File size: 19,336 Bytes
94dc344 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
# pyre-unsafe
from dataclasses import dataclass
from enum import Enum
from typing import Iterator, List, Optional, Tuple
import torch
from pytorch3d.implicitron.tools.config import registry, ReplaceableBase
from torch.utils.data import (
BatchSampler,
ConcatDataset,
DataLoader,
RandomSampler,
Sampler,
)
from .dataset_base import DatasetBase
from .dataset_map_provider import DatasetMap
from .frame_data import FrameData
from .scene_batch_sampler import SceneBatchSampler
from .utils import is_known_frame_scalar
@dataclass
class DataLoaderMap:
"""
A collection of data loaders for Implicitron.
Members:
train: a data loader for training
val: a data loader for validating during training
test: a data loader for final evaluation
"""
train: Optional[DataLoader[FrameData]]
val: Optional[DataLoader[FrameData]]
test: Optional[DataLoader[FrameData]]
def __getitem__(self, split: str) -> Optional[DataLoader[FrameData]]:
"""
Get one of the data loaders by key (name of data split)
"""
if split not in ["train", "val", "test"]:
raise ValueError(f"{split} was not a valid split name (train/val/test)")
return getattr(self, split)
class DataLoaderMapProviderBase(ReplaceableBase):
"""
Provider of a collection of data loaders for a given collection of datasets.
"""
def get_data_loader_map(self, datasets: DatasetMap) -> DataLoaderMap:
"""
Returns a collection of data loaders for a given collection of datasets.
"""
raise NotImplementedError()
@registry.register
class SimpleDataLoaderMapProvider(DataLoaderMapProviderBase):
"""
Trivial implementation of DataLoaderMapProviderBase.
If a dataset returns batches from get_eval_batches(), then
they will be what the corresponding dataloader returns,
independently of any of the fields on this class.
Otherwise, returns shuffled batches.
"""
batch_size: int = 1
num_workers: int = 0
dataset_length_train: int = 0
dataset_length_val: int = 0
dataset_length_test: int = 0
def get_data_loader_map(self, datasets: DatasetMap) -> DataLoaderMap:
"""
Returns a collection of data loaders for a given collection of datasets.
"""
return DataLoaderMap(
train=self._make_data_loader(
datasets.train,
self.dataset_length_train,
),
val=self._make_data_loader(
datasets.val,
self.dataset_length_val,
),
test=self._make_data_loader(
datasets.test,
self.dataset_length_test,
),
)
def _make_data_loader(
self,
dataset: Optional[DatasetBase],
num_batches: int,
) -> Optional[DataLoader[FrameData]]:
"""
Returns the dataloader for a dataset.
Args:
dataset: the dataset
num_batches: possible ceiling on number of batches per epoch
"""
if dataset is None:
return None
data_loader_kwargs = {
"num_workers": self.num_workers,
"collate_fn": dataset.frame_data_type.collate,
}
eval_batches = dataset.get_eval_batches()
if eval_batches is not None:
return DataLoader(
dataset,
batch_sampler=eval_batches,
**data_loader_kwargs,
)
if num_batches > 0:
num_samples = self.batch_size * num_batches
else:
num_samples = None
# sample with replacement only if a custom number of samples is specified
sampler = RandomSampler(
dataset,
replacement=num_samples is not None,
num_samples=num_samples,
)
batch_sampler = BatchSampler(sampler, self.batch_size, drop_last=True)
return DataLoader(
dataset,
batch_sampler=batch_sampler,
**data_loader_kwargs,
)
class DoublePoolBatchSampler(Sampler[List[int]]):
"""
Batch sampler for making random batches of a single frame
from one list and a number of known frames from another list.
"""
def __init__(
self,
first_indices: List[int],
rest_indices: List[int],
batch_size: int,
replacement: bool,
num_batches: Optional[int] = None,
) -> None:
"""
Args:
first_indices: indexes of dataset items to use as the first element
of each batch.
rest_indices: indexes of dataset items to use as the subsequent
elements of each batch. Not used if batch_size==1.
batch_size: The common size of any batch.
replacement: Whether the sampling of first items is with replacement.
num_batches: The number of batches in an epoch. If 0 or None,
one epoch is the length of `first_indices`.
"""
self.first_indices = first_indices
self.rest_indices = rest_indices
self.batch_size = batch_size
self.replacement = replacement
self.num_batches = None if num_batches == 0 else num_batches
if batch_size - 1 > len(rest_indices):
raise ValueError(
f"Cannot make up ({batch_size})-batches from {len(self.rest_indices)}"
)
# copied from RandomSampler
seed = int(torch.empty((), dtype=torch.int64).random_().item())
self.generator = torch.Generator()
self.generator.manual_seed(seed)
def __len__(self) -> int:
if self.num_batches is not None:
return self.num_batches
return len(self.first_indices)
def __iter__(self) -> Iterator[List[int]]:
num_batches = self.num_batches
if self.replacement:
i_first = torch.randint(
len(self.first_indices),
size=(len(self),),
generator=self.generator,
)
elif num_batches is not None:
n_copies = 1 + (num_batches - 1) // len(self.first_indices)
raw_indices = [
torch.randperm(len(self.first_indices), generator=self.generator)
for _ in range(n_copies)
]
i_first = torch.cat(raw_indices)[:num_batches]
else:
i_first = torch.randperm(len(self.first_indices), generator=self.generator)
first_indices = [self.first_indices[i] for i in i_first]
if self.batch_size == 1:
for first_index in first_indices:
yield [first_index]
return
for first_index in first_indices:
