def test_discrete_gradient():
"""Discretized spatial gradient operator."""
discr_space = Rn(1)
with pytest.raises(TypeError):
DiscreteGradient(discr_space)
# Check result of operator with explicit summation
# phantom data
data = np.array([[0., 1., 2., 3., 4.],
[1., 2., 3., 4., 5.],
[2., 3., 4., 5., 6.]])
data = np.array([[0., 1., 2., 3., 4.],
[0., 1., 2., 3., 4.],
[0., 1., 2., 3., 4.]])
# DiscreteLp Vector
discr_space = uniform_discr([0, 0], [6, 2.5], data.shape)
dom_vec = discr_space.element(data)
# computation of gradient components with helper function
dx0, dx1 = discr_space.grid.stride
df0 = finite_diff(data, axis=0, dx=dx0, zero_padding=True, edge_order=2)
df1 = finite_diff(data, axis=1, dx=dx1, zero_padding=True, edge_order=2)
# gradient
grad = DiscreteGradient(discr_space)
grad_vec = grad(dom_vec)
assert len(grad_vec) == data.ndim
assert all_equal(grad_vec[0].asarray(), df0)
assert all_equal(grad_vec[1].asarray(), df1)
# adjoint operator
ran_vec = grad.range.element([data, data ** 2])
adj_vec = grad.adjoint(ran_vec)
lhs = ran_vec.inner(grad_vec)
rhs = dom_vec.inner(adj_vec)
assert lhs != 0
assert rhs != 0
assert lhs == rhs
# higher dimensional arrays
lin_size = 3
for ndim in range(1, 6):
# DiscreteLp Vector
discr_space = uniform_discr([0.] * ndim, [lin_size] * ndim,
[lin_size] * ndim)
dom_vec = discr_space.element(ndvolume(lin_size, ndim))
# gradient
grad = DiscreteGradient(discr_space)
grad(dom_vec)
def test_discrete_divergence():
"""Discretized spatial divergence operator."""
# Invalid arguments
discr_space = Rn(1)
with pytest.raises(TypeError):
DiscreteDivergence(discr_space)
# Check result of operator with explicit summation
data = np.array([[0., 1., 2., 3., 4.],
[1., 2., 3., 4., 5.],
[2., 3., 4., 5., 6.]])
# DiscreteLp
discr_space = uniform_discr([0, 0], [6, 2.5], data.shape)
# Operator instance
div = DiscreteDivergence(discr_space)
# Apply operator
dom_vec = div.domain.element([data, data])
div_dom_vec = div(dom_vec)
# computation of divergence with helper function
dx0, dx1 = discr_space.grid.stride
df0 = finite_diff(data, axis=0, dx=dx0, zero_padding=True, edge_order=2)
df1 = finite_diff(data, axis=1, dx=dx1, zero_padding=True, edge_order=2)
assert all_equal(df0 + df1, div_dom_vec.asarray())
# Adjoint operator
adj_div = div.adjoint
ran_vec = div.range.element(data ** 2)
adj_div_ran_vec = adj_div(ran_vec)
# Adjoint condition
lhs = ran_vec.inner(div_dom_vec)
rhs = dom_vec.inner(adj_div_ran_vec)
assert lhs != 0
assert rhs != 0
assert almost_equal(lhs, rhs)
# Higher dimensional arrays
for ndim in range(1, 6):
# DiscreteLp Vector
lin_size = 3
discr_space = uniform_discr([0.] * ndim, [lin_size] * ndim,
[lin_size] * ndim)
# Divergence
div = DiscreteDivergence(discr_space)
dom_vec = div.domain.element([ndvolume(lin_size, ndim)] * ndim)
div(dom_vec)
def stir_projector_from_file(volume_file, projection_file):
"""Create a STIR projector from given template files.
Parameters
----------
volume_file : `str`
Full file path to the STIR input file containing information on the
volume. This is usually a '.hv' file. For STIR reasons,
a '.v' file is also needed.
projection_file : `str`
Full file path to the STIR input file with information on the
projection data. This is usually a '.hs' file. For STIR reasons,
a '.s' file is also needed.
