1
|
|
|
from __future__ import division |
2
|
|
|
from collections import OrderedDict |
3
|
|
|
from functools import partial |
4
|
|
|
import gzip |
5
|
|
|
import io |
6
|
|
|
import os |
7
|
|
|
import logging |
8
|
|
|
import os.path |
9
|
|
|
|
10
|
|
|
import h5py |
11
|
|
|
import numpy |
12
|
|
|
from picklable_itertools.extras import equizip |
13
|
|
|
from PIL import Image |
14
|
|
|
from scipy.io.matlab import loadmat |
15
|
|
|
from six.moves import zip, xrange |
|
|
|
|
16
|
|
|
import zmq |
17
|
|
|
|
18
|
|
|
from fuel.converters.base import check_exists, progress_bar |
19
|
|
|
from fuel.datasets import H5PYDataset |
20
|
|
|
from fuel.utils.formats import tar_open |
21
|
|
|
from fuel.utils.parallel import producer_consumer |
22
|
|
|
from fuel import config |
23
|
|
|
|
24
|
|
|
log = logging.getLogger(__name__) |
|
|
|
|
25
|
|
|
|
26
|
|
|
DEVKIT_ARCHIVE = 'ILSVRC2010_devkit-1.0.tar.gz' |
27
|
|
|
DEVKIT_META_PATH = 'devkit-1.0/data/meta.mat' |
28
|
|
|
DEVKIT_VALID_GROUNDTRUTH_PATH = ('devkit-1.0/data/' |
29
|
|
|
'ILSVRC2010_validation_ground_truth.txt') |
30
|
|
|
PATCH_IMAGES_TAR = 'patch_images.tar' |
31
|
|
|
TEST_GROUNDTRUTH = 'ILSVRC2010_test_ground_truth.txt' |
32
|
|
|
TRAIN_IMAGES_TAR = 'ILSVRC2010_images_train.tar' |
33
|
|
|
VALID_IMAGES_TAR = 'ILSVRC2010_images_val.tar' |
34
|
|
|
TEST_IMAGES_TAR = 'ILSVRC2010_images_test.tar' |
35
|
|
|
IMAGE_TARS = (TRAIN_IMAGES_TAR, VALID_IMAGES_TAR, TEST_IMAGES_TAR, |
36
|
|
|
PATCH_IMAGES_TAR) |
37
|
|
|
PUBLIC_FILES = TEST_GROUNDTRUTH, DEVKIT_ARCHIVE |
38
|
|
|
ALL_FILES = PUBLIC_FILES + IMAGE_TARS |
39
|
|
|
|
40
|
|
|
|
41
|
|
|
@check_exists(required_files=ALL_FILES) |
42
|
|
|
def convert_ilsvrc2010(directory, output_directory, |
43
|
|
|
output_filename='ilsvrc2010.hdf5', |
44
|
|
|
shuffle_seed=config.default_seed): |
45
|
|
|
"""Converter for data from the ILSVRC 2010 competition. |
46
|
|
|
|
47
|
|
|
Source files for this dataset can be obtained by registering at |
48
|
|
|
[ILSVRC2010WEB]. |
49
|
|
|
|
50
|
|
|
Parameters |
51
|
|
|
---------- |
52
|
|
|
input_directory : str |
53
|
|
|
Path from which to read raw data files. |
54
|
|
|
output_directory : str |
55
|
|
|
Path to which to save the HDF5 file. |
56
|
|
|
output_filename : str, optional |
57
|
|
|
The output filename for the HDF5 file. Default: 'ilsvrc2010.hdf5'. |
58
|
|
|
shuffle_seed : int or sequence, optional |
59
|
|
|
Seed for a random number generator used to shuffle the order |
60
|
|
|
of the training set on disk, so that sequential reads will not |
61
|
|
|
be ordered by class. |
62
|
|
|
|
63
|
|
|
.. [ILSVRC2010WEB] http://image-net.org/challenges/LSVRC/2010/index |
64
|
|
|
|
65
|
|
|
""" |
66
|
|
|
devkit_path = os.path.join(directory, DEVKIT_ARCHIVE) |
67
|
|
|
test_groundtruth_path = os.path.join(directory, TEST_GROUNDTRUTH) |
68
|
|
|
train, valid, test, patch = [os.path.join(directory, fn) |
69
|
|
|
for fn in IMAGE_TARS] |
70
|
|
|
n_train, valid_groundtruth, test_groundtruth, wnid_map = \ |
71
|
|
|
prepare_metadata(devkit_path, test_groundtruth_path) |
72
|
|
|
n_valid, n_test = len(valid_groundtruth), len(test_groundtruth) |
73
|
|
|
output_path = os.path.join(output_directory, output_filename) |
74
|
|
|
|
75
|
|
|
