tf_privacy.GenericDPAdamOptimizer

View source on GitHub

Differentially private subclass of class tf.keras.optimizers.legacyAdam.

tf_privacy.GenericDPAdamOptimizer(
    dp_sum_query: tf_privacy.DPQuery,
    num_microbatches: Optional[int] = None,
    gradient_accumulation_steps: int = 1,
    *args,
    **kwargs
)

You can use this as a differentially private replacement for tf.keras.optimizers.legacyAdam. This optimizer implements a differentiallyy private version of the stochastic gradient descent optimizer cls using the chosen dp_query.DPQuery instance.

When instantiating this optimizer, you need to supply several DP-related arguments followed by the standard arguments for Adam.

Examples:

# Create optimizer.
gaussian_query = gaussian_query.GaussianSumQuery(
    l2_norm_clip=1.0, noise_multiplier=0.5, num_microbatches=1
)
opt = DPKerasAdam(dp_sum_query=gaussian_query, <standard arguments>)

When using the optimizer, be sure to pass in the loss as a rank-one tensor with one entry for each example.

The optimizer can be used directly via its minimize method, or through a Keras Model.

# Compute loss as a tensor by using tf.losses.Reduction.NONE.
# Compute vector of per-example loss rather than its mean over a minibatch.
loss = tf.keras.losses.CategoricalCrossentropy(
    from_logits=True, reduction=tf.losses.Reduction.NONE)

# Use optimizer in a Keras model.
opt.minimize(loss, var_list=[var])

# Compute loss as a tensor by using tf.losses.Reduction.NONE.
# Compute vector of per-example loss rather than its mean over a minibatch.
loss = tf.keras.losses.CategoricalCrossentropy(
    from_logits=True, reduction=tf.losses.Reduction.NONE)

# Use optimizer in a Keras model.
model = tf.keras.Sequential(...)
model.compile(optimizer=opt, loss=loss, metrics=['accuracy'])
model.fit(...)

In DP-SGD training, a larger batch size typically helps to achieve better privacy/utility tradeoff. However there is typically a maximum batch size imposed by hardware. This optimizer can emulate large batch sizes on hardware with limited memory by accumulating gradients for several steps before actually applying them to update model weights. Constructor argument gradient_accumulation_steps controls the number of steps for which gradients are accumulated before updating the model weights.

Below is an example which demonstrates how to use this feature:

# Create optimizer which will be accumulating gradients for 4 steps.
# and then performing an update of model weights.
gaussian_query = gaussian_query.GaussianSumQuery(
    l2_norm_clip=1.0, noise_multiplier=0.5, num_microbatches=1
)
opt = DPKerasAdam(dp_sum_query=gaussian_query,
                       num_microbatches=1,
                       gradient_accumulation_steps=4,
                       <standard arguments>)

# Use optimizer in a regular way.
# First three calls to opt.minimize won't update model weights and will
# only accumulate gradients. Model weights will be updated on the fourth
# call to opt.minimize
opt.minimize(loss, var_list=[var])

Note that when using this feature,

1. effective batch size is gradient_accumulation_steps * one_step_batch_size where one_step_batch_size is the size of the batch passed to single step of the optimizer. Thus user may have to adjust learning rate, weight decay and possibly other training hyperparameters accordingly.
2. effective noise (the noise to be used for privacy computation) is noise_multiplier * sqrt(gradient_accumulation_steps), as the optimizer adds noise of self._noise_multiplier to every step. Thus user may have to adjust the noise_multiplier or the privacy computation. Additionally, user may need to adjust the batch size in the data generator, or the number of calls to the data generator, depending on the training framework used. For example, when using Keras model.fit(...) with a user-defined data generator, one may need to make the data generator return one_step_batch_size examples each time, and scale the steps_per_epoch by gradient_accumulation_steps. This is because the data generator is called steps_per_epoch times per epoch, and one call only returns one_step_batch_size (instead of effective_batch_size) examples now.

Methods

add_slot

add_slot(
    var, slot_name, initializer='zeros', shape=None
)

Add a new slot variable for var.

A slot variable is an additional variable associated with var to train. It is allocated and managed by optimizers, e.g. Adam.

add_weight

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

apply_gradients

View source

apply_gradients(
    *args, **kwargs
)

DP-SGD version of base class method.

from_config

@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.

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.

get_gradients

View source

get_gradients(
    loss, params
)

DP-SGD version of base class method.

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()

Returns the current weights of the optimizer.

The weights of an optimizer are its state (ie, variables). This function returns the weight values associated with this optimizer as a list of Numpy arrays. The first value is always the iterations count of the optimizer, followed by the optimizer's state variables in the order they were created. The returned list can in turn be used to load state into similarly parameterized optimizers.

For example, the RMSprop optimizer for this simple model returns a list of three values-- the iteration count, followed by the root-mean-square value of the kernel and bias of the single Dense layer:

opt = tf.keras.optimizers.legacy.RMSprop()
m = tf.keras.models.Sequential([tf.keras.layers.Dense(10)])
m.compile(opt, loss='mse')
data = np.arange(100).reshape(5, 20)
labels = np.zeros(5)
results = m.fit(data, labels)  # Training.
len(opt.get_weights())
3

minimize

minimize(
    loss, var_list, grad_loss=None, name=None, 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.

set_weights

set_weights(
    weights
)

Set the weights of the optimizer.

The weights of an optimizer are its state (ie, variables). This function takes the weight values associated with this optimizer as a list of Numpy arrays. The first value is always the iterations count of the optimizer, followed by the optimizer's state variables in the order they are created. The passed values are used to set the new state of the optimizer.

For example, the RMSprop optimizer for this simple model takes a list of three values-- the iteration count, followed by the root-mean-square value of the kernel and bias of the single Dense layer:

opt = tf.keras.optimizers.legacy.RMSprop()
m = tf.keras.models.Sequential([tf.keras.layers.Dense(10)])
m.compile(opt, loss='mse')
data = np.arange(100).reshape(5, 20)
labels = np.zeros(5)
results = m.fit(data, labels)  # Training.
new_weights = [np.array(10), np.ones([20, 10]), np.zeros([10])]
opt.set_weights(new_weights)
opt.iterations
<tf.Variable 'RMSprop/iter:0' shape=() dtype=int64, numpy=10>

variables

variables()

Returns variables of this Optimizer based on the order created.