#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from functools import reduce
from typing import Dict, Optional, Tuple
import pandas as pd
from ax.core.arm import Arm
from ax.core.batch_trial import BatchTrial
from ax.core.data import Data
from ax.core.experiment import Experiment
from ax.core.generator_run import GeneratorRun
from ax.core.objective import MultiObjective, Objective, ScalarizedObjective
from ax.core.optimization_config import (
OptimizationConfig,
MultiObjectiveOptimizationConfig,
)
from ax.core.outcome_constraint import OutcomeConstraint
from ax.core.trial import Trial
from ax.core.types import ComparisonOp
from ax.core.types import TModelPredictArm, TParameterization
from ax.exceptions.core import UnsupportedError
from ax.modelbridge.array import ArrayModelBridge
from ax.modelbridge.cross_validation import (
assess_model_fit,
compute_diagnostics,
cross_validate,
)
from ax.modelbridge.generation_strategy import GenerationStrategy
from ax.modelbridge.modelbridge_utils import (
predicted_pareto_frontier as predicted_pareto,
observed_pareto_frontier as observed_pareto,
)
from ax.modelbridge.multi_objective_torch import MultiObjectiveTorchModelBridge
from ax.modelbridge.registry import get_model_from_generator_run, Models
from ax.utils.common.logger import get_logger
from ax.utils.common.typeutils import not_none, checked_cast
from ax.utils.stats.statstools import relativize_data
from numpy import NaN
logger = get_logger(__name__)
[docs]def get_best_raw_objective_point_with_trial_index(
experiment: Experiment, optimization_config: Optional[OptimizationConfig] = None
) -> Tuple[int, TParameterization, Dict[str, Tuple[float, float]]]:
"""Given an experiment, identifies the arm that had the best raw objective,
based on the data fetched from the experiment.
Args
experiment: Experiment, on which to identify best raw objective arm.
optimization_config: Optimization config to use in absence or in place of
the one stored on the experiment.
Returns:
Tuple of parameterization and a mapping from metric name to a tuple of
the corresponding objective mean and SEM.
"""
# pyre-ignore [16]
if isinstance(experiment.optimization_config.objective, MultiObjective):
logger.warning(
"get_best_raw_objective_point is deprecated for multi-objective "
"optimization. This method will return an arbitrary point on the "
"pareto frontier."
)
opt_config = optimization_config or experiment.optimization_config
assert opt_config is not None, (
"Cannot identify the best point without an optimization config, but no "
"optimization config was provided on the experiment or as an argument."
)
dat = experiment.fetch_data()
if dat.df.empty:
raise ValueError("Cannot identify best point if experiment contains no data.")
objective = opt_config.objective
if isinstance(objective, ScalarizedObjective):
best_row = _get_best_row_for_scalarized_objective(
df=dat.df, objective=objective
)
else:
best_row = _get_best_feasible_row_for_single_objective(
df=dat.df,
optimization_config=opt_config,
status_quo=experiment.status_quo,
)
best_arm = experiment.arms_by_name[best_row["arm_name"]]
best_trial_index = int(best_row["trial_index"])
objective_rows = dat.df.loc[
(dat.df["arm_name"] == best_arm.name)
& (dat.df["trial_index"] == best_trial_index)
]
vals = {
row["metric_name"]: (row["mean"], row["sem"])
for _, row in objective_rows.iterrows()
}
return best_trial_index, not_none(best_arm).parameters, vals
[docs]def get_best_raw_objective_point(
experiment: Experiment, optimization_config: Optional[OptimizationConfig] = None
) -> Tuple[TParameterization, Dict[str, Tuple[float, float]]]:
_, parameterization, vals = get_best_raw_objective_point_with_trial_index(
experiment=experiment, optimization_config=optimization_config
)
return parameterization, vals
def _gr_to_prediction_with_trial_index(
idx: int, gr: GeneratorRun
) -> Optional[Tuple[int, TParameterization, Optional[TModelPredictArm]]]:
if gr.best_arm_predictions is None:
return None
best_arm, best_arm_predictions = gr.best_arm_predictions
if best_arm is None:
return None
return idx, best_arm.parameters, best_arm_predictions
def _raw_values_to_model_predict_arm(
values: Dict[str, Tuple[float, float]]
) -> TModelPredictArm:
return (
{k: v[0] for k, v in values.items()}, # v[0] is mean
{k: {k: v[1] * v[1]} for k, v in values.items()}, # v[1] is sem
)
[docs]def get_best_from_model_predictions_with_trial_index(
experiment: Experiment,
) -> Optional[Tuple[int, TParameterization, Optional[TModelPredictArm]]]:
"""Given an experiment, returns the best predicted parameterization and corresponding
prediction based on the most recent Trial with predictions. If no trials have
predictions returns None.
Only some models return predictions. For instance GPEI does while Sobol does not.
