States and transitions#
The transition entry of a state factor gives the distribution of its next value given its
parents: its own previous value, other state factors, and control nodes. It accepts:
a table over discrete levels:
ConditionalCategorical, orConditionalDirichletfor a table that is learned;a linear map with Gaussian noise, optionally one map per level of a discrete parent (a switching linear model):
LinearGaussianTransition;a nonlinear map with Gaussian noise, which the map may scale per coordinate by the state and the controls:
NonlinearGaussianTransition;a neural conditional:
MLPConditionalfor a continuous state,MLPCategoricalfor a categorical code;a deterministic map that is computed forward and carried:
Deterministic;a reference to a shared table:
Parameter(see Shared parameters, emissions and recognition models).
The distribution families themselves are listed under Distributions.
Continuous and deterministic transitions#
|
|
|
|
A deterministic node: a pure map, optionally with trainable parameters. |