Module: tfl.kronecker_factored_lattice_lib
bookmark_border Stay organized with collections Save and categorize content based on your preferences.

Algorithm implementations required for Kronecker-Factored Lattice layer.

Functions

assert_constraints(...): Asserts that weights satisfy constraints.

bias_initializer(...): Initializes bias depending on output_min and output_max.

custom_reduce_prod(...): tf.reduce_prod(t, axis) with faster custom gradient.

default_init_params(...): Returns default initialization bounds depending on layer output bounds.

evaluate_with_hypercube_interpolation(...): Evaluates a Kronecker-Factored Lattice using hypercube interpolation.

finalize_scale_constraints(...): Clips scale to strictly satisfy all constraints.

finalize_weight_constraints(...): Approximately projects weights to strictly satisfy all constraints.

kfl_random_monotonic_initializer(...): Returns a uniformly random sampled monotonic weight tensor.

scale_initializer(...): Initializes scale depending on output_min and output_max.

verify_hyperparameters(...): Verifies that all given hyperparameters are consistent.