Skip to main content

Tests PyPI Docs: stable PyTorch License: MIT

bartorch

bartorch puts BART, the Berkeley Advanced Reconstruction Toolbox, inside your Python process. You hand it PyTorch tensors, on the CPU or on a CUDA device, and BART works on them where they are: no .cfl files, no bart executable, no subprocess. What comes back is a tensor, so a BART reconstruction can sit inside any PyTorch code, including the training loop of an unrolled network.

You use BART in one of three ways, from the most ready-made to the most flexible:

  • Tools. Every BART command is a Python function: bt.ecalib(kspace), bt.phantom(128), bt.traj(...). If you know BART's command line, you already know them.
  • Apps. BART's reconstructions (pics, moba, mobafit, POCSENSE) are functions that take k-space and return an image.
  • Operators and solvers. You write the encoding yourself, A = P F S or anything else, compose it with @ and +, and solve it with one of BART's iterations. This is where a reconstruction BART has no command for lives, and it is differentiable.

The three are the same library underneath: apps.pics is an encoding operator and a solver, built for you.

bartorch architecture

How it works

  1. You pass a tensor. A C-order tensor of shape (coils, z, y, x) holds the same bytes as a BART array of dimensions [x, y, z, coils], so bartorch reverses the shape and hands BART the tensor's memory. Nothing is copied.
  2. BART runs in the same process, compiled into the package. When BART reports an error, you get a Python exception (bartorch.BartError) with BART's message, and your interpreter keeps running.
  3. A few parts of BART are swapped for faster ones: every non-uniform FFT is FINUFFT (cuFINUFFT on a card), and the FFT, BLAS and LAPACK come from MKL, PyTorch's own library or SciPy. The numbers are BART's everywhere else.
  4. A CUDA tensor stays on its card. The operators and solvers run there directly, and a tool runs on the card when that has been checked to be safe; otherwise its result is copied back to the card for you.
  5. An operator applied to a tensor that requires a gradient is recorded by autograd, and its backward pass is its adjoint. You can train through a reconstruction.

Tools: BART's commands

import bartorch.tools as bt

kspace = bt.phantom(128, coils=8, kspace=True)  # (coils, z, y, x)
compressed = bt.cc(kspace, p=4)
maps = bt.ecalib(compressed, maps=1)
  • Options are keyword arguments, under BART's long name or its one-letter flag.
  • Where BART takes a bitmask of dimensions, you pass tensor axes: bartorch.fft(x, axes=(-2, -1)).
  • Every command BART has is here, as a function or, for the few with a reason not to be, listed as such in the documentation.

Apps: BART's reconstructions

from bartorch import apps, priors

image = apps.pics(compressed, maps, regularizers=priors.Wavelet((-1, -2), 0.005))
  • A regularization term is an object from bartorch.priors, not a -R string; its axes are tensor axes.
  • apps.pics gives the same bits as BART's pics command, and the documentation's tests hold it to that.

Operators and solvers: custom reconstruction

import torch

from bartorch import linop, optim, priors

A = linop.CartesianSense(maps, (128, 128), pattern)  # A = P F S
image = optim.FISTA(priors.Wavelet((-1, -2), 0.005), maxiter=100)(kspace, A)

B = A @ linop.Diagonal(phase)  # one more factor, still one BART operator
loss = (B(x) - kspace).abs().pow(2).sum()
loss.backward()  # x.grad is 2 B^H (B x - y)
  • A(x) applies the operator, A.H(y) its adjoint, and A.H @ A its normal. A composition is handed to BART as one operator, so a solver iterating on it never comes back to Python.
  • A.plan tells you how BART will run what you composed: which transform, which factors were fused, and whether it fell back to a slower path.
  • An operator you write in Python (a forward and an adjoint) joins a composition like any other.

Command line

The bartorch command takes the same arguments as bart, so a script written for BART runs with the name changed and no BART installation:

bartorch phantom -k -s 8 kspace
bartorch ecalib -m 1 kspace maps
bartorch pics -R W:3:0:0.005 kspace maps image

Installation

pip install bartorch
pip install "bartorch[cu12]"  # with CUDA 12; use cu13 for CUDA 13

Wheels are built for Linux x86-64, macOS arm64 and Windows x86-64; the CUDA build is Linux x86-64 only. The other extras are listed under Optional dependencies, GPU builds under Installation, and source builds in the developer guide.

