hyper_skopt.py

"""Hyper optimization using scikit-optimize.
"""
from .hyper import register_hyper_optlib


def convert_param_to_skopt(param, name):
    from skopt.space import Real, Integer, Categorical

    if param["type"] == "BOOL":
        return Categorical([False, True], name=name)
    if param["type"] == "INT":
        return Integer(low=param["min"], high=param["max"], name=name)
    if param["type"] == "STRING":
        return Categorical(param["options"], name=name)
    if param["type"] == "FLOAT":
        return Real(low=param["min"], high=param["max"], name=name)
    if param["type"] == "FLOAT_EXP":
        return Real(
            low=param["min"],
            high=param["max"],
            base=10,
            prior="log-uniform",
            name=name,
        )
    else:
        raise ValueError("Didn't understand space {}.".format(param))


def get_methods_space(methods):
    from skopt.space import Categorical

    return [Categorical(methods)]


def convert_to_skopt_space(method, space):
    return [
        convert_param_to_skopt(param, name=name)
        for name, param in space[method].items()
    ]


def skopt_init_optimizers(
    self,
    methods,
    space,
    sampler="et",
    method_sampler="et",
    sampler_opts=None,
    method_sampler_opts=None,
):
    """Initialize the ``skopt`` optimizer.

    Parameters
    ----------
    space : dict[str, dict[str, dict]]
        The search space.
    sampler :  str, optional
        The regressor to use to optimize each method's search space, see
        https://scikit-optimize.github.io/stable/modules/generated/skopt.Optimizer.html#skopt.Optimizer
        .
    method_sampler : str, optional
        Meta-optimizer to use to select which overall method to use.
    """
    from skopt.optimizer import Optimizer

    sampler_opts = {} if sampler_opts is None else dict(sampler_opts)
    method_sampler_opts = (
        {} if method_sampler_opts is None else dict(method_sampler_opts)
    )

    if method_sampler is None:
        method_sampler = sampler

    self._method_chooser = Optimizer(
        get_methods_space(methods),
        base_estimator=method_sampler,
        **method_sampler_opts,
    )

    skopt_spaces = {m: convert_to_skopt_space(m, space) for m in methods}
    self._param_names = {m: [p.name for p in skopt_spaces[m]] for m in methods}
    self._optimizers = {
        m: Optimizer(skopt_spaces[m], base_estimator=sampler, **sampler_opts)
        for m in methods
    }


def skopt_get_setting(self):
    """Find the next parameters to test."""
    # params = self._optimizer.ask()
    # return params

    import warnings

    with warnings.catch_warnings():
        warnings.filterwarnings("ignore", module="skopt")
        warnings.filterwarnings("ignore", module="sklearn")
        method = self._method_chooser.ask()
        params = self._optimizers[method[0]].ask()
        names = self._param_names[method[0]]

    return {
        "method_token": method,
        "method": method[0],
        "params_token": params,
        "params": dict(zip(names, params)),
    }


def skopt_report_result(self, setting, trial, score):
    """Report the result of a trial to the ``chocolate`` optimizer."""
    import warnings

    with warnings.catch_warnings():
        warnings.filterwarnings("ignore", module="skopt")
        warnings.filterwarnings("ignore", module="sklearn")

        self._method_chooser.tell(setting["method_token"], score)
        self._optimizers[setting["method"]].tell(
            setting["params_token"], score
        )


register_hyper_optlib(
    "skopt",
    skopt_init_optimizers,
    skopt_get_setting,
    skopt_report_result,
)