
) -> tuple[dict, np.ndarray]:
X_train, X_test = standardize(X[train_idx], X[test_idx])
scores = ridge_predict(X_train, Y[train_idx], X_test, l2)
pred = (scores >= 0.5).astype(np.float32)
empty = np.where(pred.sum(axis=1) == 0)[0]
if len(empty):
pred[empty, np.argmax(scores[empty], axis=1)] = 1.0
metrics = multilabel_metrics(Y[test_idx], pred)
metrics.update({
"num_objects": int(Y.shape[1]),
"num_train": int(len(train_idx)),
"num_test": int(len(test_idx)),
})
return metrics, pred
def frame_centers(windows: list[dict]) -> np.ndarray:
return np.asarray([int(row["center_frame"]) for row in windows], dtype=np.int64)
def labels_from_windows(windows: list[dict], key: str) -> np.ndarray:
return np.asarray([str(row.get(key, "") or "") for row in windows], dtype=object)
def transition_labels_from_boundaries(suite_dir: Path, centers: np.ndarray, tolerance_frames: int = 10) -> np.ndarray:
boundaries_path = suite_dir / "transition_detection/true_boundaries.csv"
if not boundaries_path.exists():
raise FileNotFoundError(boundaries_path)
rows = read_csv(boundaries_path)
boundary_frames = np.asarray([int(row.get("boundary_frame") or row.get("frame")) for row in rows], dtype=np.int64)
labels = np.zeros(len(centers), dtype=np.int64)
for i, center in enumerate(centers):
if len(boundary_frames) and np.min(np.abs(boundary_frames - center)) <= tolerance_frames:
labels[i] = 1
return np.asarray(["transition" if x else "steady" for x in labels], dtype=object)
def task_target(
task: str,
X: np.ndarray,
windows: list[dict],
manifest: list[dict],
suite_dir: Path,
future_offset_windows: int,
object_targets: dict | None = None,
) -> dict:
centers = frame_centers(windows)
n = len(windows)
train_idx, test_idx = chronological_split(n, 0.30)
all_idx = np.arange(n, dtype=np.int64)
if task == "timeline_action":
return {"kind": "classification", "labels": labels_from_windows(windows, "action_label"), "rows": all_idx}
if task == "timeline_subtask":
return {"kind": "classification", "labels": labels_from_windows(windows, "subtask_label"), "rows": all_idx}
if task == "transition_detection":
return {"kind": "classification", "labels": transition_labels_from_boundaries(suite_dir, centers), "rows": all_idx}
if task == "next_action":
rows = np.arange(0, n - future_offset_windows, dtype=np.int64)
labels = labels_from_windows(windows, "action_label")[rows + future_offset_windows]
return {"kind": "classification", "labels": labels, "rows": rows, "target_variant": "future action label from windows.csv"}
if task == "contact_prediction":
contacts = block_indices(manifest, ["body_contacts"])
rows = all_idx
labels = np.where(np.abs(X[:, contacts]).sum(axis=1) > 1e-8, "contact", "no_contact")
return {
"kind": "classification",
"labels": labels,
"rows": rows,
"target_source_blocks": "body_contacts",
"target_variant": "contact proxy derived from body_contacts feature block",
}
if task == "hand_trajectory_forecast":
rows = np.arange(0, n - future_offset_windows, dtype=np.int64)
hand = block_indices(manifest, ["hand_left_joints", "hand_right_joints"])
target = X[rows + future_offset_windows][:, hand]
return {
"kind": "regression",
"target": target,
"rows": rows,
"target_source_blocks": "future hand_left_joints|hand_right_joints",
"target_variant": "future hand feature vector from shared_windows.npz",
}
if task == "caption_grounding":
rows = all_idx
text = block_indices(manifest, ["caption_objects_interaction_text"])
return {"kind": "retrieval", "target": X[:, text], "rows": rows, "target_source_blocks": "caption_objects_interaction_text"}
if task in {"cross_modal_retrieval", "modality_reconstruction"}:
rows = all_idx
visual = block_indices(manifest, ["depth_confidence", "video_"])
return {"kind": "retrieval" if task == "cross_modal_retrieval" else "regression", "target": X[:, visual], "rows": rows, "target_source_blocks": "depth_confidence|video_*"}
if task == "temporal_order":
pairs = []
labels = []
for i in range(n - 1):
pairs.append((i, i + 1))
labels.append("forward")
pairs.append((i + 1, i))
labels.append("reversed")
