ott.solvers.linear.acceleration.Momentum

class ott.solvers.linear.acceleration.Momentum(start=0, error_threshold=inf, value=1.0, inner_iterations=1)

Momentum for Sinkhorn updates.

Can be either constant [Thibault et al., 2021] or adaptive [Lehmann et al., 2021].

Parameters:

• start (int)

• error_threshold (float)

• value (float)

• inner_iterations (int)

Methods

lehmann(state)	Momentum formula [Lehmann et al., 2021], eq.
weight(state, iteration)	Compute momentum term if needed, using previously seen errors.

Attributes

error_threshold
inner_iterations
start
value