300 lines
10 KiB
Python
300 lines
10 KiB
Python
from __future__ import division
|
|
import math
|
|
|
|
import numpy as np
|
|
import torch
|
|
from mmcv.runner import get_dist_info
|
|
from torch.utils.data import DistributedSampler as _DistributedSampler
|
|
from torch.utils.data import Sampler
|
|
|
|
|
|
class DistributedSampler(_DistributedSampler):
|
|
|
|
def __init__(self,
|
|
dataset,
|
|
num_replicas=None,
|
|
rank=None,
|
|
shuffle=True,
|
|
replace=False):
|
|
super().__init__(dataset, num_replicas=num_replicas, rank=rank)
|
|
self.shuffle = shuffle
|
|
self.replace = replace
|
|
self.unif_sampling_flag = False
|
|
|
|
def __iter__(self):
|
|
# deterministically shuffle based on epoch
|
|
if not self.unif_sampling_flag:
|
|
self.generate_new_list()
|
|
else:
|
|
self.unif_sampling_flag = False
|
|
return iter(self.indices[self.rank * self.num_samples:(self.rank + 1) *
|
|
self.num_samples])
|
|
|
|
def generate_new_list(self):
|
|
if self.shuffle:
|
|
g = torch.Generator()
|
|
g.manual_seed(self.epoch)
|
|
if self.replace:
|
|
indices = torch.randint(
|
|
low=0,
|
|
high=len(self.dataset),
|
|
size=(len(self.dataset), ),
|
|
generator=g).tolist()
|
|
else:
|
|
indices = torch.randperm(
|
|
len(self.dataset), generator=g).tolist()
|
|
else:
|
|
indices = torch.arange(len(self.dataset)).tolist()
|
|
|
|
# add extra samples to make it evenly divisible
|
|
indices += indices[:(self.total_size - len(indices))]
|
|
assert len(indices) == self.total_size
|
|
self.indices = indices
|
|
|
|
def set_uniform_indices(self, labels, num_classes):
|
|
self.unif_sampling_flag = True
|
|
assert self.shuffle, "Using uniform sampling, the indices must be shuffled."
|
|
np.random.seed(self.epoch)
|
|
assert (len(labels) == len(self.dataset))
|
|
N = len(labels)
|
|
size_per_label = int(N / num_classes) + 1
|
|
indices = []
|
|
images_lists = [[] for i in range(num_classes)]
|
|
for i, l in enumerate(labels):
|
|
images_lists[l].append(i)
|
|
for i, l in enumerate(images_lists):
|
|
if len(l) == 0:
|
|
continue
|
|
indices.extend(
|
|
np.random.choice(
|
|
l, size_per_label, replace=(len(l) <= size_per_label)))
|
|
indices = np.array(indices)
|
|
np.random.shuffle(indices)
|
|
indices = indices[:N].astype(np.int).tolist()
|
|
|
|
# add extra samples to make it evenly divisible
|
|
assert len(indices) <= self.total_size, \
|
|
"{} vs {}".format(len(indices), self.total_size)
|
|
indices += indices[:(self.total_size - len(indices))]
|
|
assert len(indices) == self.total_size, \
|
|
"{} vs {}".format(len(indices), self.total_size)
|
|
self.indices = indices
|
|
|
|
|
|
class GroupSampler(Sampler):
|
|
|
|
def __init__(self, dataset, samples_per_gpu=1):
|
|
assert hasattr(dataset, 'flag')
|
|
self.dataset = dataset
|
|
self.samples_per_gpu = samples_per_gpu
|
|
self.flag = dataset.flag.astype(np.int64)
|
|
self.group_sizes = np.bincount(self.flag)
|
|
self.num_samples = 0
|
|
for i, size in enumerate(self.group_sizes):
|
|
self.num_samples += int(np.ceil(
|
|
size / self.samples_per_gpu)) * self.samples_per_gpu
|
|
|
|
def __iter__(self):
|
|
indices = []
|
|
for i, size in enumerate(self.group_sizes):
|
|
if size == 0:
|
|
continue
|
|
indice = np.where(self.flag == i)[0]
|
|
assert len(indice) == size
|
|
np.random.shuffle(indice)
|
|
num_extra = int(np.ceil(size / self.samples_per_gpu)
|
|
) * self.samples_per_gpu - len(indice)
|
|
indice = np.concatenate(
|
|
[indice, np.random.choice(indice, num_extra)])
|
|
indices.append(indice)
|
|
indices = np.concatenate(indices)
|
|
indices = [
|
|
indices[i * self.samples_per_gpu:(i + 1) * self.samples_per_gpu]
|
|
for i in np.random.permutation(
|
|
range(len(indices) // self.samples_per_gpu))
|
|
]
|
|
indices = np.concatenate(indices)
|
|
indices = indices.astype(np.int64).tolist()
|
|
assert len(indices) == self.num_samples
|
|
return iter(indices)
|
|
|
|
def __len__(self):
|
|
return self.num_samples
|
|
|
|
|
|
class DistributedGroupSampler(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.
