def process(self, out=None):
DATA = self.get_input()
pad = False
if len(DATA.shape) == 2:
#for 2D cases
pad = True
data_temp = numpy.expand_dims(DATA.as_array(), axis=0)
else:
data_temp = DATA.as_array()
rec_id, arr_out = astra.create_backprojection3d_gpu(
data_temp, self.proj_geom, self.vol_geom)
astra.data3d.delete(rec_id)
if pad is True:
arr_out = numpy.squeeze(arr_out, axis=0)
if out is None:
out = ImageData(arr_out,
deep_copy=False,
geometry=self.volume_geometry.copy(),
suppress_warning=True)
return out
else:
out.fill(arr_out)
def process(self):
DATA = self.get_input()
IM = ImageData(geometry=self.volume_geometry,
dimension_labels=self.output_axes_order)
rec_id, IM.array = astra.create_backprojection3d_gpu(
DATA.as_array(), self.proj_geom, self.vol_geom)
astra.data3d.delete(rec_id)
# Scaling of 3D ASTRA backprojector, works both parallel and cone.
scaling = 1 / self.volume_geometry.voxel_size_x**2
return scaling * IM
def backward(ctx, grad_out):
astra.set_gpu_index([grad_out.device.index],
memory=10 * 1024 * 1024 * 1024)
grad_input = gradProjSize = gradVectors = None
if ctx.needs_input_grad[0]:
# permute here to handle opposite data layouts
bproj_id, bproj_data = astra.create_backprojection3d_gpu(
grad_out.squeeze(1).permute(2, 0, 1).data.cpu().numpy(),
ctx.proj_geom, ctx.vol_geom)
astra.data3d.delete(bproj_id)
grad_input = Tensor(bproj_data).cuda(non_blocking=True).permute(
2, 1, 0)
return grad_input, gradProjSize, gradVectors
def single_test_adjoint3D(self, geom_type, proj_type):
for vg in VolumeGeometries(True, True):
for pg in ProjectionGeometries(geom_type):
for i in range(5):
X = np.random.random(astra.geom_size(vg))
Y = np.random.random(astra.geom_size(pg))
proj_id, fX = astra.create_sino3d_gpu(X, pg, vg)
bp_id, fTY = astra.create_backprojection3d_gpu(Y, pg, vg)
astra.data3d.delete([proj_id, bp_id])
da = np.dot(fX.ravel(), Y.ravel())
db = np.dot(X.ravel(), fTY.ravel())
m = np.abs(da - db)
TOL = 1e-1
if m / da >= TOL:
print(vg)
print(pg)
print(m / da, da / db, da, db)
self.assertTrue(m / da < TOL)
import numpy as np
# Set up multi-GPU usage.
# This only works for 3D GPU forward projection and back projection.
astra.astra.set_gpu_index([0,1])
# Optionally, you can also restrict the amount of GPU memory ASTRA will use.
# The line commented below sets this to 1GB.
#astra.astra.set_gpu_index([0,1], memory=1024*1024*1024)
vol_geom = astra.create_vol_geom(1024, 1024, 1024)
angles = np.linspace(0, np.pi, 1024,False)
proj_geom = astra.create_proj_geom('parallel3d', 1.0, 1.0, 1024, 1024, angles)
# Create a simple hollow cube phantom
cube = np.zeros((1024,1024,1024))
cube[128:895,128:895,128:895] = 1
cube[256:767,256:767,256:767] = 0
# Create projection data from this
proj_id, proj_data = astra.create_sino3d_gpu(cube, proj_geom, vol_geom)
# Backproject projection data
bproj_id, bproj_data = astra.create_backprojection3d_gpu(proj_data, proj_geom, vol_geom)
# Clean up. Note that GPU memory is tied up in the algorithm object,
# and main RAM in the data objects.
astra.data3d.delete(proj_id)
astra.data3d.delete(bproj_id)
def backprojOS(self, proj_data, no_os):
"""Applying back-projection to a specific subset"""
rec_id, object3D = astra.create_backprojection3d_gpu(
proj_data, self.proj_geom_OS[no_os], self.vol_geom)
astra.data3d.delete(rec_id)
return object3D
def backproj(self, proj_data):
"""Applying backprojection"""
rec_id, object3D = astra.create_backprojection3d_gpu(
proj_data, self.proj_geom, self.vol_geom)
astra.data3d.delete(rec_id)
