tfp.optimizer.StochasticGradientLangevinDynamics

An optimizer module for stochastic gradient Langevin dynamics.

tfp.optimizer.StochasticGradientLangevinDynamics(
    learning_rate,
    preconditioner_decay_rate=0.95,
    data_size=1,
    burnin=25,
    diagonal_bias=1e-08,
    name=None,
    parallel_iterations=10
)

Used in the notebooks

Used in the tutorials
Fitting Dirichlet Process Mixture Model Using Preconditioned Stochastic Gradient Langevin Dynamics

This implements the preconditioned Stochastic Gradient Langevin Dynamics optimizer [(Li et al., 2016)][1]. The optimization variable is regarded as a sample from the posterior under Stochastic Gradient Langevin Dynamics with noise rescaled in each dimension according to RMSProp.

Note: If a prior is included in the loss, it should be scaled by 1/data_size, where data_size is the number of points in the data set. I.e., it should be divided by the data_size term described below.

Examples

Optimizing energy of a 3D-Gaussian distribution

This example demonstrates that for a fixed step size SGLD works as an approximate version of MALA (tfp.mcmc.MetropolisAdjustedLangevinAlgorithm).

import tensorflow as tf
import tensorflow_probability as tfp
import numpy as np

tfd = tfp.distributions
dtype = np.float32

with tf.Session(graph=tf.Graph()) as sess:
  # Set up random seed for the optimizer
  tf.random.set_seed(42)
  true_mean = dtype([0, 0, 0])
  true_cov = dtype([[1, 0.25, 0.25], [0.25, 1, 0.25], [0.25, 0.25, 1]])
  # Loss is defined through the Cholesky decomposition
  chol = tf.linalg.cholesky(true_cov)

  var_1 = tf.Variable(name='var_1', initial_value=[1., 1.])
  var_2 = tf.Variable(name='var_2', initial_value=[1.])

  def loss_fn():
    var = tf.concat([var_1, var_2], axis=-1)
    loss_part = tf.linalg.cholesky_solve(chol, var[..., tf.newaxis])
    return tf.linalg.matvec(loss_part, var, transpose_a=True)

  # Set up the learning rate with a polynomial decay
  step = tf.Variable(0, dtype=tf.int64)
  starter_learning_rate = .3
  end_learning_rate = 1e-4
  decay_steps = 1e4
  learning_rate = tf.compat.v1.train.polynomial_decay(
      starter_learning_rate,
      step,
      decay_steps,
      end_learning_rate,
      power=1.)

  # Set up the optimizer
  optimizer_kernel = tfp.optimizer.StochasticGradientLangevinDynamics(
      learning_rate=learning_rate, preconditioner_decay_rate=0.99)
  optimizer_kernel.iterations = step
  optimizer = optimizer_kernel.minimize(loss_fn, var_list=[var_1, var_2])

  # Number of training steps
  training_steps = 5000
  # Record the steps as and treat them as samples
  samples = [np.zeros([training_steps, 2]), np.zeros([training_steps, 1])]
  sess.run(tf.compat.v1.global_variables_initializer())
  for step in range(training_steps):
    sess.run(optimizer)
    sample = [sess.run(var_1), sess.run(var_2)]
    samples[0][step, :] = sample[0]
    samples[1][step, :] = sample[1]

  samples_ = np.concatenate(samples, axis=-1)
  sample_mean = np.mean(samples_, 0)
  print('sample mean', sample_mean)

Args: learning_rate: Scalar float-like Tensor. The base learning rate for the optimizer. Must be tuned to the specific function being minimized. preconditioner_decay_rate: Scalar float-like Tensor. The exponential decay rate of the rescaling of the preconditioner (RMSprop). (This is "alpha" in Li et al. (2016)). Should be smaller than but nearly 1 to approximate sampling from the posterior. (Default: 0.95) data_size: Scalar int-like Tensor. The effective number of points in the data set. Assumes that the loss is taken as the mean over a minibatch. Otherwise if the sum was taken, divide this number by the batch size. If a prior is included in the loss function, it should be normalized by data_size. Default value: 1. burnin: Scalar int-like Tensor. The number of iterations to collect gradient statistics to update the preconditioner before starting to draw noisy samples. (Default: 25) diagonal_bias: Scalar float-like Tensor. Term added to the diagonal of the preconditioner to prevent the preconditioner from degenerating. (Default: 1e-8) name: Python str describing ops managed by this function. (Default: "StochasticGradientLangevinDynamics") parallel_iterations: the number of coordinates for which the gradients of the preconditioning matrix can be computed in parallel. Must be a positive integer.

Raises
InvalidArgumentError	If preconditioner_decay_rate is a Tensor not in (0,1].
NotImplementedError	If eager execution is enabled.

References

[1]: Chunyuan Li, Changyou Chen, David Carlson, and Lawrence Carin. Preconditioned Stochastic Gradient Langevin Dynamics for Deep Neural Networks. In Association for the Advancement of Artificial Intelligence, 2016. https://arxiv.org/abs/1512.07666

Args
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.

Attributes
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. Check tf.clip_by_global_norm for more details.
iterations	Variable. The number of training steps this Optimizer has run.
variable_scope	Variable scope of all calls to tf.get_variable.
weights	Returns variables of this Optimizer based on the order created.

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.

Args
var	a Variable object.
slot_name	name of the slot variable.
initializer	initializer of the slot variable
shape	(Optional) shape of the slot variable. If not set, it will default to the shape of var.

Returns
A slot variable.

add_weight

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

apply_gradients

apply_gradients(
    grads_and_vars, name=None, experimental_aggregate_gradients=True
)

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.

Example:

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

Args
grads_and_vars	List of (gradient, variable) pairs.
name	Optional name for the returned operation. When None, uses the name passed to the Optimizer constructor. Defaults to None.
experimental_aggregate_gradients	Whether to sum gradients from different replicas in the presence of tf.distribute.Strategy. If False, it's user responsibility to aggregate the gradients. Default to True.

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.
RuntimeError	If called in a cross-replica context.

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.

Args
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 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.

Returns
Python dictionary.

get_gradients

get_gradients(
    loss, params
)

Returns gradients of loss with respect to params.

Should be used only in legacy v1 graph mode.

Args
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()

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

Returns
Weights values as a list of numpy arrays.

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.

Args
loss	Tensor or callable. If a callable, loss should take no arguments and return the value to minimize. If a Tensor, the tape argument must be passed.
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) str. Name for the returned operation.
tape	(Optional) tf.GradientTape. If loss is provided as a Tensor, the tape that computed the loss must be provided.

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
)

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>

Args
weights	weight values as a list of numpy arrays.

variables

variables()

Returns variables of this Optimizer based on the order created.