ott.initializers.linear.initializers.SubsampleInitializer

ott.initializers.linear.initializers.SubsampleInitializer#

class ott.initializers.linear.initializers.SubsampleInitializer(subsample_n_x, subsample_n_y=None, **kwargs)[source]#

Subsample initializer [Thornton and Cuturi, 2022].

Subsample each PointCloud, then compute Sinkhorn potential from the subsampled approximations and use this potential to initialize Sinkhorn potentials/scalings for the original problem.

Parameters:
  • subsample_n_x (int) – number of points to subsample from the first measure in PointCloud.

  • subsample_n_y (Optional[int]) – number of points to subsample from the second measure in PointCloud. If None, use subsample_n_x.

  • kwargs (Any) – Keyword arguments for Sinkhorn.

Methods

init_dual_a(ot_prob, lse_mode[, rng])

Initialize Sinkhorn potential/scaling f_u.

init_dual_b(ot_prob, lse_mode[, rng])

Initialize Sinkhorn potential/scaling g_v.