tf.keras.optimizers.Optimizer

Abstract optimizer base class.

tf.keras.optimizers.Optimizer(
    name,
    weight_decay=0,
    clipnorm=None,
    clipvalue=None,
    global_clipnorm=None,
    use_ema=False,
    ema_momentum=0.99,
    ema_overwrite_frequency=None,
    jit_compile=True,
    **kwargs
)

This class supports distributed training. If you want to implement your own optimizer, please subclass this class instead of _BaseOptimizer.

name String. The name to use for momentum accumulator weights created by the optimizer.
weight_decay Float, defaults to None. If set, weight decay is applied.
clipnorm Float. If set, the gradient of each weight is individually clipped so that its norm is no higher than this value.
clipvalue Float. If set, the gradient of each weight is clipped to be no higher than this value.
global_clipnorm Float. If set, the gradient of all weights is clipped so that their global norm is no higher than this value.
use_ema Boolean, defaults to False. If True, exponential moving average (EMA) is applied. EMA consists of computing an exponential moving average of the weights of the model (as the weight values change after each training batch), and periodically overwriting the weights with their moving average.
ema_momentum Float, defaults to 0.99. Only used if use_ema=True. This is the momentum to use when computing the EMA of the model's weights: new_average = ema_momentum * old_average + (1 - ema_momentum) * current_variable_value.
ema_overwrite_frequency Int or None, defaults to None. Only used if use_ema=True. Every ema_overwrite_frequency steps of iterations, we overwrite the model variable by its moving average. If None, the optimizer does not overwrite model variables in the middle of training, and you need to explicitly overwrite the variables at the end of training by calling optimizer.finalize_variable_values() (which updates the model variables in-place). When using the built-in fit() training loop, this happens automatically after the last epoch, and you don't need to do anything.
jit_compile Boolean, defaults to True. If True, the optimizer will use XLA compilation. If no GPU device is found, this flag will be ignored.
mesh optional tf.experimental.dtensor.Mesh instance. When provided, the optimizer will be run in DTensor mode, e.g. state tracking variable will be a DVariable, and aggregation/reduction will happen in the global DTensor context.
**kwargs keyword arguments only used for backward compatibility.

Usage

# Create an optimizer with the desired parameters.
opt = keras.optimizers.SGD(learning_rate=0.1)
var1, var2 = tf.Variable(1.0), tf.Variable(2.0)
# `loss` is a callable that takes no argument and returns the value
# to minimize.
loss = lambda: 3 * var1 * var1 + 2 * var2 * var2
# Call minimize to update the list of variables.
opt.minimize(loss, var_list=[var1, var2])

Processing gradients before applying them

Calling minimize() takes care of both computing the gradients and applying them to the variables. If you want to process the gradients before applying them you can instead use the optimizer in three steps:

  1. Compute the gradients with tf.GradientTape.
  2. Process the gradients as you wish.
  3. Apply the processed gradients with apply_gradients().

Example:

# Create an optimizer.
opt = tf.keras.optimizers.experimental.SGD(learning_rate=0.1)
var1, var2 = tf.Variable(1.0), tf.Variable(2.0)

# Compute the gradients for a list of variables.
with tf.GradientTape() as tape:
  loss = 3 * var1 * var1 + 2 * var2 * var2
grads = tape.gradient(loss, [var1, var2])

# Process the gradients.
grads[0] = grads[0] + 1

# Ask the optimizer to apply the gradients on variables.
opt.apply_gradients(zip(grads, [var1, var2]))

Dynamic learning rate

Dynamic learning rate can be achieved by setting learning rate as a built-in or customized tf.keras.optimizers.schedules.LearningRateSchedule.

Example:

var = tf.Variable(np.random.random(size=(1,)))
learning_rate = tf.keras.optimizers.schedules.ExponentialDecay(
  initial_learning_rate=.01, decay_steps=20, decay_rate=.1)
opt = tf.keras.optimizers.experimental.SGD(learning_rate=learning_rate)
loss = lambda: 3 * var
opt.minimize(loss, var_list=[var])

Gradients clipping

Users can clip the gradients before applying to variables by setting clipnorm, clipvalue and global_clipnorm. Notice that clipnorm and global_clipnorm can only have one being set.

Example:

opt = tf.keras.optimizers.experimental.SGD(learning_rate=1, clipvalue=1)
var1, var2 = tf.Variable(2.0), tf.Variable(2.0)
with tf.GradientTape() as tape:
  loss = 2 * var1 + 2 * var2
grads = tape.gradient(loss, [var1, var2])
print([grads[0].numpy(), grads[1].numpy()])
[2.0, 2.0]
opt.apply_gradients(zip(grads, [var1, var2]))
# Without clipping, we should get [0, 0], but as gradients are clipped
# to have max value 1, we get [1.0, 1.0].
print([var1.numpy(), var2.numpy()])
[1.0, 1.0]

Using weight decay.