with h5py.File(output_path, 'w') as f: |
76
|
|
|
log.info('Creating HDF5 datasets...') |
77
|
|
|
prepare_hdf5_file(f, n_train, n_valid, n_test) |
78
|
|
|
log.info('Processing training set...') |
79
|
|
|
process_train_set(f, train, patch, n_train, wnid_map, shuffle_seed) |
80
|
|
|
log.info('Processing validation set...') |
81
|
|
|
process_other_set(f, 'valid', valid, patch, valid_groundtruth, n_train) |
82
|
|
|
log.info('Processing test set...') |
83
|
|
|
process_other_set(f, 'test', test, patch, test_groundtruth, |
84
|
|
|
n_train + n_valid) |
85
|
|
|
log.info('Done.') |
86
|
|
|
|
87
|
|
|
return (output_path,) |
88
|
|
|
|
89
|
|
|
|
90
|
|
|
def fill_subparser(subparser): |
91
|
|
|
"""Sets up a subparser to convert the ILSVRC2010 dataset files. |
92
|
|
|
|
93
|
|
|
Parameters |
94
|
|
|
---------- |
95
|
|
|
subparser : :class:`argparse.ArgumentParser` |
96
|
|
|
Subparser handling the `ilsvrc2010` command. |
97
|
|
|
|
98
|
|
|
""" |
99
|
|
|
subparser.add_argument( |
100
|
|
|
"--shuffle-seed", help="Seed to use for randomizing order of the " |
101
|
|
|
"training set on disk.", |
102
|
|
|
default=config.default_seed, type=int, required=False) |
103
|
|
|
return convert_ilsvrc2010 |
104
|
|
|
|
105
|
|
|
|
106
|
|
|
def prepare_metadata(devkit_archive, test_groundtruth_path): |
107
|
|
|
"""Extract dataset metadata required for HDF5 file setup. |
108
|
|
|
|
109
|
|
|
Parameters |
110
|
|
|
---------- |
111
|
|
|
devkit_archive : str or file-like object |
112
|
|
|
The filename or file-handle for the gzipped TAR archive |
113
|
|
|
containing the ILSVRC2010 development kit. |
114
|
|
|
test_groundtruth_path : str or file-like object |
115
|
|
|
The filename or file-handle for the text file containing |
116
|
|
|
the ILSVRC2010 test set ground truth. |
117
|
|
|
|
118
|
|
|
Returns |
119
|
|
|
------- |
120
|
|
|
n_train : int |
121
|
|
|
The number of examples in the training set. |
122
|
|
|
valid_groundtruth : ndarray, 1-dimensional |
123
|
|
|
An ndarray containing the validation set groundtruth in terms of |
124
|
|
|
0-based class indices. |
125
|
|
|
test_groundtruth : ndarray, 1-dimensional |
126
|
|
|
An ndarray containing the test groundtruth in terms of 0-based |
127
|
|
|
class indices. |
128
|
|
|
wnid_map : dict |
129
|
|
|
A dictionary that maps WordNet IDs to 0-based class indices. |
130
|
|
|
|
131
|
|
|
""" |
132
|
|
|
# Read what's necessary from the development kit. |
133
|
|
|
synsets, cost_matrix, raw_valid_groundtruth = read_devkit(devkit_archive) |
|
|
|
|
134
|
|
|
|
135
|
|
|
# Mapping to take WordNet IDs to our internal 0-999 encoding. |
136
|
|
|
wnid_map = dict(zip((s.decode('utf8') for s in synsets['WNID']), |
137
|
|
|
xrange(1000))) |
138
|
|
|
|
139
|
|
|
# Map the 'ILSVRC2010 ID' to our zero-based ID. |
140
|
|
|
ilsvrc_id_to_zero_based = dict(zip(synsets['ILSVRC2010_ID'], |
141
|
|
|
xrange(len(synsets)))) |
142
|
|
|
|
143
|
|
|
# Map the validation set groundtruth to 0-999 labels. |
144
|
|
|
valid_groundtruth = [ilsvrc_id_to_zero_based[id_] |
145
|
|
|
for id_ in raw_valid_groundtruth] |
146
|
|
|
|
147
|
|
|
# Raw test data groundtruth, ILSVRC2010 IDs. |
148
|
|
|
raw_test_groundtruth = numpy.loadtxt(test_groundtruth_path, |
149
|
|
|
dtype=numpy.int16) |
150
|
|
