File size: 16,822 Bytes
491eded |
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 |
"""
This file (`data_utils.py`) provides utility functions and classes for data handling in deep learning models.
It includes tools for moving tensors to specific devices, load-balancing utilities for distributed training,
and custom samplers for PyTorch DataLoaders that support resumable training and balanced data distribution.
Key components:
- Recursive device transfer functionality
- Load balancing utilities for distributing data across processes
- Cyclical iteration through data loaders
- Custom resumable samplers for distributed training
"""
from typing import *
import math
import torch
import numpy as np
from torch.utils.data import Sampler, Dataset, DataLoader, DistributedSampler
import torch.distributed as dist
def recursive_to_device(
data: Any,
device: torch.device,
non_blocking: bool = False,
) -> Any:
"""
Recursively move all tensors in a data structure to a device.
This function traverses nested data structures (lists, tuples, dictionaries)
and moves any PyTorch tensor to the specified device.
Args:
data: The data structure containing tensors to be moved
device: The target device (CPU, GPU) to move tensors to
non_blocking: If True, allows asynchronous copy to device if possible
Returns:
The same data structure with all tensors moved to the specified device
"""
if hasattr(data, "to"):
# print("Moving data to device")
# print(data)
return data.to(device, non_blocking=non_blocking)
elif isinstance(data, (list, tuple)):
# print("list or tuple detected")
return type(data)(recursive_to_device(d, device, non_blocking) for d in data)
elif isinstance(data, dict):
# print("dict detected")
return {k: recursive_to_device(v, device, non_blocking) for k, v in data.items()}
else:
# print(f"{type(data)} detected")
return data
def load_balanced_group_indices(
load: List[int],
num_groups: int,
equal_size: bool = False,
) -> List[List[int]]:
"""
Split indices into groups with balanced load.
This function distributes indices across groups to achieve balanced workload.
It uses a greedy algorithm that assigns each index to the group with the
minimum current load.
Args:
load: List of load values for each index
num_groups: Number of groups to split indices into
equal_size: If True, each group will have the same number of elements
Returns:
List of lists, where each inner list contains indices assigned to a group
"""
if equal_size:
group_size = len(load) // num_groups
indices = np.argsort(load)[::-1] # Sort indices by load in descending order
groups = [[] for _ in range(num_groups)]
group_load = np.zeros(num_groups)
for idx in indices:
min_group_idx = np.argmin(group_load)
groups[min_group_idx].append(idx)
if equal_size and len(groups[min_group_idx]) == group_size:
group_load[min_group_idx] = float('inf') # Mark group as full
else:
group_load[min_group_idx] += load[idx]
return groups
def cycle(data_loader: DataLoader) -> Iterator:
"""
Creates an infinite iterator over a data loader.
This function wraps a data loader to cycle through it repeatedly,
handling epoch tracking for various sampler types.
Args:
data_loader: The DataLoader to cycle through
Returns:
An iterator that indefinitely yields batches from the data loader
"""
while True:
for data in data_loader:
if isinstance(data_loader.sampler, ResumableSampler):
data_loader.sampler.idx += data_loader.batch_size # Update position for resumability
yield data
if isinstance(data_loader.sampler, DistributedSampler):
data_loader.sampler.epoch += 1 # Update epoch for DistributedSampler
if isinstance(data_loader.sampler, ResumableSampler):
data_loader.sampler.epoch += 1 # Update epoch for ResumableSampler
data_loader.sampler.idx = 0 # Reset position index
class ResumableSampler(Sampler):
"""
Distributed sampler that is resumable.
This sampler extends PyTorch's Sampler to support resuming training from
a specific point. It tracks the current position (idx) and epoch to
enable checkpointing and resuming.
Args:
dataset: Dataset used for sampling.
rank (int, optional): Rank of the current process within :attr:`num_replicas`.
By default, :attr:`rank` is retrieved from the current distributed
group.
shuffle (bool, optional): If ``True`` (default), sampler will shuffle the
indices.
seed (int, optional): random seed used to shuffle the sampler if
:attr:`shuffle=True`. This number should be identical across all
processes in the distributed group. Default: ``0``.
drop_last (bool, optional): if ``True``, then the sampler will drop the
tail of the data to make it evenly divisible across the number of
replicas. If ``False``, the sampler will add extra indices to make
the data evenly divisible across the replicas. Default: ``False``.
