layer_dense_flipout: Densely-connected layer class with Flipout estimator.

Description

This layer implements the Bayesian variational inference analogue to a dense layer by assuming the kernel and/or the bias are drawn from distributions.

Usage

layer_dense_flipout(
  object,
  units,
  activation = NULL,
  activity_regularizer = NULL,
  trainable = TRUE,
  kernel_posterior_fn = tfp$layers$util$default_mean_field_normal_fn(),
  kernel_posterior_tensor_fn = function(d) d %>% tfd_sample(),
  kernel_prior_fn = tfp$layers$util$default_multivariate_normal_fn,
  kernel_divergence_fn = function(q, p, ignore) tfd_kl_divergence(q, p),
  bias_posterior_fn = tfp$layers$util$default_mean_field_normal_fn(is_singular = TRUE),
  bias_posterior_tensor_fn = function(d) d %>% tfd_sample(),
  bias_prior_fn = NULL,
  bias_divergence_fn = function(q, p, ignore) tfd_kl_divergence(q, p),
  seed = NULL,
  ...
)

Value

a Keras layer

Arguments

object

What to compose the new Layer instance with. Typically a Sequential model or a Tensor (e.g., as returned by layer_input()). The return value depends on object. If object is:

missing or NULL, the Layer instance is returned.
a Sequential model, the model with an additional layer is returned.
a Tensor, the output tensor from layer_instance(object) is returned.

units

integer dimensionality of the output space

activation

Activation function. Set it to None to maintain a linear activation.

activity_regularizer

Regularizer function for the output.

trainable

Whether the layer weights will be updated during training.

kernel_posterior_fn

Function which creates tfd$Distribution instance representing the surrogate posterior of the kernel parameter. Default value: default_mean_field_normal_fn().

kernel_posterior_tensor_fn

Function which takes a tfd$Distribution instance and returns a representative value. Default value: function(d) d %>% tfd_sample().

kernel_prior_fn

Function which creates tfd$Distribution instance. See default_mean_field_normal_fn docstring for required parameter signature. Default value: tfd_normal(loc = 0, scale = 1).

kernel_divergence_fn

Function which takes the surrogate posterior distribution, prior distribution and random variate sample(s) from the surrogate posterior and computes or approximates the KL divergence. The distributions are tfd$Distribution-like instances and the sample is a Tensor.

bias_posterior_fn

Function which creates a tfd$Distribution instance representing the surrogate posterior of the bias parameter. Default value: default_mean_field_normal_fn(is_singular = TRUE) (which creates an instance of tfd_deterministic).

bias_posterior_tensor_fn

Function which takes a tfd$Distribution instance and returns a representative value. Default value: function(d) d %>% tfd_sample().

bias_prior_fn

Function which creates tfd instance. See default_mean_field_normal_fn docstring for required parameter signature. Default value: NULL (no prior, no variational inference)

bias_divergence_fn

seed

scalar integer which initializes the random number generator. Default value: NULL (i.e., use global seed).

...

Additional keyword arguments passed to the keras::layer_dense constructed by this layer.

Details

By default, the layer implements a stochastic forward pass via sampling from the kernel and bias posteriors,

kernel, bias ~ posterior
outputs = activation(matmul(inputs, kernel) + bias)

It uses the Flipout estimator (Wen et al., 2018), which performs a Monte Carlo approximation of the distribution integrating over the kernel and bias. Flipout uses roughly twice as many floating point operations as the reparameterization estimator but has the advantage of significantly lower variance.

The arguments permit separate specification of the surrogate posterior (q(W|x)), prior (p(W)), and divergence for both the kernel and bias distributions.

Upon being built, this layer adds losses (accessible via the losses property) representing the divergences of kernel and/or bias surrogate posteriors and their respective priors. When doing minibatch stochastic optimization, make sure to scale this loss such that it is applied just once per epoch (e.g. if kl is the sum of losses for each element of the batch, you should pass kl / num_examples_per_epoch to your optimizer).

References