File size: 13,558 Bytes
8fcf7c9 | 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 | from __future__ import annotations
import csv
import json
import os
import tempfile
import time
from collections import Counter, defaultdict
from dataclasses import asdict, dataclass
from datetime import UTC, datetime
from pathlib import Path
from typing import Callable, Sequence
from sponsor_detection.data.profile import (
REQUIRED_COLUMNS,
EligibilityPolicy,
eligibility_rejection_flags,
sha256_file,
write_json_atomic,
)
SMART_SEGMENT_ANNOTATION_SCHEMA_VERSION = 1
NORMALIZATION_POLICY_VERSION = "smart-segment-annotations/1"
DEDUPLICATION_POLICY_VERSION = "preserve/1"
SOURCE_NAME = "sponsorblock_mirror"
# Within-category normalization is opt-in. ``preserve`` keeps one canonical row
# per mirror annotation so raw counts stay reproducible; cross-category overlaps
# are always preserved regardless of policy.
DEDUPLICATION_POLICIES = ("preserve",)
def _parquet_schema(source: dict[str, object], policy: EligibilityPolicy, categories):
try:
import pyarrow as pa
except ImportError as error:
raise RuntimeError(
"Parquet extraction requires the data extra: pip install -e '.[data]'"
) from error
metadata = {
b"smart_segments.schema_version": str(
SMART_SEGMENT_ANNOTATION_SCHEMA_VERSION
).encode(),
b"smart_segments.normalization_policy": NORMALIZATION_POLICY_VERSION.encode(),
b"smart_segments.deduplication_policy": DEDUPLICATION_POLICY_VERSION.encode(),
b"smart_segments.categories": json.dumps(
sorted(categories), separators=(",", ":")
).encode(),
b"smart_segments.source_sha256": str(source.get("sha256") or "").encode(),
b"smart_segments.eligibility_policy": json.dumps(
asdict(policy), sort_keys=True, separators=(",", ":")
).encode(),
}
return pa.schema(
[
("video_id", pa.string()),
("segment_id", pa.string()),
("category", pa.string()),
("start_ms", pa.int64()),
("end_ms", pa.int64()),
("votes", pa.int32()),
("locked", pa.bool_()),
("incorrect_votes", pa.int32()),
("views", pa.int64()),
("time_submitted_ms", pa.int64()),
("video_duration_ms", pa.int64()),
("hidden", pa.bool_()),
("shadow_hidden", pa.bool_()),
("action_type", pa.string()),
("service", pa.string()),
("reputation", pa.float32()),
("is_eligible", pa.bool_()),
("rejection_flags", pa.list_(pa.string())),
("source", pa.string()),
],
metadata=metadata,
)
def _new_batch() -> dict[str, list[object]]:
return {
"video_id": [],
"segment_id": [],
"category": [],
"start_ms": [],
"end_ms": [],
"votes": [],
"locked": [],
"incorrect_votes": [],
"views": [],
"time_submitted_ms": [],
"video_duration_ms": [],
"hidden": [],
"shadow_hidden": [],
"action_type": [],
"service": [],
"reputation": [],
"is_eligible": [],
"rejection_flags": [],
"source": [],
}
def _append_annotation(
batch: dict[str, list[object]],
row: dict[str, str],
policy: EligibilityPolicy,
) -> bool:
start_time = float(row["startTime"])
end_time = float(row["endTime"])
video_duration = float(row["videoDuration"])
votes = int(row["votes"])
locked = int(row["locked"])
hidden = int(row["hidden"])
shadow_hidden = int(row["shadowHidden"])
flags = eligibility_rejection_flags(
policy,
service=row["service"],
action_type=row["actionType"],
start_time=start_time,
end_time=end_time,
votes=votes,
hidden=hidden,
shadow_hidden=shadow_hidden,
video_duration=video_duration,
)
batch["video_id"].append(row["videoID"])
batch["segment_id"].append(row["UUID"])
batch["category"].append(row["category"])
batch["start_ms"].append(round(start_time * 1000))
batch["end_ms"].append(round(end_time * 1000))
batch["votes"].append(votes)
batch["locked"].append(locked != 0)
batch["incorrect_votes"].append(int(row["incorrectVotes"]))
