AndersonAccelerationConfig#

class deepinv.optim.AndersonAccelerationConfig(history_size=10, beta=0.9, eps=0.1, full_backprop=False)[source]#

Configuration parameters for Anderson acceleration of a fixed-point algorithm.

Parameters:

history_size (int) – Number of past iterates used in Anderson acceleration.
beta (float) – Momentum coefficient in Anderson acceleration.
eps (float) – Regularization parameter for Anderson acceleration.
full_backprop (bool) – Compute backpropagation through all iterates of Anderson acceleration instead of the last iterate only. Default: False.