File size: 14,752 Bytes
2ea2835 | 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 | """Metadata-only matched-video dataset with explicitly lazy teacher I/O."""
from __future__ import annotations
from dataclasses import dataclass
import os
from pathlib import Path
import time
from typing import Any, Callable, Mapping, Sequence
import numpy as np
import torch
from torch.utils.data import Dataset
from utils.matched_video_feature_cache import (
MATCHED_PYAV_LAST64_DECODER_CONTRACT,
MATCHED_VJEPA2_L15_LAST64_FEATURE_CONTRACT,
MatchedVideoFeatureCache,
)
from utils.matched_video_manifest import (
MatchedVideoManifestEntry,
load_matched_video_manifest,
)
@dataclass(frozen=True)
class DecodedMatchedVideo:
"""The final contiguous RGB frame window in V-JEPA2 input range."""
video01: torch.Tensor
frame_indices: torch.Tensor
source_frame_count: int
def _frames_to_video01(frames: Sequence[Any]) -> torch.Tensor:
tensors: list[torch.Tensor] = []
shape: tuple[int, int, int] | None = None
for frame in frames:
array = frame.to_ndarray(format="rgb24")
if not isinstance(array, np.ndarray):
raise TypeError("PyAV frame.to_ndarray did not return ndarray")
if array.ndim != 3 or int(array.shape[2]) != 3:
raise ValueError(
f"decoded RGB frame must be [H,W,3], got {array.shape}"
)
if array.dtype != np.uint8:
raise ValueError(f"decoded RGB frame must be uint8, got {array.dtype}")
current_shape = tuple(int(value) for value in array.shape)
if shape is None:
shape = current_shape
elif current_shape != shape:
raise ValueError(
"video changes resolution inside selected frame window: "
f"{current_shape} != {shape}"
)
tensors.append(
torch.from_numpy(np.ascontiguousarray(array)).permute(2, 0, 1)
)
if not tensors:
raise ValueError("decoded frame window is empty")
return torch.stack(tensors).to(dtype=torch.float32).div_(255.0)
def decode_matched_video_last_contiguous_pyav(
video_path: str | os.PathLike,
*,
decoder_contract: Mapping[str, Any] | None = None,
) -> DecodedMatchedVideo:
"""Decode exactly frames 17..80 from a contractually 81-frame MP4."""
contract = dict(
decoder_contract
if decoder_contract is not None
else MATCHED_PYAV_LAST64_DECODER_CONTRACT
)
if contract.get("backend") != "pyav":
raise ValueError("decoder contract backend must be pyav")
if contract.get("selection") != "last_contiguous":
raise ValueError("decoder contract selection must be last_contiguous")
num_frames = int(contract.get("num_frames", 0))
expected_count = int(contract.get("expected_source_frame_count", 0))
if num_frames <= 0 or expected_count < num_frames:
raise ValueError("decoder contract frame geometry is invalid")
expected_bounds = [expected_count - num_frames, expected_count - 1]
if list(contract.get("expected_frame_indices", ())) != expected_bounds:
raise ValueError(
"decoder contract expected_frame_indices do not describe last64"
)
try:
import av # Imported only for cache misses.
except ImportError as exc: # pragma: no cover
raise ImportError(
"PyAV is required for a matched-video teacher cache miss; "
"install the pinned av==13.1.0 dependency"
) from exc
expected_version = str(contract.get("backend_version", ""))
actual_version = str(getattr(av, "__version__", ""))
if actual_version != expected_version:
raise RuntimeError(
"PyAV version differs from the cache decoder contract: "
f"{actual_version!r} != {expected_version!r}"
)
path = Path(video_path).expanduser().resolve()
if not path.is_file():
raise FileNotFoundError(path)
selected: list[Any] = []
decoded_count = 0
start = expected_count - num_frames
with av.open(str(path), mode="r") as container:
streams = container.streams.video
if len(streams) != 1:
raise ValueError(
f"expected exactly one video stream in {path}, got {len(streams)}"
)
stream = streams[0]
for frame_index, frame in enumerate(container.decode(stream)):
decoded_count = frame_index + 1
if start <= frame_index < expected_count:
selected.append(frame)
if decoded_count != expected_count:
raise ValueError(
"matched video source frame count violates the last64 contract: "
f"sample={path.name}, actual={decoded_count}, "
f"expected={expected_count}"
)
if len(selected) != num_frames:
raise RuntimeError(
f"failed to decode frames {start}..{expected_count - 1} from {path}"
)
video01 = _frames_to_video01(selected)
if tuple(video01.shape[:2]) != (num_frames, 3):
raise RuntimeError(f"decoded video has invalid shape {tuple(video01.shape)}")
return DecodedMatchedVideo(
video01=video01,
frame_indices=torch.arange(start, expected_count, dtype=torch.long),
source_frame_count=decoded_count,
)
DecodeFunction = Callable[[Path, Mapping[str, Any]], DecodedMatchedVideo]
def _default_decode(
path: Path,
contract: Mapping[str, Any],
) -> DecodedMatchedVideo:
return decode_matched_video_last_contiguous_pyav(
path, decoder_contract=contract
)
class MatchedManifestVideoDataset(Dataset):
"""Return metadata eagerly; teacher cache/decode only on explicit request."""