|
|
151
|
|
|
# Map the test set groundtruth to 0-999 labels. |
152
|
|
|
test_groundtruth = [ilsvrc_id_to_zero_based[id_] |
153
|
|
|
for id_ in raw_test_groundtruth] |
154
|
|
|
|
155
|
|
|
# Ascertain the number of filenames to prepare appropriate sized |
156
|
|
|
# arrays. |
157
|
|
|
n_train = int(synsets['num_train_images'].sum()) |
158
|
|
|
log.info('Training set: {} images'.format(n_train)) |
159
|
|
|
log.info('Validation set: {} images'.format(len(valid_groundtruth))) |
160
|
|
|
log.info('Test set: {} images'.format(len(test_groundtruth))) |
161
|
|
|
n_total = n_train + len(valid_groundtruth) + len(test_groundtruth) |
162
|
|
|
log.info('Total (train/valid/test): {} images'.format(n_total)) |
163
|
|
|
return n_train, valid_groundtruth, test_groundtruth, wnid_map |
164
|
|
|
|
165
|
|
|
|
166
|
|
|
def create_splits(n_train, n_valid, n_test): |
167
|
|
|
n_total = n_train + n_valid + n_test |
168
|
|
|
tuples = {} |
169
|
|
|
tuples['train'] = (0, n_train) |
170
|
|
|
tuples['valid'] = (n_train, n_train + n_valid) |
171
|
|
|
tuples['test'] = (n_train + n_valid, n_total) |
172
|
|
|
sources = ['encoded_images', 'targets', 'filenames'] |
173
|
|
|
return OrderedDict( |
174
|
|
|
(split, OrderedDict((source, tuples[split]) for source in sources)) |
175
|
|
|
for split in ('train', 'valid', 'test') |
176
|
|
|
) |
177
|
|
|
|
178
|
|
|
|
179
|
|
|
def prepare_hdf5_file(hdf5_file, n_train, n_valid, n_test): |
180
|
|
|
"""Create datasets within a given HDF5 file. |
181
|
|
|
|
182
|
|
|
Parameters |
183
|
|
|
---------- |
184
|
|
|
hdf5_file : :class:`h5py.File` instance |
185
|
|
|
HDF5 file handle to which to write. |
186
|
|
|
n_train : int |
187
|
|
|
The number of training set examples. |
188
|
|
|
n_valid : int |
189
|
|
|
The number of validation set examples. |
190
|
|
|
n_test : int |
191
|
|
|
The number of test set examples. |
192
|
|
|
|
193
|
|
|
""" |
194
|
|
|
n_total = n_train + n_valid + n_test |
195
|
|
|
splits = create_splits(n_train, n_valid, n_test) |
196
|
|
|
hdf5_file.attrs['split'] = H5PYDataset.create_split_array(splits) |
197
|
|
|
vlen_dtype = h5py.special_dtype(vlen=numpy.dtype('uint8')) |
198
|
|
|
hdf5_file.create_dataset('encoded_images', shape=(n_total,), |
199
|
|
|
dtype=vlen_dtype) |
200
|
|
|
hdf5_file.create_dataset('targets', shape=(n_total, 1), dtype=numpy.int16) |
201
|
|
|
hdf5_file.create_dataset('filenames', shape=(n_total, 1), dtype='S32') |
202
|
|
|
|
203
|
|
|
|
204
|
|
|
def process_train_set(hdf5_file, train_archive, patch_archive, n_train, |
205
|
|
|
wnid_map, shuffle_seed=None): |
206
|
|
|
"""Process the ILSVRC2010 training set. |
207
|
|
|
|
208
|
|
|
Parameters |
209
|
|
|
---------- |
210
|
|
|
hdf5_file : :class:`h5py.File` instance |
211
|
|
|
HDF5 file handle to which to write. Assumes `features`, `targets` |
212
|
|
|
and `filenames` already exist and have first dimension larger than |
213
|
|
|
`n_train`. |
214
|
|
|
train_archive : str or file-like object |
215
|
|
|
Filename or file handle for the TAR archive of training images. |
216
|
|
|
patch_archive : str or file-like object |
217
|
|
|
Filename or file handle for the TAR archive of patch images. |
218
|
|
|
n_train : int |
219
|
|
|
The number of items in the training set. |
