ott.solvers.linear.acceleration.AndersonAcceleration#

class ott.solvers.linear.acceleration.AndersonAcceleration(memory=2, refresh_every=1, ridge_identity=0.01)[source]#

Implements Anderson acceleration for Sinkhorn.

Methods

extrapolation(xs, fxs)

Compute Anderson extrapolation from past observations.

init_maps(pb, state)

Initialize log matrix used in Anderson acceleration with nan values.

update(state, iteration, pb, lse_mode)

Anderson acceleration update.

update_history(state, pb, lse_mode)

rtype

SinkhornState

Attributes

memory

refresh_every

ridge_identity

Parameters

  • memory (int) –

  • refresh_every (int) –

  • ridge_identity (float) –

Return type

None