|
1
|
|
|
import gzip |
|
2
|
|
|
import os |
|
3
|
|
|
import struct |
|
4
|
|
|
|
|
5
|
|
|
import h5py |
|
6
|
|
|
import numpy |
|
7
|
|
|
|
|
8
|
|
|
from fuel.converters.base import fill_hdf5_file, check_exists |
|
9
|
|
|
|
|
10
|
|
|
MNIST_IMAGE_MAGIC = 2051 |
|
11
|
|
|
MNIST_LABEL_MAGIC = 2049 |
|
12
|
|
|
|
|
13
|
|
|
TRAIN_IMAGES = 'train-images-idx3-ubyte.gz' |
|
14
|
|
|
TRAIN_LABELS = 'train-labels-idx1-ubyte.gz' |
|
15
|
|
|
TEST_IMAGES = 't10k-images-idx3-ubyte.gz' |
|
16
|
|
|
TEST_LABELS = 't10k-labels-idx1-ubyte.gz' |
|
17
|
|
|
|
|
18
|
|
|
ALL_FILES = [TRAIN_IMAGES, TRAIN_LABELS, TEST_IMAGES, TEST_LABELS] |
|
19
|
|
|
|
|
20
|
|
|
|
|
21
|
|
|
@check_exists(required_files=ALL_FILES) |
|
22
|
|
|
def convert_mnist(directory, output_directory, output_filename=None, |
|
23
|
|
|
dtype=None): |
|
24
|
|
|
"""Converts the MNIST dataset to HDF5. |
|
25
|
|
|
|
|
26
|
|
|
Converts the MNIST dataset to an HDF5 dataset compatible with |
|
27
|
|
|
:class:`fuel.datasets.MNIST`. The converted dataset is |
|
28
|
|
|
saved as 'mnist.hdf5'. |
|
29
|
|
|
|
|
30
|
|
|
This method assumes the existence of the following files: |
|
31
|
|
|
`train-images-idx3-ubyte.gz`, `train-labels-idx1-ubyte.gz` |
|
32
|
|
|
`t10k-images-idx3-ubyte.gz`, `t10k-labels-idx1-ubyte.gz` |
|
33
|
|
|
|
|
34
|
|
|
It assumes the existence of the following files: |
|
35
|
|
|
|
|
36
|
|
|
* `train-images-idx3-ubyte.gz` |
|
37
|
|
|
* `train-labels-idx1-ubyte.gz` |
|
38
|
|
|
* `t10k-images-idx3-ubyte.gz` |
|
39
|
|
|
* `t10k-labels-idx1-ubyte.gz` |
|
40
|
|
|
|
|
41
|
|
|
Parameters |
|
42
|
|
|
---------- |
|
43
|
|
|
directory : str |
|
44
|
|
|
Directory in which input files reside. |
|
45
|
|
|
output_directory : str |
|
46
|
|
|
Directory in which to save the converted dataset. |
|
47
|
|
|
output_filename : str, optional |
|
48
|
|
|
Name of the saved dataset. Defaults to `None`, in which case a name |
|
49
|
|
|
based on `dtype` will be used. |
|
50
|
|
|
dtype : str, optional |
|
51
|
|
|
Either 'float32', 'float64', or 'bool'. Defaults to `None`, |
|
52
|
|
|
in which case images will be returned in their original |
|
53
|
|
|
unsigned byte format. |
|
54
|
|
|
|
|
55
|
|
|
Returns |
|
56
|
|
|
------- |
|
57
|
|
|
output_paths : tuple of str |
|
58
|
|
|
Single-element tuple containing the path to the converted dataset. |
|
59
|
|
|
|
|
60
|
|
|
""" |
|
61
|
|
|
if not output_filename: |
|
62
|
|
|
if dtype: |
|
63
|
|
|
output_filename = 'mnist_{}.hdf5'.format(dtype) |
|
64
|
|
|
else: |
|
65
|
|
|
output_filename = 'mnist.hdf5' |
|
66
|
|
|
output_path = os.path.join(output_directory, output_filename) |
|
67
|
|
|
h5file = h5py.File(output_path, mode='w') |
|
68
|
|
|
|
|
69
|
|
|
train_feat_path = os.path.join(directory, TRAIN_IMAGES) |
|
70
|
|
|
train_features = read_mnist_images(train_feat_path, dtype) |
|
71
|
|
|
train_lab_path = os.path.join(directory, TRAIN_LABELS) |
|
72
|
|
|
train_labels = read_mnist_labels(train_lab_path) |
|
73
|
|
|
test_feat_path = os.path.join(directory, TEST_IMAGES) |
|
74
|
|
|
test_features = read_mnist_images(test_feat_path, dtype) |
|
75
|
|
|
test_lab_path = os.path.join(directory, TEST_LABELS) |
|
76
|
|
|
test_labels = read_mnist_labels(test_lab_path) |
|
77
|
|
|
data = (('train', 'features', train_features), |
|
78
|
|
|
('train', 'targets', train_labels), |
|
79
|
|
|
('test', 'features', test_features), |
|
80
|
|
|
('test', 'targets', test_labels)) |
|
81
|
|
|
fill_hdf5_file(h5file, data) |
|
82
|
|
|
h5file['features'].dims[0].label = 'batch' |
|
83
|
|
|
h5file['features'].dims[1].label = 'channel' |
|
84
|
|
|
h5file['features'].dims[2].label = 'height' |
|
85
|
|
|
h5file['features'].dims[3].label = 'width' |
|
86
|
|
|
h5file['targets'].dims[0].label = 'batch' |
|
87
|
|
|
h5file['targets'].dims[1].label = 'index' |
|
88
|
|
|
|
|
89
|
|
|
h5file.flush() |
|