batch["views"].append(int(row["views"]))
batch["time_submitted_ms"].append(int(row["timeSubmitted"]))
batch["video_duration_ms"].append(
round(video_duration * 1000) if video_duration > 0 else None
)
batch["hidden"].append(hidden != 0)
batch["shadow_hidden"].append(shadow_hidden != 0)
batch["action_type"].append(row["actionType"])
batch["service"].append(row["service"])
batch["reputation"].append(float(row["reputation"]))
batch["is_eligible"].append(not flags)
batch["rejection_flags"].append(sorted(flags))
batch["source"].append(SOURCE_NAME)
return not flags
def build_smart_segment_annotations(
input_path: Path,
output_path: Path,
manifest_path: Path,
policy: EligibilityPolicy,
*,
categories: list[str],
deduplication_policy: str = "preserve",
batch_size: int = 100_000,
compression: str = "zstd",
compute_sha256: bool = True,
expected_source_sha256: str | None = None,
progress_every_rows: int = 0,
progress_callback: Callable[[int, int, float], None] | None = None,
) -> dict[str, object]:
"""Normalize the raw SponsorBlock mirror into the canonical annotation table.
Every audited row is retained with its category, timestamps, quality
metadata, eligibility verdict and provenance; ineligible rows stay in the
artifact with ``is_eligible = false``. Overlaps between different categories
are never collapsed.
"""
try:
import pyarrow as pa
import pyarrow.parquet as pq
except ImportError as error:
raise RuntimeError(
"Parquet extraction requires the data extra: pip install -e '.[data]'"
) from error
if batch_size <= 0:
raise ValueError("batch_size must be positive")
if deduplication_policy not in DEDUPLICATION_POLICIES:
raise ValueError(
f"unsupported deduplication policy: {deduplication_policy}; "
f"supported: {', '.join(DEDUPLICATION_POLICIES)}"
)
audited_categories = list(dict.fromkeys(categories))
if not audited_categories:
raise ValueError("at least one category must be audited")
category_set = set(audited_categories)
started = time.monotonic()
source_sha256 = sha256_file(input_path) if compute_sha256 else None
if expected_source_sha256 and source_sha256 and source_sha256 != expected_source_sha256:
raise ValueError("input CSV checksum differs from the expected snapshot")
source = {
"path": str(input_path),
"bytes": input_path.stat().st_size,
"sha256": source_sha256,
}
schema = _parquet_schema(source, policy, audited_categories)
output_path.parent.mkdir(parents=True, exist_ok=True)
descriptor, temporary_name = tempfile.mkstemp(
dir=output_path.parent, prefix=f".{output_path.name}.", suffix=".tmp"
)
os.close(descriptor)
temporary_path = Path(temporary_name)
total_rows = 0
audited_rows = 0
eligible_rows = 0
malformed_rows = 0
by_category: Counter[str] = Counter()
eligible_by_category: Counter[str] = Counter()
eligibility_rejections: dict[str, Counter[str]] = {}
batch = _new_batch()
try:
with input_path.open("r", encoding="utf-8", newline="") as source_file:
reader = csv.DictReader(source_file)
missing_columns = sorted(REQUIRED_COLUMNS.difference(reader.fieldnames or []))
if missing_columns:
raise ValueError(f"missing required columns: {', '.join(missing_columns)}")
with pq.ParquetWriter(
temporary_path, schema, compression=compression, use_dictionary=True
) as writer:
for row in reader:
total_rows += 1
category = row.get("category") or ""
if category not in category_set:
continue
audited_rows += 1
by_category[category] += 1
try:
eligible = _append_annotation(batch, row, policy)
except (KeyError, TypeError, ValueError):
malformed_rows += 1
continue
if eligible:
eligible_rows += 1
eligible_by_category[category] += 1
else:
rejection_counter = eligibility_rejections.setdefault(
