Newer
Older
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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
"""
Simple dataset loaders.
For more datasets and more comprehensive loaders, you may turn to dedicated
libraries like `fuel`.
"""
import gzip
import hashlib
import os
import pickle
from typing import *
import idx2numpy
import numpy as np
from ..typing_ import *
from ..utils import CacheDir, validate_enum_arg
__all__ = ['load_mnist', 'load_fashion_mnist', 'load_cifar10', 'load_cifar100']
_MNIST_LIKE_FILE_NAMES = {
'train_x': 'train-images-idx3-ubyte.gz',
'train_y': 'train-labels-idx1-ubyte.gz',
'test_x': 't10k-images-idx3-ubyte.gz',
'test_y': 't10k-labels-idx1-ubyte.gz',
}
_MNIST_URI_PREFIX = 'http://yann.lecun.com/exdb/mnist/'
_MNIST_FILE_MD5 = {
'train_x': 'f68b3c2dcbeaaa9fbdd348bbdeb94873',
'train_y': 'd53e105ee54ea40749a09fcbcd1e9432',
'test_x': '9fb629c4189551a2d022fa330f9573f3',
'test_y': 'ec29112dd5afa0611ce80d1b7f02629c',
}
_FASHION_MNIST_URI_PREFIX = 'http://fashion-mnist.s3-website.eu-central-1.' \
'amazonaws.com/'
_FASHION_MNIST_FILE_MD5 = {
'train_x': '8d4fb7e6c68d591d4c3dfef9ec88bf0d',
'train_y': '25c81989df183df01b3e8a0aad5dffbe',
'test_x': 'bef4ecab320f06d8554ea6380940ec79',
'test_y': 'bb300cfdad3c16e7a12a480ee83cd310',
}
_CIFAR_10_URI = 'https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz'
_CIFAR_10_MD5 = 'c58f30108f718f92721af3b95e74349a'
_CIFAR_10_CONTENT_DIR = 'cifar-10-batches-py'
_CIFAR_100_URI = 'https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz'
_CIFAR_100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85'
_CIFAR_100_CONTENT_DIR = 'cifar-100-python'
def _validate_x_shape(shape, default_shape):
shape = tuple(int(v) for v in shape)
default_shape = tuple(int(v) for v in default_shape)
value_size = int(np.prod(default_shape))
if np.prod(shape) != value_size:
raise ValueError(f'`x_shape` does not product to {value_size}: {shape}')
return shape
def load_mnist_like(uri_prefix: str,
file_md5: Dict[str, str],
cache_name: str,
x_shape: Sequence[int] = (28, 28),
x_dtype: ArrayDType = np.uint8,
y_dtype: ArrayDType = np.int32
) -> Tuple[XYArrayTuple, XYArrayTuple]:
"""
Load an MNIST-like dataset as NumPy arrays.
Args:
uri_prefix: Common prefix of the URIs in `remote_files`.
file_md5: The remote file MD5 hash sums, a dict of
`{'train_x': ..., 'train_y': ..., 'test_x': ..., 'test_y': ...}`,
where each value is the md5 sum.
cache_name: Name of the cache directory.
x_shape: Reshape each digit into this shape.
x_dtype: Cast each digit into this data type.
y_dtype: Cast each label into this data type.
Returns:
The ``(train_x, train_y), (test_x, test_y)`` arrays.
"""
def _fetch_array(array_name):
uri = uri_prefix + _MNIST_LIKE_FILE_NAMES[array_name]
md5 = file_md5[array_name]
path = CacheDir(cache_name).download(
uri, hasher=hashlib.md5(), expected_hash=md5)
with gzip.open(path, 'rb') as f:
return idx2numpy.convert_from_file(f)
# check arguments
x_shape = _validate_x_shape(x_shape, (28, 28))
# load data
train_x = _fetch_array('train_x').astype(x_dtype)
train_y = _fetch_array('train_y').astype(y_dtype)
test_x = _fetch_array('test_x').astype(x_dtype)
test_y = _fetch_array('test_y').astype(y_dtype)
assert(len(train_x) == len(train_y) == 60000)
assert(len(test_x) == len(test_y) == 10000)
# change shape
train_x = train_x.reshape([len(train_x)] + list(x_shape))
test_x = test_x.reshape([len(test_x)] + list(x_shape))
return (train_x, train_y), (test_x, test_y)
def load_mnist(x_shape: Sequence[int] = (28, 28),
x_dtype: ArrayDType = np.uint8,
y_dtype: ArrayDType = np.int32
) -> Tuple[XYArrayTuple, XYArrayTuple]:
"""
Load an MNIST dataset as NumPy arrays.
Args:
x_shape: Reshape each digit into this shape.
x_dtype: Cast each digit into this data type.
y_dtype: Cast each label into this data type.
Returns:
The ``(train_x, train_y), (test_x, test_y)`` arrays.
