File size: 20,090 Bytes
78ac017 | 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 | #!/usr/bin/env python3
"""Build a source-disjoint, pixel-grounded VST prefix-32 semantic pack.
The upstream VST metadata provides cumulative count states. A recent-only
window cannot in general support those labels, so this compiler exposes a
uniform causal prefix ending strictly before the query time. It emits only
plain semantic answers (open integer and balanced MCQ); no agent action or
private state is ever an actor target.
"""
from __future__ import annotations
import argparse
import collections
import concurrent.futures
import hashlib
import io
import json
import math
import os
import shutil
import tempfile
from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Any
from PIL import Image
from data_pipeline.temporal_scope.semantic_mcq import (
DIRECT_PROMPT_PREFIX,
DIRECT_PROMPT_SUFFIX,
rebalance_answer_slot,
)
from viss_core.data_v0.canonical import (
attach_content_sha256,
canonical_json_line,
file_sha256,
stable_id,
write_jsonl,
)
from viss_core.data_v0.schema import validate_row
FORMAT = "streameva.vst_prefix32_semantic.v1"
PREFLIGHT_FORMAT = "streameva.vst_count_metadata_preflight.v1"
OPTION_LETTERS = "ABCD"
TARGET_COUNTS = tuple(range(11))
PREFIX_FRAME_MAX_PIXELS = 128 * 28 * 28
PREFIX_FRAME_MAX_SIDE = 504
class VSTPrefixSemanticError(ValueError):
"""Raised when a prefix-32 semantic candidate violates its contract."""
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--selection", type=Path, required=True)
parser.add_argument("--preflight", type=Path, required=True)
parser.add_argument("--media-root", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--frames", type=int, default=32)
parser.add_argument("--train-per-count", type=int, default=92)
parser.add_argument("--holdout-per-count", type=int, default=5)
parser.add_argument("--workers", type=int, default=16)
parser.add_argument("--seed", type=int, default=173)
args = parser.parse_args()
print(json.dumps(materialize(args), indent=2, sort_keys=True))
def select_available_states(
rows: Sequence[Mapping[str, Any]],
*,
media_root: Path,
train_per_count: int,
holdout_per_count: int,
seed: int,
) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
"""Select balanced, source-disjoint train and holdout states."""
if train_per_count <= 0 or holdout_per_count <= 0:
raise VSTPrefixSemanticError("train/holdout quotas must be positive")
by_count: dict[int, list[dict[str, Any]]] = {
value: [] for value in TARGET_COUNTS
}
seen_videos: set[str] = set()
seen_ids: set[str] = set()
for raw in rows:
row = dict(raw)
answer = int(row.get("answer", -1))
relative = Path(str(row.get("source_video_relative") or ""))
candidate_id = str(row.get("candidate_id") or "")
if (
answer not in by_count
or not candidate_id
or relative.is_absolute()
or ".." in relative.parts
or relative.suffix.casefold() != ".mp4"
):
raise VSTPrefixSemanticError("selection row is malformed")
if candidate_id in seen_ids or relative.as_posix() in seen_videos:
raise VSTPrefixSemanticError("selection IDs/videos must be unique")
seen_ids.add(candidate_id)
seen_videos.add(relative.as_posix())
source = (media_root / relative).resolve()
if not source.is_file() or source.is_symlink():
continue
row["source_video"] = str(source)
by_count[answer].append(row)
train: list[dict[str, Any]] = []
holdout: list[dict[str, Any]] = []
required = train_per_count + holdout_per_count
for answer in TARGET_COUNTS:
values = by_count[answer]
values.sort(
key=lambda row: hashlib.sha256(
f"{seed}|{answer}|{row['candidate_id']}".encode()
).digest()
)
if len(values) < required:
raise VSTPrefixSemanticError(
f"count {answer} has {len(values)} available videos; need {required}"
)
for row in values[:holdout_per_count]:
value = dict(row)
value["split"] = "policy_holdout"
holdout.append(value)
for row in values[holdout_per_count:required]:
value = dict(row)
value["split"] = "train"
train.append(value)
return (
sorted(train, key=lambda row: str(row["candidate_id"])),
sorted(holdout, key=lambda row: str(row["candidate_id"])),
)
def uniform_causal_indices(
*, source_total: int, source_fps: float, query_time: float, frames: int
) -> tuple[list[int], list[float]]:
"""Return exactly ``frames`` uniform indices whose timestamps are causal."""
