ott.initializers.linear.initializers.SortingInitializer

ott.initializers.linear.initializers.SortingInitializer#

class ott.initializers.linear.initializers.SortingInitializer(vectorized_update=True, tolerance=0.01, max_iter=100)[source]#

Sorting initializer [Thornton and Cuturi, 2022].

Solve non-regularized OT problem via sorting, then compute potential through iterated minimum on C-transform and use this potential to initialize regularized potential.

Parameters:

  • vectorized_update (bool) – Whether to use vectorized loop.

  • tolerance (float) – DualSort convergence threshold.

  • max_iter (int) – Max DualSort steps.

Methods

init_fu(ot_prob, lse_mode[, rng, init_f])

Apply DualSort algorithm.

init_gv(ot_prob, lse_mode[, rng])

Initialize Sinkhorn potential/scaling g_v.