neuralqx.optimizer.types module

class OptaxSchedule(*args, **kwargs)

Bases: Protocol

A Protocol so we can type the schedule callable that Optax returns.

class ExponentialDecay(init_value, transition_steps, decay_rate, staircase=True, end_value=None)

Bases: object

staircase: bool = True

end_value: float | None = None

class CosineDecay(init_value, decay_steps, alpha=0.0)

Bases: object

alpha: float = 0.0

class LinearDecay(init_value, end_value, transition_steps, transition_begin=0)

Bases: object

transition_begin: int = 0

class Adam(b1=0.9, b2=0.999, eps=1e-08)

Bases: object

b1: float = 0.9

b2: float = 0.999

eps: float = 1e-08

class SGD(momentum=0.0, nesterov=False)

Bases: object

momentum: float = 0.0

nesterov: bool = False

Momentum

alias of SGD

class Adagrad(eps=1e-07, initial_accumulator_value=0.1)

Bases: object

eps: float = 1e-07

initial_accumulator_value: float = 0.1

class RMSProp(decay=0.9, eps=1e-07, centered=False)

Bases: object

decay: float = 0.9

eps: float = 1e-07

centered: bool = False

class AdaBelief(b1=0.9, b2=0.999, eps=1e-16, eps_root=1e-16)

Bases: object

b1: float = 0.9

b2: float = 0.999

eps: float = 1e-16

eps_root: float = 1e-16

class AMSGrad(b1=0.9, b2=0.999, eps=1e-08, eps_root=0.0, mu_dtype=None)

Bases: object

b1: float = 0.9

b2: float = 0.999

eps: float = 1e-08

eps_root: float = 0.0

mu_dtype: Any | None = None

class Yogi(b1=0.9, b2=0.999, eps=0.001)

Bases: object

b1: float = 0.9

b2: float = 0.999

eps: float = 0.001