if (
source_total <= 0
or not math.isfinite(source_fps)
or source_fps <= 0
or not math.isfinite(query_time)
or query_time <= 0
or frames <= 1
):
raise VSTPrefixSemanticError("invalid video sampling metadata")
# ceil(t * fps) - 1 is strictly earlier than t, including exact boundaries.
last = min(source_total - 1, max(0, math.ceil(query_time * source_fps) - 1))
if last == 0:
indices = [0] * frames
else:
indices = [
int(round(position * last / (frames - 1))) for position in range(frames)
]
timestamps = [index / source_fps for index in indices]
if any(timestamp >= query_time + 1e-9 for timestamp in timestamps):
raise VSTPrefixSemanticError("prefix sampler exposed a future frame")
return indices, timestamps
def project_semantic_rows(
row: Mapping[str, Any],
*,
relative_paths: Sequence[str],
frame_timestamps: Sequence[float],
source_sha256: str,
selection_sha256: str,
) -> list[dict[str, Any]]:
"""Project one grounded count state into open and balanced-MCQ rows."""
if len(relative_paths) != 32 or len(frame_timestamps) != 32:
raise VSTPrefixSemanticError("prefix projection requires exactly 32 frames")
answer = int(row["answer"])
if answer not in TARGET_COUNTS:
raise VSTPrefixSemanticError("count answer is outside 0--10")
candidate_id = str(row["candidate_id"])
split = str(row["split"])
if split not in {"train", "policy_holdout"}:
raise VSTPrefixSemanticError("invalid semantic split")
images = [
{
"type": "image",
"image": path,
"frame_id": stable_id(
"vstprefixframe", candidate_id, index, timestamp, digest_length=20
),
}
for index, (path, timestamp) in enumerate(
zip(relative_paths, frame_timestamps, strict=True)
)
]
parent = {
"kind": "vst_count_metadata_state",
"id": stable_id("vstprefixsource", candidate_id, digest_length=20),
"sha256": selection_sha256,
}
group_id = stable_id("vstprefixgroup", candidate_id, digest_length=20)
question = " ".join(str(row["question"]).split())
answer_text = str(answer)
open_row = attach_content_sha256(
{
"schema": "streameva.sft_row.v1",
"uid": stable_id("sftvstprefixopen", candidate_id, digest_length=20),
"component": "preservation",
"group_id": group_id,
"split": split,
"decision_loss_profile": "full",
"messages": [
{
"role": "user",
"content": [
*images,
{
"type": "text",
"text": (
f"{question}\nAnswer with one integer from 0 to 10 "
"and no other text."
),
},
],
"loss": False,
},
{
"role": "assistant",
"content": [{"type": "text", "text": answer_text}],
"loss": True,
},
],
"sample_loss_weight": 1.0,
"parents": [parent],
}
)
distractors = [value for value in TARGET_COUNTS if value != answer]
distractors.sort(
key=lambda value: (
abs(value - answer),
hashlib.sha256(f"{candidate_id}|{value}".encode()).digest(),
)
)
options, mcq_answer = rebalance_answer_slot(
[answer_text, *[str(value) for value in distractors[:3]]],
answer="A",
answer_slot=int(hashlib.sha256(candidate_id.encode()).hexdigest(), 16) % 4,
seed=f"{candidate_id}|prefix32-mcq-v1",
)
option_text = "\n".join(
f"{letter}. {value}"
for letter, value in zip(OPTION_LETTERS, options, strict=True)
)
mcq_prompt = (
f"{DIRECT_PROMPT_PREFIX}\n\nQuestion: {question}\n\n"
f"Options:\n{option_text}\n\n{DIRECT_PROMPT_SUFFIX}"
)
mcq_row = attach_content_sha256(
{
"schema": "streameva.sft_row.v1",
"uid": stable_id("sftvstprefixmcq", candidate_id, digest_length=20),
"component": "preservation",
"group_id": group_id,
"split": split,
"decision_loss_profile": "full",
"messages": [
{
"role": "user",
"content": [*images, {"type": "text", "text": mcq_prompt}],
"loss": False,
},
{
"role": "assistant",
"content": [{"type": "text", "text": mcq_answer}],
"loss": True,
},
],
"sample_loss_weight": 1.0,
"parents": [parent],
}
)
for output in (open_row, mcq_row):
validate_row(output)