TModelPredictArm is of the form:
({metric_name: mean}, {metric_name_1: {metric_name_2: cov_1_2}})
Args:
experiment: Experiment, on which to identify best raw objective arm.
Returns:
Tuple of parameterization and model predictions for it.
"""
# pyre-ignore [16]
if isinstance(experiment.optimization_config.objective, MultiObjective):
logger.warning(
"get_best_from_model_predictions is deprecated for multi-objective "
"optimization configs. This method will return an arbitrary point on "
"the pareto frontier."
)
for idx, trial in sorted(
experiment.trials.items(), key=lambda x: x[0], reverse=True
):
gr = None
if isinstance(trial, Trial):
gr = trial.generator_run
elif isinstance(trial, BatchTrial):
if len(trial.generator_run_structs) > 0:
# In theory batch_trial can have >1 gr, grab the first
gr = trial.generator_run_structs[0].generator_run
if gr is not None and gr.best_arm_predictions is not None: # pragma: no cover
data = experiment.lookup_data()
try:
model = get_model_from_generator_run(
generator_run=gr, experiment=experiment, data=data
)
except ValueError:
return _gr_to_prediction_with_trial_index(idx, gr)
# If model is not ArrayModelBridge, just use the best arm frmo the
# last good generator run
if not isinstance(model, ArrayModelBridge):
return _gr_to_prediction_with_trial_index(idx, gr)
# Check to see if the model is worth using
cv_results = cross_validate(model=model)
diagnostics = compute_diagnostics(result=cv_results)
assess_model_fit_results = assess_model_fit(diagnostics=diagnostics)
objective_name = experiment.optimization_config.objective.metric.name
# If model fit is bad use raw results
if (
objective_name
in assess_model_fit_results.bad_fit_metrics_to_fisher_score
):
logger.warning(
"Model fit is poor; falling back on raw data for best point."
)
if not _is_all_noiseless(df=data.df, metric_name=objective_name):
logger.warning(
"Model fit is poor and data on objective metric "
+ f"{objective_name} is noisy; interpret best points "
+ "results carefully."
)
return _get_best_poor_model_fit(experiment=experiment)
res = model.model_best_point()
if res is None:
return _gr_to_prediction_with_trial_index(idx, gr)
best_arm, best_arm_predictions = res
return idx, not_none(best_arm).parameters, best_arm_predictions
return None
[docs]def get_best_from_model_predictions(
experiment: Experiment,
) -> Optional[Tuple[TParameterization, Optional[TModelPredictArm]]]:
res = get_best_from_model_predictions_with_trial_index(experiment=experiment)
if res is None:
return None
_, parameterization, vals = res
return parameterization, vals
[docs]def get_best_parameters_with_trial_index(
experiment: Experiment,
use_model_predictions: bool = True,
) -> Optional[Tuple[int, TParameterization, Optional[TModelPredictArm]]]:
"""Given an experiment, identifies the best arm.
First attempts according to do so with models used in optimization and
its corresponding predictions if available. Falls back to the best raw
objective based on the data fetched from the experiment.
TModelPredictArm is of the form:
({metric_name: mean}, {metric_name_1: {metric_name_2: cov_1_2}})
Args:
experiment: Experiment, on which to identify best raw objective arm.
use_model_predictions: Whether to extract the best point using
model predictions or directly observed values. If ``True``,
the metric means and covariances in this method's output will
also be based on model predictions and may differ from the
observed values.
Returns:
Tuple of parameterization and model predictions for it.
"""
# pyre-ignore [16]
if isinstance(experiment.optimization_config.objective, MultiObjective):
logger.warning(
"get_best_parameters is deprecated for multi-objective optimization. "
"This method will return an arbitrary point on the pareto frontier."
)
# Find latest trial which has a generator_run attached and get its predictions
if use_model_predictions:
model_predictions = get_best_from_model_predictions_with_trial_index(
experiment=experiment
)
if model_predictions is not None: # pragma: no cover
return model_predictions
logger.info(
"Could not use model predictions to identify best point, will use raw "
"objective values."
)
# Could not find through model, default to using raw objective.
try:
(
trial_index,
parameterization,
values,
) = get_best_raw_objective_point_with_trial_index(experiment=experiment)
except ValueError:
return None
return (
trial_index,
parameterization,
_raw_values_to_model_predict_arm(values),
)
[docs]def get_best_parameters(
experiment: Experiment,
use_model_predictions: bool = True,
) -> Optional[Tuple[TParameterization, Optional[TModelPredictArm]]]:
res = get_best_parameters_with_trial_index(
experiment=experiment, use_model_predictions=use_model_predictions
)
if res is None:
return None
_, parameterization, vals = res
return parameterization, vals
[docs]def get_pareto_optimal_parameters(
experiment: Experiment,
generation_strategy: GenerationStrategy,
use_model_predictions: bool = True,
) -> Optional[Dict[int, Tuple[TParameterization, TModelPredictArm]]]:
"""Identifies the best parameterizations tried in the experiment so far,
using model predictions if ``use_model_predictions`` is true and using
observed values from the experiment otherwise. By default, uses model
predictions to account for observation noise.
