excessive memory use of rFerns in parallel use

[tobigithub]

method rFerns 0.9353 0.9021 58.45 Random Ferns in caret uses 2 Gbyte per parallel client, so 32 clients require at least 64 GByte RAM. Method works and converges, just memory use seems excessive (plus speed).

[mbq]

rFerns author here; this is because both rFerns model size and training time scales with 2^depth, and caret scans the entire 1-16 range of its possible values. Going over 12 is simply impractical in most use cases, even if it brings some minor accuracy gain.

[tobigithub]

Thank you. Tobias