Roadmap Shogun 2015 hack

Shogun hack 2015 Ideas

This is a list of topics to address when all Shogun developers get together. We aim to have a structured list of items, with links to existing or newly created github issues. Feel free to merge this with content from the other Wiki pages or discussions.

Usability.

• A meta-langauge for Shogun example: #2508.
• Website updates (Kevin had some ideas?)
• Developer information (wiki or website?) #2533
• Manifest "What is Shogun? What tries Shogun to be?"
• Notebooks of things that are not yet covered
• Binary packages for various OS
• Ubuntu LTS #2520
• OSX #2521
• Improve [Shogun development guidelines](Shogun development guidelines)
• Scheduling a separate release for "easier installation": packages and scripts for Ubuntu/Debian/MacOSC https://github.com/shogun-toolbox/shogun/milestones/Shogun%203.4.1

Efficiency & Clean-ups

• Efficient and flexible data structure for n-dimensional array #2123 #2530
• Modularise Shogun to reduce compile memory requirements #2437 (some prototype is https://github.com/lisitsyn/aer)
• Serialisation via external library (Viktor?)
• General polishing of base classes #2113(Thoralf?)
• DPointers (again https://github.com/lisitsyn/aer)

Computing Backends.

• Parallelism backend interface
• Batch cluster backend (PBS/SGE,SLURM,etc) #1622
• Pthread/openMP backend #1623
• Director classes to use Shogun as scheduler from modular interface #1624
• Add dependencies between jobs #2524
• Graphlab backend #2525
• Zookeeper coordination?
• Populate internal Linear algebra interface
• Add matrix factorisations #2526
• Add linear solves #2527
• Stan for autodiff and MCMC #1875 #1998 #1929
• Vowpal Wabbit update

SHOGUN goes large-scale

• Investigate how non-toy examples scale in SHOGUN.
• Increase awareness of benchmarking and profiling:
• Identify hotspots using selected examples and right tools (gprof?).
• Create minimal example(s!) which covers hotspot and benchmark.
• Start optimizing minimal example(s!) and benchmark again.
• Developers should know how to benchmark and profile: Small document to describe this?
• If you know examples that scale better in other frameworks, please document them here, so we can investigate:
• ...
• Better support of sparse-datastructures and online-algorithms.
• Interface with MOE for better model selection