# Consider using this class in a program which sets the seed. This use
# of randperm means that rerunning with a higher batch_size
# results in batches whose first elements as the first run.
i_rest = torch.randperm(
len(self.rest_indices),
generator=self.generator,
)[: self.batch_size - 1]
yield [first_index] + [self.rest_indices[i] for i in i_rest]
class BatchConditioningType(Enum):
"""
Ways to add conditioning frames for the val and test batches.
SAME: Use the corresponding dataset for all elements of val batches
without regard to frame type.
TRAIN: Use the corresponding dataset for the first element of each
batch, and the training dataset for the extra conditioning
elements. No regard to frame type.
KNOWN: Use frames from the corresponding dataset but separate them
according to their frame_type. Each batch will contain one UNSEEN
frame followed by many KNOWN frames.
"""
SAME = "same"
TRAIN = "train"
KNOWN = "known"
@registry.register
class SequenceDataLoaderMapProvider(DataLoaderMapProviderBase):
"""
Default implementation of DataLoaderMapProviderBase.
If a dataset returns batches from get_eval_batches(), then
they will be what the corresponding dataloader returns,
independently of any of the fields on this class.
If conditioning is not required, then the batch size should
be set as 1, and most of the fields do not matter.
If conditioning is required, each batch will contain one main
frame first to predict and the, rest of the elements are for
conditioning.
If images_per_seq_options is left empty, the conditioning
frames are picked according to the conditioning type given.
This does not have regard to the order of frames in a
scene, or which frames belong to what scene.
If images_per_seq_options is given, then the conditioning types
must be SAME and the remaining fields are used.
Members:
batch_size: The size of the batch of the data loader.
num_workers: Number of data-loading threads in each data loader.
dataset_length_train: The number of batches in a training epoch. Or 0 to mean
an epoch is the length of the training set.
dataset_length_val: The number of batches in a validation epoch. Or 0 to mean
an epoch is the length of the validation set.
dataset_length_test: The number of batches in a testing epoch. Or 0 to mean
an epoch is the length of the test set.
train_conditioning_type: Whether the train data loader should use
only known frames for conditioning.
Only used if batch_size>1 and train dataset is
present and does not return eval_batches.
val_conditioning_type: Whether the val data loader should use
training frames or known frames for conditioning.
Only used if batch_size>1 and val dataset is
present and does not return eval_batches.
test_conditioning_type: Whether the test data loader should use
training frames or known frames for conditioning.
Only used if batch_size>1 and test dataset is
present and does not return eval_batches.
images_per_seq_options: Possible numbers of frames sampled per sequence in a batch.
If a conditioning_type is KNOWN or TRAIN, then this must be left at its initial
value. Empty (the default) means that we are not careful about which frames
come from which scene.
sample_consecutive_frames: if True, will sample a contiguous interval of frames
in the sequence. It first sorts the frames by timestimps when available,
otherwise by frame numbers, finds the connected segments within the sequence
of sufficient length, then samples a random pivot element among them and
ideally uses it as a middle of the temporal window, shifting the borders
where necessary. This strategy mitigates the bias against shorter segments
and their boundaries.
consecutive_frames_max_gap: if a number > 0, then used to define the maximum
difference in frame_number of neighbouring frames when forming connected
segments; if both this and consecutive_frames_max_gap_seconds are 0s,
the whole sequence is considered a segment regardless of frame numbers.
consecutive_frames_max_gap_seconds: if a number > 0.0, then used to define the
maximum difference in frame_timestamp of neighbouring frames when forming
connected segments; if both this and consecutive_frames_max_gap are 0s,
the whole sequence is considered a segment regardless of frame timestamps.
"""
batch_size: int = 1
num_workers: int = 0
dataset_length_train: int = 0
dataset_length_val: int = 0
dataset_length_test: int = 0
train_conditioning_type: BatchConditioningType = BatchConditioningType.SAME
val_conditioning_type: BatchConditioningType = BatchConditioningType.SAME
test_conditioning_type: BatchConditioningType = BatchConditioningType.KNOWN
images_per_seq_options: Tuple[int, ...] = ()
sample_consecutive_frames: bool = False
consecutive_frames_max_gap: int = 0
consecutive_frames_max_gap_seconds: float = 0.1
def get_data_loader_map(self, datasets: DatasetMap) -> DataLoaderMap:
"""
Returns a collection of data loaders for a given collection of datasets.
