Instructions to use XXXpilar/Fastfill-test with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use XXXpilar/Fastfill-test with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
File size: 16,159 Bytes
e6f80f3 | 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 | """Publish a new evidence-qualified revision; never repair parent geometry.
Only frozen source-derived ``faces`` constraints lacking semantic-yaw evidence
are omitted. Explicit user requirements survive. Below-floor target positions
are diagnostics, not clamped labels. This module intentionally uses stdlib only.
"""
from __future__ import annotations
import argparse
from collections import Counter
from copy import deepcopy
import hashlib
from itertools import zip_longest
import json
import math
import os
from pathlib import Path
import shutil
import tempfile
SPLITS = ("train", "validation", "test")
INPUT_FILES = tuple(f"{split}.jsonl" for split in SPLITS) + ("manifest.json", "rejections.jsonl")
OUTPUT_FILES = tuple(name for name in INPUT_FILES if name != "manifest.json")
POLICY = "source-derived-facing-evidence-v1"
BUILDER = "immutable-parent-review-v1"
FLOOR_TOLERANCE_M = 1e-4
LEGACY_CONSTRAINT_SOURCE = "frozen_sparse_legacy_request"
RAW_SOURCE_ROOT = Path("/Volumes/harddisk/3D_Room_Collections")
REVIEW_FIELDS = ("review_parent_scene_id", "review_policy", "review_source_qualification",
"review_omitted_constraints", "review_geometry_diagnostics")
def _digest(path):
value = hashlib.sha256()
with Path(path).open("rb") as stream:
for block in iter(lambda: stream.read(1024 * 1024), b""):
value.update(block)
return value.hexdigest()
def _reject_constant(value):
raise ValueError(f"nonfinite JSON number: {value}")
def _read_json(value):
return json.loads(value, parse_constant=_reject_constant)
def _parent_inputs(root):
for name in INPUT_FILES:
if not (root / name).is_file():
raise FileNotFoundError(f"required parent input missing: {root / name}")
raw = (root / "manifest.json").read_bytes()
manifest = _read_json(raw)
if not isinstance(manifest, dict) or manifest.get("schema_version") != "fastfill.v2":
raise ValueError("parent manifest must identify fastfill.v2 data")
hashes = {name: _digest(root / name) for name in INPUT_FILES if name != "manifest.json"}
return manifest, {**hashes, "manifest.json": hashlib.sha256(raw).hexdigest()}
def _recheck(root, hashes):
if any(_digest(root / name) != expected for name, expected in hashes.items()):
raise ValueError("parent input or manifest changed during review")
def _new_output(parent_root, output):
if Path(output).is_symlink():
raise FileExistsError(f"immutable review output already exists as a symlink: {output}")
parent, target = Path(parent_root).resolve(), Path(output).resolve()
for protected in (parent, RAW_SOURCE_ROOT.resolve()):
if target == protected or protected in target.parents or target in protected.parents:
raise ValueError("output must be outside source data and its ancestors")
if target.exists():
raise FileExistsError(f"immutable review output already exists: {target}")
return parent, target
def _complete(mask, dimensions):
return mask is True or (isinstance(mask, list) and len(mask) == dimensions
and all(value is True for value in mask))
def _mask(sample, field, index, dimensions):
values = sample["validity"].get(field, [])
return index < len(values) and _complete(values[index], dimensions)
def _finite(value):
return isinstance(value, (int, float)) and not isinstance(value, bool) and math.isfinite(value)
def _indices(sample):
if not isinstance(sample, dict) or sample.get("schema_version") != "fastfill.v2":
raise ValueError("parent rows must use fastfill.v2 schema")
condition, target = sample.get("condition", {}), sample.get("target", {})
if not isinstance(condition, dict) or not isinstance(target, dict):
raise ValueError("condition and target must be objects")
collections = (condition.get("objects"), target.get("objects"))
ids = []
for objects in collections:
if not isinstance(objects, list) or any(not isinstance(obj, dict)
or not isinstance(obj.get("id"), str) or not obj["id"] for obj in objects):
raise ValueError("request and target objects require nonempty string IDs")
names = [obj["id"] for obj in objects]
if len(names) != len(set(names)):
raise ValueError("duplicate request or target ID")
ids.append(names)
if set(ids[0]) != set(ids[1]):
raise ValueError("target IDs must equal requested IDs exactly once")
validity = sample.get("validity")
if not isinstance(validity, dict) or any(not isinstance(validity.get(field, []), list)
or len(validity.get(field, [])) > len(ids[1]) for field in ("position", "size", "yaw")):
raise ValueError("invalid geometry validity rows")
if not isinstance(condition.get("room"), dict) or not isinstance(condition.get("constraints", []), list):
raise ValueError("room and constraints must be structured fields")
provenance = sample.get("provenance")
if not isinstance(provenance, dict) or any(not isinstance(provenance.get(key), str)
or not provenance[key] for key in ("scene_id", "legacy_uid")):
raise ValueError("parent provenance requires stable scene_id and legacy_uid")
evidence = provenance.get("field_evidence", [])
if (not isinstance(evidence, list) or len(evidence) > len(ids[1])
or any(not isinstance(row, dict) for row in evidence)):
raise ValueError("field evidence must follow target object indices")
if any(key in provenance for key in REVIEW_FIELDS):
raise ValueError("parent already contains review provenance")
return {name: index for index, name in enumerate(ids[1])}
def _trusted_yaw(sample, index):
"""Facing needs a semantic front: a valid yaw whose symmetry order is 1 (axis-only mod-pi yaw is not evidence)."""
