def load_svhn(zero_centre=False, greyscale=False, val=False, extra=False):
#
#
# Load SVHN
#
#
print('Loading SVHN...')
if val:
d_svhn = svhn.SVHN(n_val=10000)
else:
d_svhn = svhn.SVHN(n_val=0)
if extra:
d_extra = svhn.SVHNExtra()
else:
d_extra = None
d_svhn.train_X = d_svhn.train_X[:]
d_svhn.val_X = d_svhn.val_X[:]
d_svhn.test_X = d_svhn.test_X[:]
d_svhn.train_y = d_svhn.train_y[:]
d_svhn.val_y = d_svhn.val_y[:]
d_svhn.test_y = d_svhn.test_y[:]
if extra:
d_svhn.train_X = np.append(d_svhn.train_X, d_extra.X[:], axis=0)
d_svhn.train_y = np.append(d_svhn.train_y, d_extra.y[:], axis=0)
if greyscale:
d_svhn.train_X = rgb2grey_tensor(d_svhn.train_X)
d_svhn.val_X = rgb2grey_tensor(d_svhn.val_X)
d_svhn.test_X = rgb2grey_tensor(d_svhn.test_X)
if zero_centre:
d_svhn.train_X = d_svhn.train_X * 2.0 - 1.0
d_svhn.val_X = d_svhn.val_X * 2.0 - 1.0
d_svhn.test_X = d_svhn.test_X * 2.0 - 1.0
print(
'SVHN: train: X.shape={}, y.shape={}, val: X.shape={}, y.shape={}, test: X.shape={}, y.shape={}'
.format(d_svhn.train_X.shape, d_svhn.train_y.shape, d_svhn.val_X.shape,
d_svhn.val_y.shape, d_svhn.test_X.shape, d_svhn.test_y.shape))
print('SVHN: train: X.min={}, X.max={}'.format(d_svhn.train_X.min(),
d_svhn.train_X.max()))
d_svhn.n_classes = 10
return d_svhn
def test_train_test_split():
from batchup.datasets import svhn
from batchup.tests.dataset_test_helpers import sample_hashes
ds = svhn.SVHN(n_val=0)
train_h = sample_hashes(ds.train_X_u8)
test_h = sample_hashes(ds.test_X_u8)
assert set(train_h).intersection(set(test_h)) == set()
def test_val_0():
from batchup.datasets import svhn
ds = svhn.SVHN(n_val=0)
assert ds.train_X.shape == (73257, 3, 32, 32)
assert ds.train_X.dtype == np.float32
assert ds.train_y.shape == (73257, )
assert ds.train_y.dtype == np.int32
assert ds.val_X.shape == (0, 3, 32, 32)
assert ds.val_X.dtype == np.float32
assert ds.val_y.shape == (0, )
assert ds.val_y.dtype == np.int32
assert ds.test_X.shape == (26032, 3, 32, 32)
assert ds.test_X.dtype == np.float32
assert ds.test_y.shape == (26032, )
assert ds.test_y.dtype == np.int32