"""
return DataLoaderMap(
train=self._make_data_loader(
datasets.train,
self.dataset_length_train,
datasets.train,
self.train_conditioning_type,
),
val=self._make_data_loader(
datasets.val,
self.dataset_length_val,
datasets.train,
self.val_conditioning_type,
),
test=self._make_data_loader(
datasets.test,
self.dataset_length_test,
datasets.train,
self.test_conditioning_type,
),
)
def _make_data_loader(
self,
dataset: Optional[DatasetBase],
num_batches: int,
train_dataset: Optional[DatasetBase],
conditioning_type: BatchConditioningType,
) -> Optional[DataLoader[FrameData]]:
"""
Returns the dataloader for a dataset.
Args:
dataset: the dataset
num_batches: possible ceiling on number of batches per epoch
train_dataset: the training dataset, used if conditioning_type==TRAIN
conditioning_type: source for padding of batches
"""
if dataset is None:
return None
data_loader_kwargs = {
"num_workers": self.num_workers,
"collate_fn": dataset.frame_data_type.collate,
}
eval_batches = dataset.get_eval_batches()
if eval_batches is not None:
return DataLoader(
dataset,
batch_sampler=eval_batches,
**data_loader_kwargs,
)
scenes_matter = len(self.images_per_seq_options) > 0
if scenes_matter and conditioning_type != BatchConditioningType.SAME:
raise ValueError(
f"{conditioning_type} cannot be used with images_per_seq "
+ str(self.images_per_seq_options)
)
if self.batch_size == 1 or (
not scenes_matter and conditioning_type == BatchConditioningType.SAME
):
return self._simple_loader(dataset, num_batches, data_loader_kwargs)
if scenes_matter:
assert conditioning_type == BatchConditioningType.SAME
batch_sampler = SceneBatchSampler(
dataset,
self.batch_size,
num_batches=len(dataset) if num_batches <= 0 else num_batches,
images_per_seq_options=self.images_per_seq_options,
sample_consecutive_frames=self.sample_consecutive_frames,
consecutive_frames_max_gap=self.consecutive_frames_max_gap,
consecutive_frames_max_gap_seconds=self.consecutive_frames_max_gap_seconds,
)
return DataLoader(
dataset,
batch_sampler=batch_sampler,
**data_loader_kwargs,
)
if conditioning_type == BatchConditioningType.TRAIN:
return self._train_loader(
dataset, train_dataset, num_batches, data_loader_kwargs
)
assert conditioning_type == BatchConditioningType.KNOWN
return self._known_loader(dataset, num_batches, data_loader_kwargs)
def _simple_loader(
self,
dataset: DatasetBase,
num_batches: int,
data_loader_kwargs: dict,
) -> DataLoader[FrameData]:
"""
Return a simple loader for frames in the dataset.
This is equivalent to
Dataloader(dataset, batch_size=self.batch_size, **data_loader_kwargs)
except that num_batches is fixed.
Args:
dataset: the dataset
num_batches: possible ceiling on number of batches per epoch
data_loader_kwargs: common args for dataloader
"""
if num_batches > 0:
num_samples = self.batch_size * num_batches
replacement = True
else:
num_samples = None
replacement = False
sampler = RandomSampler(
dataset, replacement=replacement, num_samples=num_samples
)
batch_sampler = BatchSampler(sampler, self.batch_size, drop_last=True)
return DataLoader(
dataset,
batch_sampler=batch_sampler,
**data_loader_kwargs,
)
def _train_loader(
self,
dataset: DatasetBase,
train_dataset: Optional[DatasetBase],
num_batches: int,
data_loader_kwargs: dict,
) -> DataLoader[FrameData]:
"""
Return the loader for TRAIN conditioning.
Args:
dataset: the dataset
train_dataset: the training dataset
num_batches: possible ceiling on number of batches per epoch
data_loader_kwargs: common args for dataloader
"""
if train_dataset is None:
raise ValueError("No training data for conditioning.")
length = len(dataset)
first_indices = list(range(length))
rest_indices = list(range(length, length + len(train_dataset)))
sampler = DoublePoolBatchSampler(
first_indices=first_indices,
rest_indices=rest_indices,
batch_size=self.batch_size,
replacement=True,
num_batches=num_batches,
)
return DataLoader(
ConcatDataset([dataset, train_dataset]),
batch_sampler=sampler,
**data_loader_kwargs,
)
def _known_loader(
self,
dataset: DatasetBase,
num_batches: int,
data_loader_kwargs: dict,
) -> DataLoader[FrameData]:
"""
Return the loader for KNOWN conditioning.
Args:
dataset: the dataset
num_batches: possible ceiling on number of batches per epoch
data_loader_kwargs: common args for dataloader
"""
first_indices, rest_indices = [], []
for idx in range(len(dataset)):
frame_type = dataset[idx].frame_type
assert isinstance(frame_type, str)
if is_known_frame_scalar(frame_type):
rest_indices.append(idx)
else:
first_indices.append(idx)
sampler = DoublePoolBatchSampler(
first_indices=first_indices,
rest_indices=rest_indices,
batch_size=self.batch_size,
replacement=True,
num_batches=num_batches,
)
return DataLoader(
dataset,
batch_sampler=sampler,
**data_loader_kwargs,
)
|