File size: 20,336 Bytes
9264c1c | 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 | import random
from pathlib import Path
from typing import Dict, List, Optional, Sequence, Tuple, Union
import cv2
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset
VIDEO_EXTENSIONS = {".mp4", ".mov", ".avi", ".mkv", ".webm"}
class ShortVideoError(ValueError):
pass
def orientation_aware_size(
source_width: int,
source_height: int,
landscape_height: int,
landscape_width: int,
) -> Tuple[int, int]:
if min(source_width, source_height, landscape_height, landscape_width) <= 0:
raise ValueError(
"Source and target dimensions must be positive, got "
f"source={source_height}x{source_width}, "
f"target={landscape_height}x{landscape_width}"
)
if source_height > source_width:
return landscape_width, landscape_height
return landscape_height, landscape_width
def load_scene_names(scene_list_path: Optional[str]) -> Optional[List[str]]:
if scene_list_path is None:
return None
path = Path(scene_list_path).expanduser().resolve()
names = []
if path.is_file():
for line in path.read_text(encoding="utf-8").splitlines():
value = line.strip()
if value and not value.startswith("#"):
names.append(Path(value).stem)
elif path.is_dir():
names = [
child.stem
for child in sorted(path.iterdir())
if not child.name.startswith(".")
]
else:
raise FileNotFoundError(f"Scene list file/directory does not exist: {path}")
names = list(dict.fromkeys(name for name in names if name))
if not names:
raise ValueError(f"Scene list contains no valid names: {path}")
return names
def discover_video_triplets(
dataset_root: str,
include_substrings: Optional[Sequence[str]] = None,
exclude_substring: Optional[str] = None,
allow_empty_after_filter: bool = False,
) -> List[Tuple[Path, Path, Path]]:
root = Path(dataset_root).expanduser().resolve()
folder_names = ("BG", "MASK", "FG_BG")
folders = {name: root / name for name in folder_names}
missing_folders = [str(path) for path in folders.values() if not path.is_dir()]
if missing_folders:
raise FileNotFoundError(f"Missing dataset folders: {missing_folders}")
indexed: Dict[str, Dict[str, Path]] = {}
for folder_name, folder in folders.items():
files = {
path.relative_to(folder).as_posix(): path
for path in folder.rglob("*")
if (
path.is_file()
and path.suffix.lower() in VIDEO_EXTENSIONS
and (
not include_substrings
or any(value in path.name for value in include_substrings)
)
and (
not exclude_substring
or exclude_substring not in path.name
)
)
}
if not files and not allow_empty_after_filter:
raise ValueError(f"No supported videos found in {folder}")
indexed[folder_name] = files
expected = set(indexed["BG"])
errors = []
for folder_name in ("MASK", "FG_BG"):
actual = set(indexed[folder_name])
missing = sorted(expected - actual)
extra = sorted(actual - expected)
if missing or extra:
errors.append(f"{folder_name}: missing={missing[:10]}, extra={extra[:10]}")
if errors:
raise ValueError("BG/MASK/FG_BG filenames do not match exactly. " + " | ".join(errors))
triplets = [
(indexed["BG"][name], indexed["MASK"][name], indexed["FG_BG"][name])
for name in sorted(expected)
]
if not triplets and not allow_empty_after_filter:
raise ValueError(
f"No video triplets remain after filtering {root}: "
f"include={list(include_substrings) if include_substrings else None}, "
f"exclude={exclude_substring!r}"
)
return triplets
def muse_temporal_union(mask: torch.Tensor, temporal_ratio: int = 4) -> torch.Tensor:
if mask.ndim != 5 or mask.shape[1] < 1:
raise ValueError(f"Expected mask features [B,C,T,H,W], got {tuple(mask.shape)}")
if temporal_ratio < 1:
raise ValueError(f"temporal_ratio must be positive, got {temporal_ratio}")
if mask.shape[2] == 1 or temporal_ratio == 1:
return mask
first = mask[:, :, :1]
remaining = mask[:, :, 1:]
pad = (-remaining.shape[2]) % temporal_ratio
if pad:
remaining = F.pad(remaining, (0, 0, 0, 0, 0, pad), value=0)
