View source on GitHub
|
Contains logic for defining custom inverses for functions.
Automatic inversion works for only a certain class of functions (see
core.inverse documentation for more details). For example, an
autoregressive neural network will likely use masked weight matrices in order
to be invertible but the automatic inversion is not aware of autoregressive
masked matrices (yet!). Furthermore, we may want more numerically stable
inverses for functions like softmax or sigmoid.
This module provides a custom_inverse decorator for Python functions that
enables overriding the default programmatic inversion. See custom_inverse
for further documentation.
Classes
class NonInvertibleError: Raised by a custom inverse definition when values are unknown.
Functions
custom_inverse(...): Decorates a function to enable defining a custom inverse.