orders = sample["validity"].get("yaw_symmetry_order", [])
if not _mask(sample, "yaw", index, 1) or index < len(orders) and orders[index] != 1:
return False
angle = sample["target"]["objects"][index].get("yaw_rad")
if not _finite(angle) or not -math.pi <= angle < math.pi:
raise ValueError("valid yaw label must be finite and wrapped to [-pi, pi)")
evidence = sample["provenance"].get("field_evidence", [])
field = evidence[index] if index < len(evidence) else {}
recorded = field.get("recorded_source_evidence", {})
if not isinstance(recorded, dict):
raise ValueError("recorded source evidence must be an object")
return "front_unknown_reason" not in recorded
def _floor_diagnostics(sample):
room = sample["condition"]["room"]
known = room.get("floor_known") is True
floor = room.get("floor_z_m") if known else None
if known and not _finite(floor):
raise ValueError("known floor requires a finite floor_z_m")
conflicts = []
for index, target in enumerate(sample["target"]["objects"]):
if not known or not _mask(sample, "position", index, 3):
continue
position = target.get("bottom_center_m")
if not isinstance(position, list) or len(position) != 3 or not all(map(_finite, position)):
raise ValueError("valid position label must be finite XYZ")
if position[2] < floor - FLOOR_TOLERANCE_M:
conflicts.append({"object_id": target["id"], "bottom_minus_floor_m": position[2] - floor})
return {"known_floor": known, "floor_z_m": floor, "tolerance_m": FLOOR_TOLERANCE_M,
"below_known_floor": conflicts,
"minimum_bottom_minus_floor_m": min((c["bottom_minus_floor_m"] for c in conflicts), default=None),
"labels_modified": False}
def review_sample(parent):
"""Pure derivation: only omit unverified frozen legacy facing constraints."""
indices = _indices(parent)
legacy = parent["provenance"].get("constraint_source") == LEGACY_CONSTRAINT_SOURCE
kept, omitted = [], []
for constraint in parent["condition"].get("constraints", []):
if not isinstance(constraint, dict):
raise ValueError("constraint must be an object")
if legacy and constraint.get("type") == "faces":
object_id = constraint.get("object_id")
if object_id not in indices:
raise ValueError("source-derived facing refers to an unknown request ID")
if not _trusted_yaw(parent, indices[object_id]):
omitted.append({"constraint": deepcopy(constraint),
"reason": "unverified_source_derived_facing"})
continue
kept.append(deepcopy(constraint))
result = deepcopy(parent)
if "constraints" in parent["condition"]:
result["condition"] = {**result["condition"], "constraints": kept}
result["provenance"] = {**result["provenance"],
"review_parent_scene_id": parent["provenance"]["scene_id"], "review_policy": POLICY,
"review_source_qualification": "Frozen legacy constraints are source-derived annotations, not newly "
"synthesized user requirements; unverified source-derived facing is omitted. "
"Explicit/nonlegacy requirements and existing exchangeability groups are preserved.",
"review_omitted_constraints": omitted, "review_geometry_diagnostics": _floor_diagnostics(parent)}
return result
def _rows(root, split, seen):
with (root / f"{split}.jsonl").open("rb") as stream:
for number, raw in enumerate(stream, 1):
if not raw.strip():
continue
try:
sample = _read_json(raw)
provenance = sample["provenance"]
uid = provenance["legacy_uid"]
if not isinstance(uid, str) or not uid or uid in seen or provenance.get("split") != split:
raise ValueError("duplicate/missing UID or provenance split disagrees with its file")
seen.add(uid)
except (ValueError, KeyError, TypeError) as exc:
raise ValueError(f"invalid row {split}.jsonl:{number}: {exc}") from exc
yield sample
def _empty_counts():
return {key: Counter() for key in ("split_samples", "split_objects", "source_split_samples",
"source_omitted_faces", "source_floor_conflict_objects", "totals")}
def _count(counts, sample, split):
source = sample["provenance"].get("source", "unknown")
omitted = len(sample["provenance"]["review_omitted_constraints"])
conflicts = len(sample["provenance"]["review_geometry_diagnostics"]["below_known_floor"])
counts["split_samples"][split] += 1
counts["split_objects"][split] += len(sample["target"]["objects"])
counts["source_split_samples"][f"{source}:{split}"] += 1
for key, value in (("omitted_faces", omitted), ("rows_with_omitted_faces", bool(omitted)),
("floor_conflict_objects", conflicts), ("rows_with_floor_conflicts", bool(conflicts))):
counts["totals"][key] += value
if omitted:
counts["source_omitted_faces"][source] += omitted
if conflicts:
counts["source_floor_conflict_objects"][source] += conflicts
def _summary(counts):
result = {key: dict(value) for key, value in counts.items() if key != "totals"}
for key in ("split_samples", "split_objects"):
result[key] = {split: counts[key][split] for split in SPLITS}
return {**result, **counts["totals"], "samples_written": sum(counts["split_samples"].values()),
"objects_written": sum(counts["split_objects"].values())}
def _write_records(parent, staging):
seen, counts = set(), _empty_counts()
for split in SPLITS:
with (staging / f"{split}.jsonl").open("xb") as output:
for record in _rows(parent, split, seen):
child = review_sample(record)
output.write((json.dumps(child, ensure_ascii=False, allow_nan=False,
separators=(",", ":")) + "\n").encode())
_count(counts, child, split)
shutil.copyfile(parent / "rejections.jsonl", staging / "rejections.jsonl")
return _summary(counts)
def build_reviewed(parent_root, output):
"""Stream all rows and atomically publish to a new immutable directory."""