NOTE: The format of this method's output is as follows:
{ trial_index --> (parameterization, (means, covariances) }, where means
are a dictionary of form { metric_name --> metric_mean } and covariances
are a nested dictionary of form
{ one_metric_name --> { another_metric_name: covariance } }.
Args:
experiment: Experiment, from which to find Pareto-optimal arms.
generation_strategy: Generation strategy containing the modelbridge.
use_model_predictions: Whether to extract the Pareto frontier using
model predictions or directly observed values. If ``True``,
the metric means and covariances in this method's output will
also be based on model predictions and may differ from the
observed values.
Returns:
``None`` if it was not possible to extract the Pareto frontier,
otherwise a mapping from trial index to the tuple of:
- the parameterization of the arm in that trial,
- two-item tuple of metric means dictionary and covariance matrix
(model-predicted if ``use_model_predictions=True`` and observed
otherwise).
"""
# Validate aspects of the experiment: that it is a MOO experiment and
# that the current model can be used to produce the Pareto frontier.
if not not_none(experiment.optimization_config).is_moo_problem:
raise UnsupportedError(
"Please use `get_best_parameters` for single-objective problems."
)
moo_optimization_config = checked_cast(
MultiObjectiveOptimizationConfig, experiment.optimization_config
)
if moo_optimization_config.outcome_constraints:
# TODO[drfreund]: Test this flow and remove error.
raise NotImplementedError(
"Support for outcome constraints is currently under development."
)
# Extract or instantiate modelbridge to use for Pareto frontier extraction.
mb = generation_strategy.model
if mb is None or not isinstance(mb, MultiObjectiveTorchModelBridge):
logger.info(
"Can only extract a Pareto frontier using a multi-objective model bridge"
f", but currently used model bridge is: {mb} of type {type(mb)}. Will "
"use `Models.MOO` instead to extract Pareto frontier."
)
mb = checked_cast(
MultiObjectiveTorchModelBridge,
Models.MOO(
experiment=experiment, data=checked_cast(Data, experiment.lookup_data())
),
)
else:
# Make sure the model is up-to-date with the most recent data.
generation_strategy._fit_or_update_current_model(data=None)
# If objective thresholds are not specified in optimization config, extract
# the inferred ones if possible or infer them anew if not.
objective_thresholds_override = None
if not moo_optimization_config.objective_thresholds:
lgr = generation_strategy.last_generator_run
if lgr and lgr.gen_metadata and "objective_thresholds" in lgr.gen_metadata:
objective_thresholds_override = lgr.gen_metadata["objective_thresholds"]
objective_thresholds_override = mb.infer_objective_thresholds(
search_space=experiment.search_space,
optimization_config=experiment.optimization_config,
fixed_features=None,
)
logger.info(
f"Using inferred objective thresholds: {objective_thresholds_override}, "
"as objective thresholds were not specified as part of the optimization "
"configuration on the experiment."
)
# Extract the Pareto frontier and format it as follows:
# { trial_index --> (parameterization, (means, covariances) }
pareto_util = predicted_pareto if use_model_predictions else observed_pareto
pareto_optimal_observations = pareto_util(
modelbridge=mb, objective_thresholds=objective_thresholds_override
)
return {
int(not_none(obs.features.trial_index)): (
obs.features.parameters,
(obs.data.means_dict, obs.data.covariance_matrix),
)
for obs in pareto_optimal_observations
}
def _get_best_row_for_scalarized_objective(
df: pd.DataFrame,
objective: ScalarizedObjective,
) -> pd.DataFrame:
df = df.copy()
# First, add a weight column, setting 0.0 if the metric is not part
# of the objective
metric_to_weight = {
m.name: objective.weights[i] for i, m in enumerate(objective.metrics)
}
df["weight"] = df["metric_name"].apply(lambda x: metric_to_weight.get(x) or 0.0)
# Now, calculate the weighted linear combination via groupby,
# filtering out NaN for missing data
df["weighted_mean"] = df["mean"] * df["weight"]
groupby_df = (
df[["arm_name", "trial_index", "weighted_mean"]]
.groupby(["arm_name", "trial_index"], as_index=False)
.sum(min_count=1)
.dropna()
)
if groupby_df.empty:
raise ValueError("No data has been logged for scalarized objective.")
