#!/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 typing import Any
import numpy as np
from ax.core.observation import ObservationFeatures
from ax.modelbridge import ModelBridge
[docs]
def predict_at_point(
model: ModelBridge,
obsf: ObservationFeatures,
metric_names: set[str],
scalarized_metric_config: list[dict[str, Any]] | None = None,
) -> tuple[dict[str, float], dict[str, float]]:
"""Make a prediction at a point.
Returns mean and standard deviation in format expected by plotting.
Args:
model: ModelBridge
obsf: ObservationFeatures for which to predict
metric_names: Limit predictions to these metrics.
scalarized_metric_config: An optional list of dicts specifying how to aggregate
multiple metrics into a single scalarized metric. For each dict, the key is
the name of the new scalarized metric, and the value is a dictionary mapping
each metric to its weight. e.g.
{"name": "metric1:agg", "weight": {"metric1_c1": 0.5, "metric1_c2": 0.5}}.
Returns:
A tuple containing
- Map from metric name to prediction.
- Map from metric name to standard error.
"""
f_pred, cov_pred = model.predict([obsf])
mean_pred_dict = {metric_name: pred[0] for metric_name, pred in f_pred.items()}
cov_pred_dict = {}
for metric_name1, pred_dict in cov_pred.items():
cov_pred_dict[metric_name1] = {}
for metric_name2, pred in pred_dict.items():
cov_pred_dict[metric_name1][metric_name2] = pred[0]
y_hat = {}
se_hat = {}
for metric_name in f_pred:
if metric_name in metric_names:
y_hat[metric_name] = mean_pred_dict[metric_name]
se_hat[metric_name] = np.sqrt(cov_pred_dict[metric_name][metric_name])
if scalarized_metric_config is not None:
for agg_metric in scalarized_metric_config:
agg_metric_name = agg_metric["name"]
if agg_metric_name in metric_names:
agg_metric_weight_dict = agg_metric["weight"]
pred_mean, pred_var = _compute_scalarized_outcome(
mean_dict=mean_pred_dict,
cov_dict=cov_pred_dict,
agg_metric_weight_dict=agg_metric_weight_dict,
)
y_hat[agg_metric_name] = pred_mean
se_hat[agg_metric_name] = np.sqrt(pred_var)
return y_hat, se_hat
[docs]
def predict_by_features(
model: ModelBridge,
label_to_feature_dict: dict[int, ObservationFeatures],
metric_names: set[str],
) -> dict[int, dict[str, tuple[float, float]]]:
"""Predict for given data points and model.
Args:
model: Model to be used for the prediction
metric_names: Names of the metrics, for which to retrieve predictions.
label_to_feature_dict: Mapping from an int label to
a Parameterization. These data points are predicted.
Returns:
A mapping from an int label to a mapping of metric names to tuples
of predicted metric mean and SEM, of form:
{ trial_index -> { metric_name: ( mean, SEM ) } }.
"""
predictions_dict = {} # Store predictions to return
for label in label_to_feature_dict:
try:
y_hat, se_hat = predict_at_point(
model=model,
obsf=label_to_feature_dict[label],
metric_names=metric_names,
)
except NotImplementedError:
raise NotImplementedError(
"The model associated with the current generation strategy "
"step is not one that can be used for predicting values. "
"For example, this may be the Sobol generator associated with the "
"initialization step where quasi-random points are generated. "
"Try again by calling the `AxClient.create_experiment()` "
"method with the `choose_generation_strategy_kwargs="
'{"num_initialization_trials": 0}` parameter if you are looking '
"to use a generation strategy without an initialization step that "
"proceeds straight to the Bayesian optimization step, but note "
"that performance of Bayesian optimization can be suboptimal if "
"search space is not sampled well in the initialization phase."
)
predictions_dict[label] = {
metric: (
y_hat[metric],
se_hat[metric],
)
for metric in metric_names
}
return predictions_dict
def _compute_scalarized_outcome(
mean_dict: dict[str, float],
cov_dict: dict[str, dict[str, float]],
agg_metric_weight_dict: dict[str, float],
) -> tuple[float, float]:
"""Compute the mean and variance of a scalarized outcome.
Args:
mean_dict: Dictionary of means of individual metrics. e.g.
{"metric1": 0.1, "metric2": 0.2}
cov_dict: Dictionary of covariances of each metric pair. e.g.
{"metric1": {"metric1": 0.25, "metric2": 0.0}, "metric2": {"metric1": 0.0,
"metric2": 0.1}}
agg_metric_weight_dict. Dictionary of scalarization weights of the scalarized
outcome. e.g. {"name": "metric1:agg", "weight": {"metric1_c1": 0.5,
"metric1_c2": 0.5}}
"""
pred_mean = 0
pred_var = 0
component_metrics = list(agg_metric_weight_dict.keys())
for i, metric_name in enumerate(component_metrics):
weight = agg_metric_weight_dict[metric_name]
pred_mean += weight * mean_dict[metric_name]
pred_var += (weight**2) * cov_dict[metric_name][metric_name]
# include cross-metric covariance
for metric_name2 in component_metrics[(i + 1) :]:
weight2 = agg_metric_weight_dict[metric_name2]
pred_var += 2 * weight * weight2 * cov_dict[metric_name][metric_name2]
return pred_mean, pred_var