Weight decay in certain scenarios can boost the model's performance. Keras has built-in support for weight decay in all optimizers. Users can apply weight decay by setting weight_decay argument.

opt = tf.keras.optimizers.experimental.SGD(1, weight_decay=0.004)
grads, var1, var2 = tf.zeros(()), tf.Variable(2.0), tf.Variable(2.0)
# You can exclude variables from weight decay, in this case we
# exclude `var2`.
opt.exclude_from_weight_decay(var_list=[var2])
opt.apply_gradients(zip([grads, grads], [var1, var2]))
print([var1.numpy(), var2.numpy()])
[1.992, 2.0]

Using exponential moving average.

Empirically it has been found that using the exponential moving average (EMA) of the trained parameters of a deep network achieves a better performance than using its trained parameters directly. Keras optimizers allows users to compute this moving average and overwrite the model variables at desired time.

Example:

# Create an SGD optimizer with EMA on. `ema_momentum` controls the decay
# rate of the moving average. `ema_momentum=1` means no decay and the stored
# moving average is always model variable's initial value before training.
# Reversely, `ema_momentum=0` is equivalent to not using EMA.
# `ema_overwrite_frequency=3` means every 3 iterations, we overwrite the
# trainable variables with their moving average values.
opt = tf.keras.optimizers.experimental.SGD(
    learning_rate=1,
    use_ema=True,
    ema_momentum=0.5,
    ema_overwrite_frequency=3)
var1, var2 = tf.Variable(2.0), tf.Variable(2.0)
with tf.GradientTape() as tape:
  loss = var1 + var2
grads = tape.gradient(loss, [var1, var2])
# First iteration: [var1, var2] = [1.0, 1.0]
opt.apply_gradients(zip(grads, [var1, var2]))
print([var1, var2])

# Second iteration: [var1, var2] = [0.0, 0.0]
opt.apply_gradients(zip(grads, [var1, var2]))
print([var1, var2])

# Third iteration, without EMA, we should see [var1, var2] = [-1.0, -1.0],
# but overwriting results in [var1, var2] = [-0.125, -0.125]. The full
# calculation for the moving average of var1 is:
# var1=2*0.5**3+1*(1-0.5)*0.5**2+0*(1-0.5)*0.5**1+(-1)*(1-0.5)=-0.125.
opt.apply_gradients(zip(grads, [var1, var2]))
print([var1, var2])

When optimizer is constructed with use_ema=True, in custom training loop, users can explicitly call finalize_variable_values() to overwrite trainable variables with their EMA values. finalize_variable_values() is by default called at the end of model.fit().

Use with tf.distribute.Strategy

This optimizer class is tf.distribute.Strategy aware, which means it automatically sums gradients across all replicas. To aggregate gradients yourself, call apply_gradients with skip_gradients_aggregation set to True. This is useful if you need to process aggregated gradients.

# This example is not runnable, it consists of dummy code for simple
# tutorial.
strategy = tf.distribute.experimental.TPUStrategy()

with strategy.scope():
  opt = tf.keras.optimizers.experimental.SGD()
  model = magic_function_that_returns_model()
  gradients = magic_function_that_returns_gradients()
  # Custom logic to aggregate gradients.
  gradients = strategy.reduce("SUM", gradients, axis=None)
  opt.apply_gradients(zip(gradients, model.trainable_variables),
      skip_gradients_aggregation=True)

Creating a custom optimizer

If you intend to create your own optimization algorithm, please inherit from this class and override the following methods:

  • build: Create your optimizer-related variables, such as momentums in SGD optimizer.
  • update_step: Implement your optimizer's updating logic.
  • get_config: serialization of the optimizer, include all hyper parameters.

Your optimizer would automatically be compatible with tensorflow distributed training if you subclass optimizer_experimental.Optimizer.

iterations The number of training steps this optimizer has run.

By default, iterations would be incremented by one every time apply_gradients() is called.

learning_rate

variables Returns variables of this optimizer.

Methods

add_variable

View source

add_variable(
    shape, dtype=None, initializer='zeros', name=None
)

Create an optimizer variable.

Args
shape A list of integers, a tuple of integers, or a 1-D Tensor of type int32. Defaults to scalar if unspecified.
dtype The DType of the optimizer variable to be created. Defaults to tf.keras.backend.floatx if unspecified.
initializer string or callable. Initializer instance.
name The name of the optimizer variable to be created.

Returns
An optimizer variable, in the format of tf.Variable.

add_variable_from_reference

View source

add_variable_from_reference(
    model_variable, variable_name, shape=None, initial_value=None
)

Create an optimizer variable from model variable.

Create an optimizer variable based on the information of model variable. For example, in SGD optimizer momemtum, for each model variable, a corresponding momemtum variable is created of the same shape and dtype.

Args
model_variable tf.Variable. The corresponding model variable to the optimizer variable to be created.
variable_name String. The name prefix of the optimizer variable to be created. The create variables name will follow the pattern {variable_name}/{model_variable.name}, e.g., momemtum/dense_1.
shape List or Tuple, defaults to None. The shape of the optimizer variable to be created. If None, the created variable will have the same shape as model_variable.
initial_value A Tensor, or Python object convertible to a Tensor, defaults to None. The initial value of the optimizer variable, if None, the initial value will be default to 0.

