[Feat] Support conditionals in Pipeline definition

[eddiebergman]

Just a suggestion, it's a bit redundant but hopefully clear in terms of how it would work

space = {
    "a": HP(
        values=[1, 2, 3, 4]
        conditions=[
            Restrict(
                to=[2, 3],
                when=(
                    HP.Ref("b", values=["red"])
                    | Hp.Ref("b", values=["blue"])
                    & Hp.Ref("b", values=["red"])
                 )
             ),
            Active(when=HP.Ref("b", values=["red", "blue"])
        ]
     ),
    "b": ["red", "green", "blue"]
}
# 'a' is inactive when 'b' is "green" but can be in [2,3] when 'b' in ["red", "blue"]

[LennartPurucker]

A workaround for activating HPs could be the following:

space = {
    "a": [1, 2, 3, 4],
    "b": ["red", "green", "blue"],
    "c": [True, False],
    "d": (0, 1)
}
config = {
    "conditional_meta_info": {
        # Only activate a when b is red or blue
        "a": ("b", ["red", "blue"]),
        # only activate b when c is true (hence, also only activate a when c is True)
        "b": ("c", True),
        # only activate d when b is blue and c is True
        "d": [("b", "blue"), ("c", True)]
    }
}

# ...build space, sample from space...

input_pipeline = pipeline.configure(trial.config)

for config_key, conditions in input_pipeline.tail.config.pop("conditional_meta_info").items():
    list_check = lambda c_v: (c_v if isinstance(c_v, list) else [c_v])
    conditions = list_check(conditions)

    if any(
            # 1. Check if relevant key already deleted or not part of the config, if so we
            #   assume the condition is not met.
            (condition_key not in input_pipeline.tail.config) or
            # 2. Check if condition is not met directly.
            (input_pipeline.tail.config[condition_key] not in list_check(condition_values))
            # - check this for all possible conditions
            for condition_key, condition_values in conditions
    ):
        #  If any of the above for any condition, delete the config key;
        #   removing it from the pipeline.
        del input_pipeline.tail.config[config_key]

We build a space with all HPs and remove deactivated HPs post hoc.