Shogun 4.1.0 - Tajinohi no Agatamori

This is a new feature and cleanup release.

Features:

• Added GEMPLP for approximate inference to the structured output framework [Jiaolong Xu].
• Effeciency improvements of the FITC framework for GP inference (FITC_Laplce, FITC, VarDTC) [Wu Lin].
• Added optimisation of inducing variables in sparse GP inference [Wu Lin].
• Added optimisation methods for GP inference (Newton, Cholesky, LBFGS, ...) [Wu Lin].
• Added Automatic Relevance Determination (ARD) kernel functionality for variational GP inference [Wu Lin].
• Updated Notebook for variational GP inference [Wu Lin].
• New framework for stochastic optimisation (L1/2 loss, mirror descent, proximal gradients, adagrad, SVRG, RMSProp, adadelta, ...) [Wu Lin].
• New Shogun meta-language for automatically generating code listings in all target languages [Esben Sörig].
• Added periodic kernel [Esben Sörig].
• Add gradient output functionality in Neural Nets [Sanuj Sharma].

Bugfixes:

• Fixes for java_modular build using OpenJDK [Björn Esser].
• Catch uncaught exceptions in Neural Net code [Khaled Nasr].
• Fix build of modular interfaces with SWIG 3.0.5 on MacOSX [Björn Esser].
• Fix segfaults when calling delete[] twice on SGMatrix-instances [Björn Esser].
• Fix for building with full-hardening-(CXX|LD)FLAGS [Björn Esser].
• Patch SWIG to fix a problem with SWIG and Python >= 3.5 [Björn Esser].
• Add modshogun.rb: make sure narray is loaded before modshogun.so [Björn Esser].
• set working-dir properly when running R (#2654) [Björn Esser].

Cleanup, efficiency updates, and API Changes:

• Added GPU based dot-products to linalg [Rahul De].
• Added scale methods to linalg [Rahul De].
• Added element wise products to linalg [Rahul De].
• Added element-wise unary operators in linalg [Rahul De].
• Dropped parameter migration framework [Heiko Strathmann].
• Disabled Python integration tests by default [Sergey Lisitsyn, Heiko Strathmann].