a2aa306b49
* fix optimizer docstring * fix comments
170 lines
6.1 KiB
Python
170 lines
6.1 KiB
Python
from keras_core import operations as ops
|
|
from keras_core.api_export import keras_core_export
|
|
from keras_core.optimizers import optimizer
|
|
|
|
|
|
@keras_core_export(["keras_core.optimizers.AdamW"])
|
|
class AdamW(optimizer.Optimizer):
|
|
"""Optimizer that implements the AdamW algorithm.
|
|
|
|
AdamW optimization is a stochastic gradient descent method that is based on
|
|
adaptive estimation of first-order and second-order moments with an added
|
|
method to decay weights per the techniques discussed in the paper,
|
|
'Decoupled Weight Decay Regularization' by
|
|
[Loshchilov, Hutter et al., 2019](https://arxiv.org/abs/1711.05101).
|
|
|
|
According to
|
|
[Kingma et al., 2014](http://arxiv.org/abs/1412.6980),
|
|
the underying Adam method is "*computationally
|
|
efficient, has little memory requirement, invariant to diagonal rescaling of
|
|
gradients, and is well suited for problems that are large in terms of
|
|
data/parameters*".
|
|
|
|
Args:
|
|
learning_rate: A float, a
|
|
`keras_core.optimizers.schedules.LearningRateSchedule` instance, or
|
|
a callable that takes no arguments and returns the actual value to
|
|
use. The learning rate. Defaults to 0.001.
|
|
beta_1: A float value or a constant float tensor, or a callable
|
|
that takes no arguments and returns the actual value to use. The
|
|
exponential decay rate for the 1st moment estimates.
|
|
Defaults to 0.9.
|
|
beta_2: A float value or a constant float tensor, or a callable
|
|
that takes no arguments and returns the actual value to use. The
|
|
exponential decay rate for the 2nd moment estimates.
|
|
Defaults to 0.999.
|
|
epsilon: A small constant for numerical stability. This epsilon is
|
|
"epsilon hat" in the Kingma and Ba paper (in the formula just
|
|
before Section 2.1), not the epsilon in Algorithm 1 of the paper.
|
|
Defaults to 1e-7.
|
|
amsgrad: Boolean. Whether to apply AMSGrad variant of this algorithm
|
|
from the paper "On the Convergence of Adam and beyond".
|
|
Defaults to `False`.
|
|
{{base_optimizer_keyword_args}}
|
|
|
|
References:
|
|
|
|
- [Loshchilov et al., 2019](https://arxiv.org/abs/1711.05101)
|
|
- [Kingma et al., 2014](http://arxiv.org/abs/1412.6980) for `adam`
|
|
- [Reddi et al., 2018](
|
|
https://openreview.net/pdf?id=ryQu7f-RZ) for `amsgrad`.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
learning_rate=0.001,
|
|
weight_decay=0.004,
|
|
beta_1=0.9,
|
|
beta_2=0.999,
|
|
epsilon=1e-7,
|
|
amsgrad=False,
|
|
clipnorm=None,
|
|
clipvalue=None,
|
|
global_clipnorm=None,
|
|
use_ema=False,
|
|
ema_momentum=0.99,
|
|
ema_overwrite_frequency=None,
|
|
name="adamw",
|
|
):
|
|
super().__init__(
|
|
learning_rate=learning_rate,
|
|
name=name,
|
|
clipnorm=clipnorm,
|
|
clipvalue=clipvalue,
|
|
global_clipnorm=global_clipnorm,
|
|
use_ema=use_ema,
|
|
ema_momentum=ema_momentum,
|
|
ema_overwrite_frequency=ema_overwrite_frequency,
|
|
)
|
|
self.weight_decay = weight_decay
|
|
self.beta_1 = beta_1
|
|
self.beta_2 = beta_2
|
|
self.epsilon = epsilon
|
|
self.amsgrad = amsgrad
|
|
|
|
if self.weight_decay is None:
|
|
raise ValueError(
|
|
"Argument `weight_decay` must be a float. Received: "
|
|
"weight_decay=None"
|
|
)
|
|
|
|
def build(self, var_list):
|
|
"""Initialize optimizer variables.
|
|
|
|
AdamW optimizer has 3 types of variables: momentums, velocities and
|
|
velocity_hat (only set when amsgrad is applied),
|
|
|
|
Args:
|
|
var_list: list of model variables to build AdamW variables on.
|
|
"""
|
|
if self.built:
|
|
return
|
|
super().build(var_list)
|
|
self._momentums = []
|
|
self._velocities = []
|
|
for var in var_list:
|
|
self._momentums.append(
|
|
self.add_variable_from_reference(
|
|
reference_variable=var, name="m"
|
|
)
|
|
)
|
|
self._velocities.append(
|
|
self.add_variable_from_reference(
|
|
reference_variable=var, name="v"
|
|
)
|
|
)
|
|
if self.amsgrad:
|
|
self._velocity_hats = []
|
|
for var in var_list:
|
|
self._velocity_hats.append(
|
|
self.add_variable_from_reference(
|
|
reference_variable=var, name="vhat"
|
|
)
|
|
)
|
|
|
|
def update_step(self, gradient, variable, learning_rate):
|
|
"""Update step given gradient and the associated model variable."""
|
|
beta_1_power = None
|
|
beta_2_power = None
|
|
lr = ops.cast(learning_rate, variable.dtype)
|
|
gradient = ops.cast(gradient, variable.dtype)
|
|
local_step = ops.cast(self.iterations + 1, variable.dtype)
|
|
beta_1_power = ops.power(
|
|
ops.cast(self.beta_1, variable.dtype), local_step
|
|
)
|
|
beta_2_power = ops.power(
|
|
ops.cast(self.beta_2, variable.dtype), local_step
|
|
)
|
|
|
|
m = self._momentums[self._get_variable_index(variable)]
|
|
v = self._velocities[self._get_variable_index(variable)]
|
|
|
|
alpha = lr * ops.sqrt(1 - beta_2_power) / (1 - beta_1_power)
|
|
|
|
m.assign(m + (gradient - m) * (1 - self.beta_1))
|
|
v.assign(v + (ops.square(gradient) - v) * (1 - self.beta_2))
|
|
if self.amsgrad:
|
|
v_hat = self._velocity_hats[self._get_variable_index(variable)]
|
|
v_hat.assign(ops.maximum(v_hat, v))
|
|
v = v_hat
|
|
variable.assign(variable - (m * alpha) / (ops.sqrt(v) + self.epsilon))
|
|
|
|
def get_config(self):
|
|
config = super().get_config()
|
|
|
|
config.update(
|
|
{
|
|
"weight_decay": self.weight_decay,
|
|
"beta_1": self.beta_1,
|
|
"beta_2": self.beta_2,
|
|
"epsilon": self.epsilon,
|
|
"amsgrad": self.amsgrad,
|
|
}
|
|
)
|
|
return config
|
|
|
|
|
|
AdamW.__doc__ = AdamW.__doc__.replace(
|
|
"{{base_optimizer_keyword_args}}", optimizer.base_optimizer_keyword_args
|
|
)
|