parent, target = _new_output(parent_root, output)
manifest, hashes = _parent_inputs(parent)
target.parent.mkdir(parents=True, exist_ok=True)
with tempfile.TemporaryDirectory(prefix=f".{target.name}-", dir=target.parent) as temporary:
staging = Path(temporary)
result = {"schema_version": "fastfill.v2", "builder": BUILDER, "review_policy": POLICY,
"parent_root": str(parent), "parent_builder": manifest.get("builder"),
"parent_front_policy": manifest.get("front_policy"),
"front_policy": manifest.get("front_policy"), "yaw_policy": manifest.get("yaw_policy"), "parent_sha256": hashes,
"split_rule": "inherit every parent UID and split without re-splitting or filtering",
"labels_modified": False, "source_data_modified": False,
"implementation_sha256": _digest(__file__), **_write_records(parent, staging),
"output_sha256": {name: _digest(staging / name) for name in OUTPUT_FILES}}
(staging / "manifest.json").write_text(json.dumps(result, indent=2, allow_nan=False) + "\n")
_recheck(parent, hashes)
if target.exists() or target.is_symlink():
raise FileExistsError("immutable review output appeared during derivation")
os.rename(staging, target)
return result
def verify_reviewed(data_root, parent_root):
"""Check fingerprints, every deterministic row, preserved identities and totals."""
child, parent = Path(data_root).resolve(), Path(parent_root).resolve()
_, parent_hashes = _parent_inputs(parent)
manifest, child_hashes = _parent_inputs(child)
if (manifest.get("builder") != BUILDER or manifest.get("review_policy") != POLICY
or manifest.get("parent_sha256") != parent_hashes):
raise ValueError("review manifest policy or parent fingerprints do not match")
expected_hashes = {name: child_hashes[name] for name in OUTPUT_FILES}
if manifest.get("output_sha256") != expected_hashes:
raise ValueError("review output hashes do not match manifest")
if child_hashes["rejections.jsonl"] != parent_hashes["rejections.jsonl"]:
raise ValueError("parent rejections were changed")
parent_seen, child_seen, counts = set(), set(), _empty_counts()
for split in SPLITS:
pairs = zip_longest(_rows(parent, split, parent_seen), _rows(child, split, child_seen))
for number, (original, observed) in enumerate(pairs, 1):
if original is None or observed != review_sample(original):
raise ValueError(f"reviewed row differs from deterministic parent revision: {split}:{number}")
_count(counts, observed, split)
summary = _summary(counts)
if any(manifest.get(key) != value for key, value in summary.items()):
raise ValueError("review manifest counts do not match verified rows")
if manifest.get("labels_modified") is not False or manifest.get("source_data_modified") is not False:
raise ValueError("review manifest must declare unchanged labels and parent data")
_recheck(parent, parent_hashes)
_recheck(child, child_hashes)
return {"verified": True, **summary, "manifest_sha256": child_hashes["manifest.json"]}
def main(argv=None):
parser = argparse.ArgumentParser(description=__doc__)
commands = parser.add_subparsers(dest="command", required=True)
for command in ("build", "verify"):
sub = commands.add_parser(command)
sub.add_argument("--parent-root", type=Path, required=True)
sub.add_argument("--output" if command == "build" else "--data-root", type=Path, required=True)
args = vars(parser.parse_args(argv))
command = args.pop("command")
result = build_reviewed(**args) if command == "build" else verify_reviewed(**args)
print(json.dumps(result, indent=2, allow_nan=False))
if __name__ == "__main__":
main()
|