def russ_features(which_set, restrict_instances):
raw = FoveatedNORB(which_set=which_set,
restrict_instances=None,
one_hot=True)
X = np.load("/data/lisa/data/norb_small/ruslan_binarized/%s_X.npy" %
which_set)
assert raw.X.shape[0] == X.shape[0], (raw.X.shape[0], X.shape[0])
raw.X = X
y = np.load("/data/lisa/data/norb_small/ruslan_binarized/%s_Y.npy" %
which_set)
print y[0:5, :]
print raw.y[0:5, :]
assert np.allclose(y, raw.y)
if restrict_instances is not None:
raw.restrict_instances(restrict_instances)
rval = BinarizerDataset(raw)
rval.X = raw.X
return rval
def setUp(self):
try:
self.train = FoveatedNORB(which_set='train')
self.test = FoveatedNORB(which_set='test')
self.train_one_hot = self.train.convert_to_one_hot()
self.test_one_hot = self.test.convert_to_one_hot()
except (NoDataPathError, NotInstalledError):
raise SkipTest()
def test_foveated_norb(self):
# Test that the FoveatedNORB class can be instantiated
norb_train = FoveatedNORB(which_set="train",
scale=1,
restrict_instances=[4, 6, 7, 8],
one_hot=1)
class TestFoveatedNORB(unittest.TestCase):
def setUp(self):
try:
self.train = FoveatedNORB(which_set='train')
self.test = FoveatedNORB(which_set='test')
self.train_one_hot = self.train.convert_to_one_hot()
self.test_one_hot = self.test.convert_to_one_hot()
except (NoDataPathError, NotInstalledError):
raise SkipTest()
def test_one_hot(self):
train_non_zeros = numpy.transpose(numpy.nonzeros(self.train_one_hot))
test_non_zeros = numpy.transpose(numpy.nonzeros(self.test_one_hot))
self.assertEquals(len(train_non_zeros), 1)
self.assertEquals(len(test_non_zeros), 1)
self.assertEquals(self.train_one_hot[train_non_zeros[0]], 1)
self.assertEquals(self.test_one_hot[test_non_zeros[0]], 1)
def test_restrict_instances(self):
pass
def russ_features(which_set, restrict_instances):
raw = FoveatedNORB(which_set=which_set, restrict_instances=None, one_hot=True)
X = np.load("/data/lisa/data/norb_small/ruslan_binarized/%s_X.npy" % which_set)
assert raw.X.shape[0] == X.shape[0], (raw.X.shape[0], X.shape[0])
raw.X = X
y = np.load("/data/lisa/data/norb_small/ruslan_binarized/%s_Y.npy" % which_set)
print y[0:5, :]
print raw.y[0:5, :]
assert np.allclose(y, raw.y)
if restrict_instances is not None:
raw.restrict_instances(restrict_instances)
rval = BinarizerDataset(raw)
rval.X = raw.X
return rval
__author__ = "Ian Goodfellow"
"""
A script for sequentially stepping through FoveatedNORB, viewing each image
and its label.
"""
import numpy as np
from pylearn2.datasets.norb_small import FoveatedNORB
from pylearn2.gui.patch_viewer import PatchViewer
from pylearn2.utils import get_choice
print('Use test set?')
choices = {'y': 'test', 'n': 'train'}
which_set = choices[get_choice(choices)]
dataset = FoveatedNORB(which_set=which_set, center=True)
topo = dataset.get_topological_view()
b, r, c, ch = topo.shape
assert ch == 2
pv = PatchViewer((1, 2), (r, c), is_color=False)
i = 0
while True:
patch = topo[i, :, :, :]
patch = patch / np.abs(patch).max()
pv.add_patch(patch[:,:,1], rescale=False)
def test_FoveatedNORB():
"""
This function tests the FoveatedNORB class. In addition to the shape and
datatype of X and y member of the returned object, it also checks the
scale of data while passing different parameters to the constructor.
"""
skip_if_no_data()
data = FoveatedNORB('train')
datamin = data.X.min()
datamax = data.X.max()
assert data.X.shape == (24300, 8976)
assert data.X.dtype == 'float32'
assert data.y.shape == (24300, )
assert data.y_labels == 5
assert data.get_topological_view().shape == (24300, 96, 96, 2)
data = FoveatedNORB('train', center=True)
assert data.X.min() == datamin - 127.5
assert data.X.max() == datamax - 127.5
data = FoveatedNORB('train', center=True, scale=True)
assert numpy.all(data.X <= 1.)
assert numpy.all(data.X >= -1.)
data = FoveatedNORB('train', scale=True)
assert numpy.all(data.X <= 1.)
assert numpy.all(data.X >= 0.)
data = FoveatedNORB('test')
assert data.X.shape == (24300, 8976)
assert data.X.dtype == 'float32'
assert data.y.shape == (24300, )
assert data.y_labels == 5
assert data.get_topological_view().shape == (24300, 96, 96, 2)
A script for sequentially stepping through FoveatedNORB, viewing each image
and its label.
"""
import numpy as np
from six.moves import input
from pylearn2.datasets.norb_small import FoveatedNORB
from pylearn2.gui.patch_viewer import PatchViewer
from pylearn2.utils import get_choice
print('Use test set?')
choices = {'y': 'test', 'n': 'train'}
which_set = choices[get_choice(choices)]
dataset = FoveatedNORB(which_set=which_set, center=True)
topo = dataset.get_topological_view()
b, r, c, ch = topo.shape
assert ch == 2
pv = PatchViewer((1, 2), (r, c), is_color=False)
i = 0
while True:
patch = topo[i, :, :, :]
patch = patch / np.abs(patch).max()
pv.add_patch(patch[:, :, 1], rescale=False)
from pylearn2.datasets.norb_small import FoveatedNORB
dataset = FoveatedNORB(which_set='train')
from pylearn2.datasets.preprocessing import Standardize
standardize = Standardize(global_mean=True, global_std=True)
standardize.apply(dataset, can_fit=True)
from pylearn2.utils import serial
serial.save("norb_prepro_global.pkl", standardize)