return {"kind": "pair_classification", "pairs": np.asarray(pairs, dtype=np.int64), "labels": np.asarray(labels, dtype=object)}
if task == "misalignment_detection":
shift = 8
pairs = []
labels = []
for i in range(n - shift):
pairs.append((i, i))
labels.append("aligned")
pairs.append((i, i + shift))
labels.append("shifted")
return {"kind": "pair_classification", "pairs": np.asarray(pairs, dtype=np.int64), "labels": np.asarray(labels, dtype=object)}
if task == "object_relevance":
if object_targets is None:
return {
"kind": "not_available",
"reason": "raw annotation.hdf5 was not provided, so full-train object relevance labels could not be exported",
}
rows = all_idx
return {
"kind": "multilabel",
"target": object_targets["Y"],
"rows": rows,
"target_source_blocks": "caption_objects_interaction_text",
"target_variant": "object sets exported from annotation.hdf5 caption_frame_info_map",
}
raise KeyError(task)
def target_overlap(group_idx: np.ndarray, target_info: dict, manifest: list[dict]) -> bool:
blocks = str(target_info.get("target_source_blocks", ""))
if not blocks:
return False
target_idx: list[int] = []
for part in blocks.split("|"):
part = part.strip()
if not part:
continue
prefix = part[:-1] if part.endswith("*") else part
target_idx.extend(block_indices(manifest, [prefix]).tolist())
if not target_idx:
return False
return bool(np.intersect1d(group_idx, np.asarray(target_idx, dtype=np.int64)).size)
def pair_features(X: np.ndarray, pairs: np.ndarray, group_idx: np.ndarray, visual_idx: np.ndarray | None = None, task: str = "") -> np.ndarray:
left = X[pairs[:, 0]][:, group_idx]
right_source = X[pairs[:, 1]]
if task == "misalignment_detection" and visual_idx is not None:
right = right_source[:, visual_idx]
else:
right = right_source[:, group_idx]
diff = right[:, : min(left.shape[1], right.shape[1])] - left[:, : min(left.shape[1], right.shape[1])]
return np.concatenate([left, right, diff], axis=1).astype(np.float32)
def run_modality_ablation(
X: np.ndarray,
windows: list[dict],
manifest: list[dict],
suite_dir: Path,
out_dir: Path,
args: argparse.Namespace,
object_targets: dict | None = None,
) -> list[dict]:
groups = modality_groups(manifest)
visual_idx = block_indices(manifest, ["depth_confidence", "video_"])
contact_idx = block_indices(manifest, ["body_contacts"])
rows: list[dict] = []
for task in TASKS:
info = task_target(task, X, windows, manifest, suite_dir, args.future_offset_windows, object_targets)
for group_name, group_idx_raw in groups.items():
row = {
"task": task,
"task_display": TASK_DISPLAY[task],
"modality_group": group_name,
"modality_display": GROUP_DISPLAY[group_name],
"status": "computed",
"score": "",
"primary_metric": "",
"primary_metric_value": "",
"target_variant": info.get("target_variant", ""),
"target_source_overlap": "",
"reason": "",
}
if info["kind"] == "not_available":
row.update({"status": "not_computed", "reason": info["reason"]})
rows.append(row)
continue
group_idx = group_idx_raw
if task == "contact_prediction":
group_idx = np.setdiff1d(group_idx_raw, contact_idx)
if len(group_idx) == 0:
row.update({"status": "not_computed", "reason": "input group would contain only contact target-source features"})
rows.append(row)
continue
try:
if info["kind"] == "classification":
data_rows = np.asarray(info["rows"], dtype=np.int64)
labels = np.asarray(info["labels"], dtype=object)
train_local, test_local = chronological_split(len(data_rows), args.test_fraction)
metrics, _ = fit_classification(
X[data_rows][:, group_idx],
labels,
train_local,
test_local,
args.ridge_l2,
)
if metrics.get("status") == "not_computed":
row.update(metrics)
else:
row.update(metrics)
row["primary_metric"] = "macro_f1"
row["primary_metric_value"] = metrics["macro_f1"]
row["score"] = metrics["macro_f1"]
elif info["kind"] == "regression":
data_rows = np.asarray(info["rows"], dtype=np.int64)
target = np.asarray(info["target"], dtype=np.float32)
train_local, test_local = chronological_split(len(data_rows), args.test_fraction)