220
|
|
|
wnid_map : dict |
221
|
|
|
A dictionary mapping WordNet IDs to class indices. |
222
|
|
|
shuffle_seed : int or sequence, optional |
223
|
|
|
Seed for a NumPy random number generator that permutes the |
224
|
|
|
training set on disk. If `None`, no permutation is performed |
225
|
|
|
(this is the default). |
226
|
|
|
|
227
|
|
|
""" |
228
|
|
|
producer = partial(train_set_producer, train_archive=train_archive, |
229
|
|
|
patch_archive=patch_archive, wnid_map=wnid_map) |
230
|
|
|
consumer = partial(image_consumer, hdf5_file=hdf5_file, |
231
|
|
|
num_expected=n_train, shuffle_seed=shuffle_seed) |
232
|
|
|
producer_consumer(producer, consumer) |
233
|
|
|
|
234
|
|
|
|
235
|
|
|
def _write_to_hdf5(hdf5_file, index, image_filename, image_data, |
236
|
|
|
class_index): |
237
|
|
|
hdf5_file['filenames'][index] = image_filename.encode('ascii') |
238
|
|
|
hdf5_file['encoded_images'][index] = image_data |
239
|
|
|
hdf5_file['targets'][index] = class_index |
240
|
|
|
|
241
|
|
|
|
242
|
|
|
def train_set_producer(socket, train_archive, patch_archive, wnid_map): |
243
|
|
|
"""Load/send images from the training set TAR file or patch images. |
244
|
|
|
|
245
|
|
|
Parameters |
246
|
|
|
---------- |
247
|
|
|
socket : :class:`zmq.Socket` |
248
|
|
|
PUSH socket on which to send loaded images. |
249
|
|
|
train_archive : str or file-like object |
250
|
|
|
Filename or file handle for the TAR archive of training images. |
251
|
|
|
patch_archive : str or file-like object |
252
|
|
|
Filename or file handle for the TAR archive of patch images. |
253
|
|
|
wnid_map : dict |
254
|
|
|
A dictionary that maps WordNet IDs to 0-based class indices. |
255
|
|
|
Used to decode the filenames of the inner TAR files. |
256
|
|
|
|
257
|
|
|
""" |
258
|
|
|
patch_images = extract_patch_images(patch_archive, 'train') |
259
|
|
|
num_patched = 0 |
260
|
|
|
with tar_open(train_archive) as tar: |
261
|
|
|
for inner_tar_info in tar: |
262
|
|
|
with tar_open(tar.extractfile(inner_tar_info.name)) as inner: |
263
|
|
|
wnid = inner_tar_info.name.split('.')[0] |
264
|
|
|
class_index = wnid_map[wnid] |
265
|
|
|
filenames = sorted(info.name for info in inner |
266
|
|
|
if info.isfile()) |
267
|
|
|
images_gen = (load_from_tar_or_patch(inner, filename, |
268
|
|
|
patch_images) |
269
|
|
|
for filename in filenames) |
270
|
|
|
pathless_filenames = (os.path.split(fn)[-1] |
271
|
|
|
for fn in filenames) |
272
|
|
|
stream = equizip(pathless_filenames, images_gen) |
273
|
|
|
for image_fn, (image_data, patched) in stream: |
274
|
|
|
if patched: |
275
|
|
|
num_patched += 1 |
276
|
|
|
socket.send_pyobj((image_fn, class_index), zmq.SNDMORE) |
277
|
|
|
socket.send(image_data) |
278
|
|
|
if num_patched != len(patch_images): |
279
|
|
|
raise ValueError('not all patch images were used') |
280
|
|
|
|
281
|
|
|
|
282
|
|
|
def image_consumer(socket, hdf5_file, num_expected, shuffle_seed=None, |
283
|
|
|
offset=0): |
284
|
|
|
"""Fill an HDF5 file with incoming images from a socket. |
285
|
|
|
|
286
|
|
|
Parameters |
287
|
|
|
---------- |
288
|
|
|
socket : :class:`zmq.Socket` |
289
|
|
|
PULL socket on which to receive images. |
290
|
|
|
hdf5_file : :class:`h5py.File` instance |
291
|
|
|
