bartorch

bartorch#

bartorch runs BART, the Berkeley Advanced Reconstruction Toolbox, inside the Python process on torch tensors: BART’s applications as functions, its linear and nonlinear operators as composable objects, and its iterative solvers as classes. It computes nothing BART does not, except where it substitutes a faster implementation: FINUFFT for every non-Cartesian transform, and a SENSE operator that applies its coils in batches.

Install#

pip install bartorch
pip install "bartorch[mkl]"          # Linux: MKL for BART's BLAS, LAPACK and FFT
pip install "bartorch[cufinufft]"    # non-Cartesian transforms on a CUDA device

Wheels are built for Linux x86_64 and macOS on Apple silicon; elsewhere pip builds from source, which needs clang or GCC 14+ and CMake. The installation guide covers CUDA builds and platform notes.

Where to start#

Task

Start here

Reconstruct with a BART application

bartorch.tools: pics, nlinv, ecalib

Build an encoding and solve it

bartorch.linop and bartorch.optim

Transform, filter or register arrays

the functions in bartorch

Fit a torch signal model

bartorch.nlop and bartorch.optim.IRGNM

Quickstart#

import bartorch
import bartorch.tools as bt
from bartorch import linop, optim, prox

kspace = bt.phantom(128, coils=8, kspace=True)        # (8, 1, 128, 128)
maps = bt.ecalib(kspace, maps=1)

# BART's own application ...
image = bt.pics(kspace, maps, regularizers=prox.Wavelet((-1, -2), 0.005), solver="fista")

# ... or the same problem assembled from an operator and a solver.
# pics also scales the data first; see optim.data_scaling.
A = linop.CartesianSense(maps.squeeze(1), (8, 128, 128))
x = optim.FISTA(prox.Wavelet((-1, -2), 0.005), maxiter=50)(kspace, A)

spectrum = bartorch.fft(bt.phantom(128), axes=(-2, -1), unitary=True)

Shapes are C order, so the last axis is the one BART calls the first, and wherever BART takes a bitmask or a dimension number a function here takes axis indices; pics takes bartorch.prox terms where BART takes -R strings.

License#

MIT. BART is distributed under its own BSD license (external/bart/LICENSE); pocketfft (BSD-3) and BlocksRuntime (MIT) are vendored under external/ with their licenses.