def fetch(self, download_if_missing):
with lockfile.FileLock(self.home()) as lock:
try:
open(self.home(self.NAME + '_inputs.npy')).close()
open(self.home(self.NAME + '_labels.npy')).close()
except IOError:
if download_if_missing:
try:
amat_test = AMat(self.test_amat())
except IOError:
logger.info('Failed to read %s, downloading %s' % (
self.test_amat(),
os.path.join(self.BASE_URL, self.REMOTE)))
if not os.path.exists(self.home()):
os.makedirs(self.home())
utils.download_and_extract(
os.path.join(self.BASE_URL, self.REMOTE),
self.home(),
verbose=False,
sha1=self.SHA1)
amat_test = AMat(self.test_amat())
amat_train = AMat(self.train_amat())
n_inputs = 28**2
n_train = self.descr['n_train']
n_valid = self.descr['n_valid']
n_test = self.descr['n_test']
assert amat_train.all.shape[0] == n_train + n_valid
assert amat_test.all.shape[0] == n_test
assert amat_train.all.shape[1] == amat_test.all.shape[1]
assert amat_train.all.shape[1] == n_inputs + 1
allmat = np.vstack((amat_train.all, amat_test.all))
inputs = np.reshape(
allmat[:, :n_inputs].astype('float32'),
(-1, 28, 28))
labels = allmat[:, n_inputs].astype('int32')
assert np.all(labels == allmat[:, n_inputs])
assert np.all(labels < self.descr['n_classes'])
np.save(self.home(self.NAME + '_inputs.npy'), inputs)
np.save(self.home(self.NAME + '_labels.npy'), labels)
# clean up the .amat files we downloaded
os.remove(self.test_amat())
os.remove(self.train_amat())
else:
raise
def fetch(self, download_if_missing):
with lockfile.FileLock(self.home()) as lock:
try:
open(self.home(self.NAME + '_inputs.npy')).close()
open(self.home(self.NAME + '_labels.npy')).close()
except IOError:
if download_if_missing:
try:
amat_test = AMat(self.test_amat())
except IOError:
logger.info('Failed to read %s, downloading %s' %
(self.test_amat(),
os.path.join(self.BASE_URL, self.REMOTE)))
if not os.path.exists(self.home()):
os.makedirs(self.home())
utils.download_and_extract(os.path.join(
self.BASE_URL, self.REMOTE),
self.home(),
verbose=False,
sha1=self.SHA1)
amat_test = AMat(self.test_amat())
amat_train = AMat(self.train_amat())
n_inputs = 28**2
n_train = self.descr['n_train']
n_valid = self.descr['n_valid']
n_test = self.descr['n_test']
assert amat_train.all.shape[0] == n_train + n_valid
assert amat_test.all.shape[0] == n_test
assert amat_train.all.shape[1] == amat_test.all.shape[1]
assert amat_train.all.shape[1] == n_inputs + 1
allmat = np.vstack((amat_train.all, amat_test.all))
inputs = np.reshape(allmat[:, :n_inputs].astype('float32'),
(-1, 28, 28))
labels = allmat[:, n_inputs].astype('int32')
assert np.all(labels == allmat[:, n_inputs])
assert np.all(labels < self.descr['n_classes'])
np.save(self.home(self.NAME + '_inputs.npy'), inputs)
np.save(self.home(self.NAME + '_labels.npy'), labels)
# clean up the .amat files we downloaded
os.remove(self.test_amat())
os.remove(self.train_amat())
else:
raise
def fetch(self, download_if_missing):
try:
open(self.home(self.NAME + '_inputs.npy')).close()
open(self.home(self.NAME + '_labels.npy')).close()
except IOError:
if download_if_missing:
all_amat_filename = self.level_amat(self.LEVEL)
try:
amat_all = AMat(all_amat_filename)
except IOError:
logger.info('Failed to read %s, downloading %s' %
(all_amat_filename,
os.path.join(self.BASE_URL, self.REMOTE)))
if not os.path.exists(self.home()):
os.makedirs(self.home())
utils.download_and_extract(os.path.join(
self.BASE_URL, self.REMOTE),
self.home(),
verbose=False,
sha1=self.SHA1)
amat_all = AMat(all_amat_filename)