Xin_train, Xin_test = standardize(X[data_rows[train_local]][:, group_idx], X[data_rows[test_local]][:, group_idx])
Y_train, Y_test = standardize(info["target"][train_local], info["target"][test_local])
pred = ridge_predict(Xin_train, Y_train, Xin_test, args.ridge_l2)
metrics = regression_metrics(Y_test, pred)
row.update(metrics)
row["primary_metric"] = "mae"
row["primary_metric_value"] = metrics["mae"]
row["score"] = 1.0 / (1.0 + metrics["mae"])
elif info["kind"] == "multilabel":
data_rows = np.asarray(info["rows"], dtype=np.int64)
target = np.asarray(info["target"], dtype=np.float32)
train_local, test_local = chronological_split(len(data_rows), args.test_fraction)
metrics, _ = fit_multilabel(X[data_rows][:, group_idx], target, train_local, test_local, args.ridge_l2)
row.update(metrics)
row["primary_metric"] = "micro_f1"
row["primary_metric_value"] = metrics["micro_f1"]
row["score"] = metrics["micro_f1"]
elif info["kind"] == "retrieval":
data_rows = np.asarray(info["rows"], dtype=np.int64)
target = np.asarray(info["target"], dtype=np.float32)
train_local, test_local = chronological_split(len(data_rows), args.test_fraction)
Xin_train, Xin_test = standardize(X[data_rows[train_local]][:, group_idx], X[data_rows[test_local]][:, group_idx])
Y_train, Y_test = standardize(target[train_local], target[test_local])
pred = ridge_predict(Xin_train, Y_train, Xin_test, args.ridge_l2)
metrics = retrieval_metrics(pred, Y_test)
row.update(metrics)
row["primary_metric"] = "mrr"
row["primary_metric_value"] = metrics["mrr"]
row["score"] = metrics["mrr"]
elif info["kind"] == "pair_classification":
pairs = np.asarray(info["pairs"], dtype=np.int64)
labels = np.asarray(info["labels"], dtype=object)
train_local, test_local = chronological_split(len(pairs), args.test_fraction)
feats = pair_features(X, pairs, group_idx, visual_idx=visual_idx, task=task)
metrics, _ = fit_classification(feats, labels, train_local, test_local, args.ridge_l2)
if metrics.get("status") == "not_computed":
row.update(metrics)
else:
row.update(metrics)
row["primary_metric"] = "macro_f1"
row["primary_metric_value"] = metrics["macro_f1"]
row["score"] = metrics["macro_f1"]
else:
raise ValueError(info["kind"])
row["target_source_overlap"] = str(target_overlap(group_idx, info, manifest)).lower()
except Exception as exc: # keep failed pairs visible instead of silently dropping them
row.update({"status": "not_computed", "reason": f"{type(exc).__name__}: {exc}"})
rows.append(row)
write_csv(out_dir / "modality_ablation/ablation_metrics.csv", rows)
write_json(
out_dir / "modality_ablation/ablation_summary.json",
{
"description": "Compact ridge-head ablation over real shared_windows.npz feature blocks.",
"num_rows": len(rows),
"num_computed": sum(1 for r in rows if r.get("status") == "computed"),
"num_not_computed": sum(1 for r in rows if r.get("status") != "computed"),
"groups": {name: int(len(idx)) for name, idx in groups.items()},
"tasks": TASKS,
"object_relevance_labels": "annotation.hdf5" if object_targets is not None else "not_available",
},
)
render_ablation_svg(rows, out_dir / "modality_ablation/ablation_matrix.svg")
write_modality_report(rows, out_dir / "modality_ablation/MODALITY_ABLATION_REPORT.md")
return rows
def color_for_score(score: float | None) -> str:
if score is None or math.isnan(score):
return "#20251f"
score = max(0.0, min(1.0, score))
r = int(37 + (154 - 37) * (1.0 - score))
g = int(72 + (224 - 72) * score)
b = int(54 + (101 - 54) * score)
return f"#{r:02x}{g:02x}{b:02x}"
def render_ablation_svg(rows: list[dict], path: Path) -> None:
groups = list(GROUP_DISPLAY.keys())
tasks = TASKS
cell_w, cell_h = 132, 34
left, top = 300, 98
width = left + cell_w * len(groups) + 44
height = top + cell_h * len(tasks) + 86
by_key = {(r["task"], r["modality_group"]): r for r in rows}
parts = [
f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}">',
'<rect width="100%" height="100%" fill="#10160f"/>',
'<text x="28" y="38" fill="#eef5e8" font-family="Inter, Arial" font-size="24" font-weight="700">Single-Episode Modality Ablation Matrix</text>',
