Datasets:
File size: 17,601 Bytes
e1ced61 | 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 | """GQA-v1.2 source adapter backed only by native questions and scene graphs.
The released question files are mappings keyed by question id. Each question
contains a native dependency-indexed ``semantic`` program and points to an image
whose released scene graph supplies object names, attributes, relations, and
bounding boxes. This adapter joins those records deterministically and never
uses a VLM or the source answer to infer a program operand.
"""
from __future__ import annotations
import json
from collections.abc import Iterator, Mapping
from pathlib import Path
from typing import TYPE_CHECKING, Any
from ..hashing import canonical_json_hash
from ..ingest.base import (
AnswerType,
ImageStore,
IngestError,
NormalizedItem,
Policy,
infer_open_answer_type,
make_item,
)
from .base import AdapterError, CertificateTier, RawItem, World, store_images
if TYPE_CHECKING:
from ..executors.gqa import GQASemanticProgram
SOURCE = "gqa"
WORLD_SCHEMA = "gqa_scene_graph_v1"
_HIDDEN_OBJECTS_KEY = "_gqa_hidden_object_ids"
_VALID_SPLITS = frozenset({"train", "val", "validation", "test", "testdev", "challenge"})
class GQAAdapter:
"""Deterministic adapter for GQA-v1.2 native JSON files."""
def __init__(
self,
raw_dir: Path,
store: ImageStore,
*,
revision: str,
images_dir: Path | None = None,
questions_path: Path | None = None,
scene_graphs_path: Path | None = None,
) -> None:
self.raw_dir = Path(raw_dir)
self.store = store
self.revision = revision
self.images_dir = Path(images_dir) if images_dir is not None else self.raw_dir / "images"
self.questions_path = Path(questions_path) if questions_path is not None else None
self.scene_graphs_path = Path(scene_graphs_path) if scene_graphs_path is not None else None
@classmethod
def is_materialized(cls, raw_dir: Path, split: str) -> bool:
if split not in _VALID_SPLITS:
return False
root = Path(raw_dir)
return (
_first_existing(root, _question_names(split)) is not None
and _first_existing(root, _scene_graph_names(split)) is not None
and (root / "images").is_dir()
)
@classmethod
def materialize(
cls,
raw_dir: Path,
split: str,
*,
source_config: Mapping[str, Any],
expected_sha256: Mapping[str, str] | None = None,
) -> Path:
"""Validate an already materialized native GQA tree.
GQA's multi-gigabyte image/question archives are intentionally not
downloaded implicitly. Callers materialize the pinned source snapshot
out of band, then this method performs the same fail-closed check used by
the ingest CLI. ``source_config`` and ``expected_sha256`` are accepted
for protocol parity; hashes belong in the surrounding resource manifest.
"""
del source_config, expected_sha256
root = Path(raw_dir)
if not cls.is_materialized(root, split):
raise AdapterError(
f"{SOURCE}: native {split!r} questions, scene graphs, and images "
f"are not materialized under {root}"
)
return root
def iter_base_items(self, split: str) -> Iterator[RawItem]:
from ..executors.gqa import compile_gqa_program
if split not in _VALID_SPLITS:
raise AdapterError(f"{SOURCE}: unsupported split {split!r}")
questions = _load_mapping(self._questions_file(split), label="questions")
scene_graphs = _load_mapping(self._scene_graphs_file(split), label="scene graphs")
for question_id in sorted(str(key) for key in questions):
row = questions.get(question_id)
if not isinstance(row, Mapping):
raise AdapterError(f"{SOURCE}/{question_id}: question row must be an object")
image_id = str(row.get("imageId", "")).strip()
if not image_id:
raise AdapterError(f"{SOURCE}/{question_id}: imageId missing/empty")
scene_graph = scene_graphs.get(image_id)
if not isinstance(scene_graph, Mapping):
raise AdapterError(f"{SOURCE}/{question_id}: scene graph {image_id!r} missing")