90
|
|
|
h5file.close() |
|
91
|
|
|
|
|
92
|
|
|
return (output_path,) |
|
93
|
|
|
|
|
94
|
|
|
|
|
95
|
|
|
def fill_subparser(subparser): |
|
96
|
|
|
"""Sets up a subparser to convert the MNIST dataset files. |
|
97
|
|
|
|
|
98
|
|
|
Parameters |
|
99
|
|
|
---------- |
|
100
|
|
|
subparser : :class:`argparse.ArgumentParser` |
|
101
|
|
|
Subparser handling the `mnist` command. |
|
102
|
|
|
|
|
103
|
|
|
""" |
|
104
|
|
|
subparser.add_argument( |
|
105
|
|
|
"--dtype", help="dtype to save to; by default, images will be " + |
|
106
|
|
|
"returned in their original unsigned byte format", |
|
107
|
|
|
choices=('float32', 'float64', 'bool'), type=str, default=None) |
|
108
|
|
|
return convert_mnist |
|
109
|
|
|
|
|
110
|
|
|
|
|
111
|
|
|
def read_mnist_images(filename, dtype=None): |
|
112
|
|
|
"""Read MNIST images from the original ubyte file format. |
|
113
|
|
|
|
|
114
|
|
|
Parameters |
|
115
|
|
|
---------- |
|
116
|
|
|
filename : str |
|
117
|
|
|
Filename/path from which to read images. |
|
118
|
|
|
|
|
119
|
|
|
dtype : 'float32', 'float64', or 'bool' |
|
120
|
|
|
If unspecified, images will be returned in their original |
|
121
|
|
|
unsigned byte format. |
|
122
|
|
|
|
|
123
|
|
|
Returns |
|
124
|
|
|
------- |
|
125
|
|
|
images : :class:`~numpy.ndarray`, shape (n_images, 1, n_rows, n_cols) |
|
126
|
|
|
An image array, with individual examples indexed along the |
|
127
|
|
|
first axis and the image dimensions along the second and |
|
128
|
|
|
third axis. |
|
129
|
|
|
|
|
130
|
|
|
Notes |
|
131
|
|
|
----- |
|
132
|
|
|
If the dtype provided was Boolean, the resulting array will |
|
133
|
|
|
be Boolean with `True` if the corresponding pixel had a value |
|
134
|
|
|
greater than or equal to 128, `False` otherwise. |
|
135
|
|
|
|
|
136
|
|
|
If the dtype provided was a float dtype, the values will be mapped to |
|
137
|
|
|
the unit interval [0, 1], with pixel values that were 255 in the |
|
138
|
|
|
original unsigned byte representation equal to 1.0. |
|
139
|
|
|
|
|
140
|
|
|
""" |
|
141
|
|
|
with gzip.open(filename, 'rb') as f: |
|
142
|
|
|
magic, number, rows, cols = struct.unpack('>iiii', f.read(16)) |
|
143
|
|
|
if magic != MNIST_IMAGE_MAGIC: |
|
144
|
|
|
raise ValueError("Wrong magic number reading MNIST image file") |
|
145
|
|
|
array = numpy.frombuffer(f.read(), dtype='uint8') |
|
146
|
|
|
array = array.reshape((number, 1, rows, cols)) |
|
147
|
|
|
if dtype: |
|
148
|
|
|
dtype = numpy.dtype(dtype) |
|
149
|
|
|
|
|
150
|
|
|
if dtype.kind == 'b': |
|
151
|
|
|
# If the user wants Booleans, threshold at half the range. |
|
152
|
|
|
array = array >= 128 |
|
153
|
|
|
elif dtype.kind == 'f': |
|
154
|
|
|
# Otherwise, just convert. |
|
155
|
|
|
array = array.astype(dtype) |
|
156
|
|
|
array /= 255. |
|
157
|
|
|
else: |
|
158
|
|
|
raise ValueError("Unknown dtype to convert MNIST to") |
|
159
|
|
|
return array |
|
160
|
|
|
|
|
161
|
|
|
|
|
162
|
|
|
def read_mnist_labels(filename): |
|
163
|
|
|
"""Read MNIST labels from the original ubyte file format. |
|
164
|
|
|
|
|
165
|
|
|
Parameters |
|
166
|
|
|
---------- |
|
167
|
|
|
filename : str |
|
168
|
|
|
Filename/path from which to read labels. |
|
169
|
|
|
|
|
170
|
|
|
Returns |
|
171
|
|
|
------- |
|
172
|
|
|
labels : :class:`~numpy.ndarray`, shape (nlabels, 1) |
|
173
|
|
|
A one-dimensional unsigned byte array containing the |
|
174
|
|
|
labels as integers. |
|
175
|
|
|
|
|
176
|
|
|
""" |
|
177
|
|
|
with gzip.open(filename, 'rb') as f: |
|
178
|
|
|
magic, _ = struct.unpack('>ii', f.read(8)) |
|
179
|
|
|
if magic != MNIST_LABEL_MAGIC: |
|
180
|
|
|
raise ValueError("Wrong magic number reading MNIST label file") |
|
181
|
|
|
array = numpy.frombuffer(f.read(), dtype='uint8') |
|
182
|
|
|
array = array.reshape(array.size, 1) |
|
183
|
|
|
return array |
|
184
|
|
|
|
mila-udem /
fuel