# at this point self.home() contains not only the
# all_amat_filename, but it also contains the amats for all
# the other levels too.
#
# This for loop transfers each amat to the dataset folder where
# it belongs.
for level in range(1, 7):
if level == self.LEVEL:
continue
if not os.path.exists(self.level_amat(level)):
continue
other = BaseNoise(level)
try:
# try loading the other one
other.fetch(download_if_missing=False)
# if that worked, then delete just-downloaded amat
# required for the other's build_meta
os.remove(self.level_amat(level))
except IOError:
# assuming this was because fetch failed,
# move the amat for the other level into the
# home folder of the other dataset.
# next time the other dataset is fetched,
# it will load the amat, save a npy, and delete the
# amat.
if not os.path.exists(other.home()):
os.makedirs(other.home())
os.rename(self.level_amat(level),
other.level_amat(level))
# now carry on loading as usual
n_inputs = 28**2
n_train = self.descr['n_train']
n_valid = self.descr['n_valid']
n_test = self.descr['n_test']
assert amat_all.all.shape[0] == n_train + n_valid + n_test
assert amat_all.all.shape[1] == n_inputs + 1
inputs = np.reshape(
amat_all.all[:, :n_inputs].astype('float32'), (-1, 28, 28))
labels = amat_all.all[:, n_inputs].astype('int32')
assert np.all(labels == amat_all.all[:, n_inputs])
assert np.all(labels < self.descr['n_classes'])
np.save(self.home(self.NAME + '_inputs.npy'), inputs)
np.save(self.home(self.NAME + '_labels.npy'), labels)
# clean up the .amat files we downloaded
os.remove(all_amat_filename)
else:
raise
def fetch(self, download_if_missing):
try:
open(self.home(self.NAME + '_inputs.npy')).close()
open(self.home(self.NAME + '_labels.npy')).close()
except IOError:
if download_if_missing:
all_amat_filename = self.level_amat(self.LEVEL)
try:
amat_all = AMat(all_amat_filename)
except IOError:
logger.info('Failed to read %s, downloading %s' % (
all_amat_filename,
os.path.join(self.BASE_URL, self.REMOTE)))
if not os.path.exists(self.home()):
os.makedirs(self.home())
utils.download_and_extract(
os.path.join(self.BASE_URL, self.REMOTE),
self.home(),
verbose=False,
sha1=self.SHA1)
amat_all = AMat(all_amat_filename)
# at this point self.home() contains not only the
# all_amat_filename, but it also contains the amats for all
# the other levels too.
#
# This for loop transfers each amat to the dataset folder where
# it belongs.
for level in range(1, 7):
if level == self.LEVEL:
continue
if not os.path.exists(self.level_amat(level)):
continue
other = BaseNoise(level)
try:
# try loading the other one
other.fetch(download_if_missing=False)
# if that worked, then delete just-downloaded amat
# required for the other's build_meta
os.remove(self.level_amat(level))
except IOError:
# assuming this was because fetch failed,
# move the amat for the other level into the
# home folder of the other dataset.
# next time the other dataset is fetched,
# it will load the amat, save a npy, and delete the
# amat.
if not os.path.exists(other.home()):
os.makedirs(other.home())
os.rename(
self.level_amat(level),
other.level_amat(level))
# now carry on loading as usual
n_inputs = 28**2
n_train = self.descr['n_train']
n_valid = self.descr['n_valid']
n_test = self.descr['n_test']
assert amat_all.all.shape[0] == n_train + n_valid + n_test
assert amat_all.all.shape[1] == n_inputs + 1
inputs = np.reshape(
amat_all.all[:, :n_inputs].astype('float32'),
(-1, 28, 28))
labels = amat_all.all[:, n_inputs].astype('int32')
assert np.all(labels == amat_all.all[:, n_inputs])
assert np.all(labels < self.descr['n_classes'])
np.save(self.home(self.NAME + '_inputs.npy'), inputs)
np.save(self.home(self.NAME + '_labels.npy'), labels)
# clean up the .amat files we downloaded
os.remove(all_amat_filename)
else:
raise
def prepare_data(self):
if not os.path.exists(self.data_dir):
download_and_extract(self.url, self.data_dir)