return object3D
# Set up multi-GPU usage.
# This only works for 3D GPU forward projection and back projection.
astra.set_gpu_index([0, 1])
# Optionally, you can also restrict the amount of GPU memory ASTRA will use.
# The line commented below sets this to 1GB.
#astra.set_gpu_index([0,1], memory=1024*1024*1024)
vol_geom = astra.create_vol_geom(1024, 1024, 1024)
angles = np.linspace(0, np.pi, 1024, False)
proj_geom = astra.create_proj_geom('parallel3d', 1.0, 1.0, 1024, 1024, angles)
# Create a simple hollow cube phantom
cube = np.zeros((1024, 1024, 1024))
cube[128:895, 128:895, 128:895] = 1
cube[256:767, 256:767, 256:767] = 0
# Create projection data from this
proj_id, proj_data = astra.create_sino3d_gpu(cube, proj_geom, vol_geom)
# Backproject projection data
bproj_id, bproj_data = astra.create_backprojection3d_gpu(
proj_data, proj_geom, vol_geom)
# Clean up. Note that GPU memory is tied up in the algorithm object,
# and main RAM in the data objects.
astra.data3d.delete(proj_id)
astra.data3d.delete(bproj_id)
def recon_multi_part(algon,proj_data,proj_data_low,det_move,angles,x_sub_boundaries,y_sub_boundaries,z_sub_boundaries,vol_size,det_spacing_x,det_spacing_y,x_subvol_width,y_subvol_width,z_subvol_width,n_x,n_y,n_z,nr_iterations_high,source_origin, origin_det,relaxation):
# this part performs the reconstruction of the subvolume.
pro_proj = proj_data - proj_data_low # we only want to have the projection data that comes from the sub volume
det_row_count= pro_proj.shape[0]
det_col_count=pro_proj.shape[2]
vectors_ROI = np.zeros((len(angles), 12)) # this vector will contain the information of the source and detector position.
move_x,move_y,move_z=how_to_move(x_sub_boundaries,y_sub_boundaries,z_sub_boundaries,vol_size,x_subvol_width,y_subvol_width,z_subvol_width,n_x,n_y,n_z)
for i in range(0,len(angles)):
# source
vectors_ROI[i,0] = np.sin(angles[i]) * (source_origin) +move_x
vectors_ROI[i,1] = -np.cos(angles[i]) * (source_origin) + move_y
vectors_ROI[i,2] = move_z
# center of detector
vectors_ROI[i,3] = -np.sin(angles[i]) * (origin_det) + move_x
vectors_ROI[i,4] = np.cos(angles[i]) * (origin_det) + move_y
vectors_ROI[i,5] = move_z +det_move
# vector from detector pixel 0 to 1
vectors_ROI[i,6] = np.cos(angles[i]) *det_spacing_x
vectors_ROI[i,7] = np.sin(angles[i])*det_spacing_x
vectors_ROI[i,8] = 0
# vector from detector pixel (0,0) to (1,0)
vectors_ROI[i,9] = 0
vectors_ROI[i,10] = 0
vectors_ROI[i,11] = det_spacing_y
#ROI=cube[:,:,z_sub_boundaries[n_z,0]:z_sub_boundaries[n_z+1,0]]
# mlab.figure()
# mlab.contour3d(ROI)
width_z=z_sub_boundaries[n_z+1,2]-z_sub_boundaries[n_z,1]
width_y=y_sub_boundaries[n_y+1,2]-y_sub_boundaries[n_y,1]
width_x=x_sub_boundaries[n_x+1,2]-x_sub_boundaries[n_x,1]
vol_geom_ROI= astra.create_vol_geom(width_y.astype(int), width_z.astype(int), width_x.astype(int))
proj_geom_ROI = astra.create_proj_geom('cone_vec', det_row_count, det_col_count, vectors_ROI)
if algon[0:4] != 'EGEN':
# Create a data object for the reconstruction
rec_id_ROI = astra.data3d.create('-vol', vol_geom_ROI)
