axtreme.utils.modelbridge_utils

This builds of the utils provided in ax.modelbridge.modelbridge_utils.

Functions

observations_to_arrays(param_names, ...)

Helper method to convert Ax results (Observation) to Numpy arrays of a standard format.
axtreme.utils.modelbridge_utils.observations_to_arrays(param_names: list[str], outcomes: list[str], observations: list[Observation]) → tuple[ndarray[Any, dtype[float64]], ndarray[Any, dtype[float64]], ndarray[Any, dtype[float64]]]

Helper method to convert Ax results (Observation) to Numpy arrays of a standard format.

This is useful e.g. when converting the output of ax.model_bridge.base.ModelBridge.get_training_data into “x” and “y” and “covariance” arrays.

Parameters:

  • param_names – List of the parameter names to report, and in what order.

  • outcomes – list of the metric/target names to report, in this order. TODO(sw): work out if the order ever changes. If not, we should remove this.

  • observations – The list of observations to convert.

Returns:

An (n_observation, n_features) array of features. - f: An (n_observation, m_metrics) array of prediction means. - cov: An (n_observation, m_metrics, m_metrics) array of covariance between metric per x point.

Return type:

  • features

Example

>>> import numpy as np
>>> from ax.core.observation import Observation, ObservationData, ObservationFeatures
>>> X = np.array([[1, 2], [3, 4], [5, 6]])
>>> Y = np.array([[10], [20], [30]])
>>> of = [ObservationFeatures(parameters={"x1": x[0], "x2": x[1]}) for x in X]
>>> od = [ObservationData(metric_names=["y"], means=y, covariance=np.eye(1)) for y in Y]
>>> observations = [Observation(features=f, data=d) for f, d in zip(of, od)]
>>> features, f, cov = observations_to_arrays(param_names=["x1", "x2"], outcomes=["y"], observations=observations)
>>> assert np.array_equal(features, X)
>>> assert np.array_equal(f, Y)
>>> # There is only one output per x point. This will have perfect covariance with itself.
>>> assert np.array_equal(cov, np.ones([3, 1, 1]))