def jacobian_determinant(disp):
"""
jacobian determinant of a displacement field.
NB: to compute the spatial gradients, we use np.gradient.
Parameters:
disp: 3D displacement field of size [nb_dims, *vol_shape]
Returns:
jacobian determinant (matrix)
"""
# check inputs
volshape = disp.shape[1:]
nb_dims = len(volshape)
assert len(volshape) in (2, 3), 'flow has to be 2D or 3D'
# compute grid
grid_lst = nd.volsize2ndgrid(volshape)
grid = np.stack(grid_lst, 0)
# compute gradients
[xFX, xFY, xFZ] = np.gradient(grid[0] - disp[0])
[yFX, yFY, yFZ] = np.gradient(grid[1] - disp[1])
[zFX, zFY, zFZ] = np.gradient(grid[2] - disp[2])
jac_det = np.zeros(grid[0].shape)
for i in range(grid.shape[1]):
for j in range(grid.shape[2]):
for k in range(grid.shape[3]):
jac_mij = [[xFX[i, j, k], xFY[i, j, k], xFZ[i, j, k]],
[yFX[i, j, k], yFY[i, j, k], yFZ[i, j, k]],
[zFX[i, j, k], zFY[i, j, k], zFZ[i, j, k]]]
jac_det[i, j, k] = np.linalg.det(jac_mij)
return jac_det
def jacobian_determinant(disp):
"""
jacobian determinant of a displacement field.
NB: to compute the spatial gradients, we use np.gradient.
Parameters:
disp: 2D or 3D displacement field of size [*vol_shape, nb_dims],
where vol_shape is of len nb_dims
Returns:
jacobian determinant (scalar)
"""
# check inputs
volshape = disp.shape[:-1]
nb_dims = len(volshape)
assert len(volshape) in (2, 3), 'flow has to be 2D or 3D'
# compute grid
grid_lst = nd.volsize2ndgrid(volshape)
grid = np.stack(grid_lst, len(volshape))
# compute gradients
J = np.gradient(disp + grid)
# 3D glow
if nb_dims == 3:
dx = J[0]
dy = J[1]
dz = J[2]
# compute jacobian components
Jdet0 = dx[...,
0] * (dy[..., 1] * dz[..., 2] - dy[..., 2] * dz[..., 1])
Jdet1 = dx[...,
1] * (dy[..., 0] * dz[..., 2] - dy[..., 2] * dz[..., 0])
Jdet2 = dx[...,
2] * (dy[..., 0] * dz[..., 1] - dy[..., 1] * dz[..., 0])
return Jdet0 - Jdet1 + Jdet2
else: # must be 2
dfdx = J[0]
dfdy = J[1]
return dfdx[..., 0] * dfdy[..., 1] - dfdy[..., 0] * dfdx[..., 1]
def vol_prior_hack(
*args,
proc_vol_fn=None,
proc_seg_fn=None,
prior_type='location', # file-static, file-gen, location
prior_file=None, # prior filename
prior_feed='input', # input or output
patch_stride=1,
patch_size=None,
batch_size=1,
collapse_2d=None,
extract_slice=None,
force_binary=False,
nb_input_feats=1,
verbose=False,
vol_rand_seed=None,
**kwargs):
"""
"""
# prepare the vol_seg
gen = vol_seg_hack(*args,
**kwargs,
proc_vol_fn=None,
proc_seg_fn=None,
collapse_2d=collapse_2d,
extract_slice=extract_slice,
force_binary=force_binary,
verbose=verbose,
patch_size=patch_size,
patch_stride=patch_stride,
batch_size=batch_size,
vol_rand_seed=vol_rand_seed,
nb_input_feats=nb_input_feats)
# get prior
if prior_type == 'location':
prior_vol = nd.volsize2ndgrid(vol_size)
prior_vol = np.transpose(prior_vol, [1, 2, 3, 0])
prior_vol = np.expand_dims(prior_vol, axis=0) # reshape for model
elif prior_type == 'file': # assumes a npz filename passed in prior_file
with timer.Timer('loading prior', True):
data = np.load(prior_file)
prior_vol = data['prior'].astype('float16')
else: # assumes a volume
with timer.Timer('astyping prior', verbose):
prior_vol = prior_file
if not (prior_vol.dtype == 'float16'):
prior_vol = prior_vol.astype('float16')
if force_binary:
nb_labels = prior_vol.shape[-1]
prior_vol[:, :, :, 1] = np.sum(prior_vol[:, :, :, 1:nb_labels], 3)
prior_vol = np.delete(prior_vol, range(2, nb_labels), 3)
