Python tensorflow.python.ops.state_ops.assign_sub() Examples
The following are 30
code examples of tensorflow.python.ops.state_ops.assign_sub().
You can vote up the ones you like or vote down the ones you don't like,
and go to the original project or source file by following the links above each example.
You may also want to check out all available functions/classes of the module
tensorflow.python.ops.state_ops
, or try the search function
.
.
Example #1
| Source File: adamax.py From iaf with MIT License | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
if var.dtype.base_dtype == tf.float16:
eps = 1e-7 # Can't use 1e-8 due to underflow -- not sure if it makes a big difference.
else:
eps = 1e-8
v = self.get_slot(var, "v")
v_t = v.assign(beta1_t * v + (1. - beta1_t) * grad)
m = self.get_slot(var, "m")
m_t = m.assign(tf.maximum(beta2_t * m + eps, tf.abs(grad)))
g_t = v_t / m_t
var_update = state_ops.assign_sub(var, lr_t * g_t)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #2
| Source File: optimizer.py From BERT with Apache License 2.0 | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
# the following equations given in [1]
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_t = state_ops.assign(m, beta1_t * m + (1. - beta1_t) * grad, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_t = state_ops.assign(v, beta2_t * v + (1. - beta2_t) * tf.square(grad), use_locking=self._use_locking)
v_prime = self.get_slot(var, "v_prime")
v_t_prime = state_ops.assign(v_prime, tf.maximum(v_prime, v_t))
var_update = state_ops.assign_sub(var,
lr_t * m_t / (tf.sqrt(v_t_prime) + epsilon_t),
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, v_t_prime])
# keras Nadam update rule
Example #3
| Source File: opt.py From EMNLP2018_NLI with GNU General Public License v3.0 | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
if var.dtype.base_dtype == tf.float16:
eps = 1e-7
# Can't use 1e-8 due to underflow -- not sure if it makes a big difference.
else:
eps = 1e-8
v = self.get_slot(var, "v")
v_t = v.assign(beta1_t * v + (1. - beta1_t) * grad)
m = self.get_slot(var, "m")
m_t = m.assign(tf.maximum(beta2_t * m + eps, tf.abs(grad)))
g_t = v_t / m_t
var_update = state_ops.assign_sub(var, lr_t * g_t)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #4
| Source File: optimizer.py From tensorflow-DSMM with MIT License | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
# the following equations given in [1]
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_t = state_ops.assign(m, beta1_t * m + (1. - beta1_t) * grad, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_t = state_ops.assign(v, beta2_t * v + (1. - beta2_t) * tf.square(grad), use_locking=self._use_locking)
v_prime = self.get_slot(var, "v_prime")
v_t_prime = state_ops.assign(v_prime, tf.maximum(v_prime, v_t))
var_update = state_ops.assign_sub(var,
lr_t * m_t / (tf.sqrt(v_t_prime) + epsilon_t),
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, v_t_prime])
# keras Nadam update rule
Example #5
| Source File: utilities.py From EMNLP2018_NLI with GNU General Public License v3.0 | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
if var.dtype.base_dtype == tf.float16:
eps = 1e-7
# Can't use 1e-8 due to underflow -- not sure if it makes a big difference.
else:
eps = 1e-8
v = self.get_slot(var, "v")
v_t = v.assign(beta1_t * v + (1. - beta1_t) * grad)
m = self.get_slot(var, "m")
m_t = m.assign(tf.maximum(beta2_t * m + eps, tf.abs(grad)))
g_t = v_t / m_t
var_update = state_ops.assign_sub(var, lr_t * g_t)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #6
| Source File: models.py From aletheia with MIT License | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
if var.dtype.base_dtype == tf.float16:
eps = 1e-7 # Can't use 1e-8 due to underflow -- not sure if it makes a big difference.
else:
eps = 1e-8
v = self.get_slot(var, "v")
v_t = v.assign(beta1_t * v + (1. - beta1_t) * grad)
m = self.get_slot(var, "m")
m_t = m.assign(tf.maximum(beta2_t * m + eps, tf.abs(grad)))
g_t = v_t / m_t
var_update = state_ops.assign_sub(var, lr_t * g_t)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #7
| Source File: adamirror.py From HyperGAN with MIT License | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
if var.dtype.base_dtype == tf.float16:
eps = 1e-7 # Can't use 1e-8 due to underflow -- not sure if it makes a big difference.