return (
groupby_df.loc[groupby_df["weighted_mean"].idxmin()]
if objective.minimize
else groupby_df.loc[groupby_df["weighted_mean"].idxmax()]
)
def _get_best_row_for_single_objective(
df: pd.DataFrame, objective: Objective
) -> pd.Series:
objective_name = objective.metric.name
objective_rows = df.loc[df["metric_name"] == objective_name]
if objective_rows.empty:
raise ValueError(f'No data has been logged for objective "{objective_name}".')
return (
objective_rows.loc[objective_rows["mean"].idxmin()]
if objective.minimize
else objective_rows.loc[objective_rows["mean"].idxmax()]
)
def _filter_feasible_rows(
df: pd.DataFrame,
optimization_config: OptimizationConfig,
status_quo: Optional[Arm],
) -> pd.DataFrame:
"""Filter out arms that do not satisfy outcome constraints
Looks at all arm data collected and removes rows corresponding to arms in
which one or more of their associated metrics' 95% confidence interval
falls outside of any outcome constraint's bounds (i.e. we are 95% sure the
bound is not satisfied).
"""
if len(optimization_config.outcome_constraints) < 1:
return df
name = df["metric_name"]
# When SEM is NaN we should treat it as if it were 0
sems = not_none(df["sem"].fillna(0))
# Bounds computed for 95% confidence interval on Normal distribution
lower_bound = df["mean"] - sems * 1.96
upper_bound = df["mean"] + sems * 1.96
# Only compute relativization if some constraints are relative
rel_df = None
rel_lower_bound = None
rel_upper_bound = None
if status_quo is not None and any(
oc.relative for oc in optimization_config.outcome_constraints
):
# relativize_data expects all arms to come from the same trial, we need to
# format the data as if it was.
to_relativize = df.copy()
to_relativize["trial_index"] = 0
rel_df = relativize_data(
data=Data(to_relativize), status_quo_name=status_quo.name
).df.append(
{
"arm_name": "status_quo",
"metric_name": status_quo.name,
"mean": 0,
"sem": 0,
},
ignore_index=True,
)
rel_sems = not_none(rel_df["sem"].fillna(0))
rel_lower_bound = rel_df["mean"] - rel_sems * 1.96
rel_upper_bound = rel_df["mean"] + rel_sems * 1.96
# Nested function from OC -> Mask for consumption in later map/reduce from
# [OC] -> Mask. Constraint relativity is handled inside so long as relative bounds
# are set in surrounding closure (which will occur in proper experiment setup).
def oc_mask(oc: OutcomeConstraint) -> pd.Series:
name_match_mask = name == oc.metric.name
if oc.relative:
if rel_lower_bound is None or rel_upper_bound is None:
logger.warning(
f"No status quo provided; relative constraint {oc} ignored."
)
return pd.Series(True, index=df.index)
observed_lower_bound = rel_lower_bound
observed_upper_bound = rel_upper_bound
else:
observed_lower_bound = lower_bound
observed_upper_bound = upper_bound
# Return True if metrics are different, or whether the confidence
# interval is entirely not within the bound
if oc.op == ComparisonOp.GEQ:
return ~name_match_mask | observed_upper_bound > oc.bound
else:
return ~name_match_mask | observed_lower_bound < oc.bound
mask = reduce(
lambda left, right: left & right,
map(oc_mask, optimization_config.outcome_constraints),
)
bad_arm_names = (
df[~mask]["arm_name"].tolist()
if rel_df is None
else rel_df[~mask]["arm_name"].tolist()
)
feasible = df.loc[df["arm_name"].apply(lambda x: x not in bad_arm_names)]
if feasible.empty:
raise ValueError(
"No points satisfied all outcome constraints within 95 percent"
+ "confidence interval"
)
return feasible
def _get_best_feasible_row_for_single_objective(
df: pd.DataFrame,
optimization_config: OptimizationConfig,
status_quo: Optional[Arm],
) -> pd.Series:
return _get_best_row_for_single_objective(
df=_filter_feasible_rows(
df=df, optimization_config=optimization_config, status_quo=status_quo
),
objective=optimization_config.objective,
)
def _is_all_noiseless(df: pd.DataFrame, metric_name: str) -> bool:
"""Noiseless is defined as SEM = 0 or SEM = NaN on a given metric (usually
the objective).
"""
name_mask = df["metric_name"] == metric_name
df_metric_arms_sems = df[name_mask]["sem"]
return ((df_metric_arms_sems == 0) | df_metric_arms_sems == NaN).all()
def _get_best_poor_model_fit(
experiment: Experiment,
) -> Optional[Tuple[int, TParameterization, Optional[TModelPredictArm]]]:
try:
(
trial_index,
parameterization,
values,
) = get_best_raw_objective_point_with_trial_index(experiment=experiment)
except ValueError as err:
logger.error(
f"Encountered error while trying to identify the best point: {err}"
)
return None
return (
trial_index,
parameterization,
_raw_values_to_model_predict_arm(values),
)