|
|
"""
|
|
|
|
def __init__(self,
|
|
dataset,
|
|
samples_per_gpu=1,
|
|
num_replicas=None,
|
|
rank=None):
|
|
_rank, _num_replicas = get_dist_info()
|
|
if num_replicas is None:
|
|
num_replicas = _num_replicas
|
|
if rank is None:
|
|
rank = _rank
|
|
self.dataset = dataset
|
|
self.samples_per_gpu = samples_per_gpu
|
|
self.num_replicas = num_replicas
|
|
self.rank = rank
|
|
self.epoch = 0
|
|
|
|
assert hasattr(self.dataset, 'flag')
|
|
self.flag = self.dataset.flag
|
|
self.group_sizes = np.bincount(self.flag)
|
|
|
|
self.num_samples = 0
|
|
for i, j in enumerate(self.group_sizes):
|
|
self.num_samples += int(
|
|
math.ceil(self.group_sizes[i] * 1.0 / self.samples_per_gpu /
|
|
self.num_replicas)) * self.samples_per_gpu
|
|
self.total_size = self.num_samples * self.num_replicas
|
|
|
|
def __iter__(self):
|
|
# deterministically shuffle based on epoch
|
|
g = torch.Generator()
|
|
g.manual_seed(self.epoch)
|
|
|
|
indices = []
|
|
for i, size in enumerate(self.group_sizes):
|
|
if size > 0:
|
|
indice = np.where(self.flag == i)[0]
|
|
assert len(indice) == size
|
|
indice = indice[list(torch.randperm(int(size),
|
|
generator=g))].tolist()
|
|
extra = int(
|
|
math.ceil(
|
|
size * 1.0 / self.samples_per_gpu / self.num_replicas)
|
|
) * self.samples_per_gpu * self.num_replicas - len(indice)
|
|
# pad indice
|
|
tmp = indice.copy()
|
|
for _ in range(extra // size):
|
|
indice.extend(tmp)
|
|
indice.extend(tmp[:extra % size])
|
|
indices.extend(indice)
|
|
|
|
assert len(indices) == self.total_size
|
|
|
|
indices = [
|
|
indices[j] for i in list(
|
|
torch.randperm(
|
|
len(indices) // self.samples_per_gpu, generator=g))
|
|
for j in range(i * self.samples_per_gpu, (i + 1) *
|
|
self.samples_per_gpu)
|
|
]
|
|
|
|
# subsample
|
|
offset = self.num_samples * self.rank
|
|
indices = indices[offset:offset + self.num_samples]
|
|
assert len(indices) == self.num_samples
|
|
|
|
return iter(indices)
|
|
|
|
def __len__(self):
|
|
return self.num_samples
|
|
|
|
def set_epoch(self, epoch):
|
|
self.epoch = epoch
|
|
|
|
|
|
class DistributedGivenIterationSampler(Sampler):
|
|
|
|
def __init__(self,
|
|
dataset,
|
|
total_iter,
|
|
batch_size,
|
|
num_replicas=None,
|
|
rank=None,
|
|
last_iter=-1):
|
|
rank, world_size = get_dist_info()
|
|
assert rank < world_size
|
|
self.dataset = dataset
|
|
self.total_iter = total_iter
|
|
self.batch_size = batch_size
|
|
self.world_size = world_size
|
|
self.rank = rank
|
|
self.last_iter = last_iter
|
|
|
|
self.total_size = self.total_iter * self.batch_size
|
|
|
|
self.indices = self.gen_new_list()
|
|
|
|
def __iter__(self):
|
|
return iter(self.indices[(self.last_iter + 1) * self.batch_size:])
|
|
|
|
def set_uniform_indices(self, labels, num_classes):
|
|
np.random.seed(0)
|
|
assert (len(labels) == len(self.dataset))
|
|
N = len(labels)
|
|
size_per_label = int(N / num_classes) + 1
|
|
indices = []
|
|
images_lists = [[] for i in range(num_classes)]
|
|
for i, l in enumerate(labels):
|
|
images_lists[l].append(i)
|
|
for i, l in enumerate(images_lists):
|
|
if len(l) == 0:
|
|
continue
|
|
indices.extend(
|
|
np.random.choice(
|
|
l, size_per_label, replace=(len(l) <= size_per_label)))
|
|
indices = np.array(indices)
|
|
np.random.shuffle(indices)
|
|
indices = indices[:N].astype(np.int)
|
|
# repeat
|
|
all_size = self.total_size * self.world_size
|
|
indices = indices[:all_size]
|
|
num_repeat = (all_size - 1) // indices.shape[0] + 1
|
|
indices = np.tile(indices, num_repeat)
|
|
indices = indices[:all_size]
|
|
np.random.shuffle(indices)
|
|
# slice
|
|
beg = self.total_size * self.rank
|
|
indices = indices[beg:beg + self.total_size]
|
|
assert len(indices) == self.total_size
|
|
# set
|
|
self.indices = indices
|
|
|
|
def gen_new_list(self):
|
|
|
|
# each process shuffle all list with same seed, and pick one piece according to rank
|
|
np.random.seed(0)
|
|
|
|
all_size = self.total_size * self.world_size
|
|
indices = np.arange(len(self.dataset))
|
|
indices = indices[:all_size]
|
|
num_repeat = (all_size - 1) // indices.shape[0] + 1
|
|
indices = np.tile(indices, num_repeat)
|
|
indices = indices[:all_size]
|
|
|
|
np.random.shuffle(indices)
|
|
beg = self.total_size * self.rank
|
|
indices = indices[beg:beg + self.total_size]
|
|
|
|
assert len(indices) == self.total_size
|
|
|
|
return indices
|
|
|
|
def __len__(self):
|
|
# note here we do not take last iter into consideration, since __len__
|
|
# should only be used for displaying, the correct remaining size is
|
|
# handled by dataloader
|
|
#return self.total_size - (self.last_iter+1)*self.batch_size
|
|
return self.total_size
|
|
|
|
def set_epoch(self, epoch):
|
|
pass
|