SMAC3 is a powerful tool for automated algorithm configuration and parameter tuning. It is based on Bayesian optimization and utilizes racing mechanisms and random forests.
This directory contains an example with three different variants on how SMAC3 can specifically be used for tuning parameters of algorithms implemented in Julia. This can be achieved in different ways.
The main script is tuning.py, and the variant to be used can be provided as
command line argument:
-
This variant uses the Python->Julia interface provided by the Python package
pyjulia. The julia functionfcan here be called from SMAC3 via a simple Python wrapper function. A limitation of this approach is that no parallelization via multithreading or multiprocessing is possible. -
Here, for each call of the function to tune, a new Julia process is started. While this approach allows for parallelization, also using a compute cluster, it can be quite inefficient when the function to tune is comparably fast w.r.t. Julia's startup and pre-compile time.
-
Here, the main Python script spawns a fixed number of Julia "server" subprocesses realized by
julia_server.jl, which accept function calls, perform them, and return the result. The communication is done via pipes, respectively stdin/stdout. In this way multiprocessing is utilized, but the overhead of starting a new Julia process for each function call is avoided.
Note that the demo needs to be started from the Tuning subdirectory.
Ensure that Python is installed with the packages julia, smac and ConfigSpace.