File size: 11,635 Bytes
35bef58 | 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 | import torch
from torch.utils.data import Sampler
from accelerate import Accelerator
import math
import os
import random
from dataclasses import dataclass
from psi.utils import initialize_overwatch
overwatch = initialize_overwatch(__name__)
class BatchMixtureSampler(Sampler):
"""
Multi-node safe dataset mixture sampler.
Example: datasets = [ds1, ds2], ratio = [4, 1]
"""
def __init__(self, dataset_lens, mixture_ratios, num_samples_per_epoch, batch_size, seed=42):
self.dataset_lens = dataset_lens
self.weights = torch.tensor(mixture_ratios, dtype=torch.double)
self.weights /= self.weights.sum()
self.num_samples = num_samples_per_epoch
self.batch_size = batch_size
self.seed = seed
# Distributed settings
self.rank = int(os.environ.get("RANK", 0))
self.world_size = int(os.environ.get("WORLD_SIZE", 1))
# each rank gets this many samples
self.num_samples_rank = math.ceil(self.num_samples / self.world_size)
self.epoch = 0
self.batch_counts = self._build_batch_counts()
def _build_batch_counts(self):
positive_ids = [i for i, weight in enumerate(self.weights.tolist()) if weight > 0]
if not positive_ids:
raise ValueError("BatchMixtureSampler received no positive mixture ratios")
counts = torch.zeros(len(self.dataset_lens), dtype=torch.long)
if self.batch_size >= len(positive_ids):
counts[positive_ids] = 1
remaining = self.batch_size - len(positive_ids)
if remaining > 0:
positive_weights = self.weights[positive_ids]
positive_weights = positive_weights / positive_weights.sum()
raw_extra = positive_weights * remaining
extra_counts = torch.floor(raw_extra).long()
counts[positive_ids] += extra_counts
extra_left = remaining - int(extra_counts.sum().item())
if extra_left > 0:
frac = raw_extra - extra_counts.double()
for frac_index in torch.argsort(frac, descending=True)[:extra_left].tolist():
counts[positive_ids[frac_index]] += 1
else:
for dataset_id in torch.argsort(self.weights, descending=True)[:self.batch_size].tolist():
counts[dataset_id] += 1
return counts.tolist()
def set_epoch(self, epoch):
self.epoch = epoch
def __iter__(self):
total_batches = len(self) * self.world_size
for batch_idx in range(total_batches):
if batch_idx % self.world_size != self.rank:
continue
g = torch.Generator()
g.manual_seed(self.seed + self.epoch * total_batches + batch_idx)
batch = []
for dataset_id, count in enumerate(self.batch_counts):
if count <= 0:
continue
dataset_len = int(self.dataset_lens[dataset_id])
rand_indices = torch.randint(
low=0,
high=dataset_len,
size=(count,),
generator=g,
).tolist()
batch.extend((dataset_id, sample_id) for sample_id in rand_indices)
shuffle_order = torch.randperm(len(batch), generator=g).tolist()
yield [batch[i] for i in shuffle_order]
def __len__(self):
return math.ceil(self.num_samples_rank / self.batch_size)
class NoReplacementBatchMixtureSampler(Sampler):
"""
Batch mixture sampler that walks each dataset through a shuffled queue.
Unlike BatchMixtureSampler, sample ids are not drawn with replacement inside
an epoch. When one dataset queue is exhausted, a new shuffled pass for that
dataset starts. This keeps task ratios fixed at the batch level while making
every active task traverse its own samples before repeating them.
"""
def __init__(self, dataset_lens, mixture_ratios, num_samples_per_epoch, batch_size, seed=42):
self.dataset_lens = [int(length) for length in dataset_lens]
self.weights = torch.tensor(mixture_ratios, dtype=torch.double)
self.weights /= self.weights.sum()
self.num_samples = int(num_samples_per_epoch)
self.batch_size = int(batch_size)
self.seed = int(seed)
self.rank = int(os.environ.get("RANK", 0))
self.world_size = int(os.environ.get("WORLD_SIZE", 1))
self.epoch = 0
self.batch_counts = self._build_batch_counts()
self.num_batches_rank = self._build_num_batches_rank()
def _build_batch_counts(self):
positive_ids = [i for i, weight in enumerate(self.weights.tolist()) if weight > 0]
if not positive_ids:
raise ValueError("NoReplacementBatchMixtureSampler received no positive mixture ratios")
counts = torch.zeros(len(self.dataset_lens), dtype=torch.long)
if self.batch_size >= len(positive_ids):
counts[positive_ids] = 1
remaining = self.batch_size - len(positive_ids)
if remaining > 0:
positive_weights = self.weights[positive_ids]
positive_weights = positive_weights / positive_weights.sum()
raw_extra = positive_weights * remaining
extra_counts = torch.floor(raw_extra).long()
counts[positive_ids] += extra_counts
extra_left = remaining - int(extra_counts.sum().item())
if extra_left > 0:
frac = raw_extra - extra_counts.double()
for frac_index in torch.argsort(frac, descending=True)[:extra_left].tolist():
counts[positive_ids[frac_index]] += 1