Returns
-------
projector : `ForwardProjectorByBinWrapper`
A STIR forward projector.
"""
volume = stir.FloatVoxelsOnCartesianGrid.read_from_file(volume_file)
proj_data_in = stir.ProjData.read_from_file(projection_file)
proj_data = stir.ProjDataInMemory(proj_data_in.get_exam_info(),
proj_data_in.get_proj_data_info())
origin = volume.get_origin()
grid_spacing = volume.get_grid_spacing()
grid_shape = [
volume.get_z_size(),
volume.get_y_size(),
volume.get_x_size()
]
min_corner = [origin[1], origin[2], origin[3]]
max_corner = [
origin[1] + grid_spacing[1] * grid_shape[0],
origin[2] + grid_spacing[2] * grid_shape[1],
origin[3] + grid_spacing[3] * grid_shape[2]
]
# reverse to handle STIR bug? See:
# https://github.com/UCL/STIR/issues/7
recon_sp = uniform_discr(min_corner,
max_corner,
grid_shape,
dtype='float32')
# TODO: set correct projection space. Currently, a default grid with
# stride (1, 1, 1) is used.
proj_shape = proj_data.to_array().shape()
data_sp = uniform_discr([0, 0, 0], proj_shape, proj_shape, dtype='float32')
return ForwardProjectorByBinWrapper(recon_sp, data_sp, volume, proj_data)
def test_discrete_gradient_cuda():
"""Discretized spatial gradient operator using CUDA."""
# Check result of operator with explicit summation
# phantom data
data = np.array([[0., 1., 2., 3., 4.],
[1., 2., 3., 4., 5.],
[2., 3., 4., 5., 6.]])
# DiscreteLp Vector
discr_space = uniform_discr([0, 0], [6, 2.5], data.shape, impl='cuda')
dom_vec = discr_space.element(data)
# computation of gradient components with helper function
dx0, dx1 = discr_space.grid.stride
df0 = finite_diff(data, axis=0, dx=dx0, zero_padding=True, edge_order=2)
df1 = finite_diff(data, axis=1, dx=dx1, zero_padding=True, edge_order=2)
# gradient
grad = DiscreteGradient(discr_space)
grad_vec = grad(dom_vec)
assert len(grad_vec) == data.ndim
assert all_equal(grad_vec[0].asarray(), df0)
assert all_equal(grad_vec[1].asarray(), df1)
# adjoint operator
ran_vec = grad.range.element([data, data ** 2])
adj_vec = grad.adjoint(ran_vec)
lhs = ran_vec.inner(grad_vec)
rhs = dom_vec.inner(adj_vec)
assert lhs != 0
assert rhs != 0
assert lhs == rhs
def test_discr_part_deriv_cuda():
"""Discretized partial derivative using CUDA."""
# phantom data
data = np.array([0., 1., 2., 3., 4., 16., 25., 36.])
# explicit calculation of finite difference
dfe = np.zeros_like(data)
# interior: second-order accurate differences
dfe[1:-1] = (data[2:] - data[:-2]) / 2.0
# boundary: second-order accurate central differences with zero padding
dfe[0] = data[1] / 2.0
dfe[-1] = -data[-2] / 2.0
# discretized space using CUDA
discr_space = uniform_discr(0, data.size, data.shape, impl='cuda')
# operator
partial = DiscretePartDeriv(discr_space, zero_padding=True)
# discretized space vector
discr_vec = partial.domain.element(data)
# apply operator
partial_vec = partial(discr_vec)
assert all_equal(partial_vec, dfe)
def __init__(self, min_pt=None, max_pt=None, return_val='mu_normed'):
"""Construct the ellipses dataset.
Parameters
----------
return_val : str, optional
Values to return. Options are
``'mu_normed'``
Normalized linear attenuation, values are in [0, 1].
``'mu'``
Linear attenuation in m^-1, values are in ~[-0.46, 81.4].
``'hu'``
Values in Hounsfield unit in [-1024, 3071].
The default is ``'mu_normed'``.
min_pt : [int, int], optional
Minimum values of the lp space. Default: [-181, -181].
max_pt : [int, int], optional
Maximum values of the lp space. Default: [181, 181].