'<text x="28" y="66" fill="#a7b5a3" font-family="Inter, Arial" font-size="13">Scores are recomputed from shared_windows.npz; gray cells are intentionally not computed.</text>',
]
for j, group in enumerate(groups):
x = left + j * cell_w + 6
parts.append(
f'<text x="{x}" y="{top - 18}" fill="#cbd8c8" font-family="Inter, Arial" font-size="12" transform="rotate(-25 {x} {top - 18})">{html.escape(GROUP_DISPLAY[group])}</text>'
)
for i, task in enumerate(tasks):
y = top + i * cell_h
parts.append(f'<text x="28" y="{y + 22}" fill="#e7efe2" font-family="Inter, Arial" font-size="13">{html.escape(TASK_DISPLAY[task])}</text>')
for j, group in enumerate(groups):
x = left + j * cell_w
row = by_key.get((task, group), {})
if row.get("status") == "computed" and row.get("score") != "":
score = float(row["score"])
fill = color_for_score(score)
label = f'{score:.2f}'
label_fill = "#061006" if score > 0.62 else "#edf5e7"
else:
fill = "#222720"
label = "n/a"
label_fill = "#7b8678"
parts.append(f'<rect x="{x}" y="{y}" width="{cell_w - 8}" height="{cell_h - 7}" rx="5" fill="{fill}" stroke="#34402f" stroke-width="1"/>')
parts.append(f'<text x="{x + (cell_w - 8) / 2}" y="{y + 19}" text-anchor="middle" fill="{label_fill}" font-family="Inter, Arial" font-size="12" font-weight="700">{label}</text>')
parts.extend(
[
f'<text x="28" y="{height - 34}" fill="#a7b5a3" font-family="Inter, Arial" font-size="12">Metric: macro-F1 / MRR / 1/(1+MAE), depending on task type. See CSV for raw values and overlap flags.</text>',
"</svg>",
]
)
write_text(path, "\n".join(parts))
def write_modality_report(rows: list[dict], path: Path) -> None:
computed = [r for r in rows if r.get("status") == "computed" and r.get("score") != ""]
by_task: dict[str, list[dict]] = {t: [] for t in TASKS}
for row in computed:
by_task[row["task"]].append(row)
lines = [
"# Single-Episode Modality Ablation Report",
"",
"This diagnostic reruns compact ridge heads on the exported one-episode feature matrix. It is useful for checking which real feature blocks can support each task on this episode, not for estimating dataset-wide generalization.",
"",
"No synthetic labels are introduced. Derived proxy targets are marked in `target_variant`, and feature groups that overlap with the target source are marked in `target_source_overlap`.",
"",
"## Best Computed Group Per Task",
"",
]
for task in TASKS:
task_rows = by_task.get(task, [])
if not task_rows:
reasons = sorted({r.get("reason", "") for r in rows if r["task"] == task and r.get("reason")})
lines.append(f"- {TASK_DISPLAY[task]}: not computed ({'; '.join(reasons)})")
continue
best = max(task_rows, key=lambda r: float(r["score"]))
line = (
f"- {TASK_DISPLAY[task]}: {best['modality_display']} score={float(best['score']):.4f}, "
f"{best['primary_metric']}={float(best['primary_metric_value']):.4f}, target overlap={best['target_source_overlap']}"
)
if best["target_source_overlap"] == "true":
no_overlap = [r for r in task_rows if r.get("target_source_overlap") == "false"]
if no_overlap:
alt = max(no_overlap, key=lambda r: float(r["score"]))
line += (
f"; best non-overlap: {alt['modality_display']} score={float(alt['score']):.4f}, "
f"{alt['primary_metric']}={float(alt['primary_metric_value']):.4f}"
)
lines.append(line)
lines.extend(
[
"",
"## Files",
"",
"- `ablation_metrics.csv`: every task/modality pair, including not-computed rows and reasons.",
"- `ablation_matrix.svg`: compact heatmap for manual inspection.",
"- `ablation_summary.json`: group dimensions and computed/not-computed counts.",
]
)
write_text(path, "\n".join(lines) + "\n")
def run_timeline_overlay(suite_dir: Path, windows: list[dict], out_dir: Path) -> list[dict]:
overlay_tasks = [
"timeline_action",
"timeline_subtask",
"transition_detection",
"next_action",
"contact_prediction",
"object_relevance",
]
window_by_index = {int(row["window_index"]): row for row in windows}
rows: list[dict] = []
for task in overlay_tasks:
pred_path = suite_dir / task / "predictions.csv"
if not pred_path.exists():
rows.append({"task": task, "status": "not_available", "reason": f"missing {pred_path}"})