else:
eps = 1e-8
v = self.get_slot(var, "v")
v_t = v.assign(beta2_t * v + (1. - beta2_t) * tf.square(grad))
m = self.get_slot(var, "m")
m_t = m.assign( beta1_t * m + (1. - beta1_t) * grad )
v_t_hat = tf.div(v_t, 1. - beta2_t)
m_t_hat = tf.div(m_t, 1. - beta1_t)
g_t = tf.div( m_t, tf.sqrt(v_t)+eps )
g_t_1 = self.get_slot(var, "g")
g_t = g_t_1.assign( g_t )
var_update = state_ops.assign_sub(var, 2. * lr_t * g_t - lr_t * g_t_1) #Adam would be lr_t * g_t
return control_flow_ops.group(*[var_update, m_t, v_t, g_t])
Example #8
| Source File: adamax.py From Awesome-GANs with MIT License | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
if var.dtype.base_dtype == tf.float16:
eps = 1e-7 # Can't use 1e-8 due to underflow -- not sure if it makes a big difference.
else:
eps = 1e-8
v = self.get_slot(var, "v")
v_t = v.assign(beta1_t * v + (1. - beta1_t) * grad)
m = self.get_slot(var, "m")
m_t = m.assign(tf.maximum(beta2_t * m + eps, tf.abs(grad)))
g_t = v_t / m_t
var_update = state_ops.assign_sub(var, lr_t * g_t)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #9
| Source File: optimizer.py From tensorflow-XNN with MIT License | 6 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
# the following equations given in [1]
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_t = state_ops.assign(m, beta1_t * m + (1. - beta1_t) * grad, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_t = state_ops.assign(v, beta2_t * v + (1. - beta2_t) * tf.square(grad), use_locking=self._use_locking)
v_prime = self.get_slot(var, "v_prime")
v_t_prime = state_ops.assign(v_prime, tf.maximum(v_prime, v_t))
var_update = state_ops.assign_sub(var,
lr_t * m_t / (tf.sqrt(v_t_prime) + epsilon_t),
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, v_t_prime])
# keras Nadam update rule
Example #10
| Source File: variables.py From Serverless-Deep-Learning-with-TensorFlow-and-AWS-Lambda with MIT License | 5 votes |
|
def assign_sub(self, delta, use_locking=False):
"""Subtracts a value from this variable.
This is essentially a shortcut for `assign_sub(self, delta)`.
Args:
delta: A `Tensor`. The value to subtract from this variable.
use_locking: If `True`, use locking during the operation.
Returns:
A `Tensor` that will hold the new value of this variable after
the subtraction has completed.
"""
return state_ops.assign_sub(self._variable, delta, use_locking=use_locking)
Example #11
| Source File: optimizer.py From tensorflow-XNN with MIT License | 5 votes |
|
def _apply_dense(self, grad, var):
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta_t = math_ops.cast(self._beta_t, var.dtype.base_dtype)
alpha_t = math_ops.cast(self._alpha_t, var.dtype.base_dtype)
eps = 1e-7 # cap for moving average
m = self.get_slot(var, "m")
m_t = m.assign(tf.maximum(beta_t * m + eps, tf.abs(grad)))
var_update = state_ops.assign_sub(var, lr_t * grad * (1.0 + alpha_t * tf.sign(grad) * tf.sign(m_t)))
# Create an op that groups multiple operations
# When this op finishes, all ops in input have finished
return control_flow_ops.group(*[var_update, m_t])
Example #12
| Source File: backend.py From Serverless-Deep-Learning-with-TensorFlow-and-AWS-Lambda with MIT License | 5 votes |
|
def update_sub(x, decrement):
"""Update the value of `x` by subtracting `decrement`.
Arguments:
x: A Variable.
decrement: A tensor of same shape as `x`.
Returns:
The variable `x` updated.
