CGLS_CUDA

This is a GPU implementation of the Conjugate Gradient Least Squares (CGLS) algorithm for 2D data sets. It takes projection data and an initial reconstruction as input, and returns the reconstruction after a specified number of iterations.

The internal state of the CGLS algorithm is reset every time astra.algorithm.run/astra_mex_algorithm('iterate') is called. This implies that running CGLS for N iterations and then running it for another N iterations may yield different results from running it 2N iterations at once.

Supported geometries: parallel, parallel_vec, fanflat, fanflat_vec.

Configuration options

Name

Description

ProjectionDataId

Projection data object ID.

ReconstructionDataId

ID of data object to store the result. The content of this data is used as the initial reconstruction.

option.ReconstructionMaskId

If specified, data object ID of a volume-data-sized volume to be used as a mask.

option.DetectorSuperSampling

During forward projection, each detector element will be subdivided by this factor along each dimension. This should only be used if detector elements are larger than the pixels in the volume (default: 1).

option.PixelSuperSampling

During backprojection, each pixel in the volume will be subdivided by this factor along each dimension. This should only be used if pixels in the volume are larger than the detector elements (default: 1).

option.GPUindex

The index of the GPU to use (default: 0).

Example

import astra
import matplotlib.pyplot as plt
import numpy as np

# Create geometries and projector
N = 256
N_ANGLES = 180
det_spacing = 1.0
angles = np.linspace(0, np.pi, N_ANGLES)
proj_geom = astra.create_proj_geom('parallel', det_spacing, N, angles)
vol_geom = astra.create_vol_geom(N, N)
projector_id = astra.create_projector('cuda', proj_geom, vol_geom)

# Generate phantom image
phantom_id, phantom = astra.data2d.shepp_logan(vol_geom)

# Create forward projection
sinogram_id, sinogram = astra.create_sino(phantom_id, projector_id)

# Reconstruct
recon_id = astra.data2d.create('-vol', vol_geom)
cfg = astra.astra_dict('CGLS_CUDA')
cfg['ProjectionDataId'] = sinogram_id
cfg['ReconstructionDataId'] = recon_id
algorithm_id = astra.algorithm.create(cfg)

astra.algorithm.run(algorithm_id, iterations=100)

reconstruction = astra.data2d.get(recon_id)
plt.imshow(reconstruction, cmap='gray')

# Clean up
astra.data2d.delete([sinogram_id, recon_id, phantom_id])
astra.projector.delete(projector_id)
astra.algorithm.delete(algorithm_id)

Extra features

CGLS_CUDA supports astra.algorithm.get_res_norm() / astra_mex_algorithm('get_res_norm') command to get the 2-norm of the residual for the reconstruction (the square root of the sum of squares of differences between the input and the projection of the result).