moscot.backends.ott.SinkhornSolver¶

class moscot.backends.ott.SinkhornSolver(jit=True, rank=-1, epsilon=0.0, initializer=None, initializer_kwargs=mappingproxy({}), **kwargs)[source]¶

Solver for the linear problem.

The (Kantorovich relaxed) OT problem is defined by two distributions in the same space. The aim is to obtain a probabilistic map from the source distribution to the target distribution such that the (weighted) sum of the distances between coupled data point in the source and the target distribution is minimized.

Parameters:

jit (bool) – Whether to jit() the solver.
rank (int) – Rank of the solver. If -1, use Sinkhorn [Cuturi, 2013], otherwise, use LRSinkhorn [Scetbon et al., 2021].
epsilon (float) – Additional epsilon regularization for the low-rank approach.
initializer (Union[Literal['default', 'gaussian', 'sorting'], Literal['random', 'rank2', 'k-means', 'generalized-k-means'], None]) – Initializer for Sinkhorn or LRSinkhorn, depending on the rank.
initializer_kwargs (Mapping[str, Any]) – Keyword arguments for the initializer.
kwargs (Any) – Keyword arguments for Sinkhorn or LRSinkhorn, depending on the rank.

Methods

__call__([xy, x, y, tags, device])

Solve an optimal transport problem.

Attributes

`is_low_rank`	Whether the `solver` is low-rank.
`problem_kind`	Problem kind this solver handles.
`rank`	Rank of the `solver`.
`solver`	`ott` solver.
`xy`	Geometry defining the linear term.