continue
for pred in read_csv(pred_path):
try:
idx = int(pred["window_index"])
except KeyError:
continue
window = window_by_index.get(idx, {})
true_value = pred.get("true_label") or pred.get("true_objects") or pred.get("true") or ""
pred_value = pred.get("predicted_label") or pred.get("predicted_objects") or pred.get("predicted") or ""
if "correct" in pred and pred["correct"] != "":
correct = int(float(pred["correct"]))
else:
correct = int(str(true_value) == str(pred_value))
rows.append(
{
"task": task,
"task_display": TASK_DISPLAY[task],
"status": "observed_prediction",
"window_index": idx,
"start_frame": pred.get("start_frame") or window.get("start_frame", ""),
"end_frame": pred.get("end_frame") or window.get("end_frame", ""),
"center_frame": pred.get("center_frame") or window.get("center_frame", ""),
"true_value": true_value,
"predicted_value": pred_value,
"confidence": pred.get("confidence", ""),
"correct": correct,
}
)
write_csv(out_dir / "timeline_overlay/timeline_overlay.csv", rows)
render_timeline_svg(rows, windows, suite_dir, out_dir / "timeline_overlay/timeline_overlay.svg")
write_timeline_report(rows, out_dir / "timeline_overlay/TIMELINE_OVERLAY_REPORT.md")
return rows
def render_timeline_svg(rows: list[dict], windows: list[dict], suite_dir: Path, path: Path) -> None:
tasks = ["timeline_action", "timeline_subtask", "transition_detection", "next_action", "contact_prediction", "object_relevance"]
min_frame = min(int(r["start_frame"]) for r in windows)
max_frame = max(int(r["end_frame"]) for r in windows)
left, top = 260, 84
row_h, plot_w = 48, 1100
width = left + plot_w + 38
height = top + row_h * len(tasks) + 92
by_task: dict[str, list[dict]] = {t: [] for t in tasks}
for row in rows:
if row.get("status") == "observed_prediction":
by_task[row["task"]].append(row)
def x_for(frame: int) -> float:
return left + (frame - min_frame) / max(1, max_frame - min_frame) * plot_w
parts = [
f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}">',
'<rect width="100%" height="100%" fill="#10160f"/>',
'<text x="28" y="38" fill="#eef5e8" font-family="Inter, Arial" font-size="24" font-weight="700">Held-Out Timeline Prediction Overlay</text>',
'<text x="28" y="64" fill="#a7b5a3" font-family="Inter, Arial" font-size="13">Bars are existing real prediction rows aligned back to the exported episode timeline.</text>',
]
boundaries_path = suite_dir / "transition_detection/true_boundaries.csv"
boundary_frames = []
if boundaries_path.exists():
boundary_frames = [int(r.get("boundary_frame") or r.get("frame")) for r in read_csv(boundaries_path)]
for i, task in enumerate(tasks):
y = top + i * row_h
parts.append(f'<text x="28" y="{y + 25}" fill="#e7efe2" font-family="Inter, Arial" font-size="13">{html.escape(TASK_DISPLAY[task])}</text>')
parts.append(f'<line x1="{left}" y1="{y + 20}" x2="{left + plot_w}" y2="{y + 20}" stroke="#2a3428" stroke-width="18" stroke-linecap="round"/>')
for row in by_task[task]:
if not row.get("start_frame") or not row.get("end_frame"):
continue
x1 = x_for(int(float(row["start_frame"])))
x2 = max(x1 + 2, x_for(int(float(row["end_frame"]))))
fill = "#8ee06a" if int(row["correct"]) else "#e46b5f"
parts.append(f'<rect x="{x1:.2f}" y="{y + 11}" width="{x2 - x1:.2f}" height="18" rx="3" fill="{fill}" opacity="0.86"/>')
for frame in boundary_frames:
x = x_for(frame)
parts.append(f'<line x1="{x:.2f}" y1="{y + 4}" x2="{x:.2f}" y2="{y + 36}" stroke="#d8e887" stroke-width="1.2" opacity="0.70"/>')
parts.extend(
[
f'<text x="{left}" y="{height - 42}" fill="#8ee06a" font-family="Inter, Arial" font-size="12">green = exact/correct prediction</text>',
f'<text x="{left + 220}" y="{height - 42}" fill="#e46b5f" font-family="Inter, Arial" font-size="12">red = mismatch</text>',
f'<text x="{left + 390}" y="{height - 42}" fill="#d8e887" font-family="Inter, Arial" font-size="12">vertical lines = real transition boundaries</text>',
"</svg>",
]
)
write_text(path, "\n".join(parts))
def write_timeline_report(rows: list[dict], path: Path) -> None:
observed = [r for r in rows if r.get("status") == "observed_prediction"]
lines = [
"# Timeline Prediction Overlay Report",
"",
"This report aligns existing prediction CSV files to the exported episode timeline. It does not rerun training.",
"",
"## Task-Level Correctness",
"",
]
for task in ["timeline_action", "timeline_subtask", "transition_detection", "next_action", "contact_prediction", "object_relevance"]:
task_rows = [r for r in observed if r["task"] == task]
if not task_rows:
lines.append(f"- {TASK_DISPLAY[task]}: no prediction rows found")
continue
correct = sum(int(r["correct"]) for r in task_rows)
lines.append(f"- {TASK_DISPLAY[task]}: {correct}/{len(task_rows)} correct ({correct / len(task_rows):.4f})")
lines.extend(
[
"",
"## Files",
"",
"- `timeline_overlay.csv`: prediction rows with frame positions.",
"- `timeline_overlay.svg`: visual overlay across the episode.",
]
)
write_text(path, "\n".join(lines) + "\n")
def run_alignment_stress(
X: np.ndarray,
manifest: list[dict],
windows: list[dict],
out_dir: Path,
args: argparse.Namespace,
) -> list[dict]:
groups = modality_groups(manifest)
stress_groups = {
"motion_capture": groups["motion_capture"],
"pose_slam": groups["pose_slam"],
"inertial": groups["inertial"],
"language": groups["language"],
"motion_pose_inertial": np.unique(np.concatenate([groups["motion_capture"], groups["pose_slam"], groups["inertial"]])),
}
if "audio" in groups:
stress_groups["audio"] = groups["audio"]
target_idx = block_indices(manifest, ["depth_confidence", "video_"])
n = X.shape[0]
train_idx, test_idx = chronological_split(n, args.test_fraction)
shifts = [-40, -20, -10, -5, 0, 5, 10, 20, 40]
rows: list[dict] = []
stride = int(windows[1]["start_frame"]) - int(windows[0]["start_frame"]) if len(windows) > 1 else 1
for group, q_idx in stress_groups.items():
q_train, q_test_all = standardize(X[train_idx][:, q_idx], X[test_idx][:, q_idx])
t_train, t_test_all = standardize(X[train_idx][:, target_idx], X[test_idx][:, target_idx])
projector_pred_all = ridge_predict(q_train, t_train, q_test_all, args.ridge_l2)
for shift in shifts:
valid = []
for local_i in range(len(test_idx)):
shifted_local = local_i + shift
if 0 <= shifted_local < len(test_idx):
valid.append((local_i, shifted_local))
if not valid:
continue
original = np.asarray([a for a, _ in valid], dtype=np.int64)
shifted = np.asarray([b for _, b in valid], dtype=np.int64)
pred = projector_pred_all[shifted]
target = t_test_all[original]
metrics = retrieval_metrics(pred, target)
row = {
"query_group": group,
"query_display": GROUP_DISPLAY[group],
"target_group": "depth_plus_video",
"shift_windows": shift,
"shift_frames": int(shift * stride),
"status": "derived_perturbation",
**metrics,
}
rows.append(row)
write_csv(out_dir / "alignment_stress/alignment_shift_metrics.csv", rows)
render_alignment_svg(rows, out_dir / "alignment_stress/alignment_shift_curves.svg")
write_json(
out_dir / "alignment_stress/alignment_stress_summary.json",
{
"description": "Real feature windows are deliberately time-shifted at evaluation time to test cross-modal alignment sensitivity.",
"target_group": "depth_confidence + video_*",
"status_meaning": "derived_perturbation means the features are real but the time shift is an explicit diagnostic perturbation.",
"num_rows": len(rows),
},
)
write_alignment_report(rows, out_dir / "alignment_stress/ALIGNMENT_STRESS_REPORT.md")
return rows
def render_alignment_svg(rows: list[dict], path: Path) -> None:
groups = sorted({r["query_group"] for r in rows})
width, height = 1020, 520
left, top, plot_w, plot_h = 94, 80, 810, 330
shifts = sorted({int(r["shift_windows"]) for r in rows})
if not shifts:
write_text(path, "<svg xmlns=\"http://www.w3.org/2000/svg\"></svg>")
return
min_shift, max_shift = min(shifts), max(shifts)
max_mrr = max(float(r["mrr"]) for r in rows) if rows else 1.0
max_mrr = max(max_mrr, 1e-6)
palette = ["#8ee06a", "#d8e887", "#7bd3ff", "#f0a45e", "#cba8ff"]
def x_for(shift: int) -> float:
return left + (shift - min_shift) / max(1, max_shift - min_shift) * plot_w