"""
return state_ops.assign_sub(x, decrement)
Example #13
| Source File: optimization_gpu.py From BERT-multi-gpu with Apache License 2.0 | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
learning_rate_t = math_ops.cast(
self.learning_rate_t, var.dtype.base_dtype)
beta_1_t = math_ops.cast(self.beta_1_t, var.dtype.base_dtype)
beta_2_t = math_ops.cast(self.beta_2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self.epsilon_t, var.dtype.base_dtype)
weight_decay_rate_t = math_ops.cast(
self.weight_decay_rate_t, var.dtype.base_dtype)
m = self.get_slot(var, 'm')
v = self.get_slot(var, 'v')
m_t = state_ops.assign(m, m * beta_1_t,
use_locking=self._use_locking)
m_scaled_g_values = grad * (1 - beta_1_t)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
v_scaled_g_values = (grad * grad) * (1 - beta_2_t)
v_t = state_ops.assign(v, v * beta_2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
update = m_t / (math_ops.sqrt(v_t) + epsilon_t)
if self._do_use_weight_decay(var.name):
update += weight_decay_rate_t * var
update_with_lr = learning_rate_t * update
var_update = state_ops.assign_sub(var,
update_with_lr,
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #14
| Source File: RAdam.py From RAdam-Tensorflow with MIT License | 5 votes |
|
def _resource_apply_dense(self, grad, var):
step, beta1_power, beta2_power = self._get_beta_accumulators()
step = math_ops.cast(step, var.dtype.base_dtype)
beta1_power = math_ops.cast(beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
sma_inf = 2.0 / (1.0 - beta2_t) - 1.0
sma_t = sma_inf - 2.0 * step * beta2_power / (1.0 - beta2_power)
m = self.get_slot(var, "m")
m_t = state_ops.assign(m, beta1_t * m + (1.0 - beta1_t) * grad, use_locking=self._use_locking)
mhat_t = m_t / (1.0 - beta1_power)
v = self.get_slot(var, "v")
v_t = state_ops.assign(v, beta2_t * v + (1.0 - beta2_t) * math_ops.square(grad), use_locking=self._use_locking)
vhat_t = math_ops.sqrt(v_t / ((1.0 - beta2_power) + epsilon_t))
r_t = math_ops.sqrt( ((sma_t - 4.0) * (sma_t - 2.0) * sma_inf) / ((sma_inf - 4.0) * (sma_inf - 2.0) * sma_t) )
var_t = tf.cond(sma_t >= 5.0, lambda : r_t * mhat_t / (vhat_t + epsilon_t), lambda : mhat_t)
if self._weight_decay > 0.0:
var_t += math_ops.cast(self._weight_decay_t, var.dtype.base_dtype) * var
var_update = state_ops.assign_sub(var, lr_t * var_t, use_locking=self._use_locking)
updates = [var_update, m_t, v_t]
return control_flow_ops.group(*updates)
Example #15
| Source File: adam.py From Serverless-Deep-Learning-with-TensorFlow-and-AWS-Lambda with MIT License | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, m * beta1_t,
use_locking=self._use_locking)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
v_sqrt = math_ops.sqrt(v_t)
var_update = state_ops.assign_sub(var,
lr * m_t / (v_sqrt + epsilon_t),
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #16
| Source File: normalization.py From Serverless-Deep-Learning-with-TensorFlow-and-AWS-Lambda with MIT License | 5 votes |
|
def _assign_moving_average(self, variable, value, one_minus_decay):
with ops.name_scope(None, 'AssignMovingAvg',
[variable, value, one_minus_decay]) as scope:
with ops.colocate_with(variable):
update_delta = math_ops.multiply(
math_ops.subtract(variable.read_value(), value),
one_minus_decay)
if isinstance(variable, resource_variable_ops.ResourceVariable):
# state_ops.assign_sub does an extra read_variable_op after the
# assign. We avoid that here.
return gen_resource_variable_ops.assign_sub_variable_op(
variable.handle, update_delta, name=scope)
else:
return state_ops.assign_sub(variable, update_delta, name=scope)
Example #17
| Source File: AMSGrad.py From AMSGrad-Tensorflow with MIT License | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, m * beta1_t, use_locking=self._use_locking)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #18
| Source File: AMSGrad.py From AMSGrad-Tensorflow with MIT License | 5 votes |
|
def _resource_apply_dense(self, grad, var):
var = var.handle
beta1_power = math_ops.cast(self._beta1_power, grad.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, grad.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, grad.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, grad.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, grad.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, grad.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m").handle
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v").handle
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat").handle
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #19
| Source File: AMSGrad.py From PhysNet with MIT License | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, m * beta1_t, use_locking=self._use_locking)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #20
| Source File: opt.py From EMNLP2018_NLI with GNU General Public License v3.0 | 5 votes |
|
def _finish(self, update_ops, name_scope):
