dda/examples/data.py at master · moskomule/dda · GitHub

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
import shutil
from copy import deepcopy
from pathlib import Path
from typing import Optional

import numpy as np
import torch
from PIL import Image
from omegaconf import OmegaConf
from torch.utils.data import DataLoader, RandomSampler
from torchvision import transforms
from torchvision.datasets import VisionDataset, CIFAR10, CIFAR100, SVHN, ImageFolder, utils as dataset_utils

from utils import RandAugment


class TinyImageNet(ImageFolder):

    def __init__(self, root, train=True, transform=None, target_transform=None, download=False):
        root = Path(root).expanduser()
        if not root.exists() and download:
            root.mkdir(exist_ok=True, parents=True)
            self.download(root)
        if root.exists() and download:
            print("Files already downloaded")
        root = root / "train" if train else root / "val"
        super(TinyImageNet, self).__init__(root, transform=transform, target_transform=target_transform)
        self.data = [self.loader(img) for img, _ in self.samples]

    @staticmethod
    def download(root: Path):
        dataset_utils.download_and_extract_archive(url="http://cs231n.stanford.edu/tiny-imagenet-200.zip",
                                                   download_root=root)
        temp_dir = "tiny-imagenet-200"
        shutil.move(str(root / temp_dir / "train"), str(root))
        shutil.move(str(root / temp_dir / "val"), str(root))
        val_dir = root / "val"
        with (val_dir / "val_annotations.txt").open() as f:
            val_anns = f.read()
        name_to_cls = {i[0]: i[1] for i in
                       [l.split() for l in val_anns.strip().split("\n")]}
        for name, cls in name_to_cls.items():
            (val_dir / cls).mkdir(exist_ok=True)
            shutil.move(str(val_dir / "images" / name), str(val_dir / cls))
        shutil.rmtree(str(root / temp_dir))

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


def new_getitem(self,
                index: int):
    img, target = self.data[index], int(self.targets[index])
    img = Image.fromarray(np.transpose(img, (1, 2, 0)))
    if self.transform is not None:
        img = self.transform(img)
    return img, target


SVHN.__getitem__ = new_getitem


class OriginalSVHN(SVHN):
    def __init__(self,
                 root,
                 train=True,
                 transform=None,
                 target_transform=None,
                 download=False):
        super(OriginalSVHN, self).__init__(root, split="train" if train else "test", transform=transform,
                                           target_transform=target_transform, download=download)
        self.targets = self.labels


class ExtraSVHN(object):
    def __new__(cls,
                root,
                train=True,
                transform=None,
                target_transform=None,
                download=False):
        if train:
            return (SVHN(root, split='train', transform=transform, download=download) +
                    SVHN(root, split='extra', transform=transform, download=download))
        else:
            return OriginalSVHN(root, train=False, transform=transform, download=download)


def get_dataloader(name: str,
                   val_size: int,
                   batch_size: int,
                   download: bool,
                   augment: Optional[OmegaConf],
                   skip_normalize: bool) -> (DataLoader, DataLoader, DataLoader, int):
    train_set, val_set, test_set, num_classes = _get_dataset(name, val_size, download, augment, skip_normalize)
    train_loader = DataLoader(train_set, batch_size=batch_size, sampler=RandomSampler(train_set, True),
                              num_workers=4, pin_memory=True)
    val_loader = DataLoader(val_set, batch_size=2 * batch_size, pin_memory=True)
    test_loader = DataLoader(test_set, batch_size=2 * batch_size, pin_memory=True)
    return train_loader, val_loader, test_loader, num_classes


def _get_dataset(name: str,
                 val_size: int,
                 download: bool,
                 augment: Optional[OmegaConf],
                 skip_normalize: bool) -> (VisionDataset, VisionDataset, VisionDataset, int):
    _datasets = {"cifar10": (CIFAR10, "~/.torch/data/cifar10",
                             [transforms.ToTensor(),
                              transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))],
                             [transforms.RandomCrop(32, padding=4, padding_mode='reflect'),
                              transforms.RandomHorizontalFlip()], 10),
                 "cifar100": (CIFAR100, "~/.torch/data/cifar100",
                              [transforms.ToTensor(),
                               transforms.Normalize((0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761))],
                              [transforms.RandomCrop(32, padding=4, padding_mode='reflect'),
                               transforms.RandomHorizontalFlip()], 100),
                 "svhn": (OriginalSVHN, "~/.torch/data/svhn",
                          [transforms.ToTensor(),
                           transforms.Normalize((0.4390, 0.4443, 0.4692), (0.1189, 0.1222, 0.1049))],
                          [transforms.RandomCrop(32, padding=4, padding_mode='reflect')], 10),
                 "extrasvhn": (ExtraSVHN, "~/.torch/data/svhn",
                               [transforms.ToTensor(),
                                transforms.Normalize((0.4390, 0.4443, 0.4692), (0.1189, 0.1222, 0.1049))],
                               [transforms.RandomCrop(32, padding=4, padding_mode='reflect')], 10),
                 "tinyimagenet": (TinyImageNet, "~/.torch/data/tinyimagenet",
                                  [transforms.ToTensor(),
                                   transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))],
                                  [transforms.Resize(40),
                                   transforms.RandomCrop(32),
                                   transforms.RandomHorizontalFlip()], 200)
                 }

    if name not in _datasets.keys():
        raise RuntimeError(f'Unknown dataset name {name}')

    dset, root, norm_transform, data_aug, num_cls = _datasets[name]
    train_set = dset(root, train=True, transform=transforms.Compose(norm_transform), download=download)
    test_set = dset(root, train=False, transform=transforms.Compose(norm_transform), download=download)
    train_set, val_set = _split_dataset(train_set, val_size)

    if augment is not None and hasattr(augment, 'name'):
        if str(augment.name).lower() == 'randaugment':
            print('Apply randaugment')
            data_aug += [RandAugment(augment.num_aug, augment.magnitude)]
        else:
            raise NotImplementedError

    if skip_normalize:
        norm_transform = norm_transform[:1]
    train_set.transform = transforms.Compose(data_aug + norm_transform)

    return train_set, val_set, test_set, num_cls


def _split_dataset(dataset: VisionDataset,
                   val_size: int) -> (VisionDataset, VisionDataset):
    indices = torch.randperm(len(dataset))
    valset = deepcopy(dataset)
    dataset.data = [dataset.data[i] for i in indices[val_size:]]
    dataset.targets = [dataset.targets[i] for i in indices[val_size:]]

    valset.data = [valset.data[i] for i in indices[:val_size]]
    valset.targets = [valset.targets[i] for i in indices[:val_size]]

    return dataset, valset