Fix API export for non-namex case (#33)
This commit is contained in:
Ian Stenbit
Francois Chollet
parent
009803fd49
commit
ce4ecc6cc2
@ -42,7 +42,7 @@ else:
|
||||
|
||||
class keras_core_export:
|
||||
def __init__(self, path):
|
||||
pass
|
||||
self.path = path
|
||||
|
||||
def __call__(self, symbol):
|
||||
register_internal_serializable(self.path, symbol)
|
||||
|
||||
@ -1,144 +0,0 @@
|
||||
from keras_core import operations as ops
|
||||
from keras_core.optimizers import optimizer
|
||||
|
||||
|
||||
class Adam(optimizer.Optimizer):
|
||||
"""Optimizer that implements the Adam algorithm.
|
||||
|
||||
Adam optimization is a stochastic gradient descent method that is based on
|
||||
adaptive estimation of first-order and second-order moments.
|
||||
|
||||
According to
|
||||
[Kingma et al., 2014](http://arxiv.org/abs/1412.6980),
|
||||
the 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 `Tensor`, floating point value, a schedule that is a
|
||||
`tf.keras.optimizers.schedules.LearningRateSchedule`, 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}}
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
learning_rate=0.001,
|
||||
beta_1=0.9,
|
||||
beta_2=0.999,
|
||||
epsilon=1e-7,
|
||||
amsgrad=False,
|
||||
weight_decay=None,
|
||||
clipnorm=None,
|
||||
clipvalue=None,
|
||||
global_clipnorm=None,
|
||||
use_ema=False,
|
||||
ema_momentum=0.99,
|
||||
ema_overwrite_frequency=None,
|
||||
name="Adam",
|
||||
**kwargs
|
||||
):
|
||||
super().__init__(
|
||||
learning_rate=learning_rate
|
||||
name=name,
|
||||
weight_decay=weight_decay,
|
||||
clipnorm=clipnorm,
|
||||
clipvalue=clipvalue,
|
||||
global_clipnorm=global_clipnorm,
|
||||
use_ema=use_ema,
|
||||
ema_momentum=ema_momentum,
|
||||
ema_overwrite_frequency=ema_overwrite_frequency,
|
||||
**kwargs
|
||||
)
|
||||
self.beta_1 = beta_1
|
||||
self.beta_2 = beta_2
|
||||
self.epsilon = epsilon
|
||||
self.amsgrad = amsgrad
|
||||
|
||||
def build(self, var_list):
|
||||
"""Initialize optimizer variables.
|
||||
|
||||
Adam 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 Adam 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(
|
||||
model_variable=var, variable_name="m"
|
||||
)
|
||||
)
|
||||
self._velocities.append(
|
||||
self.add_variable_from_reference(
|
||||
model_variable=var, variable_name="v"
|
||||
)
|
||||
)
|
||||
if self.amsgrad:
|
||||
self._velocity_hats = []
|
||||
for var in var_list:
|
||||
self._velocity_hats.append(
|
||||
self.add_variable_from_reference(
|
||||
model_variable=var, variable_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(self.learning_rate, 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(m - (m * alpha) / (ops.sqrt(v) + self.epsilon))
|
||||
|
||||
def get_config(self):
|
||||
config = super().get_config()
|
||||
config.update(
|
||||
{
|
||||
"beta_1": self.beta_1,
|
||||
"beta_2": self.beta_2,
|
||||
"epsilon": self.epsilon,
|
||||
"amsgrad": self.amsgrad,
|
||||
}
|
||||
)
|
||||
return config
|
||||
|
||||
|
||||
Adam.__doc__ = Adam.__doc__.replace(
|
||||
"{{base_optimizer_keyword_args}}", optimizer.base_optimizer_keyword_args
|
||||
)
|
||||
@ -1,45 +0,0 @@
|
||||
from keras_core import backend
|
||||
from keras_core import operations as ops
|
||||
from keras_core import testing
|
||||
from keras_core.optimizers.adam import Adam
|
||||
import numpy as np
|
||||
|
||||
class AdamTest(testing.TestCase):
|
||||
def test_config(self):
|
||||
# TODO
|
||||
pass
|
||||
|
||||
def test_weight_decay(self):
|
||||
grads, var1, var2, var3 = (
|
||||
np.zeros(()),
|
||||
backend.Variable(2.0),
|
||||
backend.Variable(2.0, name="exclude"),
|
||||
backend.Variable(2.0),
|
||||
)
|
||||
optimizer_1 = Adam(learning_rate=1, weight_decay=0.004)
|
||||
optimizer_1.apply_gradients(zip([grads], [var1]))
|
||||
|
||||
optimizer_2 = Adam(learning_rate=1, weight_decay=0.004)
|
||||
optimizer_2.exclude_from_weight_decay(var_names=["exclude"])
|
||||
optimizer_2.apply_gradients(zip([grads, grads], [var1, var2]))
|
||||
|
||||
optimizer_3 = Adam(learning_rate=1, weight_decay=0.004)
|
||||
optimizer_3.exclude_from_weight_decay(var_list=[var3])
|
||||
optimizer_3.apply_gradients(zip([grads, grads], [var1, var3]))
|
||||
|
||||
self.assertAlmostEqual(var1, 1.9760959)
|
||||
self.assertAlmostEqual(var2, 2.0)
|
||||
self.assertAlmostEqual(var3, 2.0)
|
||||
|
||||
def test_clip_norm(self):
|
||||
optimizer = Adam(clipnorm=1)
|
||||
grad = [np.array([100.0, 100.0])]
|
||||
clipped_grad = optimizer._clip_gradients(grad)
|
||||
self.assertAllClose(clipped_grad[0], [2**0.5 / 2, 2**0.5 / 2])
|
||||
|
||||
def test_clip_value(self):
|
||||
optimizer = Adam(clipvalue=1)
|
||||
grad = [np.array([100.0, 100.0])]
|
||||
clipped_grad = optimizer._clip_gradients(grad)
|
||||
self.assertAllClose(clipped_grad[0], [1.0, 1.0])
|
||||
|
||||
Loading…
Reference in New Issue
Block a user