GSoC 2016 summit discussion

Heiko Strathmann edited this page Oct 31, 2016 · 2 revisions
Clone this wiki locally

Viktor and Heiko had some really good discussions on the GSoC 2016 mentor summit, as a result of spending a week of hacking Shogun before the summit. See photo of notes on bottom.

Hacking

  • Updated the website, and automated its deployment
  • Automated the release process
  • Automate nightly binaries for ubuntu
  • Extended integration testing across all interfaces
  • fixed tons of bugs
  • ...

Next steps

This is a suggestion for a set of releases, roughly ordered in priority, with big blobs (that require a major release) grouped together. There is a parallel stream of smaller fixes below, along with a set of more vague goals for the more distant future.

5.0

  • This release was prepared at the summit

5.1

  • Bugfix release for 5.0
  • milestone
  • Mostly about fixing problems we wanted to fix at the summit, but did not find time.
  • All relatively high priority and should be done before moving on, at least partly
  • Add scala in examples (re-uses java)

6.0

  • milestone
  • Integrate tags GSoC 2016 project (requires a SG_ADD hack)
  • Integreate cereal GSoC 2016 project (requires tags)
  • Based on new cerealisation, make a hash based equals method
  • remove existing parameter framework
  • c++11 hard requirement
  • integrate linalg GSoC 2016 project

7.0

  • Constrained parameters
  • Every parameter gets a range of valid values it can take, potentially a domain
  • ...

8.0 (maybe can be merged with previous)

  • shared pointer shogun wide
  • base interface for ML (abstract classes, restructure shoguns API, depends on constrained parameters)
  • split SWIG/cmake/tests, modularise library, ship separate binaries

Other changes, independent of release, in increasing priority

  • integration of interfaces into target language system (e.g. maven for java)
  • BSD license change
  • get rid of jblas and ndarray and replace with more modern libraries
  • New typemaps and examples for js, Matlab, D

Important things for moving the project forward

  • Build a pipeline for supervised prediction, fix things on the fly
  • Pull out the interfaces from the Shogun repo (and examples, testing, build)
  • Establish affiliation (numfocus or other) for funding, give up on own eV

Distance future dreams

  • graph based representation of computation, use something as tensorflow as backend, just like linalg
  • Standard ML API with Shogun as reference implementation
  • OpenML

Photo of notes

mentor_summit_notes