# Update the power accumulators.
with ops.control_dependencies(update_ops):
with ops.colocate_with(self._beta1_power):
update_beta1 = self._beta1_power.assign(
self._beta1_power * self._beta1_t,
use_locking=self._use_locking)
update_beta2 = self._beta2_power.assign(
self._beta2_power * self._beta2_t,
use_locking=self._use_locking)
return control_flow_ops.group(*update_ops + [update_beta1, update_beta2],
name=name_scope)* (1 - beta1_t)
m_t = state_ops.assign(m, m * beta1_t, use_locking=self._use_locking)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #21
| Source File: opt.py From EMNLP2018_NLI with GNU General Public License v3.0 | 5 votes |
|
def _resource_apply_dense(self, grad, var):
var = var.handle
beta1_power = math_ops.cast(self._beta1_power, grad.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, grad.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, grad.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, grad.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, grad.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, grad.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m").handle
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v").handle
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat").handle
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #22
| Source File: opt.py From EMNLP2018_NLI with GNU General Public License v3.0 | 5 votes |
|
def _apply_dense(self, grad, var):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #23
| Source File: opt.py From EMNLP2018_NLI with GNU General Public License v3.0 | 5 votes |
|
def _resource_apply_dense(self, grad, var):
var = var.handle
beta1_power = math_ops.cast(self._beta1_power, grad.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, grad.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, grad.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, grad.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, grad.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, grad.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m").handle
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v").handle
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat").handle
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #24
| Source File: opt.py From EMNLP2018_NLI with GNU General Public License v3.0 | 5 votes |
|
def _apply_dense(self, grad, var):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #25
| Source File: AMSGrad.py From DCRNN with MIT License | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, m * beta1_t, use_locking=self._use_locking)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #26
| Source File: AMSGrad.py From DCRNN with MIT License | 5 votes |
|
def _resource_apply_dense(self, grad, var):
var = var.handle
beta1_power = math_ops.cast(self._beta1_power, grad.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, grad.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, grad.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, grad.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, grad.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, grad.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m").handle
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v").handle
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat").handle
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #27
| Source File: AMSGrad.py From DCRNN with MIT License | 5 votes |
|
def _apply_dense(self, grad, var):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad * (1 - beta1_t)
m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad * grad) * (1 - beta2_t)
v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)
# amsgrad
vhat = self.get_slot(var, "vhat")
vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
v_sqrt = math_ops.sqrt(vhat_t)
var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])
Example #28
| Source File: multi_gpu_optimizer.py From uda with Apache License 2.0 | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
learning_rate_t = math_ops.cast(
self.learning_rate_t, var.dtype.base_dtype)
beta_1_t = math_ops.cast(self.beta_1_t, var.dtype.base_dtype)
beta_2_t = math_ops.cast(self.beta_2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self.epsilon_t, var.dtype.base_dtype)
weight_decay_rate_t = math_ops.cast(
self.weight_decay_rate_t, var.dtype.base_dtype)
m = self.get_slot(var, 'm')
v = self.get_slot(var, 'v')
m_t = state_ops.assign(m, m * beta_1_t,
use_locking=self._use_locking)
m_scaled_g_values = grad * (1 - beta_1_t)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
v_scaled_g_values = (grad * grad) * (1 - beta_2_t)
v_t = state_ops.assign(v, v * beta_2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
update = m_t / (math_ops.sqrt(v_t) + epsilon_t)
if self._do_use_weight_decay(var.name):
update += weight_decay_rate_t * var
update_with_lr = learning_rate_t * update
var_update = state_ops.assign_sub(var,
update_with_lr,
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #29
| Source File: custom_optimization.py From bert-multi-gpu with Apache License 2.0 | 5 votes |
|
def _apply_sparse_shared(self, grad, var, indices, scatter_add):
learning_rate_t = math_ops.cast(
self.learning_rate_t, var.dtype.base_dtype)
beta_1_t = math_ops.cast(self.beta_1_t, var.dtype.base_dtype)
beta_2_t = math_ops.cast(self.beta_2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self.epsilon_t, var.dtype.base_dtype)
weight_decay_rate_t = math_ops.cast(
self.weight_decay_rate_t, var.dtype.base_dtype)
m = self.get_slot(var, 'm')
v = self.get_slot(var, 'v')
m_t = state_ops.assign(m, m * beta_1_t,
use_locking=self._use_locking)
m_scaled_g_values = grad * (1 - beta_1_t)
with ops.control_dependencies([m_t]):
m_t = scatter_add(m, indices, m_scaled_g_values)
v_scaled_g_values = (grad * grad) * (1 - beta_2_t)
v_t = state_ops.assign(v, v * beta_2_t, use_locking=self._use_locking)
with ops.control_dependencies([v_t]):
v_t = scatter_add(v, indices, v_scaled_g_values)
update = m_t / (math_ops.sqrt(v_t) + epsilon_t)
if self._do_use_weight_decay(var.name):
update += weight_decay_rate_t * var
update_with_lr = learning_rate_t * update
var_update = state_ops.assign_sub(var,
update_with_lr,
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t])
Example #30
| Source File: adam.py From deep_image_model with Apache License 2.0 | 5 votes |
|
def _apply_sparse(self, grad, var):
beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
# m_t = beta1 * m + (1 - beta1) * g_t
m = self.get_slot(var, "m")
m_scaled_g_values = grad.values * (1 - beta1_t)
m_t = state_ops.assign(m, m * beta1_t,
use_locking=self._use_locking)
m_t = state_ops.scatter_add(m_t, grad.indices, m_scaled_g_values,
use_locking=self._use_locking)
# v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
v = self.get_slot(var, "v")
v_scaled_g_values = (grad.values * grad.values) * (1 - beta2_t)
v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
v_t = state_ops.scatter_add(v_t, grad.indices, v_scaled_g_values,
use_locking=self._use_locking)
v_sqrt = math_ops.sqrt(v_t)
var_update = state_ops.assign_sub(var,
lr * m_t / (v_sqrt + epsilon_t),
use_locking=self._use_locking)
return control_flow_ops.group(*[var_update, m_t, v_t])