b, c, t, h, w = remaining.shape
remaining = remaining.reshape(b, c, t // temporal_ratio, temporal_ratio, h, w)
return torch.cat([first, remaining.amax(dim=3)], dim=2)
def degrade_mask(
mask: torch.Tensor,
frame_drop_probability: float = 0.7,
frame_drop_rate_min: float = 0.2,
frame_drop_rate_max: float = 0.99,
morphology_probability: float = 0.5,
morphology_kernel_sizes: Sequence[int] = (3, 5, 7),
bbox_probability: float = 0.25,
generator: Optional[torch.Generator] = None,
return_metadata: bool = False,
) -> Union[torch.Tensor, Tuple[torch.Tensor, List[Dict[str, object]]]]:
if mask.ndim != 5 or mask.shape[1] != 1:
raise ValueError(f"Expected mask [B,1,T,H,W], got {tuple(mask.shape)}")
if not 0 <= frame_drop_rate_min <= frame_drop_rate_max <= 1:
raise ValueError(
"Frame drop rates must satisfy 0 <= min <= max <= 1, got "
f"{frame_drop_rate_min}, {frame_drop_rate_max}"
)
if not morphology_kernel_sizes or any(
kernel < 1 or kernel % 2 == 0 for kernel in morphology_kernel_sizes
):
raise ValueError(
"Morphology kernel sizes must be non-empty positive odd integers, got "
f"{tuple(morphology_kernel_sizes)}"
)
degraded = (mask > 0.5).to(mask.dtype)
device = degraded.device
metadata = []
def draw() -> float:
return torch.rand((), generator=generator, device=device).item()
for batch_index in range(degraded.shape[0]):
sample = degraded[batch_index : batch_index + 1]
sample_metadata = {
"sample_index": batch_index,
"operations": [],
"input_mask_ratio": sample.float().mean().item(),
}
if draw() < frame_drop_probability:
drop_rate = frame_drop_rate_min + (
frame_drop_rate_max - frame_drop_rate_min
) * draw()
keep = (
torch.rand(
(1, 1, sample.shape[2], 1, 1),
generator=generator,
device=device,
)
>= drop_rate
).to(sample.dtype)
if keep.sum() == 0:
keep[:, :, int(draw() * sample.shape[2]) % sample.shape[2]] = 1
sample = sample * keep
sample_metadata["operations"].append("frame_dropout")
sample_metadata["sampled_frame_drop_rate"] = drop_rate
sample_metadata["actual_frame_drop_rate"] = 1.0 - keep.float().mean().item()
if draw() < morphology_probability:
kernel_index = int(draw() * len(morphology_kernel_sizes)) % len(
morphology_kernel_sizes
)
kernel = int(morphology_kernel_sizes[kernel_index])
padding = kernel // 2
if draw() < 0.5:
sample = F.max_pool3d(sample, (1, kernel, kernel), stride=1, padding=(0, padding, padding))
morphology_type = "dilation"
else:
sample = -F.max_pool3d(
-sample,
(1, kernel, kernel),
stride=1,
padding=(0, padding, padding),
)
morphology_type = "erosion"
sample_metadata["operations"].append(morphology_type)
sample_metadata["morphology_kernel"] = kernel
if draw() < bbox_probability:
boxed = torch.zeros_like(sample)
for frame_index in range(sample.shape[2]):
coordinates = torch.nonzero(sample[0, 0, frame_index] > 0.5, as_tuple=False)
if coordinates.numel() == 0:
continue
top_left = coordinates.amin(dim=0)
bottom_right = coordinates.amax(dim=0)
boxed[
0,
0,
frame_index,
top_left[0] : bottom_right[0] + 1,
top_left[1] : bottom_right[1] + 1,
] = 1
sample = boxed
sample_metadata["operations"].append("bbox_fit")
degraded[batch_index : batch_index + 1] = sample
sample_metadata["output_mask_ratio"] = sample.float().mean().item()
metadata.append(sample_metadata)
if return_metadata:
return degraded, metadata
return degraded
def derive_side_effect_mask(
foreground_background: torch.Tensor,
background: torch.Tensor,
object_mask: torch.Tensor,
difference_threshold: float = 0.05,
) -> torch.Tensor:
if foreground_background.shape != background.shape:
raise ValueError(
"FG_BG and BG must have identical shapes, got "
f"{tuple(foreground_background.shape)} and {tuple(background.shape)}"
)
if foreground_background.ndim != 5 or object_mask.ndim != 5:
raise ValueError("Expected video and mask tensors in [B,C,T,H,W] layout")
difference = (foreground_background.float() - background.float()).abs().mean(dim=1, keepdim=True)
