ribs.emitters.opt.AdamOpt¶

class ribs.emitters.opt.AdamOpt(theta0: ArrayLike, lr: float | floating = 0.001, beta1: float | floating = 0.9, beta2: float | floating = 0.999, epsilon: float | floating = 1e-08, l2_coeff: float | floating = 0.0)[source]¶

Adam optimizer.

Refer to Kingma and Ba 2014 for more information on hyperparameters.

Parameters:¶

theta0: ArrayLike¶

Initial solution. 1D array.

lr: float | floating = 0.001¶

Learning rate for the update.

beta1: float | floating = 0.9¶

Exponential decay rate for the moment estimates.

beta2: float | floating = 0.999¶

Another exponential decay rate for the moment estimates.

epsilon: float | floating = 1e-08¶

Hyperparameter for numerical stability.

l2_coeff: float | floating = 0.0¶

Coefficient for L2 regularization. Note this is not the same as “weight decay” – see this blog post <https://www.fast.ai/posts/2018-07-02-adam-weight-decay.html>_ and Loshchilov and Hutler 2019 <https://arxiv.org/abs/1711.05101>_ for more info.

Methods

reset(theta0)	Resets the solution point to a new value.
step(gradient)	Ascends the solution based on the given gradient.

Attributes

theta

The current solution point.

reset(theta0: ArrayLike) → None[source]¶

Resets the solution point to a new value.

Parameters:¶

theta0: ArrayLike¶

The new solution point. 1D array.

step(gradient: ArrayLike) → None[source]¶

Ascends the solution based on the given gradient.

Parameters:¶

gradient: ArrayLike¶

The (estimated) gradient of the current solution point. 1D array.

property theta : ndarray¶

The current solution point.