def prepare_data(self, *args, **kwargs):
mnist_test = MNIST(self.root_dir, train=False, download=True).data.type(torch.float32)
mnist_train_val = MNIST(self.root_dir, train=True, download=True).data.type(torch.float32)
np.random.seed(1612)
perm = np.random.permutation(mnist_train_val.shape[0])
mnist_train = mnist_train_val[perm[:55000], 1:, 1:]
mnist_val = mnist_train_val[perm[55000:], 1:, 1:]
mnist_test = mnist_test[:, 1:, 1:]
assert mnist_train.shape[1] == self.gt_shape
assert mnist_train.shape[2] == self.gt_shape
circle = self.__get_circle__()
mnist_train = circle * np.clip(mnist_train, 50, 255)
mnist_val = circle * np.clip(mnist_val, 50, 255)
mnist_test = circle * np.clip(mnist_test, 50, 255)
self.mean = mnist_train.mean()
self.std = mnist_train.std()
mnist_train = normalize(mnist_train, self.mean, self.std)
mnist_val = normalize(mnist_val, self.mean, self.std)
mnist_test = normalize(mnist_test, self.mean, self.std)
mnist_train *= circle
mnist_val *= circle
mnist_test *= circle
ds_factory = GroundTruthDatasetFactory(mnist_train, mnist_val, mnist_test, inner_circle=self.inner_circle)
self.gt_ds = ds_factory.build_projection_dataset(num_angles=self.num_angles,
upscale_shape=70,
impl='astra_cpu')
def prepare_data(self, *args, **kwargs):
if not exists(join(self.root_dir, 'gt_data.npz')):
wget.download('https://download.fht.org/jug/fit/Kanji_TRec_gt_data.npz',
out=join(self.root_dir, 'gt_data.npz'))
gt_data = np.load(join(self.root_dir, 'gt_data.npz'))
gt_train = torch.from_numpy(gt_data['gt_train'])
gt_val = torch.from_numpy(gt_data['gt_val'])
gt_test = torch.from_numpy(gt_data['gt_test'])
self.mean = gt_train.mean()
self.std = gt_train.std()
gt_train = normalize(gt_train, self.mean, self.std)
gt_val = normalize(gt_val, self.mean, self.std)
gt_test = normalize(gt_test, self.mean, self.std)
circle = self.__get_circle__()
gt_train *= circle
gt_val *= circle
gt_test *= circle
ds_factory = GroundTruthDatasetFactory(gt_train, gt_val, gt_test, inner_circle=self.inner_circle)
self.gt_ds = ds_factory.build_projection_dataset(num_angles=self.num_angles,
upscale_shape=133,
impl='astra_cpu')
def prepare_data(self, *args, **kwargs):
if not exists(join(self.root_dir, 'gt_data.npz')):
wget.download(
'https://cloud.mpi-cbg.de/index.php/s/7MK9vNUnq4Ndkhg/download',
out=join(self.root_dir, 'gt_data.npz'))
gt_data = np.load(join(self.root_dir, 'gt_data.npz'))
gt_train = torch.from_numpy(gt_data['gt_train'])
gt_val = torch.from_numpy(gt_data['gt_val'])
gt_test = torch.from_numpy(gt_data['gt_test'])
self.mean = gt_train.mean()
self.std = gt_train.std()
gt_train = normalize(gt_train, self.mean, self.std)
gt_val = normalize(gt_val, self.mean, self.std)
gt_test = normalize(gt_test, self.mean, self.std)
circle = self.__get_circle__()
gt_train *= circle
gt_val *= circle
gt_test *= circle
ds_factory = GroundTruthDatasetFactory(gt_train,
gt_val,
gt_test,
inner_circle=self.inner_circle)
impl = kwargs.get('impl', 'skimage')
self.gt_ds = ds_factory.build_projection_dataset(
num_angles=self.num_angles, upscale_shape=133, impl=impl)
def prepare_data(self, *args, **kwargs):
mnist_test = MNIST(self.root_dir, train=False, download=True).data.type(torch.float32)
mnist_train_val = MNIST(self.root_dir, train=True, download=True).data.type(torch.float32)
np.random.seed(1612)
perm = np.random.permutation(mnist_train_val.shape[0])
mnist_train = mnist_train_val[perm[:55000], 1:, 1:]
mnist_val = mnist_train_val[perm[55000:], 1:, 1:]