side_effect = difference > difference_threshold
return (side_effect & (object_mask > 0.5).logical_not()).to(background.dtype)
class RemoveTripletDataset(Dataset):
def __init__(
self,
dataset_root: str,
num_frames: int = 81,
frame_stride: int = 1,
height: int = 480,
width: int = 832,
seed: int = 0,
extra_dataset_root: Optional[str] = None,
extra_dataset_root2: Optional[str] = None,
extra_dataset_roots: Optional[Sequence[str]] = None,
extra_filter_key: Optional[str] = None,
extra_scene_list_path: Optional[str] = None,
):
if num_frames < 1 or (num_frames - 1) % 4 != 0:
raise ValueError(f"num_frames must be 4n+1 for Wan VAE, got {num_frames}")
if frame_stride < 1:
raise ValueError(f"frame_stride must be positive, got {frame_stride}")
if height % 16 or width % 16:
raise ValueError(f"height and width must be divisible by 16, got {height}x{width}")
self.root = Path(dataset_root).expanduser().resolve()
self.triplets = []
self.triplet_roots = []
self.triplet_sources = []
self.source_summaries = []
def add_source(
source_name: str,
root_value: str,
include_substrings: Optional[Sequence[str]] = None,
) -> None:
root = Path(root_value).expanduser().resolve()
source_triplets = discover_video_triplets(
str(root),
include_substrings=include_substrings,
allow_empty_after_filter=True,
)
self.triplets.extend(source_triplets)
self.triplet_roots.extend([root] * len(source_triplets))
self.triplet_sources.extend([source_name] * len(source_triplets))
self.source_summaries.append(
{
"name": source_name,
"root": str(root),
"count": len(source_triplets),
"include": list(include_substrings) if include_substrings else None,
"exclude": None,
}
)
if include_substrings and not source_triplets:
raise ValueError(
f"Source '{source_name}' at {root} has no triplets whose filename "
f"contains any configured scene/filter value"
)
self.extra_filter_key = extra_filter_key
self.extra_scene_list_path = extra_scene_list_path
scene_names = load_scene_names(extra_scene_list_path)
extra1_include = (
scene_names
if scene_names is not None
else [extra_filter_key]
if extra_filter_key
else None
)
add_source("main", str(self.root))
if extra_dataset_root is not None:
add_source("extra1", extra_dataset_root, include_substrings=extra1_include)
if extra_dataset_root2 is not None:
add_source("extra2", extra_dataset_root2)
for extra_index, extra_root in enumerate(extra_dataset_roots or [], start=3):
add_source(f"extra{extra_index}", extra_root)
if not self.triplets:
raise ValueError("No video triplets were found in the provided dataset roots")
self.num_frames = num_frames
self.frame_stride = frame_stride
self.height = height
self.width = width
self.seed = seed
self._warned_trimmed_indices = set()
self._invalid_indices = set()
def __len__(self) -> int:
return len(self.triplets)
@staticmethod
def _probe_video(path: Path) -> Tuple[int, int, int]:
capture = cv2.VideoCapture(str(path))
if not capture.isOpened():
raise RuntimeError(f"Failed to open video: {path}")
try:
frame_count = int(capture.get(cv2.CAP_PROP_FRAME_COUNT))
width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
finally:
capture.release()
if frame_count <= 0 or width <= 0 or height <= 0:
raise RuntimeError(
f"Invalid video metadata for {path}: frames={frame_count}, size={width}x{height}"
)
return frame_count, width, height
def _sample_indices(
self,
frame_count: int,
rng: Optional[random.Random] = None,
) -> List[int]:
required = (self.num_frames - 1) * self.frame_stride + 1
stride = self.frame_stride if frame_count >= required else 1
required = (self.num_frames - 1) * stride + 1
if frame_count < required:
raise ValueError(
f"Video has {frame_count} frames, but at least {self.num_frames} are required"
)
random_source = rng if rng is not None else random
start = random_source.randint(0, frame_count - required)
return [start + index * stride for index in range(self.num_frames)]
@staticmethod
