These repositories are small self-contained tools written in pure PyTorch, that I have found useful in many projects.
They are (relatively) stable, as backward-compatible as possible with respect to PyTorch versions, and can be used as core dependencies to higher level projects.
Note
The last package, jitfields, reimplements many of the utilities from the other core
packages, but does it directly in CUDA/C++.
The CUDA/C++ sources are compiled
just-in-time using cupy
and cppyy.
These packages underpin my research in medical image computing.
In general, my aim is to write a set of mid-level packages that specialize in various tasks (data augmentation, network architectures, modality-specific tasks, etc.).
| Package | Description | Readiness |
|---|---|---|
cornucopia |
An abundance of augmentation layers | 🟢 |
nitorch |
An (overweight and poorly maintained) package for everything neuroimaging | 🟠 |
synthsurf |
Surface-based image synthesis and PyTorch utilities for triangular surfaces | 🟠 |
synthspline |
Synthetic tubular structures (vessels, axons) for NN pretraining | 🟠 |
cassetta |
A deep learning toolbox (under early development) | 🔴 |
braindataprep |
Download, bidsify and preprocess public datasets (work-in-progress) | 🔴 |
| Package | Description | Readiness |
|---|---|---|
variational_staple |
STAPLE and variants | 🟢 |
optimal_affine |
Build optimal "subject to mean space" affines from "subject to subject" pairwise affines | 🟢 |
metrics |
A bunch of metrics | 🔴 |
| Package | Description | Readiness |
|---|---|---|
spm_mni_align |
SPM toolbox to align an image to SPM's template space | 🟠 |
multi-bias |
Fit a multi-view bias field | 🟢 |
super-resolution |
MTV-based denoising/super-resolution | 🟢 |
cmaps |
(Some) Matplotlib colormaps in Matlab | 🟢 |