def __init__(
self,
manifest_path: str | os.PathLike,
video_root: str | os.PathLike,
cache_root: str | os.PathLike,
expected_model_sha256: str,
expected_feature_contract_sha256: str,
*,
expected_source_csv_sha256: str | None = None,
expected_count: int | None = None,
feature_contract: Mapping[str, Any] | None = None,
decoder_contract: Mapping[str, Any] | None = None,
decode_fn: DecodeFunction | None = None,
) -> None:
self.manifest_path = Path(manifest_path).expanduser().resolve()
self.video_root = Path(video_root).expanduser().resolve()
if not self.video_root.is_dir():
raise NotADirectoryError(self.video_root)
self.entries, self.manifest_metadata = load_matched_video_manifest(
self.manifest_path,
expected_source_csv_sha256=expected_source_csv_sha256,
expected_count=expected_count,
)
if not self.entries:
raise ValueError("matched-video manifest is empty")
self.feature_cache = MatchedVideoFeatureCache(
cache_root,
expected_model_sha256=expected_model_sha256,
expected_feature_contract_sha256=expected_feature_contract_sha256,
feature_contract=(
feature_contract
if feature_contract is not None
else MATCHED_VJEPA2_L15_LAST64_FEATURE_CONTRACT
),
decoder_contract=(
decoder_contract
if decoder_contract is not None
else MATCHED_PYAV_LAST64_DECODER_CONTRACT
),
)
self.decoder_contract = self.feature_cache.decoder_contract
self.decode_fn = decode_fn if decode_fn is not None else _default_decode
self._cache_key_to_index: dict[str, int] = {}
self._sample_id_to_index: dict[str, int] = {}
for index, entry in enumerate(self.entries):
cache_key = self.feature_cache.key_for(entry)
if cache_key in self._cache_key_to_index:
raise ValueError(f"duplicate matched-video cache key: {cache_key}")
if entry.sample_id in self._sample_id_to_index:
raise ValueError(f"duplicate sample_id: {entry.sample_id}")
self._cache_key_to_index[cache_key] = index
self._sample_id_to_index[entry.sample_id] = index
def __len__(self) -> int:
return len(self.entries)
def _entry_for_sample(self, sample_id: str) -> MatchedVideoManifestEntry:
try:
return self.entries[self._sample_id_to_index[str(sample_id)]]
except KeyError as exc:
raise KeyError(f"unknown matched-video sample_id: {sample_id!r}") from exc
def _video_path_and_verify_stat(
self,
entry: MatchedVideoManifestEntry,
) -> Path:
path = self.video_root / entry.video_name
if not path.is_file():
raise FileNotFoundError(path)
stat = path.stat()
actual = (int(stat.st_size), int(stat.st_mtime_ns))
expected = (entry.video_size_bytes, entry.video_mtime_ns)
if actual != expected:
raise ValueError(
"matched video is stale after manifest freeze for "
f"sample_id={entry.sample_id}: actual={actual}, "
f"expected={expected}"
)
return path
def cache_key_for_sample(self, sample_id: str) -> str:
return self.feature_cache.key_for(self._entry_for_sample(sample_id))
def __getitem__(self, index: int) -> dict[str, Any]:
"""Return only prompt/sample/path metadata; never touch teacher data."""
