I have a function not supported out of the box by autodiff, and looking for ways to specify its derivative manually. Specifically, in my case it is a fourier transform from NFFT.jl, so it is possible to differentiate efficiently. Surprisingly, I could not find anything related to custom differentiable functions in the docs. Any suggestions?
Take a look at the section on user-defined functions in the JuMP docs:
http://www.juliaopt.org/JuMP.jl/0.18/nlp.html#user-defined-functions
It includes an example where the derivatives are supplied manually.