Releases · jeanfeydy/geomloss

Major update in the backend

We released KeOps v2.0 last week (release notes available here).
This minor GeomLoss update is a compatibility fix for KeOps v2.0 and PyTorch v1.10 deprecation warnings.

What should you expect for GeomLoss v0.3?

KeOps v2.0 is a major upgrade that had been brewing for a long time: it unlocks many possibilities for downstream projects such as GeomLoss. In parallel, I have now secured a permanent research position at Inria Paris and will commit to software development long-term.

I will tidy up GeomLoss over the next few weeks and finally include:

Fixes for the GitHub issues.
Proper support for images and Wasserstein barycenters.
A new, cleaner API (without breaking backward compatibility).

The long-term plan is to:

Add GeomLoss as an optional backend for Python Optimal Transport.
Maintain an open benchmarking platform inspired by ANN Benchmarks at optimal-transport-benchmarks.com.

Feel free to contact me if you would like to discuss this further :-)

Other changes

Corrected misspelling by @ashwhall in #41
"online'' routine now supports the batch dimension by @wzm2256 in #39
Updated check_shapes() to return updated variables l_x, α, l_y, β by @RaghuSpaceRajan in #22
Update sinkhorn_divergence.py by @AdrienCorenflos in #19