mehdidc/fluentopt

A flexible hyper-parameter optimization library for machine learning

fluentopt

A flexible hyper-parameter optimization library. The full documentation is available at http://fluentopt.readthedocs.io/en/latest/

Most hyper-parameter optimization libraries impose three main restrictions :

• they control the optimization loop
• they force the inputs to be represented by vectors
• the priors are very restricted, e.g gaussian, uniform or discrete uniform

the goal of fluentopt is to provide hyper-parameter optimization library where :

• the optimization loop is controlled by the user (but we will provide also helpers).
• the inputs can be represented by a python dictionary to express conditionals rather than just a list (or vector), but in case not needed they can also just be a list or a scalar. The dictionaries can contain strings, varying length lists and special objects like 'None'.
• the priors of the hyper-parameters are not restricted to some pre-defined probability distributions. Users will just provide samplers as a python function, that is, a function that takes a seed and returns a python dictionary.