Hi Aditya, thanks for getting in touch! Your contributions are more than welcome here.
we decided against doing an autodiff project this year. Here are our reasons: -we want to use an existing library, implementing this ourselves is not an option -the existing libraries do not meet our requirements. tensorlow might in the future, but for now it can only build computation grapgs from Python. stan-math is based on heavy template programming which makes boost a requirement, slows down compile time. We might use stan-math for a sub-module at some point (there is a project on that), but not as a basis for doing autodiff in Shogun. I reccomend you have a look at the large-scale GP project, which will be the first place to try autodiff in the future. The project will in fact prepare that via re-factoring the framework. It is a great project, and Wu is an extremely talented coder. The other interesting project is obviously the deep learning one, where you will have to implement all sorts of gradients. Both projects (and in fact most projects) will involve some very interesting low-level linear algegra, and indirectly provide grounds for potentially integrating an autodiff library in the future. All the best Heiko 2016-03-02 9:52 GMT+00:00 Aditya Chivukula <[email protected]>: > Hello, > > I had e-mailed earlier regarding using automatic differentiation in > Shogun. I'd like to get started on writing some base code for the same. The > getting involved page provides instructions for existing issues, so do I > create a new issue? Do the developers have any guidelines to follow for > autodiff in particular? > > Also, autodiff was a separate project on the project list page a week back > but seems have been removed. Is not no longer a candidate project for GSoC > or has it been clubbed with another project (couldn't find it)? > > Thanks! > > -- > Aditya Chivukula, > *Ubiety International* > >
