searchlight revisited

[JasperVanDenBosch]

Rethink of the searchlight interface, architecture as well as introducing a surface-based version.

datasets

• sitek (private)
• jones-michael-s (private)
• cichy16 (public)
• jsong (private)

issues

• Runtime warnings and errors when using searchlight #248
• Surface searchlight: cache indices for template spaces #292
• Add pattern descriptor to get_searchlight_RDMs output #272
• parallel searchlight does not use more than 8-10 cores efficiently #213

thoughts and notes

Performance is important for searchlight analysis. The current interface does not readily allow or isn't optimised for an approach where each chunk of data goes through multiple steps in a pipeline; but this would enable parallelisation on an HPC infrastructure. An approach with in-process parallelization using e.g. job lib brings certain complications as well.

[iancharest]

i had to create an intermediate calc_rdm_batch function to deal with the fact that calc_rdm involves recursion and with parallel, this lead to maximum recursive depth errors.