# Bind source pixels in the manifest-facing auxiliary state without leaking
# an absolute path into the strict trainer row.
del source_sha256
return [open_row, mcq_row]
def materialize(args: argparse.Namespace) -> dict[str, Any]:
selection = args.selection.resolve()
preflight_path = args.preflight.resolve()
media_root = args.media_root.resolve()
output = args.output.resolve()
if output.exists():
raise FileExistsError(output)
if args.frames != 32:
raise VSTPrefixSemanticError("this release is frozen to exactly 32 frames")
if not 1 <= args.workers <= 32:
raise VSTPrefixSemanticError("workers must lie in [1, 32]")
preflight = json.loads(preflight_path.read_text(encoding="utf-8"))
declaration = preflight.get("selection_artifact", {})
if (
preflight.get("format") != PREFLIGHT_FORMAT
or preflight.get("status") != "pass_metadata_firewall_and_selection"
or preflight.get("contracts", {}).get("basename_hits_admitted") != 0
or preflight.get("contracts", {}).get("question_hits_admitted") != 0
or declaration.get("sha256") != file_sha256(selection)
):
raise VSTPrefixSemanticError("metadata preflight is not valid or hash-bound")
rows = [
json.loads(line)
for line in selection.read_text(encoding="utf-8").splitlines()
if line.strip()
]
train, holdout = select_available_states(
rows,
media_root=media_root,
train_per_count=args.train_per_count,
holdout_per_count=args.holdout_per_count,
seed=args.seed,
)
selected = [*train, *holdout]
output.parent.mkdir(parents=True, exist_ok=True)
scratch = Path(tempfile.mkdtemp(prefix=f".{output.name}.partial.", dir=output.parent))
try:
media_out = scratch / "media"
def decode(row: Mapping[str, Any]) -> tuple[str, dict[str, Any]]:
return str(row["candidate_id"]), _decode_prefix(row, media_out, args.frames)
decoded: dict[str, dict[str, Any]] = {}
errors: list[dict[str, str]] = []
with concurrent.futures.ThreadPoolExecutor(max_workers=args.workers) as pool:
futures = {pool.submit(decode, row): row for row in selected}
for future in concurrent.futures.as_completed(futures):
row = futures[future]
try:
candidate_id, state = future.result()
decoded[candidate_id] = state
except Exception as exc: # pragma: no cover - remote media failures
errors.append(
{
"candidate_id": str(row["candidate_id"]),
"reason": f"{type(exc).__name__}: {exc}",
}
)
write_jsonl(errors, scratch / "decode_errors.jsonl")
if errors or len(decoded) != len(selected):
raise VSTPrefixSemanticError(f"prefix decode failed for {len(errors)} rows")
train_rows: list[dict[str, Any]] = []
holdout_rows: list[dict[str, Any]] = []
source_inventory = []
selection_sha = file_sha256(selection)
for row in selected:
state = decoded[str(row["candidate_id"])]
source_sha = file_sha256(str(row["source_video"]))
source_inventory.append(
{
"candidate_id": row["candidate_id"],
"source_video_relative": row["source_video_relative"],
"source_sha256": source_sha,
"split": row["split"],
"query_time": row["query_time"],
"answer": row["answer"],
"frame_indices": state["frame_indices"],
"frame_timestamps": state["frame_timestamps"],
}
)
projected = project_semantic_rows(
row,
relative_paths=state["relative_paths"],
frame_timestamps=state["frame_timestamps"],
source_sha256=source_sha,
selection_sha256=selection_sha,
)
(train_rows if row["split"] == "train" else holdout_rows).extend(projected)
write_jsonl(train_rows, scratch / "trainer/sft.jsonl", sort_key="uid")
write_jsonl(
holdout_rows, scratch / "holdout/policy_holdout.jsonl", sort_key="uid"
)
write_jsonl(
source_inventory,
scratch / "source_inventory.jsonl",
sort_key="candidate_id",
)
media_files = sorted(path for path in media_out.rglob("*.jpg") if path.is_file())
manifest = {
"format": FORMAT,
"status": "materialized_internal_research_candidate",
"training_allowed": True,
"public_release_allowed": False,
"protocol": {