HDF5 file handle to which to write. Assumes `features`, `targets` |
292
|
|
|
and `filenames` already exist and have first dimension larger than |
293
|
|
|
`sum(images_per_class)`. |
294
|
|
|
num_expected : int |
295
|
|
|
The number of items we expect to be sent over the socket. |
296
|
|
|
shuffle_seed : int or sequence, optional |
297
|
|
|
Seed for a NumPy random number generator that permutes the |
298
|
|
|
images on disk. |
299
|
|
|
offset : int, optional |
300
|
|
|
The offset in the HDF5 datasets at which to start writing |
301
|
|
|
received examples. Defaults to 0. |
302
|
|
|
|
303
|
|
|
""" |
304
|
|
|
with progress_bar('images', maxval=num_expected) as pb: |
305
|
|
|
if shuffle_seed is None: |
306
|
|
|
index_gen = iter(xrange(num_expected)) |
307
|
|
|
else: |
308
|
|
|
rng = numpy.random.RandomState(shuffle_seed) |
309
|
|
|
index_gen = iter(rng.permutation(num_expected)) |
310
|
|
|
for i, num in enumerate(index_gen): |
311
|
|
|
image_filename, class_index = socket.recv_pyobj(zmq.SNDMORE) |
312
|
|
|
image_data = numpy.fromstring(socket.recv(), dtype='uint8') |
313
|
|
|
_write_to_hdf5(hdf5_file, num + offset, image_filename, |
314
|
|
|
image_data, class_index) |
315
|
|
|
pb.update(i + 1) |
316
|
|
|
|
317
|
|
|
|
318
|
|
|
def process_other_set(hdf5_file, which_set, image_archive, patch_archive, |
319
|
|
|
groundtruth, offset): |
320
|
|
|
"""Process the validation or test set. |
321
|
|
|
|
322
|
|
|
Parameters |
323
|
|
|
---------- |
324
|
|
|
hdf5_file : :class:`h5py.File` instance |
325
|
|
|
HDF5 file handle to which to write. Assumes `features`, `targets` |
326
|
|
|
and `filenames` already exist and have first dimension larger than |
327
|
|
|
`sum(images_per_class)`. |
328
|
|
|
which_set : str |
329
|
|
|
Which set of images is being processed. One of 'train', 'valid', |
330
|
|
|
'test'. Used for extracting the appropriate images from the patch |
331
|
|
|
archive. |
332
|
|
|
image_archive : str or file-like object |
333
|
|
|
The filename or file-handle for the TAR archive containing images. |
334
|
|
|
patch_archive : str or file-like object |
335
|
|
|
Filename or file handle for the TAR archive of patch images. |
336
|
|
|
groundtruth : iterable |
337
|
|
|
Iterable container containing scalar 0-based class index for each |
338
|
|
|
image, sorted by filename. |
339
|
|
|
offset : int |
340
|
|
|
The offset in the HDF5 datasets at which to start writing. |
341
|
|
|
|
342
|
|
|
""" |
343
|
|
|
producer = partial(other_set_producer, image_archive=image_archive, |
344
|
|
|
patch_archive=patch_archive, |
345
|
|
|
groundtruth=groundtruth, which_set=which_set) |
346
|
|
|
consumer = partial(image_consumer, hdf5_file=hdf5_file, |
347
|
|
|
num_expected=len(groundtruth), offset=offset) |
348
|
|
|
producer_consumer(producer, consumer) |
349
|
|
|
|
350
|
|
|
|
351
|
|
|
def other_set_producer(socket, which_set, image_archive, patch_archive, |
352
|
|
|
groundtruth): |
353
|
|
|
"""Push image files read from the valid/test set TAR to a socket. |
354
|
|
|
|
355
|
|
|
Parameters |
356
|
|
|
---------- |
357
|
|
|
socket : :class:`zmq.Socket` |
358
|
|
|
PUSH socket on which to send images. |
359
|
|
|
which_set : str |
360
|
|
|
Which set of images is being processed. One of 'train', 'valid', |
