File size: 6,826 Bytes
4121bec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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) Facebook, Inc. and its affiliates.
import copy
import itertools
import logging
import numpy as np
import pickle
import random
import torch.utils.data as data
from torch.utils.data.sampler import Sampler

from detectron2.utils.serialize import PicklableWrapper

__all__ = ["MapDataset", "DatasetFromList", "AspectRatioGroupedDataset", "ToIterableDataset"]


class MapDataset(data.Dataset):
    """
    Map a function over the elements in a dataset.

    Args:
        dataset: a dataset where map function is applied.
        map_func: a callable which maps the element in dataset. map_func is
            responsible for error handling, when error happens, it needs to
            return None so the MapDataset will randomly use other
            elements from the dataset.
    """

    def __init__(self, dataset, map_func):
        self._dataset = dataset
        self._map_func = PicklableWrapper(map_func)  # wrap so that a lambda will work

        self._rng = random.Random(42)
        self._fallback_candidates = set(range(len(dataset)))

    def __len__(self):
        return len(self._dataset)

    def __getitem__(self, idx):
        retry_count = 0
        cur_idx = int(idx)

        while True:
            data = self._map_func(self._dataset[cur_idx])
            if data is not None:
                self._fallback_candidates.add(cur_idx)
                return data

            # _map_func fails for this idx, use a random new index from the pool
            retry_count += 1
            self._fallback_candidates.discard(cur_idx)
            cur_idx = self._rng.sample(self._fallback_candidates, k=1)[0]

            if retry_count >= 3:
                logger = logging.getLogger(__name__)
                logger.warning(
                    "Failed to apply `_map_func` for idx: {}, retry count: {}".format(
                        idx, retry_count
                    )
                )


class DatasetFromList(data.Dataset):
    """
    Wrap a list to a torch Dataset. It produces elements of the list as data.
    """

    def __init__(self, lst: list, copy: bool = True, serialize: bool = True):
        """
        Args:
            lst (list): a list which contains elements to produce.
            copy (bool): whether to deepcopy the element when producing it,
                so that the result can be modified in place without affecting the
                source in the list.
            serialize (bool): whether to hold memory using serialized objects, when
                enabled, data loader workers can use shared RAM from master
                process instead of making a copy.
        """
        self._lst = lst
        self._copy = copy
        self._serialize = serialize

        def _serialize(data):
            buffer = pickle.dumps(data, protocol=-1)
            return np.frombuffer(buffer, dtype=np.uint8)

        if self._serialize:
            logger = logging.getLogger(__name__)
            logger.info(
                "Serializing {} elements to byte tensors and concatenating them all ...".format(
                    len(self._lst)
                )
            )
            self._lst = [_serialize(x) for x in self._lst]
            self._addr = np.asarray([len(x) for x in self._lst], dtype=np.int64)
            self._addr = np.cumsum(self._addr)
            self._lst = np.concatenate(self._lst)
            logger.info("Serialized dataset takes {:.2f} MiB".format(len(self._lst) / 1024 ** 2))

    def __len__(self):
        if self._serialize:
            return len(self._addr)
        else:
            return len(self._lst)

    def __getitem__(self, idx):
        if self._serialize:
            start_addr = 0 if idx == 0 else self._addr[idx - 1].item()
            end_addr = self._addr[idx].item()
            bytes = memoryview(self._lst[start_addr:end_addr])
            return pickle.loads(bytes)
        elif self._copy:
            return copy.deepcopy(self._lst[idx])
        else:
            return self._lst[idx]


class ToIterableDataset(data.IterableDataset):
    """
    Convert an old indices-based (also called map-style) dataset
    to an iterable-style dataset.
    """

    def __init__(self, dataset, sampler):
        """
        Args:
            dataset (torch.utils.data.Dataset): an old-style dataset with ``__getitem__``
            sampler (torch.utils.data.sampler.Sampler): a cheap iterable that produces indices
                to be applied on ``dataset``.
        """
        assert not isinstance(dataset, data.IterableDataset), dataset
        assert isinstance(sampler, Sampler), sampler
        self.dataset = dataset
        self.sampler = sampler

    def __iter__(self):
        worker_info = data.get_worker_info()
        if worker_info is None or worker_info.num_workers == 1:
            for idx in self.sampler:
                yield self.dataset[idx]
        else:
            # With map-style dataset, `DataLoader(dataset, sampler)` runs the
            # sampler in main process only. But `DataLoader(ToIterableDataset(dataset, sampler))`
            # will run sampler in every of the N worker and only keep 1/N of the ids on each
            # worker. The assumption is that sampler is cheap to iterate and it's fine to discard
            # ids in workers.
            for idx in itertools.islice(
                self.sampler, worker_info.id, None, worker_info.num_workers
            ):
                yield self.dataset[idx]


class AspectRatioGroupedDataset(data.IterableDataset):
    """
    Batch data that have similar aspect ratio together.
    In this implementation, images whose aspect ratio < (or >) 1 will
    be batched together.
    This improves training speed because the images then need less padding
    to form a batch.

    It assumes the underlying dataset produces dicts with "width" and "height" keys.
    It will then produce a list of original dicts with length = batch_size,
    all with similar aspect ratios.
    """

    def __init__(self, dataset, batch_size):
        """
        Args:
            dataset: an iterable. Each element must be a dict with keys
                "width" and "height", which will be used to batch data.
            batch_size (int):
        """
        self.dataset = dataset
        self.batch_size = batch_size
        self._buckets = [[] for _ in range(2)]
        # Hard-coded two aspect ratio groups: w > h and w < h.
        # Can add support for more aspect ratio groups, but doesn't seem useful

    def __iter__(self):
        for d in self.dataset:
            w, h = d["width"], d["height"]
            bucket_id = 0 if w > h else 1
            bucket = self._buckets[bucket_id]
            bucket.append(d)
            if len(bucket) == self.batch_size:
                yield bucket[:]
                del bucket[:]