I created this tool to design parameter spaces, run functions in parallel and optimize hyper-parameters for machine learning models, it's pretty generic and a work in progress, but it gets the job done. There are possibly corner cases I didn't consider, but I'll add/fix things as I need/find them.
Yes, I just wanted a way to separate the model logic from the experiment design. So you design a parameter space in a configuration file, and either use the optimizer to find the best parameters within that space, or run every possible combination of parameters (grid seach)
There are plenty of ways to optimize models from evolutionary algorithms to bandit models, in my simple case I'm working with limited resources, so I included a wrapper to a global bayesian optimizer for sequential search. Grid search is out of the question, but the tool is still handy if I want to freeze all the parameters and vary a couple of them to study some aspect.
bonus: the parameter space file is a nice way to document my experiments --I can track exactly how each experiment was configured.
