Software

I try to make software as available as is practical. If you need other stuff from me, just send me an e-mail.

Software from the team and I

The team of wonderful people with whom I work often make code available. I have written only a tiny bit of the code below and most of the stuff I actually have written should be considered outdated. But the team produces fresh and crisp code, so I link to it here even if I did not produce it myself.

Large-scale linear algebra

We have developed a series of numerical routines for matrix-free linear algebra with Jax. These routines are essential building blocks in our Bayesian approximations. Do check out the matfree library.

VIKING

VIKING is a pretty neat large-scale variational family for Bayesian deep learning. Check out the Jax code. This also includes code for alternating projects.

Grassmann VAEs

Do you need a generative model that emit subspaces? Do Grassmann VAE provides a neat such model. Code is here.

Stochastic manifolds

The key algorithms and data structures we have used for working with the geometry of learned manifolds is available in StochMan. This code is based on pytorch, and the team has now moved on to the superiority of Jax, so we no longer maintain StochMan. See e.g. this code for something Jax based.

Diffeomorphic Deformations

The code we use for diffeomorphic deformations is available through libcpab. Through this we provide an easy-to-use interface from numpy, pytorch and tensorflow. Both CPU and GPU computations are supported as well as automatic differentiation. So using diffeomorphisms have never been easier.

Subspace Estimation

The software for computing Grassmann Averages are available at the project webpage.

Pyplot + pytorch wrapper

I got sick and tired of typing plt.plot(x.detach().cpu().numpy(), y.detach().cpu().numpy(), color="green"), so I have created a little pyplot wrapper that converts torch tensors automatically to numpy arrays. Here's the code; let me know if you have smarter ways of getting the same result!

Metric Learning

The software developed for the NeurIPS 2012 paper used for regression and dimensionality reduction using multiple metrics can be found at the project website. This is really only useful for historical purposes and should not be used for new projects.

Differential Equations

The ODE solver used both in the AISTATS 2014 and MICCAI 2014/2015 papers is being further developed at http://probabilistic-numerics.org/. Our old solver should not be used at this point as the field has produced vastly superior methods.

Pose Estimation

Most of the software used in our papers in articulated tracking / pose estimate / whatever-you-call-it is available at the project website. This is terribly arcane code; please only use it for historical purposes.