autora.experimentalist.mixture

Mixture Experimentalist Sampler

sample(conditions, temperature, samplers, params, num_samples=None)

Parameters:

Name	Type	Description	Default
conditions	Union[DataFrame, ndarray]	pool of experimental conditions to evaluate: pd.Dataframe	required
temperature	float	how random is selection of conditions (cannot be 0; (0:1) - the choices are more deterministic than the choices made wrt	required
samplers	list	tuple containing sampler functions, their names, and weights	required
for	sampler functions that return both positive and negative scores, user can provide a list with two weights	the first one will be applied to positive scores, the second one -- to the negative	required
params	dict	nested dictionary. keys correspond to the sampler function names (same as provided in samplers),	required
values	correspond to the dictionaries of function arguments (argument name	its value)	required
num_samples	Optional[int]	number of experimental conditions to select	None

Returns:

Type	Description
DataFrame	Sampled pool of experimental conditions with the scores attached to them

Source code in temp_dir/mixture/src/autora/experimentalist/mixture/__init__.py

def sample(conditions: Union[pd.DataFrame, np.ndarray], temperature: float,
                   samplers: list, params: dict,
                   num_samples: Optional[int] = None) -> pd.DataFrame:
    """

    Args:
        conditions: pool of experimental conditions to evaluate: pd.Dataframe
        temperature: how random is selection of conditions (cannot be 0; (0:1) - the choices are more deterministic than the choices made wrt
        samplers: tuple containing sampler functions, their names, and weights
        for sampler functions that return both positive and negative scores, user can provide a list with two weights: the first one will be applied to positive scores, the second one -- to the negative
        params: nested dictionary. keys correspond to the sampler function names (same as provided in samplers),
        values correspond to the dictionaries of function arguments (argument name: its value)
        num_samples: number of experimental conditions to select

    Returns:
        Sampled pool of experimental conditions with the scores attached to them
    """

    condition_pool = pd.DataFrame(conditions)

    rankings = pd.DataFrame()
    mixture_scores = np.zeros(len(condition_pool))
    ## getting rankings and weighted scores from each function
    for (function, name, weight) in samplers:

        try:
            sampler_params = params[name]
            pd_ranking = function(conditions=condition_pool, **sampler_params)
        except:
            pd_ranking = function(conditions=condition_pool)
        # sorting by index
        pd_ranking = pd_ranking.sort_index()

        if len(weight) == 1:
            weight = weight[0]

        # if only one weight is provided, use it for both negative and positive dimensions
        if isinstance(weight, float) or isinstance(weight, int):
            pd_ranking["score"] = pd_ranking["score"] * weight
        else:
            if len(pd_ranking["score"] < 0) > 0 and len(pd_ranking["score"] > 0) > 0:  # there are both positive and negative values

                pd_ranking.loc[pd_ranking["score"] > 0]["score"] = pd_ranking.loc[pd_ranking["score"] > 0]["score"] * weight[0]  # positive dimension gets the first weight
                pd_ranking.loc[pd_ranking["score"] < 0]["score"] = pd_ranking.loc[pd_ranking["score"] < 0]["score"] * weight[1]  # negative dimension gets the second weight
            else:
                pd_ranking["score"] = pd_ranking["score"] * weight[0]

        pd_ranking.rename(columns={"score": f"{name}_score"}, inplace=True)
        # sum_scores are arranged based on the original conditions_ indices
        mixture_scores = mixture_scores + pd_ranking[f"{name}_score"]

        rankings = pd.merge(rankings, pd_ranking, left_index=True, right_index=True, how="outer")

    # adjust mixture scores wrt temperature
    weighted_mixture_scores_adjusted = adjust_distribution(mixture_scores, temperature)

    if num_samples is None:
        num_samples = condition_pool.shape[0]

    condition_indices = np.random.choice(np.arange(len(condition_pool)), num_samples,
                                         p=weighted_mixture_scores_adjusted, replace=False)
    conditions_ = condition_pool.iloc[condition_indices]
    conditions_["score"] = mixture_scores

    return conditions_