Module: tf.contrib.kfac.fisher_factors

Defined in tensorflow/contrib/kfac/python/ops/fisher_factors_lib.py.

FisherFactor definitions.

Classes

class ConvDiagonalFactor: FisherFactor for a diagonal approx of a convolutional layer's Fisher.

class ConvInputKroneckerFactor: Kronecker factor for the input side of a convolutional layer.

class ConvOutputKroneckerFactor: Kronecker factor for the output side of a convolutional layer.

class DiagonalFactor: A base class for FisherFactors that use diagonal approximations.

class EmbeddingInputKroneckerFactor: FisherFactor for input to an embedding layer.

class FisherFactor: Base class for objects modeling factors of approximate Fisher blocks.

class FullFactor: FisherFactor for a full matrix representation of the Fisher of a parameter.

class FullyConnectedDiagonalFactor: FisherFactor for a diagonal approx of a fully-connected layer's Fisher.

class FullyConnectedKroneckerFactor: Kronecker factor for the input or output side of a fully-connected layer.

class NaiveDiagonalFactor: FisherFactor for a diagonal approximation of any type of param's Fisher.

Functions

append_homog(...): Appends a homogeneous coordinate to the last dimension of a Tensor.

compute_cov(...): Compute the empirical second moment of the rows of a 2D Tensor.

covariance_initializer(...)

diagonal_covariance_initializer(...)

inverse_initializer(...)

scalar_or_tensor_to_string(...)

scope_string_from_name(...)

scope_string_from_params(...): Builds a variable scope string name from the given parameters.

set_global_constants(...): Sets various global constants used by the classes in this module.