361
|
|
|
'test'. Used for extracting the appropriate images from the patch |
362
|
|
|
archive. |
363
|
|
|
image_archive : str or file-like object |
364
|
|
|
The filename or file-handle for the TAR archive containing images. |
365
|
|
|
patch_archive : str or file-like object |
366
|
|
|
Filename or file handle for the TAR archive of patch images. |
367
|
|
|
groundtruth : iterable |
368
|
|
|
Iterable container containing scalar 0-based class index for each |
369
|
|
|
image, sorted by filename. |
370
|
|
|
|
371
|
|
|
""" |
372
|
|
|
patch_images = extract_patch_images(patch_archive, which_set) |
373
|
|
|
num_patched = 0 |
374
|
|
|
with tar_open(image_archive) as tar: |
375
|
|
|
filenames = sorted(info.name for info in tar if info.isfile()) |
376
|
|
|
images = (load_from_tar_or_patch(tar, filename, patch_images) |
377
|
|
|
for filename in filenames) |
378
|
|
|
pathless_filenames = (os.path.split(fn)[-1] for fn in filenames) |
379
|
|
|
image_iterator = equizip(images, pathless_filenames, groundtruth) |
380
|
|
|
for (image_data, patched), filename, class_index in image_iterator: |
381
|
|
|
if patched: |
382
|
|
|
num_patched += 1 |
383
|
|
|
socket.send_pyobj((filename, class_index), zmq.SNDMORE) |
384
|
|
|
socket.send(image_data, copy=False) |
385
|
|
|
if num_patched != len(patch_images): |
386
|
|
|
raise Exception |
387
|
|
|
|
388
|
|
|
|
389
|
|
|
def load_from_tar_or_patch(tar, image_filename, patch_images): |
390
|
|
|
"""Do everything necessary to process an image inside a TAR. |
391
|
|
|
|
392
|
|
|
Parameters |
393
|
|
|
---------- |
394
|
|
|
tar : `TarFile` instance |
395
|
|
|
The tar from which to read `image_filename`. |
396
|
|
|
image_filename : str |
397
|
|
|
Fully-qualified path inside of `tar` from which to read an |
398
|
|
|
image file. |
399
|
|
|
patch_images : dict |
400
|
|
|
A dictionary containing filenames (without path) of replacements |
401
|
|
|
to be substituted in place of the version of the same file found |
402
|
|
|
in `tar`. |
403
|
|
|
|
404
|
|
|
Returns |
405
|
|
|
------- |
406
|
|
|
image_data : bytes |
407
|
|
|
The JPEG bytes representing either the image from the TAR archive |
408
|
|
|
or its replacement from the patch dictionary. |
409
|
|
|
patched : bool |
410
|
|
|
True if the image was retrieved from the patch dictionary. False |
411
|
|
|
if it was retrieved from the TAR file. |
412
|
|
|
|
413
|
|
|
""" |
414
|
|
|
patched = True |
415
|
|
|
image_bytes = patch_images.get(os.path.basename(image_filename), None) |
416
|
|
|
if image_bytes is None: |
417
|
|
|
patched = False |
418
|
|
|
try: |
419
|
|
|
image_bytes = tar.extractfile(image_filename).read() |
420
|
|
|
numpy.array(Image.open(io.BytesIO(image_bytes))) |
421
|
|
|
except (IOError, OSError): |
422
|
|
|
with gzip.GzipFile(fileobj=tar.extractfile(image_filename)) as gz: |
423
|
|
|
image_bytes = gz.read() |
424
|
|
|
numpy.array(Image.open(io.BytesIO(image_bytes))) |
425
|
|
|
return image_bytes, patched |
426
|
|
|
|
427
|
|
|
|
428
|
|
|
def read_devkit(f): |
429
|
|
|
"""Read relevant information from the development kit archive. |
430
|
|
|
|
431
|
|
|
Parameters |
432
|
|
|
---------- |
433
|
|
|
f : str or file-like object |
434
|
|
|