"causal_prefix_only": True,
"frames": 32,
"per_frame_max_pixels": PREFIX_FRAME_MAX_PIXELS,
"sampling": "exact_uniform_indices_over_[0,query_time)",
"visual_input_mode": "independent_image_frames",
},
"counts": {
"train_source_states": len(train),
"holdout_source_states": len(holdout),
"train_rows": len(train_rows),
"holdout_rows": len(holdout_rows),
"media_files": len(media_files),
"train_by_answer": dict(
sorted(collections.Counter(str(row["answer"]) for row in train).items())
),
"holdout_by_answer": dict(
sorted(
collections.Counter(str(row["answer"]) for row in holdout).items()
)
),
},
"inputs": {
"selection_sha256": selection_sha,
"preflight_sha256": file_sha256(preflight_path),
"builder_sha256": file_sha256(Path(__file__)),
},
"artifacts": {
"trainer/sft.jsonl": file_sha256(scratch / "trainer/sft.jsonl"),
"holdout/policy_holdout.jsonl": file_sha256(
scratch / "holdout/policy_holdout.jsonl"
),
"source_inventory.jsonl": file_sha256(
scratch / "source_inventory.jsonl"
),
"decode_errors.jsonl": file_sha256(scratch / "decode_errors.jsonl"),
},
"contracts": {
"all_labels_metadata_derived": True,
"all_pixels_strictly_before_query": True,
"benchmark_firewall_inherited": True,
"holdout_source_disjoint": True,
"no_agent_actions": True,
"no_private_state_actor_targets": True,
"source_videos_globally_unique": True,
"teacher_used_for_labels": False,
},
"license": {
"public_media_redistribution_claimed": False,
"release_mode": "recipe_plus_source_checksums",
},
}
(scratch / "manifest.json").write_bytes(canonical_json_line(manifest))
os.replace(scratch, output)
return {**manifest, "output": str(output)}
except Exception:
shutil.rmtree(scratch, ignore_errors=True)
raise
def _decode_prefix(
row: Mapping[str, Any], media_root: Path, frames: int
) -> dict[str, Any]:
from decord import VideoReader
source = Path(str(row["source_video"]))
reader = VideoReader(str(source))
try:
fps = float(reader.get_avg_fps())
total = len(reader)
indices, timestamps = uniform_causal_indices(
source_total=total,
source_fps=fps,
query_time=float(row["query_time"]),
frames=frames,
)
arrays = reader.get_batch(indices).asnumpy()
finally:
del reader
relative_paths = []
for array in arrays:
image = Image.fromarray(array)
height, width = image.height, image.width
resized_height, resized_width = _rounded_size(height, width)
if (resized_height, resized_width) != (height, width):
image = image.resize((resized_width, resized_height), Image.Resampling.LANCZOS)
buffer = io.BytesIO()
image.save(buffer, format="JPEG", quality=95, subsampling=0, optimize=False)
payload = buffer.getvalue()
digest = hashlib.sha256(payload).hexdigest()
relative = Path("media") / digest[:2] / f"{digest}.jpg"
destination = media_root.parent / relative
destination.parent.mkdir(parents=True, exist_ok=True)
try:
with destination.open("xb") as handle:
handle.write(payload)
except FileExistsError:
pass
if file_sha256(destination) != digest:
raise VSTPrefixSemanticError("decoded frame hash mismatch")
relative_paths.append(relative.as_posix())
return {
"relative_paths": relative_paths,
"frame_indices": indices,
"frame_timestamps": timestamps,
}
def _rounded_size(height: int, width: int) -> tuple[int, int]:
factor = 28
max_pixels = PREFIX_FRAME_MAX_PIXELS
max_side = PREFIX_FRAME_MAX_SIDE
scale = min(1.0, max_side / max(height, width))
if height * width * scale * scale > max_pixels:
scale = min(scale, math.sqrt(max_pixels / (height * width)))
new_height = max(factor, int(round(height * scale / factor)) * factor)
new_width = max(factor, int(round(width * scale / factor)) * factor)
while new_height * new_width > max_pixels:
if new_height >= new_width and new_height > factor:
new_height -= factor
elif new_width > factor:
new_width -= factor
else:
break
return new_height, new_width
if __name__ == "__main__":
main()
|