def y_for(mrr: float) -> float:
return top + plot_h - mrr / max_mrr * plot_h
parts = [
f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}">',
'<rect width="100%" height="100%" fill="#10160f"/>',
'<text x="28" y="38" fill="#eef5e8" font-family="Inter, Arial" font-size="24" font-weight="700">Cross-Modal Alignment Stress Test</text>',
'<text x="28" y="64" fill="#a7b5a3" font-family="Inter, Arial" font-size="13">Query features are shifted in time; the target visual window remains the original held-out window.</text>',
f'<rect x="{left}" y="{top}" width="{plot_w}" height="{plot_h}" fill="#151d14" stroke="#34402f"/>',
]
for tick in shifts:
x = x_for(tick)
parts.append(f'<line x1="{x:.2f}" y1="{top}" x2="{x:.2f}" y2="{top + plot_h}" stroke="#263024" stroke-width="1"/>')
parts.append(f'<text x="{x:.2f}" y="{top + plot_h + 24}" fill="#a7b5a3" font-family="Inter, Arial" font-size="11" text-anchor="middle">{tick}</text>')
parts.append(f'<text x="{left + plot_w / 2}" y="{height - 48}" fill="#cbd8c8" font-family="Inter, Arial" font-size="13" text-anchor="middle">shift in windows</text>')
parts.append(f'<text x="30" y="{top + plot_h / 2}" fill="#cbd8c8" font-family="Inter, Arial" font-size="13" transform="rotate(-90 30 {top + plot_h / 2})">MRR</text>')
for gi, group in enumerate(groups):
color = palette[gi % len(palette)]
group_rows = sorted([r for r in rows if r["query_group"] == group], key=lambda r: int(r["shift_windows"]))
points = [(x_for(int(r["shift_windows"])), y_for(float(r["mrr"]))) for r in group_rows]
if points:
d = " ".join(f"{x:.2f},{y:.2f}" for x, y in points)
parts.append(f'<polyline points="{d}" fill="none" stroke="{color}" stroke-width="2.4"/>')
for x, y in points:
parts.append(f'<circle cx="{x:.2f}" cy="{y:.2f}" r="4" fill="{color}"/>')
parts.append(f'<rect x="{left + plot_w + 28}" y="{top + gi * 25}" width="12" height="12" fill="{color}"/>')
parts.append(f'<text x="{left + plot_w + 46}" y="{top + 11 + gi * 25}" fill="#e7efe2" font-family="Inter, Arial" font-size="12">{html.escape(GROUP_DISPLAY.get(group, group))}</text>')
parts.append("</svg>")
write_text(path, "\n".join(parts))
def write_alignment_report(rows: list[dict], path: Path) -> None:
lines = [
"# Cross-Modal Alignment Stress Report",
"",
"This diagnostic uses real held-out feature windows, then deliberately shifts the query modality in time at evaluation. The perturbation is derived; it is not treated as observed data.",
"",
"## Zero-Shift Versus Worst Shift",
"",
]
for group in sorted({r["query_group"] for r in rows}):
group_rows = [r for r in rows if r["query_group"] == group]
zero = next((r for r in group_rows if int(r["shift_windows"]) == 0), None)
worst = min(group_rows, key=lambda r: float(r["mrr"])) if group_rows else None
if zero and worst:
lines.append(
f"- {GROUP_DISPLAY.get(group, group)}: zero-shift MRR={float(zero['mrr']):.4f}; "
f"worst shift={worst['shift_windows']} windows, MRR={float(worst['mrr']):.4f}"
)
lines.extend(
[
"",
"## Files",
"",
"- `alignment_shift_metrics.csv`: MRR/rank metrics for each query group and time shift.",
"- `alignment_shift_curves.svg`: MRR curves across time shifts.",
"- `alignment_stress_summary.json`: perturbation definition and status.",
]
)
write_text(path, "\n".join(lines) + "\n")
def build_provenance(
suite_dir: Path,
out_dir: Path,
X: np.ndarray,
starts: np.ndarray,
ends: np.ndarray,
manifest: list[dict],
summary: dict,
annotation: Path | None = None,
) -> dict:
repo_root = suite_dir.parent.parent.resolve()
input_files = [
suite_dir / "shared_windows.npz",
suite_dir / "windows.csv",
suite_dir / "feature_manifest.json",
suite_dir / "summary_report.json",
suite_dir / "transition_detection/true_boundaries.csv",
]
for task in ["timeline_action", "timeline_subtask", "transition_detection", "next_action", "contact_prediction", "object_relevance"]:
pred_path = suite_dir / task / "predictions.csv"
if pred_path.exists():
input_files.append(pred_path)
if annotation is not None and annotation.exists():
input_files.append(annotation)
provenance = {
