#%% Prepro:
proj = (proj - dark) / (flat.mean(0) - dark)
proj = -numpy.log(proj)
proj = array.raw2astra(proj)
display.display_slice(proj, title='Sinogram. What else?')
#%% Recon:
vol = numpy.zeros([50, 2000, 2000], dtype='float32')
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(meta['geometry'], vol.shape)
proj_geom = io.astra_proj_geom(meta['geometry'], proj.shape)
# This is ASTRAAA!!!
sin_id = astra.data3d.link('-sino', proj_geom, numpy.ascontiguousarray(proj))
vol_id = astra.data3d.link('-vol', vol_geom, numpy.ascontiguousarray(vol))
cfg = astra.astra_dict('FDK_CUDA')
cfg['ReconstructionDataId'] = vol_id
cfg['ProjectionDataId'] = sin_id
alg_id = astra.algorithm.create(cfg)
astra.algorithm.run(alg_id, 1)
astra.algorithm.delete(alg_id)
astra.data3d.delete(sin_id)
def backproject(projections, volume, geometry, algorithm='BP3D_CUDA'):
"""
Backproject using standard ASTRA functionality. If data array is memmap, backprojection is done in blocks to save RAM.
Args:
projections : input numpy.array (dtype = float32) with the following dimensions: [vrt, rot, hrz]
volume : output numpy.array (dtype = float32) with the following dimensions: [vrt, mag, hrz]
geometry : geometry description - one of threee types: 'simple', 'static_offsets', 'linear_offsets'
algorithm : ASTRA algorithm type ['BP3D_CUDA', 'FDK_CUDA' etc.]
"""
global settings
block_number = settings['block_number']
mode = settings['mode']
# Check if projections should be subsampled:
sam = geometry['proj_sample']
if sum(sam) > 3:
projections = projections[::sam[0], ::sam[1], ::sam[2]]
# Weight correction
prj_weight = _astra_norm_(projections, volume, geometry, algorithm)
# If algorithm is FDK we use single-block projection unless data is a memmap
if block_number == 1:
projections = _contiguous_check_(projections)
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(geometry, volume.shape)
proj_geom = io.astra_proj_geom(geometry, projections.shape)
# Progress bar:
pbar = _pbar_start_(1)
projections *= prj_weight
# ASTRA here...
_backproject_block_add_(projections, volume, proj_geom, vol_geom,
algorithm)
projections /= prj_weight
_pbar_update_(pbar)
_pbar_close_(pbar)
else:
# Here is the multi-block version:
# Initialize ASTRA volume geometry:
vol_geom = io.astra_vol_geom(geometry, volume.shape)
# Progress bar:
pbar = _pbar_start_(block_number, 'block')
# Loop over blocks:
for ii in range(block_number):
# Extract a block:
index = _block_index_(ii, block_number, projections.shape[1], mode)
if len(index) > 0:
proj_geom = io.astra_proj_geom(geometry, projections.shape,
index)
# number of projections in a block can vary a bit and FDK is not aware of that...
block = projections[:, index, :] * prj_weight * len(index)
# BP and FDK behave differently in terms of normalization:
if (algorithm == 'BP3D_CUDA'):
block *= block_number
block = _contiguous_check_(block)
# Backproject:
_backproject_block_add_(block, volume, proj_geom, vol_geom,
algorithm)
_pbar_update_(pbar)
_pbar_close_(pbar)
# ASTRA is not aware of the number of blocks:
volume /= projections.shape[1]
""" Test reading Flexray raw and writing ASTRA readable """ #%% from flexdata import io #%% Read / write a geometry file: path = '/ufs/ciacc/flexbox/al_test/90KV_no_filt/' meta = io.read_meta(path, 'flexray') io.write_toml(path + 'flexray.toml', meta) #%% Read / write raw data files: dark = io.read_tiffs(path, 'di00') flat = io.read_tiffs(path, 'io00') proj = io.read_tiffs(path, 'scan_') #%% Read geometry and convert to ASTRA: meta_1 = io.read_toml(path + 'flexray.toml') vol_geom = io.astra_vol_geom(meta['geometry'], [100, 100, 100]) proj_geom = io.astra_proj_geom(meta['geometry'], proj.shape) print(vol_geom) print(proj_geom)
def EM_step(projections, volume, geometry):
"""
A single Expecrtation Maximization step. Supports blocking and subsets.