mnist_test = mnist_test[:, 1:, 1:]
self.mean = mnist_train.mean()
self.std = mnist_train.std()
mnist_train = normalize(mnist_train, self.mean, self.std)
mnist_val = normalize(mnist_val, self.mean, self.std)
mnist_test = normalize(mnist_test, self.mean, self.std)
self.gt_ds = GroundTruthDatasetFactory(mnist_train, mnist_val, mnist_test)
def prepare_data(self, *args, **kwargs):
if not exists(join(self.root_dir, 'gt_data.npz')):
wget.download('https://cloud.mpi-cbg.de/index.php/s/Wtuy9IqUsSpjKav/download',
out=join(self.root_dir, 'gt_data.npz'))
gt_data = np.load(join(self.root_dir, 'gt_data.npz'))
gt_train = torch.from_numpy(gt_data['gt_train'])
gt_val = torch.from_numpy(gt_data['gt_val'])
gt_test = torch.from_numpy(gt_data['gt_test'])
self.mean = gt_train.mean()
self.std = gt_train.std()
gt_train = normalize(gt_train, self.mean, self.std)
gt_val = normalize(gt_val, self.mean, self.std)
gt_test = normalize(gt_test, self.mean, self.std)
self.gt_ds = GroundTruthDatasetFactory(gt_train, gt_val, gt_test)
def prepare_data(self, *args, **kwargs):
lodopab = dival.get_standard_dataset('lodopab', impl='astra_cpu')
assert self.gt_shape <= self.IMG_SHAPE, 'GT is larger than original images.'
if self.gt_shape < self.IMG_SHAPE:
crop_off = (362 - self.gt_shape) // 2
gt_train = np.array([lodopab.get_sample(i, part='train', out=(False, True))[1][crop_off:-(crop_off + 1),
crop_off:-(crop_off + 1)] for i in
range(4000)])
gt_val = np.array([lodopab.get_sample(i, part='validation', out=(False, True))[1][crop_off:-(crop_off + 1),
crop_off:-(crop_off + 1)] for i in
range(400)])
gt_test = np.array([lodopab.get_sample(i, part='test', out=(False, True))[1][crop_off:-(crop_off + 1),
crop_off:-(crop_off + 1)] for i in
range(3553)])
else:
gt_train = np.array(
[lodopab.get_sample(i, part='train', out=(False, True))[1][1:, 1:] for i in range(4000)])
gt_val = np.array(
[lodopab.get_sample(i, part='validation', out=(False, True))[1][1:, 1:] for i in range(400)])
gt_test = np.array(
[lodopab.get_sample(i, part='test', out=(False, True))[1][1:, 1:] for i in range(3553)])
gt_train = torch.from_numpy(gt_train)
gt_val = torch.from_numpy(gt_val)
gt_test = torch.from_numpy(gt_test)
assert gt_train.shape[1] == self.gt_shape
assert gt_train.shape[2] == self.gt_shape
self.mean = gt_train.mean()
self.std = gt_train.std()
gt_train = normalize(gt_train, self.mean, self.std)
gt_val = normalize(gt_val, self.mean, self.std)
gt_test = normalize(gt_test, self.mean, self.std)
circle = self.__get_circle__()
gt_train *= circle
gt_val *= circle
gt_test *= circle
ds_factory = GroundTruthDatasetFactory(gt_train, gt_val, gt_test, inner_circle=self.inner_circle)
self.gt_ds = ds_factory.build_projection_dataset(num_angles=self.num_angles,
upscale_shape=self.gt_shape + (self.gt_shape // 2 - 7),
impl='astra_cpu')
def prepare_data(self, *args, **kwargs):
if not exists(join(self.root_dir, 'gt_data.npz')):
wget.download(
'https://download.fht.org/jug/fit/CelebA_SRes_gt_data.npz',
out=join(self.root_dir, 'gt_data.npz'))
gt_data = np.load(join(self.root_dir, 'gt_data.npz'))
gt_train = torch.from_numpy(gt_data['gt_train'])
gt_val = torch.from_numpy(gt_data['gt_val'])
gt_test = torch.from_numpy(gt_data['gt_test'])
self.mean = gt_train.mean()
self.std = gt_train.std()
gt_train = normalize(gt_train, self.mean, self.std)
gt_val = normalize(gt_val, self.mean, self.std)
gt_test = normalize(gt_test, self.mean, self.std)
self.gt_ds = GroundTruthDatasetFactory(gt_train, gt_val, gt_test)