sinogram_id_ROI = astra.data3d.create('-proj3d', proj_geom_ROI, pro_proj)
if algon=='SIRT':
algon='SIRT3D_CUDA'
elif algon=='CGLS':
algon='CGLS3D_CUDA'
# Set up the parameters for a reconstruction algorithm using the GPU
cfg = astra.astra_dict(algon)
cfg['ReconstructionDataId'] = rec_id_ROI
cfg['ProjectionDataId'] = sinogram_id_ROI
cfg['option']={}
cfg['option']['MinConstraint'] = 0 #attenuation coefficient can't be negative
#cfg['option']['VoxelSuperSampling'] = np.round(float(det_spacing_x)/vol_size,decimals=2)
alg_id = astra.algorithm.create(cfg)
print "startar algo"
if algon == 'SIRT3D_CUDA' or algon=='CGLS3D_CUDA':
astra.algorithm.run(alg_id,nr_iterations_high)
elif algon=='EGEN_SART':
rec_volume=(width_x.astype(int), width_y.astype(int), width_z.astype(int))
rec_volume=np.zeros(rec_volume,dtype= np.float16)
sides=[rec_volume.shape[0],rec_volume.shape[1],rec_volume.shape[2]]
max_side=np.max(sides)
print max_side
h=relaxation*float(1)/(max_side)#0.00001 change to max size of y o z
vectors=np.matrix('0 0 0 0 0 0 0 0 0 0 0 0',dtype= np.float16)
for ii in range(0,nr_iterations_high):
angles_ind = np.linspace(0, len(angles), len(angles),False)
np.random.shuffle(angles_ind) #shuffle the projection angles to get faster convergence
for jj in angles_ind:
#proj_geom = astra.create_proj_geom('cone', det_spacing, det_spacing, det_size[1], det_size[1], angles[jj], source_to_origin_pixels,origin_to_detector_pixels)
vectors[:,:]=vectors_ROI[jj,:]
proj_geom_SART = astra.create_proj_geom('cone_vec', det_row_count, det_col_count, vectors)
sinogram_id, proj_it = astra.create_sino3d_gpu(rec_volume, proj_geom_SART,vol_geom_ROI)
proj_it=proj_it[:,0,:]
residual=np.zeros((det_row_count,1,det_col_count))
residual[:,0,:]=proj_it-pro_proj[:,jj,:]
astra.data3d.store(sinogram_id, residual)
#h=0.001
[id, rec_volume_it] = astra.create_backprojection3d_gpu(residual, proj_geom_SART, vol_geom_ROI)
rec_volume= rec_volume -h*rec_volume_it
rec_volume=rec_volume*(rec_volume>0)
astra.data3d.delete(id)
astra.data3d.delete(sinogram_id)
else:
print"algorithm is not supported"
print "s**t algo"
if algon[0:4] != 'EGEN':
# Get the result
rec_sub_vol=astra.data3d.get(rec_id_ROI)
# Clean up. Note that GPU memory is tied up in the algorithm object,
# and main RAM in the data objects.
astra.algorithm.delete(alg_id)
astra.data3d.delete(rec_id_ROI)
astra.data3d.delete(sinogram_id_ROI)
return rec_sub_vol
else:
return rec_volume
def recon_multi_part(algon,proj_data,proj_data_low,det_move,angles,x_sub_boundaries,y_sub_boundaries,z_sub_boundaries,vol_size,det_spacing_x,det_spacing_y,x_subvol_width,y_subvol_width,z_subvol_width,n_x,n_y,n_z,nr_iterations_high,source_origin, origin_det,relaxation):
# this part performs the reconstruction of the subvolume.
pro_proj = proj_data - proj_data_low # we only want to have the projection data that comes from the sub volume
det_row_count= pro_proj.shape[0]
det_col_count=pro_proj.shape[2]
vectors_ROI = np.zeros((len(angles), 12)) # this vector will contain the information of the source and detector position.