# Compile during iteration so an unsupported native operation is a
# typed hard rejection, never a silently dropped candidate.
compile_gqa_program(question_id, row)
image = self._image_bytes(image_id)
payload = {
"question_id": question_id,
**dict(row),
"scene_graph": dict(scene_graph),
}
yield RawItem(
source=SOURCE,
split=split,
source_revision=self.revision,
native_id=question_id,
payload=payload,
images={image_id: image},
)
def normalize(self, raw: RawItem) -> NormalizedItem:
from ..executors.gqa import compile_gqa_program
if raw.source != SOURCE:
raise IngestError(f"{SOURCE}: cannot normalize raw source {raw.source!r}")
row = raw.payload
question = row.get("question")
if not isinstance(question, str) or not question.strip():
raise IngestError(f"{SOURCE}/{raw.native_id}: question missing/empty")
answer = row.get("answer")
if answer is None or not str(answer).strip():
raise IngestError(f"{SOURCE}/{raw.native_id}: answer missing/empty")
image_id = str(row.get("imageId", "")).strip()
scene_graph = row.get("scene_graph")
if not image_id or not isinstance(scene_graph, Mapping):
raise IngestError(f"{SOURCE}/{raw.native_id}: joined scene graph missing")
if not raw.images:
raise IngestError(f"{SOURCE}/{raw.native_id}: image missing")
program = compile_gqa_program(raw.native_id, row)
paths, image_hashes = store_images(self.store, raw.images)
answer_text = str(answer).strip()
answer_type: AnswerType
if answer_text.casefold() in {"yes", "no"}:
answer_type = "boolean"
answer_canonical = answer_text.casefold()
else:
answer_type = infer_open_answer_type(answer_text)
answer_canonical = answer_text
policy: Policy = (
"c1_train_candidate"
if raw.split in {"train", "val", "validation"}
else "c1_certified_eval_candidate"
)
types = row.get("types")
subject = None
if isinstance(types, Mapping) and types.get("semantic"):
subject = f"gqa_{str(types['semantic']).strip()}"
extra = {
"image_id": image_id,
"semantic": [step.to_dict() for step in program.steps],
"semanticStr": program.semantic_str,
"semantic_program_sha256": program.canonical_program_sha256,
"annotations": row.get("annotations")
if isinstance(row.get("annotations"), Mapping)
else {},
"types": dict(types) if isinstance(types, Mapping) else {},
"scene_graph": dict(scene_graph),
"source_record_sha256": canonical_json_hash(
{
"question_id": raw.native_id,
"question": {key: value for key, value in row.items() if key != "scene_graph"},
"scene_graph": scene_graph,
}
),
}
return make_item(
source=SOURCE,
source_revision=raw.source_revision,
source_config="default",
source_split=raw.split,
source_native_id=raw.native_id,
question=question,
choices=[],
answer_raw=answer_text,
answer_canonical=answer_canonical,
answer_type=answer_type,
image_paths=paths,
image_sha256=image_hashes,
policy=policy,
native_row={key: value for key, value in row.items() if key != "scene_graph"},
subject=subject,
extra_provenance=extra,
)
def build_world(self, item: NormalizedItem) -> World:
scene_graph = item.provenance.get("scene_graph")
image_id = str(item.provenance.get("image_id", "")).strip()
if not isinstance(scene_graph, Mapping) or not image_id:
raise IngestError(f"{SOURCE}/{item.source_native_id}: scene graph provenance missing")
return build_gqa_world(scene_graph, image_id=image_id, question_id=item.source_native_id)
def get_or_compile_program(self, item: NormalizedItem) -> GQASemanticProgram:
from ..executors.gqa import compile_gqa_program
semantic = item.provenance.get("semantic")
if not isinstance(semantic, list):
raise IngestError(f"{SOURCE}/{item.source_native_id}: semantic provenance missing")
return compile_gqa_program(
item.source_native_id,
{
"imageId": item.provenance.get("image_id"),
"semantic": semantic,
"semanticStr": item.provenance.get("semanticStr", ""),
},
)
def official_answer(self, item: NormalizedItem) -> str | int | bool:
return item.answer_canonical
def source_certificate_tier(self, item: NormalizedItem) -> CertificateTier:
return "C1_SOURCE_NATIVE"
def _questions_file(self, split: str) -> Path:
if self.questions_path is not None:
if not self.questions_path.is_file():
raise AdapterError(f"{SOURCE}: questions file not found: {self.questions_path}")
return self.questions_path
path = _first_existing(self.raw_dir, _question_names(split))
if path is None:
raise AdapterError(f"{SOURCE}: no native questions file for split {split!r}")
return path
def _scene_graphs_file(self, split: str) -> Path:
if self.scene_graphs_path is not None:
if not self.scene_graphs_path.is_file():
raise AdapterError(
f"{SOURCE}: scene-graphs file not found: {self.scene_graphs_path}"
)
return self.scene_graphs_path
path = _first_existing(self.raw_dir, _scene_graph_names(split))
if path is None:
raise AdapterError(f"{SOURCE}: no native scene-graphs file for split {split!r}")
return path
def _image_bytes(self, image_id: str) -> bytes:
roots = (self.images_dir, self.raw_dir / "images", self.raw_dir)
for root in roots:
for suffix in (".jpg", ".jpeg", ".png"):
candidate = root / f"{image_id}{suffix}"
if candidate.is_file():
return candidate.read_bytes()
raise AdapterError(f"{SOURCE}: image {image_id!r} not found")
def build_gqa_world(
scene_graph: Mapping[str, Any],
*,
image_id: str,
question_id: str | None = None,
) -> World:
"""Validate and canonicalize one native GQA scene graph.
Object ids remain the released ids. Relations are sorted only after their
source order has been validated, making hashes stable even when the native
relation container is a mapping. Dangling relation targets are rejected.