The filename or file-handle for the gzipped TAR archive |
435
|
|
|
containing the ILSVRC2010 development kit. |
436
|
|
|
|
437
|
|
|
Returns |
438
|
|
|
------- |
439
|
|
|
synsets : ndarray, 1-dimensional, compound dtype |
440
|
|
|
See :func:`read_metadata_mat_file` for details. |
441
|
|
|
cost_matrix : ndarray, 2-dimensional, uint8 |
442
|
|
|
See :func:`read_metadata_mat_file` for details. |
443
|
|
|
raw_valid_groundtruth : ndarray, 1-dimensional, int16 |
444
|
|
|
The labels for the ILSVRC2010 validation set, |
445
|
|
|
distributed with the development kit code. |
446
|
|
|
|
447
|
|
|
""" |
448
|
|
|
with tar_open(f) as tar: |
449
|
|
|
# Metadata table containing class hierarchy, textual descriptions, etc. |
450
|
|
|
meta_mat = tar.extractfile(DEVKIT_META_PATH) |
451
|
|
|
synsets, cost_matrix = read_metadata_mat_file(meta_mat) |
452
|
|
|
|
453
|
|
|
# Raw validation data groundtruth, ILSVRC2010 IDs. Confusingly |
454
|
|
|
# distributed inside the development kit archive. |
455
|
|
|
raw_valid_groundtruth = numpy.loadtxt(tar.extractfile( |
456
|
|
|
DEVKIT_VALID_GROUNDTRUTH_PATH), dtype=numpy.int16) |
457
|
|
|
return synsets, cost_matrix, raw_valid_groundtruth |
458
|
|
|
|
459
|
|
|
|
460
|
|
|
def read_metadata_mat_file(meta_mat): |
461
|
|
|
"""Read ILSVRC2010 metadata from the distributed MAT file. |
462
|
|
|
|
463
|
|
|
Parameters |
464
|
|
|
---------- |
465
|
|
|
meta_mat : str or file-like object |
466
|
|
|
The filename or file-handle for `meta.mat` from the |
467
|
|
|
ILSVRC2010 development kit. |
468
|
|
|
|
469
|
|
|
Returns |
470
|
|
|
------- |
471
|
|
|
synsets : ndarray, 1-dimensional, compound dtype |
472
|
|
|
A table containing ILSVRC2010 metadata for the "synonym sets" |
473
|
|
|
or "synsets" that comprise the classes and superclasses, |
474
|
|
|
including the following fields: |
475
|
|
|
* `ILSVRC2010_ID`: the integer ID used in the original |
476
|
|
|
competition data. |
477
|
|
|
* `WNID`: A string identifier that uniquely identifies |
478
|
|
|
a synset in ImageNet and WordNet. |
479
|
|
|
* `wordnet_height`: The length of the longest path to |
480
|
|
|
a leaf node in the FULL ImageNet/WordNet hierarchy |
481
|
|
|
(leaf nodes in the FULL ImageNet/WordNet hierarchy |
482
|
|
|
have `wordnet_height` 0). |
483
|
|
|
* `gloss`: A string representation of an English |
484
|
|
|
textual description of the concept represented by |
485
|
|
|
this synset. |
486
|
|
|
* `num_children`: The number of children in the hierarchy |
487
|
|
|
for this synset. |
488
|
|
|
* `words`: A string representation, comma separated, |
489
|
|
|
of different synoym words or phrases for the concept |
490
|
|
|
represented by this synset. |
491
|
|
|
* `children`: A vector of `ILSVRC2010_ID`s of children |
492
|
|
|
of this synset, padded with -1. Note that these refer |
493
|
|
|
to `ILSVRC2010_ID`s from the original data and *not* |
494
|
|
|
the zero-based index in the table. |
495
|
|
|
* `num_train_images`: The number of training images for |
496
|
|
|
this synset. |
497
|
|
|
cost_matrix : ndarray, 2-dimensional, uint8 |
498
|
|
|
A 1000x1000 matrix containing the precomputed pairwise |
499
|
|
|
cost (based on distance in the hierarchy) for all |
500
|
|
|
low-level synsets (i.e. the thousand possible output |