move_x,move_y,move_z=how_to_move(x_sub_boundaries,y_sub_boundaries,z_sub_boundaries,vol_size,x_subvol_width,y_subvol_width,z_subvol_width,n_x,n_y,n_z)
for i in range(0,len(angles)):
# source
vectors_ROI[i,0] = np.sin(angles[i]) * (source_origin) +move_x
vectors_ROI[i,1] = -np.cos(angles[i]) * (source_origin) + move_y
vectors_ROI[i,2] = move_z
# center of detector
vectors_ROI[i,3] = -np.sin(angles[i]) * (origin_det) + move_x
vectors_ROI[i,4] = np.cos(angles[i]) * (origin_det) + move_y
vectors_ROI[i,5] = move_z +det_move
# vector from detector pixel 0 to 1
vectors_ROI[i,6] = np.cos(angles[i]) *det_spacing_x
vectors_ROI[i,7] = np.sin(angles[i])*det_spacing_x
vectors_ROI[i,8] = 0
# vector from detector pixel (0,0) to (1,0)
vectors_ROI[i,9] = 0
vectors_ROI[i,10] = 0
vectors_ROI[i,11] = det_spacing_y
#ROI=cube[:,:,z_sub_boundaries[n_z,0]:z_sub_boundaries[n_z+1,0]]
# mlab.figure()
# mlab.contour3d(ROI)
width_z=z_sub_boundaries[n_z+1,2]-z_sub_boundaries[n_z,1]
width_y=y_sub_boundaries[n_y+1,2]-y_sub_boundaries[n_y,1]
width_x=x_sub_boundaries[n_x+1,2]-x_sub_boundaries[n_x,1]
vol_geom_ROI= astra.create_vol_geom(width_y.astype(int), width_z.astype(int), width_x.astype(int))
proj_geom_ROI = astra.create_proj_geom('cone_vec', det_row_count, det_col_count, vectors_ROI)
if algon[0:4] != 'EGEN':
# Create a data object for the reconstruction
rec_id_ROI = astra.data3d.create('-vol', vol_geom_ROI)
sinogram_id_ROI = astra.data3d.create('-proj3d', proj_geom_ROI, pro_proj)
if algon=='SIRT':
algon='SIRT3D_CUDA'
elif algon=='CGLS':
algon='CGLS3D_CUDA'
# Set up the parameters for a reconstruction algorithm using the GPU
cfg = astra.astra_dict(algon)
cfg['ReconstructionDataId'] = rec_id_ROI
cfg['ProjectionDataId'] = sinogram_id_ROI
cfg['option']={}
cfg['option']['MinConstraint'] = 0 #attenuation coefficient can't be negative
#cfg['option']['VoxelSuperSampling'] = np.round(float(det_spacing_x)/vol_size,decimals=2)
alg_id = astra.algorithm.create(cfg)
print "startar algo"
if algon == 'SIRT3D_CUDA' or algon=='CGLS3D_CUDA':
astra.algorithm.run(alg_id,nr_iterations_high)
elif algon=='EGEN_SART':
rec_volume=(width_x.astype(int), width_y.astype(int), width_z.astype(int))
rec_volume=np.zeros(rec_volume,dtype= np.float16)
sides=[rec_volume.shape[0],rec_volume.shape[1],rec_volume.shape[2]]
max_side=np.max(sides)
print max_side
h=relaxation*float(1)/(max_side)#0.00001 change to max size of y o z
vectors=np.matrix('0 0 0 0 0 0 0 0 0 0 0 0',dtype= np.float16)
for ii in range(0,nr_iterations_high):
angles_ind = np.linspace(0, len(angles), len(angles),False)
np.random.shuffle(angles_ind) #shuffle the projection angles to get faster convergence
for jj in angles_ind:
#proj_geom = astra.create_proj_geom('cone', det_spacing, det_spacing, det_size[1], det_size[1], angles[jj], source_to_origin_pixels,origin_to_detector_pixels)
vectors[:,:]=vectors_ROI[jj,:]
proj_geom_SART = astra.create_proj_geom('cone_vec', det_row_count, det_col_count, vectors)
sinogram_id, proj_it = astra.create_sino3d_gpu(rec_volume, proj_geom_SART,vol_geom_ROI)
proj_it=proj_it[:,0,:]
residual=np.zeros((det_row_count,1,det_col_count))
residual[:,0,:]=proj_it-pro_proj[:,jj,:]
astra.data3d.store(sinogram_id, residual)
#h=0.001
[id, rec_volume_it] = astra.create_backprojection3d_gpu(residual, proj_geom_SART, vol_geom_ROI)
rec_volume= rec_volume -h*rec_volume_it
rec_volume=rec_volume*(rec_volume>0)
astra.data3d.delete(id)
astra.data3d.delete(sinogram_id)
else:
print"algorithm is not supported"
print "s**t algo"
if algon[0:4] != 'EGEN':
# Get the result
rec_sub_vol=astra.data3d.get(rec_id_ROI)