def _decode(path: Path, indices: List[int], size: Tuple[int, int], is_mask: bool) -> torch.Tensor:
target_height, target_width = size
capture = cv2.VideoCapture(str(path))
if not capture.isOpened():
raise RuntimeError(f"Failed to open video: {path}")
frames = []
next_frame_index = None
try:
for frame_index in indices:
if frame_index != next_frame_index:
capture.set(cv2.CAP_PROP_POS_FRAMES, frame_index)
ok, frame = capture.read()
next_frame_index = frame_index + 1
if not ok:
raise RuntimeError(f"{path}: failed to decode frame {frame_index}")
if is_mask:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
frame = cv2.resize(
frame,
(target_width, target_height),
interpolation=cv2.INTER_NEAREST,
)
tensor = torch.from_numpy(frame.copy()).unsqueeze(0).float().div_(255.0)
tensor = (tensor > 0.5).float()
else:
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
frame = cv2.resize(
frame,
(target_width, target_height),
interpolation=cv2.INTER_LINEAR,
)
tensor = torch.from_numpy(frame.copy()).permute(2, 0, 1).float()
tensor = tensor.div_(127.5).sub_(1.0)
frames.append(tensor)
finally:
capture.release()
return torch.stack(frames, dim=1).contiguous()
def __getitem__(
self,
index: Union[int, Tuple[int, int]],
) -> Dict[str, torch.Tensor]:
epoch = None
if isinstance(index, tuple):
if len(index) != 2:
raise ValueError(f"Expected (sample_index, epoch), got {index}")
index, epoch = (int(index[0]), int(index[1]))
else:
index = int(index)
for offset in range(len(self.triplets)):
candidate_index = (index + offset) % len(self.triplets)
if candidate_index in self._invalid_indices:
continue
try:
return self._load_item(candidate_index, epoch)
except ShortVideoError as error:
self._invalid_indices.add(candidate_index)
print(
f"[Dataset:SkipShort] index={candidate_index}, {error}",
flush=True,
)
raise RuntimeError(
f"No triplet contains at least {self.num_frames} aligned frames"
)
def _load_item(
self,
index: int,
epoch: Optional[int],
) -> Dict[str, torch.Tensor]:
background_path, mask_path, foreground_background_path = self.triplets[index]
source_root = self.triplet_roots[index]
source_name = self.triplet_sources[index]
metadata = [
self._probe_video(background_path),
self._probe_video(mask_path),
self._probe_video(foreground_background_path),
]
frame_counts = [item[0] for item in metadata]
usable_frame_count = min(frame_counts)
if usable_frame_count < self.num_frames:
raise ShortVideoError(
f"sample={background_path.name}, "
f"BG={frame_counts[0]}, MASK={frame_counts[1]}, "
f"FG_BG={frame_counts[2]}, required={self.num_frames}"
)
if len(set(frame_counts)) != 1 and index not in self._warned_trimmed_indices:
print(
"[Dataset:Trim] "
f"index={index}, sample={background_path.name}, "
f"using first {usable_frame_count} aligned frames from "
f"BG={frame_counts[0]}, MASK={frame_counts[1]}, FG_BG={frame_counts[2]}",
flush=True,
)
self._warned_trimmed_indices.add(index)
spatial_sizes = [(item[1], item[2]) for item in metadata]
if len(set(spatial_sizes)) != 1:
raise ValueError(
f"Source size mismatch for {background_path.name}: "
f"BG={spatial_sizes[0]}, MASK={spatial_sizes[1]}, FG_BG={spatial_sizes[2]}"
)
clip_rng = None
if epoch is not None:
clip_seed = (
self.seed * 6364136223846793005
+ epoch * 1442695040888963407
+ index
) % (2**63)
clip_rng = random.Random(clip_seed)
indices = self._sample_indices(usable_frame_count, rng=clip_rng)
source_width, source_height = metadata[2][1], metadata[2][2]
size = orientation_aware_size(
source_width,
source_height,
self.height,
self.width,
)
return {
"background": self._decode(background_path, indices, size, is_mask=False),
"mask": self._decode(mask_path, indices, size, is_mask=True),
"foreground_background": self._decode(
foreground_background_path,
indices,
size,
is_mask=False,
),
"sample_name": (
f"{source_name}:"
f"{background_path.relative_to(source_root / 'BG').as_posix()}"
),
"dataset_source": source_name,
}
|