"""
global settings
norm_update = settings['norm_update']
block_number = settings['block_number']
bounds = settings['bounds']
mode = settings['mode']
prj_weight = _astra_norm_(projections, volume, geometry, 'BP3D_CUDA') * 2
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(geometry, volume.shape)
# Norm update:
norm = 0
for ii in range(block_number):
# Create index slice to address projections:
index = _block_index_(ii, block_number, projections.shape[1], mode)
if index is []: break
# Extract a block:
proj_geom = io.astra_proj_geom(geometry,
projections.shape,
index=index)
# Copy data to a block or simply pass a pointer to data itself if block is one.
if (mode == 'sequential') & (block_number == 1):
block = projections
else:
block = (projections[:, index, :]).copy()
# Reserve memory for a forward projection (keep it separate):
resid = _contiguous_check_(numpy.zeros_like(block))
# Forwardproject:
_forwardproject_block_add_(resid, volume, proj_geom, vol_geom)
# Compute residual:
resid[resid < resid.max() / 100] = numpy.inf
resid = (block / resid)
# L2 norm (use the last block to update):
if norm_update:
res_pos = resid[resid > 0]
norm += res_pos.std() / res_pos.mean()
# Project
_backproject_block_mult_(resid * prj_weight * block_number, volume,
proj_geom, vol_geom, 'BP3D_CUDA')
if norm_update:
settings['norm'].append(norm / block_number)
# Apply bounds
if bounds:
numpy.clip(volume, a_min=bounds[0], a_max=bounds[1], out=volume)
def FISTA_step(projections, vol, vol_old, vol_t, t, geometry):
"""
A single FISTA step. Supports blocking and subsets.
"""
global settings
norm_update = settings['norm_update']
block_number = settings['block_number']
bounds = settings['bounds']
poisson_weight = settings['poisson_weight']
mode = settings['mode']
prj_weight = _astra_norm_(projections, vol, geometry, 'BP3D_CUDA')
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(geometry, vol.shape)
vol_old[:] = vol.copy()
t_old = t
t = (1 + numpy.sqrt(1 + 4 * t**2)) / 2
vol[:] = vol_t.copy()
norm = 0
for ii in range(block_number):
# Create index slice to address projections:
index = _block_index_(ii, block_number, projections.shape[1], mode)
if index is []: break
# Extract a block:
proj_geom = io.astra_proj_geom(geometry,
projections.shape,
index=index)
# Copy data to a block or simply pass a pointer to data itself if block is one.
if (mode == 'sequential') & (block_number == 1):
block = projections.copy()
else:
block = (projections[:, index, :]).copy()
block = numpy.ascontiguousarray(block)
# Forwardproject:
_forwardproject_block_add_(block,
vol_t,
proj_geom,
vol_geom,
negative=True)
# Take into account Poisson:
if poisson_weight:
# Some formula representing the effect of photon starvation...
block *= numpy.exp(-projections[:, index, :])
# Normalization of the backprojection (depends on ASTRA):
block *= prj_weight * block_number
# Apply ramp to reduce boundary effects:
#block = block = flexData.ramp(block, 2, 5, mode = 'linear')
#block = block = flexData.ramp(block, 0, 5, mode = 'linear')
# L2 norm (use the last block to update):
if norm_update:
norm += numpy.sqrt((block**2).mean())
# Project
_backproject_block_add_(block, vol, proj_geom, vol_geom, 'BP3D_CUDA')
vol_t[:] = vol + ((t_old - 1) / t) * (vol - vol_old)
if norm_update:
settings['norm'].append(norm / block_number)
# Apply bounds
if bounds is not None:
numpy.clip(vol, a_min=bounds[0], a_max=bounds[1], out=vol)
def L2_step(projections, volume, geometry):
"""
A single L2 minimization step. Supports blocking and subsets.
"""
global settings
norm_update = settings['norm_update']
block_number = settings['block_number']
bounds = settings['bounds']
poisson_weight = settings['poisson_weight']
mode = settings['mode']
prj_weight = _astra_norm_(projections, volume, geometry, 'BP3D_CUDA')
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(geometry, volume.shape)
norm = 0
for ii in range(block_number):
# Create index slice to address projections:
index = _block_index_(ii, block_number, projections.shape[1], mode)
if index is []: break
# Extract a block:
proj_geom = io.astra_proj_geom(geometry,
projections.shape,
index=index)
# The block will contain the discrepancy eventually (that's why we need a copy):
if (mode == 'sequential') & (block_number == 1):
block = projections.copy()
else:
block = (projections[:, index, :]).copy()
block = _contiguous_check_(block)
# Forwardproject:
_forwardproject_block_add_(block,
volume,
proj_geom,
vol_geom,
negative=True)
# Take into account Poisson:
if poisson_weight:
# Some formula representing the effect of photon starvation...
block *= numpy.exp(-projections[:, index, :])
block *= prj_weight * block_number
# Apply ramp to reduce boundary effects:
#block = array.ramp(block, 0, 5, mode = 'linear')
#block = array.ramp(block, 2, 5, mode = 'linear')
# L2 norm (use the last block to update):
if norm_update:
norm = numpy.sqrt((block**2).mean())
# Project
_backproject_block_add_(block, volume, proj_geom, vol_geom,
'BP3D_CUDA')
if norm_update:
settings['norm'].append(norm / block_number)
# Apply bounds
if bounds:
numpy.clip(volume, a_min=bounds[0], a_max=bounds[1], out=volume)
def MULTI_PWLS(projections,
volume,
geometries,
iterations=10,
student=False,
weight_power=1):
'''
Penalized Weighted Least Squares based on multiple inputs.
'''
#error log:
global settings
norm = settings['norm']
block_number = settings['block_number']
mode = settings['mode']
norm = []
fac = volume.shape[2] * geometries[0]['img_pixel'] * numpy.sqrt(2)
print('PWLS-ing in progress...')
sleep(0.5)
# Iterations:
for ii in tqdm(range(iterations)):
# Error:
L_mean = 0
#Blocks:
for jj in range(block_number):
# Volume update:
vol_tmp = numpy.zeros_like(volume)
bwp_w = numpy.zeros_like(volume)
for kk, projs in enumerate(projections):
index = _block_index_(jj, block_number, projs.shape[1], mode)
proj = numpy.ascontiguousarray(projs[:, index, :])
geom = geometries[kk]
proj_geom = io.astra_proj_geom(geom, projs.shape, index=index)
vol_geom = io.astra_vol_geom(geom, volume.shape)
prj_tmp = numpy.zeros_like(proj)
# Compute weights:
if student:
fwp_w = numpy.ones_like(proj)
else:
me = proj.max() * weight_power / 5
fwp_w = numpy.exp(-proj * weight_power / me)
#fwp_w = scipy.ndimage.morphology.grey_erosion(fwp_w, size=(3,1,3))
_backproject_block_add_(fwp_w, bwp_w, proj_geom, vol_geom,
'BP3D_CUDA')
_forwardproject_block_add_(prj_tmp, volume, proj_geom,
vol_geom)
prj_tmp = (proj - prj_tmp) * fwp_w / fac
if student:
prj_tmp = _studentst_(prj_tmp, 5)
_backproject_block_add_(prj_tmp, vol_tmp, proj_geom, vol_geom,
'BP3D_CUDA')
# Mean L for projection
L_mean += (prj_tmp**2).mean()
eps = bwp_w.max() / 1000
bwp_w[bwp_w < eps] = eps
volume += vol_tmp / bwp_w
volume[volume < 0] = 0
#print((volume<0).sum())
norm.append(L_mean / block_number / len(projections))
display.plot(numpy.array(norm), semilogy=True)
def forwardproject(projections, volume, geometry):
"""
Forwardproject using standard ASTRA functionality. If data array is memmap, projection is done in blocks to save RAM.
Args:
projections : output numpy.array (dtype = float32) with the following dimensions: [vrt, rot, hrz]
volume : input numpy.array (dtype = float32) with the following dimensions: [vrt, mag, hrz]
geometry : geometry description - one of threee types: 'simple', 'static_offsets', 'linear_offsets'
"""
global settings
block_number = settings['block_number']
mode = settings['mode']
# Check if projections should be subsampled:
sam = geometry['vol_sample']
if sum(sam) > 3:
volume = volume[sam[0], sam[1], sam[2]]
# Non-memmap case is a single block:
volume = _contiguous_check_(volume)
# Forward project will always use blocks:
if block_number == 1:
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(geometry, volume.shape)
proj_geom = io.astra_proj_geom(geometry, projections.shape)
# Progress bar:
pbar = _pbar_start_(1)
_forwardproject_block_add_(projections, volume, proj_geom, vol_geom)
_pbar_update_(pbar)
_pbar_close_(pbar)
else:
# Multi-block:
# Initialize ASTRA geometries:
vol_geom = io.astra_vol_geom(geometry, volume.shape)
# Progress bar:
pbar = _pbar_start_(unit='block', total=block_number)
# Loop over blocks:
for ii in range(block_number):
index = _block_index_(ii, block_number, projections.shape[1], mode)
if len(index) > 0:
# Extract a block:
proj_geom = io.astra_proj_geom(geometry, projections.shape,
index)
block = projections[:, index, :]
block = _contiguous_check_(block)
# Backproject:
_forwardproject_block_add_(block, volume, proj_geom, vol_geom)
projections[:, index, :] = block
_pbar_update_(pbar)
_pbar_close_(pbar)