Returns
An optimizer variable.

aggregate_gradients

View source

aggregate_gradients(
    grads_and_vars
)

Aggregate gradients on all devices.

By default, we will perform reduce_sum of gradients across devices. Users can implement their own aggregation logic by overriding this method.

Args
grads_and_vars List of (gradient, variable) pairs.

Returns
List of (gradient, variable) pairs.

apply_gradients

View source

apply_gradients(
    grads_and_vars, name=None, skip_gradients_aggregation=False, **kwargs
)

Apply gradients to variables.

Args
grads_and_vars List of (gradient, variable) pairs.
name string, defaults to None. The name of the namescope to use when creating variables. If None, self.name will be used.
skip_gradients_aggregation If true, gradients aggregation will not be performed inside optimizer. Usually this arg is set to True when you write custom code aggregating gradients outside the optimizer.
**kwargs keyword arguments only used for backward compatibility.

Returns
A tf.Variable, representing the current iteration.

Raises
TypeError If grads_and_vars is malformed.
RuntimeError If called in a cross-replica context.

build

View source

@abc.abstractmethod
build(
    var_list
)

Initialize the optimizer's variables, such as momemtum variables.

This function has to be implemented by subclass optimizers, and subclass optimizers need to call super().build(var_list).

Args
var_list List of model variables to build optimizers on. For example, SGD optimizer with momentum will store one momentum variable corresponding to each model variable.

compute_gradients

View source

compute_gradients(
    loss, var_list, tape=None
)

Compute gradients of loss on trainable variables.

Args
loss Tensor or callable. If a callable, loss should take no arguments and return the value to minimize.
var_list list or tuple of Variable objects to update to minimize loss, or a callable returning the list or tuple of Variable objects. Use callable when the variable list would otherwise be incomplete before minimize since the variables are created at the first time loss is called.
tape (Optional) tf.GradientTape. If loss is provided as a Tensor, the tape that computed the loss must be provided.

Returns
A list of (gradient, variable) pairs. Variable is always present, but gradient can be None.

exclude_from_weight_decay

View source

exclude_from_weight_decay(
    var_list=None, var_names=None
)

Exclude variables from weight decay.

This method must be called before the optimizer's build method is called. You can set specific variables to exclude out, or set a list of strings as the anchor words, if any of which appear in a variable's name, then the variable is excluded.

Args
var_list A list of tf.Variables to exclude from weight decay.
var_names A list of strings. If any string in var_names appear in the model variable's name, then this model variable is excluded from weight decay. For example, var_names=['bias'] excludes all bias variables from weight decay.

finalize_variable_values

View source

finalize_variable_values(
    var_list
)

Set the final value of model's trainable variables.

Sometimes there are some extra steps before ending the variable updates, such as overriding the model variables with its average value.

Args
var_list list of model variables.

from_config

View source

@classmethod
from_config(
    config, custom_objects=None
)

Creates an optimizer from its config.

This method is the reverse of get_config, capable of instantiating the same optimizer from the config dictionary.

Args
config A Python dictionary, typically the output of get_config.
custom_objects A Python dictionary mapping names to additional user-defined Python objects needed to recreate this optimizer.

Returns
An optimizer instance.

get_config

View source

get_config()

Returns the config of the optimizer.

An optimizer config is a Python dictionary (serializable) containing the configuration of an optimizer. The same optimizer can be reinstantiated later (without any saved state) from this configuration.

Subclass optimizer should override this method to include other hyperparameters.

Returns
Python dictionary.

load_own_variables

View source

load_own_variables(
    store
)

Set the state of this optimizer object.

minimize

View source

minimize(
    loss, var_list, tape=None
)

Minimize loss by updating var_list.

This method simply computes gradient using tf.GradientTape and calls apply_gradients(). If you want to process the gradient before applying then call tf.GradientTape and apply_gradients() explicitly instead of using this function.

Args
loss Tensor or callable. If a callable, loss should take no arguments and return the value to minimize.
var_list list or tuple of Variable objects to update to minimize loss, or a callable returning the list or tuple of Variable objects. Use callable when the variable list would otherwise be incomplete before minimize since the variables are created at the first time loss is called.
tape (Optional) tf.GradientTape.

Returns
None

save_own_variables

View source

save_own_variables(
    store
)

Get the state of this optimizer object.

set_weights

View source

set_weights(
    weights
)

Set the weights of the optimizer.

Args
weights a list of tf.Variables or numpy arrays, the target values of optimizer variables. It should have the same order as self._variables.

update_step

View source

@abc.abstractmethod
update_step(
    gradient, variable
)

Function to update variable value based on given gradients.

This method must be implemented in customized optimizers.

Args
gradient backpropagated gradient of the given variable.
variable variable whose value needs to be updated.

Returns
An Operation that applies the specified gradients.