Marco Cuturi
Marco Cuturi
This refers to issue #156. This PR introduces a few changes: - old`Euclidean` is moved to `SqEuclidean` and continues wokring with naive **2 and vdot - `Euclidean` is now the...
**Describe the bug** We chose originally to implement the squared Euclidean distance as `jnp.sum(x**2,axis=-1) + jnp.sum(y**2,axis=-1) - 2 * jnp.vdot(x,y)`. Although this works with `power=2.0` (which leaves it unchanged) this...
this PR implements an extension of the univariate solver that outputs dual vectors, regardless of input sizes / weights. the implementation uses a lineax sparse solver, instead of going through...
Given the newfound prominence of several neural approaches that go beyond the W2 / ICNN case, we are currently refactoring the ``neural`` part of OTT-JAX. This is going to break...
@JTT94 , shouldn't we consider adding the sorting initializer by default to soft-sort?
in one way or another... Use case: when sinkhorn is run, the marginal deviation is used as a stopping criterion. By default the 1-norm is used: https://github.com/ott-jax/ott/blob/3ebac6369acfe848930e1e2abb69b61ecea7e5cc/src/ott/solvers/linear/sinkhorn.py#L455 and fed into...
At the moment, the KL loss in the GW solver applies directly to the cost matrices. This is not really the way we envisioned it in the original paper http://proceedings.mlr.press/v48/peyre16.pdf,...
As suggested by @bosr, provide functions to subset transport matrices (i.e. materialize lines, or columns, or a block of `out.matrix` without requiring the materialization of the entire matrix. At this...
At the moment, only default options are considered, and we know that they are suboptimal in most cases.