category, Counter()
)
for flag in batch["rejection_flags"][-1]:
rejection_counter[flag] += 1
if len(batch["video_id"]) >= batch_size:
writer.write_table(pa.Table.from_pydict(batch, schema=schema))
batch = _new_batch()
if progress_every_rows and total_rows % progress_every_rows == 0:
if progress_callback:
progress_callback(
total_rows, audited_rows, time.monotonic() - started
)
if batch["video_id"]:
writer.write_table(pa.Table.from_pydict(batch, schema=schema))
os.replace(temporary_path, output_path)
except BaseException:
temporary_path.unlink(missing_ok=True)
raise
report = {
"schema_version": SMART_SEGMENT_ANNOTATION_SCHEMA_VERSION,
"generated_at": datetime.now(tz=UTC).isoformat(),
"normalization_policy": NORMALIZATION_POLICY_VERSION,
"deduplication_policy": deduplication_policy,
"source": source,
"eligibility_policy": asdict(policy),
"categories": audited_categories,
"counts": {
"source_rows": total_rows,
"audited_rows": audited_rows,
"written_rows": audited_rows - malformed_rows,
"eligible_rows": eligible_rows,
"ineligible_rows": audited_rows - malformed_rows - eligible_rows,
"malformed_rows": malformed_rows,
"by_category": dict(sorted(by_category.items())),
"eligible_by_category": dict(sorted(eligible_by_category.items())),
},
"rejections_by_category": {
category: dict(sorted(counter.items()))
for category, counter in sorted(eligibility_rejections.items())
},
"output": {
"path": str(output_path),
"bytes": output_path.stat().st_size,
"sha256": sha256_file(output_path) if compute_sha256 else None,
"compression": compression,
},
"build_seconds": round(time.monotonic() - started, 3),
}
write_json_atomic(manifest_path, report)
return report
@dataclass(frozen=True, slots=True)
class CanonicalAnnotation:
video_id: str
segment_id: str
category: str
start_ms: int
end_ms: int
def load_canonical_annotations(
path: Path,
*,
video_ids: set[str] | None = None,
categories: Sequence[str] | None = None,
eligible_only: bool = True,
batch_size: int = 131_072,
) -> dict[str, list[CanonicalAnnotation]]:
"""Read the canonical annotation table, grouped and sorted by video.
This is the single join point every downstream builder uses instead of
re-interpreting the raw mirror. Cross-category overlaps are returned
intact; each annotation keeps its own category identity.
"""
import pyarrow.parquet as parquet
columns = ["video_id", "segment_id", "category", "start_ms", "end_ms", "is_eligible"]
category_set = set(categories) if categories is not None else None
annotations: dict[str, list[CanonicalAnnotation]] = defaultdict(list)
file = parquet.ParquetFile(path)
available = set(file.schema_arrow.names)
missing = [column for column in columns if column not in available]
if missing:
raise ValueError(
f"canonical annotation table is missing columns: {', '.join(missing)}"
)
for batch in file.iter_batches(columns=columns, batch_size=batch_size):
values = batch.to_pydict()
for video_id, segment_id, category, start_ms, end_ms, is_eligible in zip(
*(values[column] for column in columns), strict=True
):
if video_id is None or start_ms is None or end_ms is None:
continue
if video_ids is not None and video_id not in video_ids:
continue
if category_set is not None and category not in category_set:
continue
if eligible_only and not is_eligible:
continue
annotations[str(video_id)].append(
CanonicalAnnotation(
video_id=str(video_id),
segment_id=str(segment_id) if segment_id is not None else "",
category=str(category),
start_ms=int(start_ms),
end_ms=int(end_ms),
)
)
for records in annotations.values():
records.sort(key=lambda annotation: (annotation.start_ms, annotation.end_ms))
return annotations
|