"""
self.return_val = return_val
self.shape = (362, 362)
if min_pt is None:
min_pt = [-self.shape[0] / 2, -self.shape[1] / 2]
if max_pt is None:
max_pt = [self.shape[0] / 2, self.shape[1] / 2]
space = uniform_discr(min_pt, max_pt, self.shape, dtype=np.float32)
with open(FILE_LIST_FILE, 'r') as json_file:
self.dcm_files_dict = json.load(json_file)
self.train_len = len(self.dcm_files_dict['train'])
self.validation_len = len(self.dcm_files_dict['validation'])
self.test_len = len(self.dcm_files_dict['test'])
super().__init__(space=space)
def test_discrete_divergence_cuda():
"""Discretized spatial divergence operator using CUDA."""
# Check result of operator with explicit summation
# phantom data
data = np.array([[0.0, 1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0, 5.0], [2.0, 3.0, 4.0, 5.0, 6.0]])
# DiscreteLp
discr_space = uniform_discr([0, 0], [1.5, 10], data.shape, impl="cuda")
# operator instance
div = DiscreteDivergence(discr_space)
# apply operator
dom_vec = div.domain.element([data, data])
div_dom_vec = div(dom_vec)
# computation of divergence with helper function
dx0, dx1 = discr_space.grid.stride
df0 = finite_diff(data, axis=0, dx=dx0, zero_padding=True, edge_order=2)
df1 = finite_diff(data, axis=1, dx=dx1, zero_padding=True, edge_order=2)
assert all_equal(df0 + df1, div_dom_vec.asarray())
# Adjoint operator
adj_div = div.adjoint
ran_vec = div.range.element(data ** 2)
adj_div_ran_vec = adj_div(ran_vec)
# Adjoint condition
lhs = ran_vec.inner(div_dom_vec)
rhs = dom_vec.inner(adj_div_ran_vec)
assert lhs != 0
assert rhs != 0
assert almost_equal(lhs, rhs)
def stir_projector_from_file(volume_file, projection_file):
"""Create a STIR projector from given template files.
Parameters
----------
volume_file : `str`
Full file path to the STIR input file containing information on the
volume. This is usually a '.hv' file. For STIR reasons,
a '.v' file is also needed.
projection_file : `str`
Full file path to the STIR input file with information on the
projection data. This is usually a '.hs' file. For STIR reasons,
a '.s' file is also needed.
Returns
-------
projector : `ForwardProjectorByBinWrapper`
A STIR forward projector.
"""
volume = stir.FloatVoxelsOnCartesianGrid.read_from_file(volume_file)
proj_data_in = stir.ProjData.read_from_file(projection_file)
proj_data = stir.ProjDataInMemory(proj_data_in.get_exam_info(),
proj_data_in.get_proj_data_info())
origin = volume.get_origin()
grid_spacing = volume.get_grid_spacing()
grid_shape = [volume.get_z_size(),
volume.get_y_size(),
volume.get_x_size()]
min_corner = [origin[1], origin[2], origin[3]]
max_corner = [origin[1] + grid_spacing[1] * grid_shape[0],
origin[2] + grid_spacing[2] * grid_shape[1],
origin[3] + grid_spacing[3] * grid_shape[2]]
# reverse to handle STIR bug? See:
# https://github.com/UCL/STIR/issues/7
recon_sp = uniform_discr(min_corner, max_corner, grid_shape,
dtype='float32')
# TODO: set correct projection space. Currently, a default grid with
# stride (1, 1, 1) is used.
proj_shape = proj_data.to_array().shape()
data_sp = uniform_discr([0, 0, 0], proj_shape, proj_shape, dtype='float32')
return ForwardProjectorByBinWrapper(recon_sp, data_sp, volume, proj_data)
def test_discr_part_deriv():
"""Discretized partial derivative."""
discr_space = Rn(1)
with pytest.raises(TypeError):
DiscretePartDeriv(discr_space)
# phantom data
data = np.array([[0., 1., 2., 3., 4.],
[1., 2., 3., 4., 5.],
[2., 3., 4., 5., 6.]])