# Clean up. Note that GPU memory is tied up in the algorithm object,
# and main RAM in the data objects.
astra.algorithm.delete(alg_id)
astra.data3d.delete(rec_id_ROI)
astra.data3d.delete(sinogram_id_ROI)
return rec_sub_vol
else:
return rec_volume
def make_reconsruction(proj_data,number_of_projections,vol_size_ratio,det_size,rec_volume,source_to_origin_pixels,origin_to_detector_pixels,nr_iterations,algo,relaxation,initializer):
vol_geom = astra.create_vol_geom(rec_volume)
angles = np.linspace(0, 2*np.pi, number_of_projections, False)
det_spacing=np.round((float(det_size[0])/det_size[1])/vol_size_ratio,decimals=3)
proj_geom = astra.create_proj_geom('cone', det_spacing, det_spacing, det_size[1], det_size[1], angles, source_to_origin_pixels,origin_to_detector_pixels)
#np.round(float(rec_volume[0])/new_size[0],decimals=2)
if algo[0:4] != 'EGEN':
# Create projection data from this
proj_id=4
# Create a data object for the reconstruction
rec_id = astra.data3d.create('-vol', vol_geom,initializer)
sinogram_id = astra.data3d.create('-proj3d', proj_geom, proj_data)
#Set up the parameters for a reconstruction algorithm using the GPU
cfg = astra.astra_dict(algo)
cfg['ReconstructionDataId'] = rec_id
cfg['ProjectionDataId'] = sinogram_id
cfg['option']={}
cfg['option']['MinConstraint'] = 0 #attenuation coefficient can't be negative
# Create the algorithm object from the configuration structure
alg_id = astra.algorithm.create(cfg)
print "startar algo"
if algo == 'SIRT3D_CUDA' or algo=='CGLS3D_CUDA':
astra.algorithm.run(alg_id,nr_iterations)
elif algo=='FDK_CUDA':
astra.algorithm.run(alg_id)
elif algo=='EGEN_SIRT':
if isinstance( initializer, ( int, long ) ):
rec_volume=np.zeros(rec_volume,dtype= np.float16)
else:
rec_volume=initializer
for ii in range(0,nr_iterations):
sinogram_id, proj_it = astra.create_sino3d_gpu(rec_volume, proj_geom,vol_geom)
residual=proj_it-proj_data
h=relaxation*float(1)/(rec_volume.shape[0]*number_of_projections)
[id, rec_volume_it] = astra.create_backprojection3d_gpu(residual, proj_geom, vol_geom)
rec_volume= rec_volume -h*rec_volume_it
rec_volume=rec_volume*(rec_volume>0)
astra.data3d.delete(id)
astra.data3d.delete(sinogram_id)
elif algo=='EGEN_SART':
if isinstance( initializer, ( int, long ) ):
rec_volume=np.zeros(rec_volume,dtype= np.float16)
else:
rec_volume=initializer
#print rec_volume.shape[0]
for ii in range(0,nr_iterations):
angles_ind = np.linspace(0, number_of_projections, number_of_projections,False)
np.random.shuffle(angles_ind)
for jj in angles_ind:
proj_geom = astra.create_proj_geom('cone', det_spacing, det_spacing, det_size[1], det_size[1], angles[jj], source_to_origin_pixels,origin_to_detector_pixels)
#print np.min(sub_volume_high)
sinogram_id, proj_it = astra.create_sino3d_gpu(rec_volume, proj_geom,vol_geom)
proj_it=proj_it[:,0,:]
residual=np.zeros((det_size[1],1,det_size[1]))
residual[:,0,:]=proj_it-proj_data[:,jj,:]
#print np.shape(proj_it),np.shape(residual),np.shape(proj_data[:,jj,:])
#astra.data3d.store(sinogram_id, residual)
h=relaxation*float(1)/(rec_volume.shape[0])#0.00001
[id, rec_volume_it] = astra.create_backprojection3d_gpu(residual, proj_geom, vol_geom)
rec_volume= rec_volume -h*rec_volume_it
rec_volume=rec_volume*(rec_volume>0)
astra.data3d.delete(id)
astra.data3d.delete(sinogram_id)
# pylab.figure()
# pylab.gray()
# pylab.imshow(rec_volume[:,:,128])
else:
print"algorithm is not supported"
print "s**t algo"
if algo[0:4] != 'EGEN':
# Get the result
rec = astra.data3d.get(rec_id)
# Clean up. Note that GPU memory is tied up in the algorithm object,
# and main RAM in the data objects.
astra.algorithm.delete(alg_id)
astra.data3d.delete(rec_id)
astra.data3d.delete(proj_id)
astra.data3d.delete(sinogram_id)
return rec
else:
return rec_volume