Learn more

  • Course: from k-space to an image, then parallel imaging, regularization, non-Cartesian and model-based reconstruction, and learned regularization. Every lesson opens in Colab.
  • Documentation, for every function, operator and option.

How to cite

bartorch has no publication of its own. Please cite BART (Uecker et al., doi:10.5281/zenodo.592960) and the papers behind the methods you used (ESPIRiT, compressed sensing, nonlinear inversion, ...), and report the bartorch version. Contributors and citation lists the references.

License

MIT. The embedded BART (BSD-3-Clause), FINUFFT (Apache-2.0) and the other bundled code keep their own licenses; see License and third-party notices. bartorch is an independent project, not affiliated with or endorsed by the BART developers, the PyTorch Foundation or The Linux Foundation. The logo combines BART's mark with the PyTorch logo's flame; PyTorch, the PyTorch logo and any related marks are trademarks of The Linux Foundation.

Metadata

Release files for bartorch 0.0.22

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for bartorch 0.0.22
File Size Uploaded Provenance Actions
bartorch-0.0.22.tar.gz 7.4 MB With provenance Details for bartorch-0.0.22.tar.gz

Built distributions (wheels)

Table of built distributions (wheels) for bartorch 0.0.22
File Interpreter ABI Platform Actions
bartorch-0.0.22-py3-none-win_amd64.whl Python 3 none Windows x86-64 Details for bartorch-0.0.22-py3-none-win_amd64.whl
bartorch-0.0.22-py3-none-manylinux_2_28_x86_64.whl Python 3 none Linux glibc 2.28+ x86-64 Details for bartorch-0.0.22-py3-none-manylinux_2_28_x86_64.whl
bartorch-0.0.22-py3-none-macosx_11_0_arm64.whl Python 3 none macOS 11.0+ ARM64 Details for bartorch-0.0.22-py3-none-macosx_11_0_arm64.whl

Total release size: 31.3 MB

Release files / bartorch-0.0.22.tar.gz

Download URL bartorch-0.0.22.tar.gz
Size 7.4 MB
Tags Source
SHA-256 checksum
How to use checksums
51e7604fd5b610e2c799e7a8f6609a12e6d34cafc821715c1c8dbe4d10848c2b
BLAKE2b-256 checksum
How to use checksums
67e848793e3c4436b2c89e7177db98b756f355165b4f513b2756dd6bb18100a9
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 4, 2026.

Transparency log

Release files / bartorch-0.0.22-py3-none-win_amd64.whl

Download URL bartorch-0.0.22-py3-none-win_amd64.whl
Size 7.7 MB
Tags Python 3 Windows x86-64
SHA-256 checksum
How to use checksums
9d61c52987c213c7c94078daa45751a31deac6ebb044da740893c3508389234c
BLAKE2b-256 checksum
How to use checksums
f16ed6991efe6ecc03008878807361c4ba2bc3f2fa7db9b16e118d4b28ad0635
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 4, 2026.

Transparency log

Release files / bartorch-0.0.22-py3-none-manylinux_2_28_x86_64.whl

Download URL bartorch-0.0.22-py3-none-manylinux_2_28_x86_64.whl
Size 13.2 MB
Tags Linux glibc 2.28+ x86-64 Python 3
SHA-256 checksum
How to use checksums
3fa06b7df44216a1d97d217c812dc99c878f84fbc97427d380bfedb8089e118a
BLAKE2b-256 checksum
How to use checksums
2967ae7f0c113935e57f3d65b2769346bf929c590e8950c0d3dc13cbbc64e759
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 4, 2026.

Transparency log

Release files / bartorch-0.0.22-py3-none-macosx_11_0_arm64.whl

Download URL bartorch-0.0.22-py3-none-macosx_11_0_arm64.whl
Size 3.0 MB
Tags Python 3 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
e8049d0d264bbf486c6e1bb77effea4960981fbab58e76b6818147417f33cb29
BLAKE2b-256 checksum
How to use checksums
e9c03b65cc4306e68118b1d085c32a31fdb8adb38027e01eac4a9596732b82d8
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 4, 2026.

Transparency log

Security for bartorch 0.0.22

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

All 4 files in this release have been uploaded with provenance metadata.

Select from 4 files

Attestation for bartorch-0.0.22.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 4, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page