def __init__(self, root, mode, is_gray = False):
"""Loads the STL dataset. mode should be either 'train', 'test', or
'unlabeled'
"""
if mode == 'train':
self._data, self._label = \
STL10Dataset.get_data(os.path.join(root, 'train.mat'))
elif mode == 'test':
self._data, self._label = \
STL10Dataset.get_data(os.path.join(root, 'test.mat'))
elif mode == 'unlabeled':
# h5py allows us to directly read part of the matrix, so each
# node will work on his own
matdata = h5py.File(os.path.join(root, 'unlabeled.mat'),'r')
segments = mpi.get_segments(matdata['X'].shape[1])
# read
X = matdata['X'][:, segments[mpi.RANK]:segments[mpi.RANK+1]]
X.resize(STL10Dataset._image_dim[::-1] + (X.shape[1],))
self._data = np.ascontiguousarray(np.transpose(X))
self._label = None
else:
raise ValueError, "Unrecognized mode."
if is_gray:
self._data = self._data.mean(axis=-1)
self._dim = STL10Dataset._image_dim[:2]
self._channels = 1
else:
self._dim = STL10Dataset._image_dim
self._channels = STL10Dataset._num_channels
self._prefetch = True
def testGetSegments(self):
total = 100
segments, inv = mpi.get_segments(total, True)
self.assertEqual(len(segments), mpi.SIZE + 1)
self.assertEqual(segments[0], 0)
self.assertEqual(segments[-1], total)
self.assertEqual(len(inv), total)
for i in range(total):
self.assertGreaterEqual(i, segments[inv[i]])
self.assertLess(i, segments[inv[i] + 1])
def testGetSegments(self):
total = 100
segments, inv = mpi.get_segments(total, True)
self.assertEqual(len(segments), mpi.SIZE + 1)
self.assertEqual(segments[0], 0)
self.assertEqual(segments[-1], total)
self.assertEqual(len(inv), total)
for i in range(total):
self.assertGreaterEqual(i, segments[inv[i]])
self.assertLess(i, segments[inv[i] + 1])
def __init__(self, root, mode, is_gray = False, target_size = None):
"""Loads the STL dataset. mode should be either 'train', 'test', or
'unlabeled'
"""
if mode == 'train':
self._data, self._label = \
STL10Dataset.get_data(os.path.join(root, 'train.mat'))
elif mode == 'test':
self._data, self._label = \
STL10Dataset.get_data(os.path.join(root, 'test.mat'))
elif mode == 'unlabeled':
# h5py allows us to directly read part of the matrix, so each
# node will work on his own
matdata = h5py.File(os.path.join(root, 'unlabeled.mat'),'r')
segments = mpi.get_segments(matdata['X'].shape[1])
# read
X = matdata['X'][:, segments[mpi.RANK]:segments[mpi.RANK+1]]
X.resize(STL10Dataset._image_dim[::-1] + (X.shape[1],))
self._data = np.ascontiguousarray(np.transpose(X))
self._label = None
else:
raise ValueError, "Unrecognized mode."
if is_gray:
self._data = self._data.mean(axis=-1)
self._dim = STL10Dataset._image_dim[:2]
self._channels = 1
else:
self._dim = STL10Dataset._image_dim
self._channels = STL10Dataset._num_channels
if target_size is not None:
# we often want to resize the STL dataset to some other sizes
if type(target_size) is not int:
raise TypeError, "The input target_size should be an int!"
self._dim = (target_size, target_size)
old_data = self._data
new_size = np.asarray(self._data.shape)
new_size[1:3] = target_size
self._data = np.empty(new_size)
for i in range(self._data.shape[0]):
self._data[i] = skimage.transform.resize(old_data[i],
(target_size, target_size), mode='nearest')
self._prefetch = True
def __init__(self, root, mode, is_gray=False, target_size=None):
"""Loads the STL dataset. mode should be either 'train', 'test', or
'unlabeled'
"""
if mode == 'train':
self._data, self._label = \
STL10Dataset.get_data(os.path.join(root, 'train.mat'))
elif mode == 'test':
self._data, self._label = \
STL10Dataset.get_data(os.path.join(root, 'test.mat'))
elif mode == 'unlabeled':
# h5py allows us to directly read part of the matrix, so each
# node will work on his own
matdata = h5py.File(os.path.join(root, 'unlabeled.mat'), 'r')
segments = mpi.get_segments(matdata['X'].shape[1])
# read
X = matdata['X'][:, segments[mpi.RANK]:segments[mpi.RANK + 1]]
X.resize(STL10Dataset._image_dim[::-1] + (X.shape[1], ))
self._data = np.ascontiguousarray(np.transpose(X))
self._label = None
else:
raise ValueError, "Unrecognized mode."
if is_gray:
self._data = self._data.mean(axis=-1)
self._dim = STL10Dataset._image_dim[:2]
self._channels = 1
else:
self._dim = STL10Dataset._image_dim
self._channels = STL10Dataset._num_channels
if target_size is not None:
# we often want to resize the STL dataset to some other sizes
if type(target_size) is not int:
raise TypeError, "The input target_size should be an int!"
self._dim = (target_size, target_size)
old_data = self._data
new_size = np.asarray(self._data.shape)
new_size[1:3] = target_size
self._data = np.empty(new_size)
for i in range(self._data.shape[0]):
self._data[i] = skimage.transform.resize(
old_data[i], (target_size, target_size), mode='nearest')
self._prefetch = True
