Array functions and settings

The functions of the bartorch namespace. Shapes are C order, and an axis argument is an index into a tensor’s shape.

Fourier transforms

fft	Fourier transform along axes, centred by default.
ifft	Inverse Fourier transform along axes: fft() with inverse=True.
fftshift	Move the zero frequency to the middle along axes, or back with inverse.
fftmod	Multiply by the alternating phase that centres an uncentred transform, along axes.
nufft	Non-uniform Fourier transform of an image to samples along traj.
nufft_adjoint	Adjoint of nufft(): samples along traj back to an image.

Wavelet transforms

fwt	Forward multi-level discrete wavelet transform along axes.
iwt	Wavelet synthesis along axes: coefficients from fwt() back to samples.

Thresholding

soft_thresh	Soft thresholding, x * max(1 - lamda / |x|, 0).
hard_thresh	Hard thresholding: keep x where |x| > lamda, zero elsewhere (-H).

Array utilities

resize	Crop or zero-pad input to oshape.
flip	Reverse input along axes.
circshift	Cyclic shift along axes: element i moves to (i + shift) % n, as torch.roll.
conv	Cyclic convolution of input with kernel along axes, of the input's shape.
window	Multiply by a Hamming window along each of axes, or a Hann window (-H).
median_filter	Median over each window of length consecutive elements along axis (-m, -l).
moving_average	Mean over each window of length consecutive elements along axis (-a, -l).
normalize	Divide by the l2 norm over axes, or the l1 norm (-b), per index of the other axes.
mip	Maximum (or minimum, -m) intensity projection over axes, which are kept as size one.
unwrap	Unwrap the real part along axis: remove the jumps a period of 2 * bound puts in it.
casorati	Casorati matrix of the overlapping blocks of input.
rss	Root of the sum of squared magnitudes over axes.

Interpolation

interpolate	Sample input at the voxel positions coord along axes.
warp	Pull input through a displacement field: out[p] = input(p + displacement[p]).
affine_transform	Resample input on an affinely mapped grid: out(u) = input(matrix @ [u, 1]).
fovshift	Shift the field of view of k-space data by a linear phase.

Registration

register_affine	Affine transform that maps reference onto moved, by mutual information.
register_nonrigid	Displacement field from reference into moved, by greedy SyN or TV-L1 optical flow.
estimate_shift	Sub-voxel shift that moves b onto a, from the phase of their cross-spectrum.

Image-quality metrics

nrmse	Normalized root mean square error ‖input - reference‖ / ‖reference‖.
mse	Mean squared error mean |input - reference|² over every element (--mse).
ssim	Mean structural similarity of the magnitudes (--ssim).
psnr	Peak signal-to-noise ratio of the magnitudes in decibels, averaged over images (--psnr).
roi_stat	Statistic of input over the region roi.

Coil kernels

maps_to_kernels	Coil sensitivities as k-space kernels: the centre of each map's unitary spectrum.
kernels_to_maps	The maps a kernel bank stands for: zero-padded to spatial and transformed back.

Interoperability

to_deepinv

A LinearOperator as a deepinv.physics.LinearPhysics.

Errors

BartError

A BART command or operator failed; the message is BART's.

Runtime settings

set_num_threads	Set the number of threads BART, its FFT and FINUFFT's host transforms use.
set_copy_inputs	Whether a command works on a private copy of each input, which is the default.
set_debug_level	Set BART's verbosity: 0 errors, 1 warnings, 2 info, 3 and up debug.
get_debug_level

CUDA

cuda_available	Whether BART can run on a CUDA device: built with CUDA, and a device present.
set_cuda_streams	Set the number of CUDA streams BART runs on, 1 to 8.
use_cuda_memcache	Whether BART keeps freed device memory for its own reuse.

Build information

bart_version	Version string of the embedded BART.
build_info	How the library was built: BART version, compiler, nested-function mode, CUDA.
backend_sources	The library serving each BLAS and LAPACK routine, and the FFT.