AD for Manopt

[stum...@gmail.com]

ADiGator is a free MATLAB automatic differentiation (AD) toolbox: https://sourceforge.net/projects/adigator/. Perhaps ADiGator could be integrated into Manopt to construct the gradient and Hessian of the cost function if the user cannot supply the analytical gradient and/or Hessian. Pymanopt (https://pymanopt.github.io/), a Python variant of Manopt, is able to utilize the third-party AD libraries Autograd, Theano, and TensorFlow to compute the gradient and Hessian of the cost function.

[Nicolas Boumal]

Hello stumarcus,

Thanks for your message -- we are aware of the automatic differentiation feature in pymanopt, and we also think it's great.

Somehow, we didn't find any completely convincing AD solution in matlab though. Last summer, I had some success building a prototype of AD in Manopt using ADiMat. It worked, but the issue is that to get the derivative of a function, the toolbox actually connects to a distant server over the internet (once for each function file; then it's saved). Since that projects appears no longer to be maintained, this would create an undesirable dependency of manopt with a feature that might break at any point. So we decided not to make this public.

ADiGator was also on my list of packages to try; I do not remember what my conclusions were. I'll have another look. In the meantime, don't hesitate to share your input in that possible direction.

Best,

Nicolas

[Nicolas Boumal]

It seems that ADiGator requires the variable to be a vector.  Am I mistaken?

(In Manopt, it is typical to have matrices, or even higher dimensional arrays, structures or cells as variable.)

[Stuart Rogers]

ADiGator can differentiate vectors and matrices but not higher-dimensional (such as 3-d) arrays. I don't think ADiGator can differentiate structures or cells; you'd have to ask on the ADiGator discussion forum: https://sourceforge.net/p/adigator/discussion/general/  

--
http://www.manopt.org
---
You received this message because you are subscribed to a topic in the Google Groups "Manopt" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/manopttoolbox/JoQsGHUy8OU/unsubscribe.
To unsubscribe from this group and all its topics, send an email to manopttoolbo...@googlegroups.com.
Visit this group at https://groups.google.com/group/manopttoolbox.
To view this discussion on the web visit https://groups.google.com/d/msgid/manopttoolbox/04b45966-26c2-4220-96ee-314d66d2343a%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.