tfa.optimizers.SWA

View source on GitHub

Class SWA

This class extends optimizers with Stochastic Weight Averaging (SWA).

Inherits From: AveragedOptimizerWrapper

The Stochastic Weight Averaging mechanism was proposed by Pavel Izmailov et. al in the paper Averaging Weights Leads to Wider Optima and Better Generalization. The optimizer implements averaging of multiple points along the trajectory of SGD. The optimizer expects an inner optimizer which will be used to apply the gradients to the variables and itself computes a running average of the variables every k steps (which generally corresponds to the end of a cycle when a cyclic learning rate is employed).

We also allow the specification of the number of steps averaging should first happen after. Let's say, we want averaging to happen every k steps after the first m steps. After step m we'd take a snapshot of the variables and then average the weights appropriately at step m + k, m + 2k and so on. The assign_average_vars function can be called at the end of training to obtain the averaged_weights from the optimizer.

Example of usage:

opt = tf.keras.optimizers.SGD(learning_rate)
opt = tfa.optimizers.SWA(opt, start_averaging=m, average_period=k)

__init__

View source

__init__(
    optimizer,
    start_averaging=0,
    average_period=10,
    name='SWA',
    sequential_update=True,
    **kwargs
)

Wrap optimizer with the Stochastic Weight Averaging mechanism.

Args:

  • optimizer: The original optimizer that will be used to compute and apply the gradients.
  • start_averaging: An integer. Threshold to start averaging using SWA. Averaging only occurs at start_averaging iters, must be >= 0. If start_averaging = m, the first snapshot will be taken after the mth application of gradients (where the first iteration is iteration 0).
  • average_period: An integer. The synchronization period of SWA. The averaging occurs every average_period steps. Averaging period needs to be >= 1.
  • name: Optional name for the operations created when applying gradients. Defaults to 'SWA'.
  • sequential_update: Bool. If False, will compute the moving average at the same time as the model is updated, potentially doing benign data races. If True, will update the moving average after gradient updates.
  • **kwargs: keyword arguments. Allowed to be {clipnorm, clipvalue, lr, decay}. clipnorm is clip gradients by norm; clipvalue is clip gradients by value, decay is included for backward compatibility to allow time inverse decay of learning rate. lr is included for backward compatibility, recommended to use learning_rate instead.

Properties

iterations

Variable. The number of training steps this Optimizer has run.

learning_rate

lr

weights

Returns variables of this Optimizer based on the order created.

Methods

add_slot

add_slot(
    var,
    slot_name,
    initializer='zeros'
)

Add a new slot variable for var.

add_weight

add_weight(
    name,
    shape,
    dtype=None,
    initializer='zeros',
    trainable=None,
    synchronization=tf_variables.VariableSynchronization.AUTO,
    aggregation=tf_variables.VariableAggregation.NONE
)

apply_gradients

View source

apply_gradients(
    grads_and_vars,
    name=None
)

Apply gradients to variables.

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

Args:

  • grads_and_vars: List of (gradient, variable) pairs.
  • name: Optional name for the returned operation. Default to the name passed to the Optimizer constructor.

Returns:

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

Raises:

  • TypeError: If grads_and_vars is malformed.
  • ValueError: If none of the variables have gradients.

assign_average_vars

View source

assign_average_vars(var_list)

Assign variables in var_list with their respective averages.

Args:

  • var_list: List of model variables to be assigned to their average.

Returns:

  • assign_op: The op corresponding to the assignment operation of variables to their average.

Example:

model = tf.Sequential([...])
opt = tfa.optimizers.SWA(
        tf.keras.optimizers.SGD(lr=2.0), 100, 10)
model.compile(opt, ...)
model.fit(x, y, ...)

# Update the weights to their mean before saving
opt.assign_average_vars(model.variables)

model.save('model.h5')

average_op

View source

average_op(
    var,
    average_var
)

from_config

View source

@classmethod
from_config(
    cls,
    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.

Arguments:

  • 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.

Returns:

An optimizer instance.

get_config

View source

get_config()

Returns the config of the optimimizer.

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.

Returns:

Python dictionary.

get_gradients

get_gradients(
    loss,
    params
)

Returns gradients of loss with respect to params.

Arguments:

  • loss: Loss tensor.
  • params: List of variables.

Returns:

List of gradient tensors.

Raises:

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

get_slot

get_slot(
    var,
    slot_name
)

get_slot_names

get_slot_names()

A list of names for this optimizer's slots.

get_updates

get_updates(
    loss,
    params
)

get_weights

get_weights()

minimize

minimize(
    loss,
    var_list,
    grad_loss=None,
    name=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: A callable taking no arguments which returns 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.
  • grad_loss: Optional. A Tensor holding the gradient computed for loss.
  • name: Optional name for the returned operation.

Returns:

An Operation that updates the variables in var_list. The iterations will be automatically increased by 1.

Raises:

  • ValueError: If some of the variables are not Variable objects.

set_weights

set_weights(weights)

variables

variables()

Returns variables of this Optimizer based on the order created.