def load_cifar10_class_subsets(classes,
train=True,
device="cpu",
dtype=torch.float,
minmax=(0, 1.0)):
"""
Load and return a subset of the CIFAR10 dataset, including only the
specified classes.
- classes: a list of class indices (as returned by label_class_list()
- minmax: transform the range of pixel intensity to satisfy this range.
"""
trdata = load_cifar10_dataset(train)
# ntr = trdata.train_data.shape[0]
# Pick out the specified classes
# numpy arrays
X = trdata.train_data
Y = np.array(trdata.train_labels)
# filter data according to the chosen classes
tr_inds = [Y[i] in classes for i in range(len(Y))]
Xtr = X[tr_inds]
Ytr = Y[tr_inds]
Xtr = util.linear_range_transform(Xtr, (0, 255), minmax)
Ytr = util.linear_range_transform(Ytr, (0, 255), minmax)
TXtr = torch.tensor(Xtr.transpose(0, 3, 1, 2), device=device, dtype=dtype)
TYtr = torch.tensor(Ytr, device=device, dtype=dtype)
Tr = torch.utils.data.TensorDataset(TXtr, TYtr)
return Tr
def test_linear_range_transform(self):
A = np.array([0.1, 0.8])
testing.assert_almost_equal(
np.array([1, 8]), util.linear_range_transform(A, (0, 1), (0, 10)))
testing.assert_almost_equal(
np.array([-2, 3]),
util.linear_range_transform(np.array([-0.2, 0.3]), (-1, 1),
(-10, 10)))
def forward(self, z):
out = self.l1(z)
out = out.view(out.shape[0], 128, self.init_size, self.init_size)
img = self.conv_blocks(out)
# tanh's range is [-1, 1]. Rescale the range to match what we want.
minmax = self.minmax
img = util.linear_range_transform(img, from_range=(-1, 1), to_range=minmax)
return img
def forward(self, z):
z = z.view(-1, self.latent_dim, 1, 1)
img = self.main(z)
# print('output size: {}'.format(img.shape))
# tanh's range is [-1, 1]. Rescale the range to match what we want.
minmax = self.minmax
img = util.linear_range_transform(img, from_range=(-1, 1), to_range=minmax)
# print('output size: {}'.format(img.shape))
return img
def forward(self, z):
out = self.net(z)
img = out.view(out.shape[0], self.channels, 32, 32)
# tanh's range is [-1, 1]. Rescale the range to match what we want.
minmax = self.minmax
img = util.linear_range_transform(img, from_range=(-1, 1), to_range=minmax)
assert (img >= minmax[0]).all()
assert (img <= minmax[1]).all()
# print('output size: {}'.format(img.shape))
return img
def forward(self, img):
minmax = self.minmax
# assert (img >= minmax[0]).all()
# assert (img <= minmax[1]).all()
# first normalize to [-1, 1]
img = util.linear_range_transform(img, from_range=self.minmax, to_range=(-1.0, 1.0))
out = self.model(img)
out = out.view(out.shape[0], -1)
# print(out.shape)
validity = self.adv_layer(out)
return validity
def forward(self, z):
out = self.l1(z)
out = out.view(out.shape[0], 128, self.init_size, self.init_size)
img = self.conv_blocks(out)
# tanh's range is [-1, 1]. Rescale the range to match what we want.
minmax = self.minmax
img = util.linear_range_transform(img, from_range=(-1, 1), to_range=minmax)
assert (img >= minmax[0]).all()
assert (img <= minmax[1]).all()
# print('output size: {}'.format(img.shape))
return img
def numpy_image_to_float(img, out_range=(0, 1)):
"""
Convert a numpy image (3d tensor, h x w x channels) into float format where
the output range is as given in out_range.
Return a numpy array.
"""
# http://scikit-image.org/docs/dev/api/skimage.util.html#skimage.util.img_as_float
new_img = skimage.util.img_as_float(img)
if np.any(new_img < 0):
# the range is (-1, 1)
assert np.all(new_img <= 1)
assert np.all(new_img >= -1)
# transform to the specified out_range
new_img = util.linear_range_transform(new_img, (-1, 1), out_range)
else:
# the range is (0, 1)
assert np.all(new_img <= 1)
assert np.all(new_img >= 0)
# transform to the specified out_range
new_img = util.linear_range_transform(new_img, (0, 1), out_range)
return new_img