"""
width = _positive_int(scene_graph.get("width"), "width")
height = _positive_int(scene_graph.get("height"), "height")
raw_objects = scene_graph.get("objects")
if not isinstance(raw_objects, Mapping) or not raw_objects:
raise IngestError(f"{SOURCE}/{image_id}: objects must be a non-empty mapping")
objects: dict[str, dict[str, Any]] = {}
for raw_id in sorted(raw_objects, key=str):
object_id = str(raw_id)
raw = raw_objects[raw_id]
if not isinstance(raw, Mapping):
raise IngestError(f"{SOURCE}/{image_id}: object {object_id!r} must be an object")
name = str(raw.get("name", "")).strip()
if not name:
raise IngestError(f"{SOURCE}/{image_id}: object {object_id!r} name missing")
x = _nonnegative_int(raw.get("x"), f"object {object_id} x")
y = _nonnegative_int(raw.get("y"), f"object {object_id} y")
w = _positive_int(raw.get("w"), f"object {object_id} w")
h = _positive_int(raw.get("h"), f"object {object_id} h")
if x + w > width or y + h > height:
raise IngestError(
f"{SOURCE}/{image_id}: object {object_id!r} bbox exceeds image bounds"
)
raw_attributes = raw.get("attributes", [])
if not isinstance(raw_attributes, list) or any(
not isinstance(attribute, str) for attribute in raw_attributes
):
raise IngestError(
f"{SOURCE}/{image_id}: object {object_id!r} attributes must be strings"
)
relations = _normalize_relations(raw.get("relations", []), image_id, object_id)
objects[object_id] = {
"id": object_id,
"name": name,
"x": x,
"y": y,
"w": w,
"h": h,
"attributes": list(raw_attributes),
"relations": relations,
}
known = frozenset(objects)
for source_id, obj in objects.items():
for relation in obj["relations"]:
target = str(relation["object"])
if target not in known:
raise IngestError(
f"{SOURCE}/{image_id}: relation {source_id!r}->{target!r} is dangling"
)
world: World = {
"world_schema": WORLD_SCHEMA,
"image_id": str(image_id),
"width": width,
"height": height,
"objects": objects,
_HIDDEN_OBJECTS_KEY: [],
"provenance": {
"source": SOURCE,
"image_id": str(image_id),
"question_id": question_id,
},
}
for field in ("location", "weather"):
value = scene_graph.get(field)
if isinstance(value, str) and value.strip():
world[field] = value.strip()
return world
def _normalize_relations(
raw: Any,
image_id: str,
object_id: str,
) -> list[dict[str, str]]:
values = list(raw.values()) if isinstance(raw, Mapping) else raw
if not isinstance(values, list):
raise IngestError(
f"{SOURCE}/{image_id}: object {object_id!r} relations must be a list or mapping"
)
out: list[dict[str, str]] = []
for index, relation in enumerate(values):
if not isinstance(relation, Mapping):
raise IngestError(
f"{SOURCE}/{image_id}: object {object_id!r} relation {index} must be an object"
)
name = str(relation.get("name", "")).strip()
target = str(relation.get("object", "")).strip()
if not name or not target:
raise IngestError(
f"{SOURCE}/{image_id}: object {object_id!r} relation {index} is incomplete"
)
out.append({"name": name, "object": target})
return sorted(out, key=lambda relation: (relation["name"], relation["object"]))
def _question_names(split: str) -> tuple[str, ...]:
aliases = _split_aliases(split)
names: list[str] = []
for alias in aliases:
names.extend(
(
f"{alias}_balanced_questions.json",
f"{alias}_all_questions.json",
f"{alias}_questions.json",
)
)
names.extend(("questions.json", "Questions.json"))
return tuple(names)
def _scene_graph_names(split: str) -> tuple[str, ...]:
aliases = _split_aliases(split)
names: list[str] = []
for alias in aliases:
names.extend(
(
f"{alias}_sceneGraphs.json",
f"{alias}_scene_graphs.json",
)
)
names.extend(("sceneGraphs.json", "scene_graphs.json", "Scene_graphs.json"))
return tuple(names)
def _split_aliases(split: str) -> tuple[str, ...]:
if split == "validation":
return ("validation", "val")
if split == "val":
return ("val", "validation")
return (split,)
def _first_existing(root: Path, names: tuple[str, ...]) -> Path | None:
for name in names:
path = root / name
if path.is_file():
return path
return None
def _load_mapping(path: Path, *, label: str) -> dict[str, Any]:
try:
value = json.loads(path.read_text(encoding="utf-8"))
except (OSError, UnicodeDecodeError, json.JSONDecodeError) as exc:
raise AdapterError(f"{SOURCE}: cannot read {label} at {path}: {exc}") from exc
if not isinstance(value, Mapping):
raise AdapterError(f"{SOURCE}: {label} at {path} must be a JSON object")
return {str(key): row for key, row in value.items()}
def _positive_int(value: Any, label: str) -> int:
number = _integer(value, label)
if number <= 0:
raise IngestError(f"{SOURCE}: {label} must be positive")
return number
def _nonnegative_int(value: Any, label: str) -> int:
number = _integer(value, label)
if number < 0:
raise IngestError(f"{SOURCE}: {label} must be non-negative")
return number
def _integer(value: Any, label: str) -> int:
if isinstance(value, bool) or not isinstance(value, int):
raise IngestError(f"{SOURCE}: {label} must be an integer")
return int(value)
__all__ = [
"GQAAdapter",
"SOURCE",
"WORLD_SCHEMA",
"build_gqa_world",
]
|