Module: tf.contrib.kfac.fisher_blocks

Defined in tensorflow/contrib/kfac/python/ops/

FisherBlock definitions.


class ConvDiagonalFB: FisherBlock for 2-D convolutional layers using a diagonal approx.

class ConvKFCBasicFB: FisherBlock for convolutional layers using the basic KFC approx.

class EmbeddingKFACFB: K-FAC FisherBlock for embedding layers.

class FisherBlock: Abstract base class for objects modeling approximate Fisher matrix blocks.

class FullFB: FisherBlock using a full matrix estimate (no approximations).

class FullyConnectedDiagonalFB: FisherBlock for fully-connected (dense) layers using a diagonal approx.

class FullyConnectedKFACBasicFB: K-FAC FisherBlock for fully-connected (dense) layers.

class KroneckerProductFB: A base class for blocks with separate input and output Kronecker factors.

class NaiveDiagonalFB: FisherBlock using a diagonal matrix approximation.



compute_pi_tracenorm(...): Computes the scalar constant pi for Tikhonov regularization/damping.

normalize_damping(...): Normalize damping after adjusting scale by NORMALIZE_DAMPING_POWER.

num_conv_locations(...): Returns the number of spatial locations a 2D Conv kernel is applied to.

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

Other Members