501
|
|
|
classes with training data associated). |
502
|
|
|
|
503
|
|
|
""" |
504
|
|
|
mat = loadmat(meta_mat, squeeze_me=True) |
505
|
|
|
synsets = mat['synsets'] |
506
|
|
|
cost_matrix = mat['cost_matrix'] |
507
|
|
|
new_dtype = numpy.dtype([ |
508
|
|
|
('ILSVRC2010_ID', numpy.int16), |
509
|
|
|
('WNID', ('S', max(map(len, synsets['WNID'])))), |
510
|
|
|
('wordnet_height', numpy.int8), |
511
|
|
|
('gloss', ('S', max(map(len, synsets['gloss'])))), |
512
|
|
|
('num_children', numpy.int8), |
513
|
|
|
('words', ('S', max(map(len, synsets['words'])))), |
514
|
|
|
('children', (numpy.int8, max(synsets['num_children']))), |
515
|
|
|
('num_train_images', numpy.uint16) |
516
|
|
|
]) |
517
|
|
|
new_synsets = numpy.empty(synsets.shape, dtype=new_dtype) |
518
|
|
|
for attr in ['ILSVRC2010_ID', 'WNID', 'wordnet_height', 'gloss', |
519
|
|
|
'num_children', 'words', 'num_train_images']: |
520
|
|
|
new_synsets[attr] = synsets[attr] |
521
|
|
|
children = [numpy.atleast_1d(ch) for ch in synsets['children']] |
522
|
|
|
padded_children = [ |
523
|
|
|
numpy.concatenate((c, |
524
|
|
|
-numpy.ones(new_dtype['children'].shape[0] - len(c), |
525
|
|
|
dtype=numpy.int16))) |
526
|
|
|
for c in children |
527
|
|
|
] |
528
|
|
|
new_synsets['children'] = padded_children |
529
|
|
|
return new_synsets, cost_matrix |
530
|
|
|
|
531
|
|
|
|
532
|
|
|
def extract_patch_images(f, which_set): |
533
|
|
|
"""Extracts a dict of the "patch images" for ILSVRC2010. |
534
|
|
|
|
535
|
|
|
Parameters |
536
|
|
|
---------- |
537
|
|
|
f : str or file-like object |
538
|
|
|
The filename or file-handle to the patch images TAR file. |
539
|
|
|
which_set : str |
540
|
|
|
Which set of images to extract. One of 'train', 'valid', 'test'. |
541
|
|
|
|
542
|
|
|
Returns |
543
|
|
|
------- |
544
|
|
|
dict |
545
|
|
|
A dictionary contains a mapping of filenames (without path) to a |
546
|
|
|
bytes object containing the replacement image. |
547
|
|
|
|
548
|
|
|
Notes |
549
|
|
|
----- |
550
|
|
|
Certain images in the distributed archives are blank, or display |
551
|
|
|
an "image not available" banner. A separate TAR file of |
552
|
|
|
"patch images" is distributed with the corrected versions of |
553
|
|
|
these. It is this archive that this function is intended to read. |
554
|
|
|
|
555
|
|
|
""" |
556
|
|
|
if which_set not in ('train', 'valid', 'test'): |
557
|
|
|
raise ValueError('which_set must be one of train, valid, or test') |
558
|
|
|
which_set = 'val' if which_set == 'valid' else which_set |
559
|
|
|
patch_images = {} |
560
|
|
|
with tar_open(f) as tar: |
561
|
|
|
for info_obj in tar: |
562
|
|
|
if not info_obj.name.endswith('.JPEG'): |
563
|
|
|
continue |
564
|
|
|
# Pretty sure that '/' is used for tarfile regardless of |
565
|
|
|
# os.path.sep, but I officially don't care about Windows. |
566
|
|
|
tokens = info_obj.name.split('/') |
567
|
|
|
file_which_set = tokens[-2] |
568
|
|
|
if file_which_set != which_set: |
569
|
|
|
continue |
570
|
|
|
filename = tokens[-1] |
571
|
|
|
patch_images[filename] = tar.extractfile(info_obj.name).read() |
572
|
|
|
return patch_images |
573
|
|
|
|
It is generally discouraged to redefine built-ins as this makes code very hard to read.