"""
return load_mnist_like(
_MNIST_URI_PREFIX, _MNIST_FILE_MD5, 'mnist', x_shape, x_dtype, y_dtype)
def load_fashion_mnist(x_shape: Sequence[int] = (28, 28),
x_dtype: ArrayDType = np.uint8,
y_dtype: ArrayDType = np.int32
) -> Tuple[XYArrayTuple, XYArrayTuple]:
"""
Load an MNIST dataset as NumPy arrays.
Args:
x_shape: Reshape each digit into this shape.
x_dtype: Cast each digit into this data type.
y_dtype: Cast each label into this data type.
Returns:
The ``(train_x, train_y), (test_x, test_y)`` arrays.
"""
return load_mnist_like(
_FASHION_MNIST_URI_PREFIX, _FASHION_MNIST_FILE_MD5, 'fashion_mnist',
x_shape, x_dtype, y_dtype)
def _cifar_load_batch(path, x_shape, x_dtype, y_dtype, expected_batch_label,
labels_key='labels'):
# load from file
with open(path, 'rb') as f:
d = {
k.decode('utf-8'): v
for k, v in pickle.load(f, encoding='bytes').items()
}
d['batch_label'] = d['batch_label'].decode('utf-8')
assert(d['batch_label'] == expected_batch_label)
data = np.asarray(d['data'], dtype=x_dtype)
labels = np.asarray(d[labels_key], dtype=y_dtype)
# change shape
data = data.reshape((data.shape[0], 3, 32, 32))
data = np.transpose(data, (0, 2, 3, 1))
if x_shape:
data = data.reshape([data.shape[0]] + list(x_shape))
return data, labels
def load_cifar10(x_shape: Sequence[int] = (32, 32, 3),
x_dtype: ArrayDType = np.float32,
y_dtype: ArrayDType = np.int32) -> Tuple[XYArrayTuple, XYArrayTuple]:
"""
Load the CIFAR-10 dataset as NumPy arrays.
Args:
x_shape: Reshape each digit into this shape.
x_dtype: Cast each digit into this data type.
y_dtype: Cast each label into this data type.
Returns:
The ``(train_x, train_y), (test_x, test_y)`` arrays.
"""
# check the arguments
x_shape = _validate_x_shape(x_shape, (32, 32, 3))
# fetch data
path = CacheDir('cifar').download_and_extract(
_CIFAR_10_URI, hasher=hashlib.md5(), expected_hash=_CIFAR_10_MD5)
data_dir = os.path.join(path, _CIFAR_10_CONTENT_DIR)
# load the data
train_num = 50000
train_x = np.zeros((train_num,) + x_shape, dtype=x_dtype)
train_y = np.zeros((train_num,), dtype=y_dtype)
for i in range(1, 6):
path = os.path.join(data_dir, 'data_batch_{}'.format(i))
x, y = _cifar_load_batch(
path, x_shape=x_shape, x_dtype=x_dtype, y_dtype=y_dtype,
expected_batch_label='training batch {} of 5'.format(i)
)
(train_x[(i - 1) * 10000: i * 10000, ...],
train_y[(i - 1) * 10000: i * 10000]) = x, y
path = os.path.join(data_dir, 'test_batch')
test_x, test_y = _cifar_load_batch(
path, x_shape=x_shape, x_dtype=x_dtype, y_dtype=y_dtype,
expected_batch_label='testing batch 1 of 1'
)
assert(len(test_x) == len(test_y) == 10000)
return (train_x, train_y), (test_x, test_y)
def load_cifar100(label_mode: str = 'fine',
x_shape: Sequence[int] = (32, 32, 3),
x_dtype: ArrayDType = np.float32,
y_dtype: ArrayDType = np.int32) -> Tuple[XYArrayTuple, XYArrayTuple]:
"""
Load the CIFAR-100 dataset as NumPy arrays.
Args:
label_mode: One of {"fine", "coarse"}.
x_shape: Reshape each digit into this shape.
x_dtype: Cast each digit into this data type.
y_dtype: Cast each label into this data type.
Returns:
The ``(train_x, train_y), (test_x, test_y)`` arrays.
"""
# check the arguments
label_mode = validate_enum_arg('label_mode', label_mode, ('fine', 'coarse'))
x_shape = _validate_x_shape(x_shape, (32, 32, 3))
# fetch data
path = CacheDir('cifar').download_and_extract(
_CIFAR_100_URI, hasher=hashlib.md5(), expected_hash=_CIFAR_100_MD5)
data_dir = os.path.join(path, _CIFAR_100_CONTENT_DIR)
# load the data
path = os.path.join(data_dir, 'train')
train_x, train_y = _cifar_load_batch(
path, x_shape=x_shape, x_dtype=x_dtype, y_dtype=y_dtype,
expected_batch_label='training batch 1 of 1',
labels_key='{}_labels'.format(label_mode)
)
assert(len(train_x) == len(train_y) == 50000)
path = os.path.join(data_dir, 'test')
test_x, test_y = _cifar_load_batch(
path, x_shape=x_shape, x_dtype=x_dtype, y_dtype=y_dtype,
expected_batch_label='testing batch 1 of 1',
labels_key='{}_labels'.format(label_mode)
)
assert(len(test_x) == len(test_y) == 10000)
return (train_x, train_y), (test_x, test_y)