else:
for dataset_id in torch.argsort(self.weights, descending=True)[:self.batch_size].tolist():
counts[dataset_id] += 1
return counts.tolist()
def _build_num_batches_rank(self):
requested_batches = self.num_samples // (self.batch_size * self.world_size)
feasible_batches = []
for dataset_id, count in enumerate(self.batch_counts):
if count <= 0:
continue
feasible_batches.append(self.dataset_lens[dataset_id] // (int(count) * self.world_size))
if not feasible_batches:
raise ValueError("NoReplacementBatchMixtureSampler has no positive batch counts")
max_no_replacement_batches = min(feasible_batches)
num_batches = min(requested_batches, max_no_replacement_batches)
if num_batches <= 0:
raise ValueError(
"NoReplacementBatchMixtureSampler cannot form one full distributed batch without replacement: "
f"dataset_lens={self.dataset_lens}, batch_counts={self.batch_counts}, "
f"world_size={self.world_size}, batch_size={self.batch_size}"
)
return num_batches
def set_epoch(self, epoch):
self.epoch = int(epoch)
def _reshuffle_dataset(self, dataset_id, pass_id):
dataset_len = self.dataset_lens[dataset_id]
if dataset_len <= 0:
raise ValueError(f"Dataset {dataset_id} has non-positive length: {dataset_len}")
g = torch.Generator()
g.manual_seed(self.seed + self.epoch * 1_000_003 + dataset_id * 10_007 + pass_id)
return torch.randperm(dataset_len, generator=g).tolist()
def __iter__(self):
num_batches_rank = len(self)
total_batches = num_batches_rank * self.world_size
dataset_orders = [[] for _ in self.dataset_lens]
dataset_offsets = [0 for _ in self.dataset_lens]
dataset_passes = [0 for _ in self.dataset_lens]
def take(dataset_id):
if dataset_offsets[dataset_id] >= len(dataset_orders[dataset_id]):
if dataset_passes[dataset_id] > 0:
raise RuntimeError(
f"Dataset {dataset_id} was exhausted inside a no-replacement epoch"
)
dataset_orders[dataset_id] = self._reshuffle_dataset(dataset_id, dataset_passes[dataset_id])
dataset_offsets[dataset_id] = 0
dataset_passes[dataset_id] += 1
sample_id = dataset_orders[dataset_id][dataset_offsets[dataset_id]]
dataset_offsets[dataset_id] += 1
return sample_id
for batch_idx in range(total_batches):
g = torch.Generator()
g.manual_seed(self.seed + self.epoch * total_batches + batch_idx)
batch = []
for dataset_id, count in enumerate(self.batch_counts):
for _ in range(max(0, int(count))):
batch.append((dataset_id, take(dataset_id)))
shuffle_order = torch.randperm(len(batch), generator=g).tolist()
if batch_idx % self.world_size == self.rank:
yield [batch[i] for i in shuffle_order]
def __len__(self):
return self.num_batches_rank
@dataclass
class DatasetSpec:
# dataset: torch.utils.data.Dataset
dataset_length: int
prob: float # mixture ratio
image_size: tuple[int, int] # e.g. 224, 448
tokens_per_image: int # (image_size / patch)^2
class TokenMixtureSampler(Sampler):
"""
Multi-node safe dataset mixture sampler that samples from datasets with different image resolutoion while
keeping the total number of tokens per batch (approximately) fixed.
"""
def __init__(
self,
specs: list[DatasetSpec],
tokens_per_batch: int, # per rank
num_batches_per_rank: int,
seed: int = 0,
):
self.specs = specs
self.tokens_per_batch = tokens_per_batch
self.num_batches_per_rank = num_batches_per_rank
self.seed = seed
self.probs = [s.prob for s in specs]
self.epoch = 0
# Distributed settings
self.rank = int(os.environ.get("RANK", 0))
self.world_size = int(os.environ.get("WORLD_SIZE", 1))
def set_epoch(self, epoch):
self.epoch = epoch
# overwatch.info(f"TokenMixtureSampler set to epoch {epoch}")
def __iter__(self):
# assert self.epoch is not None, "TokenMixtureSampler epoch not set. Please call set_epoch(epoch) before iterating."
# For each global batch index, synchronize dataset_id selection across all ranks
for batch_idx in range(self.num_batches_per_rank * self.world_size):
if batch_idx % self.world_size != self.rank:
continue
# Use a deterministic RNG for this global! batch to pick dataset_id and indices
batch_seed = self.seed + self.epoch + batch_idx # // self.world_size
local_rng = random.Random(batch_seed)
dataset_id = local_rng.choices(range(len(self.specs)), weights=self.probs, k=1)[0]
# print(f"[TokenMixtureSampler] rank={self.rank} seed={batch_seed} dataset_id={dataset_id}")
spec = self.specs[dataset_id]
batch_size = max(1, self.tokens_per_batch // spec.tokens_per_image)
local_batch_rng = random.Random(self.seed + self.epoch + batch_idx)
indices = [
(dataset_id, local_batch_rng.randrange(spec.dataset_length))
for _ in range(batch_size)
]
# overwatch.info(f"rank {overwatch.rank()}: {self.epoch} batch {batch_idx} dataset_id {dataset_id} batch_size {batch_size}")
yield indices
def __len__(self):
return self.num_batches_per_rank
|