nb_channels = prior_vol.shape[-1]
if extract_slice is not None:
if isinstance(extract_slice, int):
prior_vol = prior_vol[:, :, extract_slice, np.newaxis, :]
else: # assume slices
prior_vol = prior_vol[:, :, extract_slice, :]
# get the prior to have the right volume [x, y, z, nb_channels]
assert np.ndim(prior_vol) == 4 or np.ndim(
prior_vol) == 3, "prior is the wrong size"
# prior generator
if patch_size is None:
patch_size = prior_vol.shape[0:3]
assert len(patch_size) == len(patch_stride)
prior_gen = patch(
prior_vol,
[*patch_size, nb_channels],
patch_stride=[*patch_stride, nb_channels],
batch_size=batch_size,
collapse_2d=collapse_2d,
keep_vol_size=True,
infinite=True,
#variable_batch_size=True, # this
nb_labels_reshape=0)
# assert next(prior_gen) is None, "bad prior gen setup"
# generator loop
while 1:
# generate input and output volumes
input_vol = next(gen)
if verbose and np.all(input_vol.flat == 0):
print("all entries are 0")
# generate prior batch
# with timer.Timer("with send?"):
# prior_batch = prior_gen.send(input_vol.shape[0])
prior_batch = next(prior_gen)
if prior_feed == 'input':
yield ([input_vol, prior_batch], input_vol)
else:
assert prior_feed == 'output'
yield (input_vol, [input_vol, prior_batch])
def add_prior(
gen,
proc_vol_fn=None,
proc_seg_fn=None,
prior_type='location', # file-static, file-gen, location
prior_file=None, # prior filename
prior_feed='input', # input or output
patch_stride=1,
patch_size=None,
batch_size=1,
collapse_2d=None,
extract_slice=None,
force_binary=False,
verbose=False,
patch_rand=False,
patch_rand_seed=None):
"""
#
# add a prior generator to a given generator
# with the number of patches in batch matching output of gen
"""
# get prior
if prior_type == 'location':
prior_vol = nd.volsize2ndgrid(vol_size)
prior_vol = np.transpose(prior_vol, [1, 2, 3, 0])
prior_vol = np.expand_dims(prior_vol, axis=0) # reshape for model
elif prior_type == 'file': # assumes a npz filename passed in prior_file
with timer.Timer('loading prior', True):
data = np.load(prior_file)
prior_vol = data['prior'].astype('float16')
else: # assumes a volume
with timer.Timer('loading prior', True):
prior_vol = prior_file.astype('float16')
if force_binary:
nb_labels = prior_vol.shape[-1]
prior_vol[:, :, :, 1] = np.sum(prior_vol[:, :, :, 1:nb_labels], 3)
prior_vol = np.delete(prior_vol, range(2, nb_labels), 3)
nb_channels = prior_vol.shape[-1]
if extract_slice is not None:
if isinstance(extract_slice, int):
prior_vol = prior_vol[:, :, extract_slice, np.newaxis, :]
else: # assume slices
prior_vol = prior_vol[:, :, extract_slice, :]
# get the prior to have the right volume [x, y, z, nb_channels]
assert np.ndim(prior_vol) == 4 or np.ndim(
prior_vol) == 3, "prior is the wrong size"
# prior generator
if patch_size is None:
patch_size = prior_vol.shape[0:3]
assert len(patch_size) == len(patch_stride)
prior_gen = patch(prior_vol, [*patch_size, nb_channels],
patch_stride=[*patch_stride, nb_channels],
batch_size=batch_size,
collapse_2d=collapse_2d,
keep_vol_size=True,
infinite=True,
patch_rand=patch_rand,
patch_rand_seed=patch_rand_seed,
variable_batch_size=True,
nb_labels_reshape=0)
assert next(prior_gen) is None, "bad prior gen setup"
# generator loop
while 1:
# generate input and output volumes
gen_sample = next(gen)
# generate prior batch
gs_sample = _get_shape(gen_sample)
prior_batch = prior_gen.send(gs_sample)
yield (gen_sample, prior_batch)