Save the date! Google I/O returns May 18-20 Register now


Base class for Keras optimizers.

You should not use this class directly, but instead instantiate one of its subclasses such as tf.keras.optimizers.SGD, tf.keras.optimizers.Adam, etc.


# Create an optimizer with the desired parameters.
opt = tf.keras.optimizers.SGD(learning_rate=0.1)
# `loss` is a callable that takes no argument and returns the value
# to minimize.
loss = lambda: 3 * var1 * var1 + 2 * var2 * var2
# In graph mode, returns op that minimizes the loss by updating the listed
# variables.
opt_op = opt.minimize(loss, var_list=[var1, var2])
# In eager mode, simply call minimize to update the list of variables.
opt.minimize(loss, var_list=[var1, var2])

Usage in custom training loops

In Keras models, sometimes variables are created when the model is first called, instead of construction time. Examples include 1) sequential models without input shape pre-defined, or 2) subclassed models. Pass var_list as callable in these cases.


opt = tf.keras.optimizers.SGD(learning_rate=0.1)
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(num_hidden, activation='relu'))
model.add(tf.keras.layers.Dense(num_classes, activation='sigmoid'))
loss_fn = lambda: tf.keras.losses.mse(model(input), output)
var_list_fn = lambda: model.trainable_weights
for input, output in data:
  opt.minimize(loss_fn, var_list_fn)

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().


# Create an optimizer.
opt = tf.keras.optimizers.SGD(learning_rate=0.1)

# Compute the gradients for a list of variables.
with tf.GradientTape() as tape:
  loss = <call_loss_function>
vars = <list_of_variables>
grads = tape.gradient(loss, vars)

# Process the gradients, for example cap them, etc.
# capped_grads = [MyCapper(g) for g in grads]
processed_grads = [process_gradient(g) for g in grads]

# Ask the optimizer to apply the processed gradients.
opt.apply_gradients(zip(processed_grads, var_list))

Use with tf.distribute.Strategy

This optimizer class is tf.distribute.Strategy aware, which means it automatically sums gradients across all replicas. To average gradients, you divide your loss by the global batch size, which is done automatically if you use tf.keras built-in training or evaluation loops. See the reduction argument of your loss which should be set to tf.keras.losses.Reduction.SUM_OVER_BATCH_SIZE for averaging or tf.keras.losses.Reduction.SUM for not.

To aggregate gradients yourself, call apply_gradients with experimental_aggregate_gradients set to False. This is useful if you need to process aggregated gradients.

If you are not using these and you want to average gradients, you should use tf.math.reduce_sum to add up your per-example losses and then divide by the global batch size. Note that when using tf.distribute.Strategy, the first component of a tensor's shape is the replica-local batch size, which is off by a factor equal to the number of replicas being used to compute a single step. As a result, using tf.math.reduce_mean will give the wrong answer, resulting in gradients that can be many times too big.

Variable Constraints

All Keras optimizers respect variable constraints. If constraint function is passed to any variable, the constraint will be applied to the variable after the gradient has been applied to the variable. Important: If gradient is sparse tensor, variable constraint is not supported.

Thread Compatibility

The entire optimizer is currently thread compatible, not thread-safe. The user needs to perform synchronization if necessary.


Many optimizer subclasses, such as Adam and Adagrad allocate and manage additional variables associated with the variables to train. These are called Slots. Slots have names and you can ask the optimizer for the names of the slots that it uses. Once you have a slot name you can ask the optimizer for the variable it created to hold the slot value.

This can be useful if you want to log debug a training algorithm, report stats about the slots, etc.


These are arguments passed to the optimizer subclass constructor (the __init__ method), and then passed to self._set_hyper(). They can be either regular Python values (like 1.0), tensors, or callables. If they are callable, the callable will be called during apply_gradients() to get the value for the hyper parameter.

Hyperparameters can be overwritten through user code:


# Create an optimizer with the desired parameters.
opt = tf.keras.optimizers.SGD(learning_rate=0.1)
# `loss` is a callable that takes no argument and returns the value
# to minimize.
loss = lambda: 3 * var1 + 2 * var2
# In eager mode, simply call minimize to update the list of variables.
opt.minimize(loss, var_list=[var1, var2])
# update learning rate
opt.learning_rate = 0.05
opt.minimize(loss, var_list=[var1, var2])

Callable learning rate

Optimizer accepts a callable learning rate in two ways. The first way is through built-in or customized tf.keras.optimizers.schedules.LearningRateSchedule. The schedule will be called on each iteration with schedule(iteration), a tf.Variable owned by the optimizer.


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.SGD(learning_rate=learning_rate)
loss = lambda: 3 * var
opt.minimize(loss, var_list=[var])

The second way is through a callable function that does not accept any arguments.


var = tf.Variable(np.random.random(size=(1,)))
def lr_callable():
  return .1
opt = tf.keras.optimizers.SGD(learning_rate=lr_callable)
loss = lambda: 3 * var
opt.minimize(loss, var_list=[var])

Creating a custom optimizer

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

  • _resource_apply_dense (update variable given gradient tensor is dense)
  • _resource_apply_sparse (update variable given gradient tensor is sparse)
  • _create_slots (if your optimizer algorithm requires additional variables)
  • get_config (serialization of the optimizer, include all hyper parameters)

name String. The name to use for momentum accumulator weights created by the optimizer.
gradient_aggregator The function to use to aggregate gradients across devices (when using tf.distribute.Strategy). If None, defaults to summing the gradients across devices. The function should accept and return a list of (gradient, variable) tuples.
gradient_transformers Optional. List of functions to use to transform gradients before applying updates to Variables. The functions are applied after gradient_aggregator. The functions should accept and return a list of (gradient, variable) tuples.
**kwargs keyword arguments. Allowed arguments are clipvalue, clipnorm, global_clipnorm. If clipvalue (float) is set, the gradient of each weight is clipped to be no higher than this value. If clipnorm (float) is set, the gradient of each weight is individually clipped so that its norm is no higher than this value. If global_clipnorm (float) is set the gradient of all weights is clipped so that their global norm is no higher than this value.

ValueError in case of any invalid argument.

clipnorm float or None. If set, clips gradients to a maximum norm.
clipvalue float or None. If set, clips gradients to a maximum value.
global_clipnorm float or None. If set, clips gradients to a maximum norm.
iterations Variable. The number of training steps this Optimizer has run.
weights Returns variables of this Optimizer based on the order created.



View source

Add a new slot variable for var.


View source


View source

Apply gradients to variables.

This is the second part of minimize(). It returns an Operation that applies gradients.

The method sums gradients from all replicas in the presence of tf.distribute.Strategy by default. You can aggregate gradients yourself by passing experimental_aggregate_gradients=False.


grads = tape.gradient(loss, vars)
grads = tf.distribute.get_replica_context().all_reduce('sum', grads)
# Processing aggregated gradients.
optimizer.apply_gradients(zip(grads, vars),

grads_and_vars List of (gradient, variable) pairs.
name Optional name for the returned operation. Default to the name passed to the Optimizer constructor.
experimental_aggregate_gradients Whether to sum gradients from different replicas in the presense of tf.distribute.Strategy. If False, it's user responsibility to aggregate the gradients. Default to True.

An Operation that applies the specified gradients. The iterations will be automatically increased by 1.

TypeError If grads_and_vars is malformed.
ValueError If none of the variables have gradients.
RuntimeError If called in a cross-replica context.


View source

Creates an optimizer from its config.

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

config A Python dictionary, typically the output of get_config.
custom_objects A Python dictionary mapping names to additional Python objects used to create this optimizer, such as a function used for a hyperparameter.

An optimizer instance.


View source

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.

Python dictionary.


View source

Returns gradients of loss with respect to params.

Should be used only in legacy v1 graph mode.

loss Loss tensor.
params List of variables.

List of gradient tensors.

ValueError In case any gradient cannot be computed (e.g. if gradient function not implemented).


View source


View source

A list of names for this optimizer's slots.


View source