"artifact_policy": "Only existing local artifacts are consumed. Missing labels/tasks are marked not_computed instead of filled.",
"source_suite_dir": public_artifact_path(suite_dir, repo_root),
"output_dir": public_artifact_path(out_dir, repo_root),
"shared_windows_shape": [int(X.shape[0]), int(X.shape[1])],
"starts_first_last": [int(starts[0]), int(starts[-1])],
"ends_first_last": [int(ends[0]), int(ends[-1])],
"feature_blocks": manifest,
"summary_report_core": {
"num_windows": summary.get("num_windows"),
"feature_dim": summary.get("feature_dim"),
"window_frames": summary.get("window_frames"),
"stride_frames": summary.get("stride_frames"),
"annotation": public_source_reference(summary.get("annotation")),
},
"input_file_hashes": {
public_artifact_path(path, repo_root): sha256(path)
for path in input_files
if path.exists()
},
}
write_json(out_dir / "provenance.json", provenance)
return provenance
def write_index(out_dir: Path, provenance: dict, ablation_rows: list[dict], timeline_rows: list[dict], stress_rows: list[dict]) -> None:
lines = [
"# Single-Episode Diagnostics Index",
"",
"These outputs are local diagnostics built from the existing one-episode Xperience-10M artifacts. They are designed for manual verification while waiting for full multi-episode data access.",
"",
"## Generated Analyses",
"",
"- `modality_ablation/`: compact ridge-head ablations across real feature blocks.",
"- `timeline_overlay/`: existing prediction CSVs aligned to the episode timeline.",
"- `alignment_stress/`: cross-modal retrieval under explicit time-shift perturbations.",
"- `provenance.json`: input hashes, feature dimensions, and source artifact identifiers.",
"",
"## Validity Boundaries",
"",
"- This is a single-episode diagnostic, not a full Xperience-10M benchmark.",
"- Rows marked `not_computed` are intentionally left blank when train labels or valid splits are unavailable.",
"- Rows marked `derived_perturbation` use real features with deliberate time shifts for stress testing.",
"",
"## Counts",
"",
f"- Ablation rows: {len(ablation_rows)}; computed: {sum(1 for r in ablation_rows if r.get('status') == 'computed')}.",
f"- Timeline overlay rows: {sum(1 for r in timeline_rows if r.get('status') == 'observed_prediction')}.",
f"- Alignment stress rows: {len(stress_rows)}.",
f"- Shared feature shape: {provenance['shared_windows_shape'][0]} windows x {provenance['shared_windows_shape'][1]} features.",
]
write_text(out_dir / "README.md", "\n".join(lines) + "\n")
def main() -> None:
args = parse_args()
suite_dir = args.suite_dir.resolve()
out_dir = args.output_dir.resolve()
X, starts, ends, windows, manifest, summary = load_inputs(suite_dir)
object_targets = None
if args.annotation is not None:
object_targets = load_object_targets_from_annotation(args.annotation, windows, args.homie_toolkit)
write_csv(
out_dir / "object_labels/window_object_labels.csv",
object_targets["rows"],
["window_index", "start_frame", "end_frame", "center_frame", "objects", "object_count"],
)
write_json(
out_dir / "object_labels/object_vocab.json",
{
"vocab": object_targets["vocab"],
"num_objects": len(object_targets["vocab"]),
"source_annotation": object_targets["annotation"],
"source_toolkit": object_targets["toolkit"],
"source_note": object_targets["source_note"],
},
)
provenance = build_provenance(suite_dir, out_dir, X, starts, ends, manifest, summary, args.annotation)
ablation_rows = run_modality_ablation(X, windows, manifest, suite_dir, out_dir, args, object_targets)
timeline_rows = run_timeline_overlay(suite_dir, windows, out_dir)
stress_rows = run_alignment_stress(X, manifest, windows, out_dir, args)
write_index(out_dir, provenance, ablation_rows, timeline_rows, stress_rows)
print(f"Wrote diagnostics to {out_dir}")
print(f"Ablation computed rows: {sum(1 for r in ablation_rows if r.get('status') == 'computed')}/{len(ablation_rows)}")
print(f"Timeline observed rows: {sum(1 for r in timeline_rows if r.get('status') == 'observed_prediction')}")
print(f"Alignment stress rows: {len(stress_rows)}")
if __name__ == "__main__":
main()
|