On top of this, maybe we should do a version of train without any of the logging/dashboard logic, and a return of nothing (or only the final accuracy). All the pieces are in place already. After that, doing a pmap and using Distributed.jl could speed things up allowing us to do our grid search in the cluster. I will take a look at this tomorrow.

I managed to use threads in parallel with the @thyperopt macro that ships with the package. They also have the @phyperopt macro to use different workers but I didn't get that to work yet.

I maybe have not understood threads and workers - is threads simply different cpus on one node and workers maybe potentially on different nodes?