# explicit calculation of finite difference
# axis: 0
dfe0 = np.zeros_like(data)
# interior: second-order accurate differences
dfe0[1:-1, :] = (data[2:, :] - data[:-2, :]) / 2.0
# boundary: second-order accurate central differences with zero padding
dfe0[0, :] = data[1, :] / 2.0
dfe0[-1, :] = -data[-2, :] / 2.0
# axis: 1
dfe1 = np.zeros_like(data)
# interior: second-order accurate differences
dfe1[:, 1:-1] = (data[:, 2:] - data[:, :-2]) / 2.0
# boundary: second-order accurate central differences with zero padding
dfe1[:, 0] = data[:, 1] / 2.0
dfe1[:, -1] = -data[:, -2] / 2.0
# assert `dfe0` and `dfe1` do differ
assert (dfe0 != dfe1).any()
# discretized space
discr_space = uniform_discr([0, 0], [2, 1], data.shape)
# operator
partial_0 = DiscretePartDeriv(discr_space, axis=0, zero_padding=True)
partial_1 = DiscretePartDeriv(discr_space, axis=1, zero_padding=True)
# discretized space vector
vec = partial_0.domain.element(data)
# partial derivative
partial_vec_0 = partial_0(vec)
partial_vec_1 = partial_1(vec)
assert partial_vec_0 != partial_vec_1
assert all_equal(partial_vec_0.asarray(), dfe0)
assert all_equal(partial_vec_1.asarray(), dfe1)
# operator
partial_0 = DiscretePartDeriv(discr_space, axis=1, dx=0.2, edge_order=2,
zero_padding=True)
# adjoint not implemented
with pytest.raises(NotImplementedError):
partial_0.adjoint
def __init__(self,
image_size=128,
min_pt=None,
max_pt=None,
train_len=32000,
validation_len=3200,
test_len=3200,
fixed_seeds=False):
"""
Parameters
----------
image_size : int, optional
Number of pixels per image dimension. Default: ``128``.
min_pt : [int, int], optional
Minimum values of the lp space.
Default: ``[-image_size/2, -image_size/2]``.
max_pt : [int, int], optional
Maximum values of the lp space.
Default: ``[image_size/2, image_size/2]``.
train_len : int or `None`, optional
Length of training set. Default: ``32000``.
If `None`, infinitely many samples could be generated.
validation_len : int, optional
Length of training set. Default: ``3200``.
test_len : int, optional
Length of test set. Default: ``3200``.
fixed_seeds : dict or bool, optional
Seeds to use for random generation.
The values of the keys ``'train'``, ``'validation'`` and ``'test'``
are used. If a seed is `None` or omitted, it is choosen randomly.
If ``True`` is passed, the seeds
``fixed_seeds={'train': 42, 'validation': 2, 'test': 1}`` are used.
If ``False`` is passed (the default), all seeds are chosen
randomly.
"""
self.shape = (image_size, image_size)
if min_pt is None:
min_pt = [-self.shape[0] / 2, -self.shape[1] / 2]
if max_pt is None:
max_pt = [self.shape[0] / 2, self.shape[1] / 2]
space = uniform_discr(min_pt, max_pt, self.shape, dtype=np.float32)
self.train_len = train_len
self.validation_len = validation_len
self.test_len = test_len
self.random_access = False
if isinstance(fixed_seeds, bool):
if fixed_seeds:
self.fixed_seeds = {'train': 42, 'validation': 2, 'test': 1}
else:
self.fixed_seeds = {}
else:
self.fixed_seeds = fixed_seeds.copy()
super().__init__(space=space)
def uniform_discr_element(inp, space=None):
"""Generate an element of a ODL space from an array-like.
Parameters
----------
inp : array-like
The input data from which the element is generated.
space : :class:`odl.discr.DiscretizedSpace`, optional
The space which the element will belong to. If not given, a uniform
discretization space with cell size 1 centered around the origin is
generated.
"""
inp = np.asarray(inp)
if space is None:
space = uniform_discr(-np.array(inp.shape) / 2,
np.array(inp.shape) / 2, inp.shape)
element = space.element(inp)
return element
def __init__(self, min_pt=None, max_pt=None):
"""Construct the ellipses dataset.
