DEQConfig

class deepinv.optim.DEQConfig(jacobian_free=False, anderson_acceleration_backward=False, history_size_backward=5, beta_backward=1.0, eps_backward=0.0001, max_iter_backward=50)

Bases: object

Configuration parameters for Deep Equilibrium models.

Parameters:

• jacobian_free (bool) – Whether to use a Jacobian-free backward pass (see Fung et al.[1]).

• anderson_acceleration_backward (bool) – Whether to use Anderson acceleration for solving the backward equilibrium.

• history_size_backward (int) – Number of past iterates used in Anderson acceleration for the backward pass.

• beta_backward (float) – Momentum coefficient in Anderson acceleration for the backward pass.

• eps_backward (float) – Regularization parameter for Anderson acceleration in the backward pass.

• max_iter_backward (int) – Maximum number of iterations in the backward equilibrium solver.

---

References:

[1]

Samy Wu Fung, Howard Heaton, Qiuwei Li, Daniel McKenzie, Stanley Osher, and Wotao Yin. Jfb: jacobian-free backpropagation for implicit networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 6648–6656. 2022.