Skip to content

Rules for gradients of complex loss functions

Notifications You must be signed in to change notification settings

guochu/complexAD

Folders and files

NameName
Last commit message
Last commit date

Latest commit

acc5b1e · Dec 10, 2020

History

4 Commits
Mar 2, 2020
Mar 2, 2020
Mar 2, 2020
Mar 2, 2020
Dec 10, 2020
Mar 2, 2020

Repository files navigation

Examples of automatic differentiation mentioned in the paper arXiv:~. The fold src contains utility functions which are used to derive the numerical gradient by finite difference method.

  • In the file "common_funcs.jl", we show the definition of the adjoint function for the non-holomorphic function "dot", and checked the gradient.

  • In the file "arjovsky16.jl", we parameterize the W matrix using a mixture of real and complex numbers, which is used for unitary recurrent neural network in the paper Unitary Evolution Recurrent Neural Networks

  • In the file "wisdom.jl", we use complex numbers to parameterize the W matrix propose in the paper Full-Capacity Unitary Recurrent Neural Networks. Compared to the original work which split the W matrix into two matrices which contain the real and imaginary parts separately, we use a single complex matrix to represent it, and the gradient of any complex function built with W can be computed as straightforward as the real case.

About

Rules for gradients of complex loss functions

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages