File size: 7,064 Bytes
51f6859
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import itertools

import numpy as np
import torch
from mmcv.runner import get_dist_info
from torch.utils.data.sampler import Sampler

from mmdet.core.utils import sync_random_seed


class InfiniteGroupBatchSampler(Sampler):
    """Similar to `BatchSampler` warping a `GroupSampler. It is designed for
    iteration-based runners like `IterBasedRunner` and yields a mini-batch
    indices each time, all indices in a batch should be in the same group.

    The implementation logic is referred to
    https://github.com/facebookresearch/detectron2/blob/main/detectron2/data/samplers/grouped_batch_sampler.py

    Args:
        dataset (object): The dataset.
        batch_size (int): When model is :obj:`DistributedDataParallel`,
            it is the number of training samples on each GPU.
            When model is :obj:`DataParallel`, it is
            `num_gpus * samples_per_gpu`.
            Default : 1.
        world_size (int, optional): Number of processes participating in
            distributed training. Default: None.
        rank (int, optional): Rank of current process. Default: None.
        seed (int): Random seed. Default: 0.
        shuffle (bool): Whether shuffle the indices of a dummy `epoch`, it
            should be noted that `shuffle` can not guarantee that you can
            generate sequential indices because it need to ensure
            that all indices in a batch is in a group. Default: True.
    """  # noqa: W605

    def __init__(self,
                 dataset,
                 batch_size=1,
                 world_size=None,
                 rank=None,
                 seed=0,
                 shuffle=True):
        _rank, _world_size = get_dist_info()
        if world_size is None:
            world_size = _world_size
        if rank is None:
            rank = _rank
        self.rank = rank
        self.world_size = world_size
        self.dataset = dataset
        self.batch_size = batch_size
        # In distributed sampling, different ranks should sample
        # non-overlapped data in the dataset. Therefore, this function
        # is used to make sure that each rank shuffles the data indices
        # in the same order based on the same seed. Then different ranks
        # could use different indices to select non-overlapped data from the
        # same data list.
        self.seed = sync_random_seed(seed)
        self.shuffle = shuffle

        assert hasattr(self.dataset, 'flag')
        self.flag = self.dataset.flag
        self.group_sizes = np.bincount(self.flag)
        # buffer used to save indices of each group
        self.buffer_per_group = {k: [] for k in range(len(self.group_sizes))}

        self.size = len(dataset)
        self.indices = self._indices_of_rank()

    def _infinite_indices(self):
        """Infinitely yield a sequence of indices."""
        g = torch.Generator()
        g.manual_seed(self.seed)
        while True:
            if self.shuffle:
                yield from torch.randperm(self.size, generator=g).tolist()

            else:
                yield from torch.arange(self.size).tolist()

    def _indices_of_rank(self):
        """Slice the infinite indices by rank."""
        yield from itertools.islice(self._infinite_indices(), self.rank, None,
                                    self.world_size)

    def __iter__(self):
        # once batch size is reached, yield the indices
        for idx in self.indices:
            flag = self.flag[idx]
            group_buffer = self.buffer_per_group[flag]
            group_buffer.append(idx)
            if len(group_buffer) == self.batch_size:
                yield group_buffer[:]
                del group_buffer[:]

    def __len__(self):
        """Length of base dataset."""
        return self.size

    def set_epoch(self, epoch):
        """Not supported in `IterationBased` runner."""
        raise NotImplementedError


class InfiniteBatchSampler(Sampler):
    """Similar to `BatchSampler` warping a `DistributedSampler. It is designed
    iteration-based runners like `IterBasedRunner` and yields a mini-batch
    indices each time.

    The implementation logic is referred to
    https://github.com/facebookresearch/detectron2/blob/main/detectron2/data/samplers/grouped_batch_sampler.py

    Args:
        dataset (object): The dataset.
        batch_size (int): When model is :obj:`DistributedDataParallel`,
            it is the number of training samples on each GPU,
            When model is :obj:`DataParallel`, it is
            `num_gpus * samples_per_gpu`.
            Default : 1.
        world_size (int, optional): Number of processes participating in
            distributed training. Default: None.
        rank (int, optional): Rank of current process. Default: None.
        seed (int): Random seed. Default: 0.
        shuffle (bool): Whether shuffle the dataset or not. Default: True.
    """  # noqa: W605

    def __init__(self,
                 dataset,
                 batch_size=1,
                 world_size=None,
                 rank=None,
                 seed=0,
                 shuffle=True):
        _rank, _world_size = get_dist_info()
        if world_size is None:
            world_size = _world_size
        if rank is None:
            rank = _rank
        self.rank = rank
        self.world_size = world_size
        self.dataset = dataset
        self.batch_size = batch_size
        # In distributed sampling, different ranks should sample
        # non-overlapped data in the dataset. Therefore, this function
        # is used to make sure that each rank shuffles the data indices
        # in the same order based on the same seed. Then different ranks
        # could use different indices to select non-overlapped data from the
        # same data list.
        self.seed = sync_random_seed(seed)
        self.shuffle = shuffle
        self.size = len(dataset)
        self.indices = self._indices_of_rank()

    def _infinite_indices(self):
        """Infinitely yield a sequence of indices."""
        g = torch.Generator()
        g.manual_seed(self.seed)
        while True:
            if self.shuffle:
                yield from torch.randperm(self.size, generator=g).tolist()

            else:
                yield from torch.arange(self.size).tolist()

    def _indices_of_rank(self):
        """Slice the infinite indices by rank."""
        yield from itertools.islice(self._infinite_indices(), self.rank, None,
                                    self.world_size)

    def __iter__(self):
        # once batch size is reached, yield the indices
        batch_buffer = []
        for idx in self.indices:
            batch_buffer.append(idx)
            if len(batch_buffer) == self.batch_size:
                yield batch_buffer
                batch_buffer = []

    def __len__(self):
        """Length of base dataset."""
        return self.size

    def set_epoch(self, epoch):
        """Not supported in `IterationBased` runner."""
        raise NotImplementedError