Parameters
----------
min_pt : [int, int], optional
Minimum values of the lp space. Default: [-64, -64].
max_pt : [int, int], optional
Maximum values of the lp space. Default: [64, 64].
"""
self.shape = (128, 128)
if min_pt is None:
min_pt = [-self.shape[0] / 2, -self.shape[1] / 2]
if max_pt is None:
max_pt = [self.shape[0] / 2, self.shape[1] / 2]
space = uniform_discr(min_pt, max_pt, self.shape, dtype=np.float32)
self.train_len = 50000
self.validation_len = 5000
self.test_len = 5000
super().__init__(space=space)
def __init__(self,
min_pt=None,
max_pt=None,
observation_model='post-log',
min_photon_count=None,
impl='astra_cuda'):
"""
Parameters
----------
min_pt : [float, float], optional
Minimum values of the lp space. Default: ``[-0.13, -0.13]``.
max_pt : [float, float], optional
Maximum values of the lp space. Default: ``[0.13, 0.13]``.
observation_model : {'post-log', 'pre-log'}, optional
The observation model to use.
The default is ``'post-log'``.
``'post-log'``
Observations are linearly related to the normalized ground
truth via the ray transform, ``obs = ray_trafo(gt) + noise``.
Note that the scaling of the observations matches the
normalized ground truth, i.e., they are divided by the linear
attenuation of 3071 HU.
``'pre-log'``
Observations are non-linearly related to the ground truth, as
given by the Beer-Lambert law.
The model is
``obs = exp(-ray_trafo(gt * MU(3071 HU))) + noise``,
where `MU(3071 HU)` is the factor, by which the ground truth
was normalized.
min_photon_count : float, optional
Replacement value for a simulated photon count of zero.
If ``observation_model == 'post-log'``, a value greater than zero
is required in order to avoid undefined values. The default is 0.1,
both for ``'post-log'`` and ``'pre-log'`` model.
impl : {``'skimage'``, ``'astra_cpu'``, ``'astra_cuda'``},\
optional
Implementation passed to :class:`odl.tomo.RayTransform` to
construct :attr:`ray_trafo`.
"""
global DATA_PATH
NUM_ANGLES = 1000
NUM_DET_PIXELS = 513
self.shape = ((NUM_ANGLES, NUM_DET_PIXELS), (362, 362))
self.num_elements_per_sample = 2
if min_pt is None:
min_pt = MIN_PT
if max_pt is None:
max_pt = MAX_PT
domain = uniform_discr(min_pt, max_pt, self.shape[1], dtype=np.float32)
if observation_model == 'post-log':
self.post_log = True
elif observation_model == 'pre-log':
self.post_log = False
else:
raise ValueError("`observation_model` must be 'post-log' or "
"'pre-log', not '{}'".format(observation_model))
if min_photon_count is None or min_photon_count <= 1.:
self.min_photon_count = min_photon_count
else:
self.min_photon_count = 1.
warn('`min_photon_count` changed from {} to 1.'.format(
min_photon_count))
self.train_len = 35820
self.validation_len = 3522
self.test_len = 3553
self.random_access = True
while not self.check_for_lodopab():
print('The LoDoPaB-CT dataset could not be found under the '
"configured path '{}'.".format(
CONFIG['lodopab_dataset']['data_path']))
print('Do you want to download it now? (y: download, n: input '
'other path)')
download = input_yes_no()
if download:
success = download_lodopab()
if not success:
raise RuntimeError('lodopab dataset not available, '
'download failed')
else:
print('Path to LoDoPaB dataset:')
DATA_PATH = input()
set_config('lodopab_dataset/data_path', DATA_PATH)
self.geometry = odl.tomo.parallel_beam_geometry(
domain, num_angles=NUM_ANGLES, det_shape=(NUM_DET_PIXELS, ))
range_ = uniform_discr(self.geometry.partition.min_pt,
self.geometry.partition.max_pt,
self.shape[0],
dtype=np.float32)
super().__init__(space=(range_, domain))
self.ray_trafo = self.get_ray_trafo(impl=impl)