"""
def __init__(
self,
dataset: Dataset,
shuffle: bool = True,
seed: int = 0,
drop_last: bool = False,
) -> None:
self.dataset = dataset
self.epoch = 0 # Current epoch counter
self.idx = 0 # Current index position for resuming
self.drop_last = drop_last
self.world_size = dist.get_world_size() if dist.is_initialized() else 1 # Get total number of processes
self.rank = dist.get_rank() if dist.is_initialized() else 0 # Get current process rank
# Calculate number of samples per process
if self.drop_last and len(self.dataset) % self.world_size != 0:
# Split to nearest available length that is evenly divisible
# This ensures each rank receives the same amount of data
self.num_samples = math.ceil(
(len(self.dataset) - self.world_size) / self.world_size
)
else:
self.num_samples = math.ceil(len(self.dataset) / self.world_size)
self.total_size = self.num_samples * self.world_size # Total size after padding
self.shuffle = shuffle
self.seed = seed
def __iter__(self) -> Iterator:
if self.shuffle:
# Deterministically shuffle based on epoch and seed
g = torch.Generator()
g.manual_seed(self.seed + self.epoch)
indices = torch.randperm(len(self.dataset), generator=g).tolist()
else:
indices = list(range(len(self.dataset)))
if not self.drop_last:
# Add extra samples to make it evenly divisible across processes
padding_size = self.total_size - len(indices)
if padding_size <= len(indices):
indices += indices[:padding_size] # Reuse some samples from the beginning
else:
indices += (indices * math.ceil(padding_size / len(indices)))[
:padding_size
] # Repeat samples if padding_size > len(indices)
else:
# Remove tail of data to make it evenly divisible
indices = indices[: self.total_size]
assert len(indices) == self.total_size
# Subsample according to rank for distributed training
indices = indices[self.rank : self.total_size : self.world_size]
# Resume from previous state by skipping already processed indices
indices = indices[self.idx:]
return iter(indices)
def __len__(self) -> int:
return self.num_samples
def state_dict(self) -> Dict[str, int]:
"""
Returns the state of the sampler as a dictionary.
This enables saving the sampler state for checkpointing.
Returns:
Dictionary containing epoch and current index
"""
return {
'epoch': self.epoch,
'idx': self.idx,
}
def load_state_dict(self, state_dict):
"""
Loads the sampler state from a dictionary.
This enables restoring the sampler state from a checkpoint.
Args:
state_dict: Dictionary containing sampler state
"""
self.epoch = state_dict['epoch']
self.idx = state_dict['idx']
class BalancedResumableSampler(ResumableSampler):
"""
Distributed sampler that is resumable and balances the load among the processes.
This sampler extends ResumableSampler to distribute data across processes
in a load-balanced manner, ensuring that each process receives a similar
computational workload despite potentially varying sample processing times.
Args:
dataset: Dataset used for sampling. Must have 'loads' attribute.
shuffle (bool, optional): If ``True`` (default), sampler will shuffle the
indices.
seed (int, optional): random seed used to shuffle the sampler if
:attr:`shuffle=True`. This number should be identical across all
processes in the distributed group. Default: ``0``.
drop_last (bool, optional): if ``True``, then the sampler will drop the
tail of the data to make it evenly divisible across the number of
replicas. If ``False``, the sampler will add extra indices to make
the data evenly divisible across the replicas. Default: ``False``.
batch_size (int, optional): Size of mini-batches used for balancing. Default: 1.
"""
def __init__(
self,
dataset: Dataset,
shuffle: bool = True,
seed: int = 0,
drop_last: bool = False,
batch_size: int = 1,
) -> None:
assert hasattr(dataset, 'loads'), 'Dataset must have "loads" attribute to use BalancedResumableSampler'
super().__init__(dataset, shuffle, seed, drop_last)
self.batch_size = batch_size
self.loads = dataset.loads # Load values for each sample in the dataset
def __iter__(self) -> Iterator:
# print(f"[BalancedResumableSampler] Starting __iter__ for rank {self.rank}, epoch {self.epoch}")
if self.shuffle:
# Deterministically shuffle based on epoch and seed
g = torch.Generator()
g.manual_seed(self.seed + self.epoch)
# print(f"[BalancedResumableSampler] Shuffling with seed {self.seed + self.epoch}") # 0
indices = torch.randperm(len(self.dataset), generator=g).tolist()
else:
# print(f"[BalancedResumableSampler] No shuffle, using sequential indices")
indices = list(range(len(self.dataset)))
# print(indices)
# print(f"[BalancedResumableSampler] Initial indices length: {len(indices)}") # 128
if not self.drop_last:
# Add extra samples to make it evenly divisible
padding_size = self.total_size - len(indices)
# print(f"[BalancedResumableSampler] Adding padding of size {padding_size}") # 0
if padding_size <= len(indices):
indices += indices[:padding_size]
else:
indices += (indices * math.ceil(padding_size / len(indices)))[:padding_size]
else:
# Remove tail of data to make it evenly divisible
# print(f"[BalancedResumableSampler] Dropping last, trimming to {self.total_size}")
indices = indices[: self.total_size]
# print(indices)
assert len(indices) == self.total_size
# print(f"[BalancedResumableSampler] After padding/trimming, indices length: {len(indices)}") # 128
# Balance load among processes by distributing batches based on their loads
num_batches = len(indices) // (self.batch_size * self.world_size)
# print(f"[BalancedResumableSampler] Number of batches: {num_batches}") # 16
balanced_indices = []
if len(self.loads) < len(indices):
# repeat the loads to match the indices
self.loads = self.loads * (len(indices) // len(self.loads)) + self.loads[:len(indices) % len(self.loads)]
for i in range(num_batches):
start_idx = i * self.batch_size * self.world_size
end_idx = (i + 1) * self.batch_size * self.world_size
# print("start idx", start_idx) # 0
# print("end idx", end_idx) # 8
# print("batch size", self.batch_size) # 8
# print("world size", self.world_size) # 1
batch_indices = indices[start_idx:end_idx]
# print(f"[BalancedResumableSampler] Processing batch {i+1}/{num_batches}, size: {len(batch_indices)}") #1/16 8
batch_loads = [self.loads[idx] for idx in batch_indices]
groups = load_balanced_group_indices(batch_loads, self.world_size, equal_size=True)
balanced_indices.extend([batch_indices[j] for j in groups[self.rank]])
# print(f"[BalancedResumableSampler] Total balanced indices for rank {self.rank}: {len(balanced_indices)}")
# Resume from previous state
indices = balanced_indices[self.idx:]
# print(f"[BalancedResumableSampler] After resuming from idx {self.idx}, returning {len(indices)} indices")
return iter(indices)
class DuplicatedDataset(torch.utils.data.Dataset):
"""Dataset wrapper that duplicates a dataset multiple times."""
def __init__(self, dataset, repeat=1000):
"""
Initialize the duplicated dataset.
Args:
dataset: Original dataset to duplicate
repeat: Number of times to repeat the dataset
"""
self.dataset = dataset
self.repeat = repeat
self.original_length = len(dataset)
def __getitem__(self, idx):
"""Get an item from the original dataset, repeating as needed."""
return self.dataset[idx % self.original_length]
def __len__(self):
"""Return the length of the duplicated dataset."""
return self.original_length * self.repeat
def __getattr__(self, name):
"""Forward all other attribute accesses to the original dataset."""
if name == 'dataset' or name == 'repeat' or name == 'original_length':
return object.__getattribute__(self, name)
return getattr(self.dataset, name)
def save_coords_as_ply(coords, save_dir: str):
"""
Save the coordinates to a PLY file using normalization similar to voxelize.py.
Args:
file_path (str): The directory path to save the PLY file.
"""
import os
# import numpy as np
os.makedirs(save_dir, exist_ok=True) # Ensure the directory exists
# Get coordinates and convert to numpy
coords_np = coords.cpu().numpy()
# Print debug info
# print(f"Original coordinates shape: {coords_np.shape}")
# print(f"First few coordinates:\n{coords_np[:5]}")
if coords_np.shape[1] == 4:
# Extract XYZ coordinates (skip batch index at position 0)
vertices = coords_np[:, 1:4]
else:
vertices = coords_np
# Normalize coordinates to [-0.5, 0.5] like in voxelize.py
# Assuming the coordinates are in a 64³ grid
GRID_SIZE = 64
vertices = (vertices + 0.5) / GRID_SIZE - 0.5
# print(f"Normalized vertex range: min={np.min(vertices, axis=0)}, max={np.max(vertices, axis=0)}")
# Create PLY file (simplified format like in voxelize.py)
filename = os.path.join(save_dir, 'coords.ply')
try:
with open(filename, 'w') as f:
# Write header (no color, just XYZ coordinates)
f.write("ply\n")
f.write("format ascii 1.0\n")
f.write(f"element vertex {vertices.shape[0]}\n")
f.write("property float x\n")
f.write("property float y\n")
f.write("property float z\n")
f.write("end_header\n")
# Write vertices (no color)
for i in range(vertices.shape[0]):
f.write(f"{vertices[i, 0]} {vertices[i, 1]} {vertices[i, 2]}\n")
# print(f"PLY file saved to {filename} with {vertices.shape[0]} points")
# Verify file creation
# file_size = os.path.getsize(filename)
# print(f"File size: {file_size} bytes")
except Exception as e:
print(f"Error creating PLY file: {e}")
return filename |