lightningdot
copied
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Readme
Files and versions
116 lines
4.0 KiB
116 lines
4.0 KiB
""" sampler for length bucketing (batch by tokens) """
|
|
import math
|
|
import random
|
|
|
|
import torch
|
|
import horovod.torch as hvd
|
|
from torch.utils.data import Sampler
|
|
from cytoolz import partition_all
|
|
|
|
|
|
class TokenBucketSampler(Sampler):
|
|
def __init__(self, lens, bucket_size, batch_size,
|
|
droplast=False, size_multiple=8):
|
|
self._lens = lens
|
|
self._max_tok = batch_size
|
|
self._bucket_size = bucket_size
|
|
self._droplast = droplast
|
|
self._size_mul = size_multiple
|
|
|
|
def _create_ids(self):
|
|
return list(range(len(self._lens)))
|
|
|
|
def _sort_fn(self, i):
|
|
return self._lens[i]
|
|
|
|
def __iter__(self):
|
|
ids = self._create_ids()
|
|
random.shuffle(ids)
|
|
buckets = [sorted(ids[i:i+self._bucket_size],
|
|
key=self._sort_fn, reverse=True)
|
|
for i in range(0, len(ids), self._bucket_size)]
|
|
# fill batches until max_token (include padding)
|
|
batches = []
|
|
for bucket in buckets:
|
|
max_len = 0
|
|
batch_indices = []
|
|
for indices in partition_all(self._size_mul, bucket):
|
|
max_len = max(max_len, max(self._lens[i] for i in indices))
|
|
if (max_len * (len(batch_indices) + self._size_mul)
|
|
> self._max_tok):
|
|
if not batch_indices:
|
|
raise ValueError(
|
|
"max_tokens too small / max_seq_len too long")
|
|
assert len(batch_indices) % self._size_mul == 0
|
|
batches.append(batch_indices)
|
|
batch_indices = list(indices)
|
|
else:
|
|
batch_indices.extend(indices)
|
|
if not self._droplast and batch_indices:
|
|
batches.append(batch_indices)
|
|
random.shuffle(batches)
|
|
return iter(batches)
|
|
|
|
def __len__(self):
|
|
raise ValueError("NOT supported. "
|
|
"This has some randomness across epochs")
|
|
|
|
|
|
class DistributedSampler(Sampler):
|
|
"""Sampler that restricts data loading to a subset of the dataset.
|
|
|
|
It is especially useful in conjunction with
|
|
:class:`torch.nn.parallel.DistributedDataParallel`. In such case, each
|
|
process can pass a DistributedSampler instance as a DataLoader sampler,
|
|
and load a subset of the original dataset that is exclusive to it.
|
|
|
|
.. note::
|
|
Dataset is assumed to be of constant size.
|
|
|
|
Arguments:
|
|
dataset: Dataset used for sampling.
|
|
num_replicas (optional): Number of processes participating in
|
|
distributed training.
|
|
rank (optional): Rank of the current process within num_replicas.
|
|
shuffle (optional): If true (default), sampler will shuffle the indices
|
|
"""
|
|
|
|
def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
|
|
if num_replicas is None:
|
|
num_replicas = hvd.size()
|
|
if rank is None:
|
|
rank = hvd.rank()
|
|
self.dataset = dataset
|
|
self.num_replicas = num_replicas
|
|
self.rank = rank
|
|
self.epoch = 0
|
|
self.num_samples = int(math.ceil(len(self.dataset)
|
|
* 1.0 / self.num_replicas))
|
|
self.total_size = self.num_samples * self.num_replicas
|
|
self.shuffle = shuffle
|
|
|
|
def __iter__(self):
|
|
# deterministically shuffle based on epoch
|
|
g = torch.Generator()
|
|
g.manual_seed(self.epoch)
|
|
|
|
indices = list(range(len(self.dataset)))
|
|
# add extra samples to make it evenly divisible
|
|
indices += indices[:(self.total_size - len(indices))]
|
|
assert len(indices) == self.total_size
|
|
|
|
# subsample
|
|
indices = indices[self.rank:self.total_size:self.num_replicas]
|
|
|
|
if self.shuffle:
|
|
shufle_ind = torch.randperm(len(indices), generator=g).tolist()
|
|
indices = [indices[i] for i in shufle_ind]
|
|
assert len(indices) == self.num_samples
|
|
|
|
return iter(indices)
|
|
|
|
def __len__(self):
|
|
return self.num_samples
|
|
|
|
def set_epoch(self, epoch):
|
|
self.epoch = epoch
|