entry = self.entries[int(index)]
video_path = self._video_path_and_verify_stat(entry)
return {
"idx": entry.manifest_index,
"prompts": entry.caption,
"sample_id": entry.sample_id,
"teacher_cache_key": self.feature_cache.key_for(entry),
"video_name": entry.video_name,
"video_path": str(video_path),
}
def load_teacher_payload(self, sample_id: str) -> dict[str, Any]:
"""Load/decode teacher input only after suffix replay selects a sample."""
entry = self._entry_for_sample(sample_id)
video_path = self._video_path_and_verify_stat(entry)
cache_key = self.feature_cache.key_for(entry)
cached = self.feature_cache.load(entry, cache_key=cache_key)
payload: dict[str, Any] = {
"sample_id": entry.sample_id,
"teacher_cache_key": cache_key,
"video_name": entry.video_name,
"video_path": str(video_path),
}
if cached is not None:
payload.update({
"current_teacher_cache_hit": True,
"current_teacher_decode_seconds": 0.0,
"current_teacher_features": cached,
})
return payload
decode_started = time.perf_counter()
decoded = self.decode_fn(video_path, self.decoder_contract)
decode_seconds = time.perf_counter() - decode_started
if not isinstance(decoded, DecodedMatchedVideo):
raise TypeError("decode_fn must return DecodedMatchedVideo")
expected_count = int(self.decoder_contract["expected_source_frame_count"])
expected_frames = int(self.decoder_contract["num_frames"])
expected_indices = torch.arange(
expected_count - expected_frames,
expected_count,
dtype=torch.long,
)
if (
int(decoded.source_frame_count) != expected_count
or not isinstance(decoded.frame_indices, torch.Tensor)
or decoded.frame_indices.dtype != torch.long
or not torch.equal(
decoded.frame_indices.detach().cpu(), expected_indices
)
):
raise ValueError(
"decode_fn violated the strict last-contiguous 17..80 contract"
)
video = decoded.video01
if (
not isinstance(video, torch.Tensor)
or video.dtype != torch.float32
or video.ndim != 4
or tuple(video.shape[:2]) != (expected_frames, 3)
or not bool(torch.isfinite(video).all().item())
or float(video.min()) < 0.0
or float(video.max()) > 1.0
):
raise ValueError(
"decode_fn returned invalid video01; expected finite FP32 "
f"[{expected_frames},3,H,W] in [0,1]"
)
payload.update({
"current_teacher_cache_hit": False,
"current_teacher_decode_seconds": decode_seconds,
"current_teacher_frame_indices": decoded.frame_indices,
"current_teacher_source_frame_count": expected_count,
"current_teacher_video01": video.contiguous(),
})
return payload
def write_feature_cache(
self,
cache_key: str,
feature: torch.Tensor,
*,
sample_id: str,
) -> Path:
try:
index = self._cache_key_to_index[str(cache_key)]
except KeyError as exc:
raise KeyError(f"unknown teacher cache key: {cache_key!r}") from exc
entry = self.entries[index]
if entry.sample_id != str(sample_id):
raise ValueError(
"sample_id does not own teacher cache key: "
f"{sample_id!r} != {entry.sample_id!r}"
)
self._video_path_and_verify_stat(entry)
return self.feature_cache.write(
entry, feature, cache_key=str(cache_key)
)
@staticmethod
def collate_fn(batch: Sequence[Mapping[str, Any]]) -> dict[str, Any]:
return collate_matched_video_metadata(batch)
def collate_matched_video_metadata(
batch: Sequence[Mapping[str, Any]],
) -> dict[str, Any]:
"""Collate only metadata; teacher payloads intentionally have no collator."""
if not batch:
raise ValueError("cannot collate an empty matched-video batch")
forbidden = {
"current_teacher_features",
"current_teacher_video01",
"current_teacher_cache_hit",
}
for item in batch:
present = forbidden.intersection(item)
if present:
raise ValueError(
"metadata collator received eager teacher payload fields: "
f"{sorted(present)}"
)
return {
"idx": torch.tensor([int(item["idx"]) for item in batch]),
"prompts": [str(item["prompts"]) for item in batch],
"sample_ids": [str(item["sample_id"]) for item in batch],
"teacher_cache_keys": [str(item["teacher_cache_key"]) for item in batch],
"video_names": [str(item["video_name"]) for item in batch],
"video_paths": [str(item["video_path"]) for item in batch],
}
__all__ = [
"DecodedMatchedVideo",
"MatchedManifestVideoDataset",
"collate_matched_video_metadata",
"decode_matched_video_last_contiguous_pyav",
]
|