Source code for ax.modelbridge.transforms.merge_repeated_measurements

#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

# pyre-strict

from __future__ import annotations

from collections import defaultdict
from copy import deepcopy
from typing import Optional

import numpy as np
from ax.core.arm import Arm
from ax.core.observation import Observation, ObservationData, separate_observations
from ax.core.search_space import SearchSpace
from ax.modelbridge.base import ModelBridge
from ax.modelbridge.transforms.base import Transform
from ax.models.types import TConfig


[docs]class MergeRepeatedMeasurements(Transform):
    """Merge repeated measurements for to obtain one observation per arm.

    Repeated measurements are merged via inverse variance weighting (e.g. over
    different trials). This intentionally ignores the trial index and assumes
    stationarity.

    TODO: Support inverse variance weighting correlated outcomes (full covariance).

    Note: this is not reversible.
    """

    def __init__(
        self,
        search_space: Optional[SearchSpace] = None,
        observations: Optional[list[Observation]] = None,
        modelbridge: Optional[ModelBridge] = None,
        config: Optional[TConfig] = None,
    ) -> None:
        if observations is None:
            raise RuntimeError("MergeRepeatedMeasurements requires observations")
        # create a mapping of arm_key -> {metric_name: {means: [], vars: []}}
        arm_to_multi_obs: defaultdict[
            str, defaultdict[str, defaultdict[str, list[float]]]
        ] = defaultdict(lambda: defaultdict(lambda: defaultdict(list)))
        observation_features, observation_data = separate_observations(observations)
        for j, obsd in enumerate(observation_data):
            # This intentionally ignores the trial index
            key = Arm.md5hash(observation_features[j].parameters)
            # TODO: support inverse variance weighting for multivariate distributions
            # (full covariance)
            diag = np.diag(np.diag(obsd.covariance))
            if np.any(np.isnan(obsd.covariance)):
                raise NotImplementedError("All metrics must have noise observations.")
            elif ~np.all(obsd.covariance == diag):
                raise NotImplementedError(
                    "Only independent metrics are currently supported."
                )
            for i, m in enumerate(obsd.metric_names):
                arm_to_multi_obs[key][m]["means"].append(obsd.means[i])
                arm_to_multi_obs[key][m]["vars"].append(obsd.covariance[i, i])

        self.arm_to_merged: defaultdict[str, dict[str, dict[str, float]]] = defaultdict(
            dict
        )
        for k, metric_dict in arm_to_multi_obs.items():
            for m, v in metric_dict.items():
                # inverse variance weighting
                var = np.array(v["vars"])
                means = np.array(v["means"])
                noiseless = var == 0
                if np.any(noiseless):
                    noiseless_means = means[noiseless]
                    if (noiseless_means.shape[0] > 1) and (
                        not np.all(noiseless_means[1:] == noiseless_means[0])
                    ):
                        raise ValueError(
                            "All repeated arms with noiseless measurements "
                            "must have the same means."
                        )
                    self.arm_to_merged[k][m] = {
                        "mean": noiseless_means[0],
                        "var": 0.0,
                    }
                else:
                    inv_var = 1 / np.array(var)
                    inv_sum_inv_var = 1 / np.sum(inv_var)
                    weights = inv_var * inv_sum_inv_var
                    self.arm_to_merged[k][m] = {
                        "mean": np.sum(means * weights),
                        "var": inv_sum_inv_var,
                    }

[docs]    def transform_observations(
        self,
        observations: list[Observation],
    ) -> list[Observation]:
        # Transform observations
        new_observations = []
        observation_features, observation_data = separate_observations(observations)
        arm_to_merged = deepcopy(self.arm_to_merged)
        for j, obsd in enumerate(observation_data):
            key = Arm.md5hash(observation_features[j].parameters)
            # pop to ensure that the resulting observations list has one
            # observation per unique arm
            metric_dict = arm_to_merged.pop(key, None)
            if metric_dict is None:
                continue
            merged_means = np.zeros(len(obsd.metric_names))
            merged_covariance = np.zeros(
                (len(obsd.metric_names), len(obsd.metric_names))
            )
            for i, m in enumerate(obsd.metric_names):
                merged_metric = metric_dict[m]
                merged_means[i] = merged_metric["mean"]
                merged_covariance[i, i] = merged_metric["var"]
            new_obsd = ObservationData(
                metric_names=obsd.metric_names,
                means=merged_means,
                covariance=merged_covariance,
            )
            new_obs = Observation(
                features=observation_features[j],
                data=new_obsd,
                arm_name=observations[j].arm_name,
            )
            new_observations.append(new_obs)
        return new_observations