#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
from typing import Any, Dict, List, Optional, Tuple, Union
from ax.core.arm import Arm
from ax.core.data import Data
from ax.core.experiment import Experiment
from ax.core.trial import Trial
from ax.core.types import (
TEvaluationOutcome,
TModelPredictArm,
TParameterization,
TParamValue,
)
from ax.modelbridge.generation_strategy import GenerationStrategy
from ax.service.utils.best_point import (
get_best_from_model_predictions,
get_best_raw_objective_point,
)
from ax.service.utils.dispatch import choose_generation_strategy
from ax.service.utils.instantiation import make_experiment
from ax.utils.common.typeutils import not_none
try: # We don't require SQLAlchemy by default.
from ax.storage.sqa_store.structs import DBSettings
except ModuleNotFoundError: # pragma: no cover
DBSettings = None
[docs]class AxClient:
"""
Convenience handler for management of experimentation cycle through a
service-like API. External system manages scheduling of the cycle and makes
calls to this client to get next suggestion in the experiment and log back
data from the evaluation of that suggestion.
Note: `AxClient` expects to only propose 1 arm (suggestion) per trial; support
for use cases that require use of batches is coming soon.
Two custom types used in this class for convenience are `TParamValue` and
`TParameterization`. Those are shortcuts for `Union[str, bool, float, int]`
and `Dict[str, Union[str, bool, float, int]]`, respectively.
Args:
generation_strategy: Optional generation strategy. If not set, one is
intelligently chosen based on properties of search space.
db_settings: Settings for saving and reloading the underlying experiment
to a database.
enforce_sequential_optimization: Whether to enforce that when it is
reasonable to switch models during the optimization (as prescribed
by `num_arms` in generation strategy), Ax will wait for enough trials
to be completed with data to proceed. Defaults to True. If set to
False, Ax will keep generating new trials from the previous model
until enough data is gathered. Use this only if necessary;
otherwise, it is more resource-efficient to
optimize sequentially, by waiting until enough data is available to
use the next model.
"""
def __init__(
self,
generation_strategy: Optional[GenerationStrategy] = None,
db_settings: Any = None,
enforce_sequential_optimization: bool = True,
) -> None:
self.generation_strategy = generation_strategy
if db_settings and (not DBSettings or not isinstance(db_settings, DBSettings)):
raise ValueError(
"`db_settings` argument should be of type ax.storage.sqa_store."
"structs.DBSettings. To use `DBSettings`, you will need SQLAlchemy "
"installed in your environment (can be installed through pip)."
)
self.db_settings = db_settings
self._experiment: Optional[Experiment] = None
self._enforce_sequential_optimization = enforce_sequential_optimization
# Trials, for which we received data since last `GenerationStrategy.gen`,
# used to make sure that generation strategy is updated with new data.
self._updated_trials: List[int] = []
# ------------------------ Public API methods. ------------------------
[docs] def create_experiment(
self,
parameters: List[Dict[str, Union[TParamValue, List[TParamValue]]]],
name: Optional[str] = None,
objective_name: Optional[str] = None,
minimize: bool = False,
parameter_constraints: Optional[List[str]] = None,
outcome_constraints: Optional[List[str]] = None,
status_quo: Optional[TParameterization] = None,
) -> None:
"""Create a new experiment and save it if DBSettings available.
Args:
parameters: List of dictionaries representing parameters in the
experiment search space. Required elements in the dictionaries
are: "name" (name of this parameter, string), "type" (type of the
parameter: "range", "fixed", or "choice", string), and "bounds"
for range parameters (list of two values, lower bound first),
"values" for choice parameters (list of values), and "value" for
fixed parameters (single value).
objective: Name of the metric used as objective in this experiment.
This metric must be present in `raw_data` argument to `complete_trial`.
name: Name of the experiment to be created.
minimize: Whether this experiment represents a minimization problem.
parameter_constraints: List of string representation of parameter
constraints, such as "x3 >= x4" or "x3 + x4 + x5 >= 2". For sum
constraints, any number of arguments is accepted, and acceptable
operators are "<=" and ">=".
outcome_constraints: List of string representation of outcome
constraints of form "metric_name >= bound", like "m1 <= 3."
status_quo: Parameterization of the current state of the system.
If set, this will be added to each trial to be evaluated alongside
test configurations.
"""
if self.db_settings and not name:
raise ValueError( # pragma: no cover
"Must give the experiment a name if `db_settings` is not None."
)
self._experiment = make_experiment(
name=name,
parameters=parameters,
objective_name=objective_name,
minimize=minimize,
parameter_constraints=parameter_constraints,
outcome_constraints=outcome_constraints,
status_quo=status_quo,
)
if self.generation_strategy is None:
self.generation_strategy = choose_generation_strategy(
search_space=self._experiment.search_space,
enforce_sequential_optimization=self._enforce_sequential_optimization,
)
self._save_experiment_if_possible()
[docs] def get_next_trial(self) -> Tuple[TParameterization, int]:
"""
Generate trial with the next set of parameters to try in the iteration process.
Note: Service API currently supports only 1-arm trials.
Returns:
Tuple of trial parameterization, trial index
"""
# NOTE: Could move this into complete_trial to save latency on this call.
trial = self._suggest_new_trial()
trial.mark_dispatched()
self._updated_trials = []
self._save_experiment_if_possible()
return not_none(trial.arm).parameters, trial.index
[docs] def complete_trial(
self,
trial_index: int,
# acceptable `raw_data` argument formats:
# 1) {metric_name -> (mean, standard error)}
# 2) (mean, standard error) and we assume metric name == objective name
# 3) only the mean, and we assume metric name == objective name and
# standard error == 0
raw_data: TEvaluationOutcome,
metadata: Optional[Dict[str, str]] = None,
) -> None:
"""
Completes the trial with given metric values and adds optional metadata
to it.
Args:
trial_index: Index of trial within the experiment.
raw_data: Evaluation data for the trial. Can be a mapping from
metric name to a tuple of mean and SEM, just a tuple of mean and
SEM if only one metric in optimization, or just the mean if there
is no SEM.
metadata: Additional metadata to track about this run.
"""
assert isinstance(
trial_index, int
), f"Trial index must be an int, got: {trial_index}." # pragma: no cover
trial = self.experiment.trials[trial_index]
if not isinstance(trial, Trial):
raise NotImplementedError(
"Batch trial functionality is not yet available through Service API."
)
if metadata is not None:
trial._run_metadata = metadata
if isinstance(raw_data, dict):
evaluations = {not_none(trial.arm).name: raw_data}
elif isinstance(raw_data, tuple):
evaluations = {
not_none(trial.arm).name: {
self.experiment.optimization_config.objective.metric.name: raw_data
}
}
elif isinstance(raw_data, float) or isinstance(raw_data, int):
evaluations = {
not_none(trial.arm).name: {
self.experiment.optimization_config.objective.metric.name: (
raw_data,
0.0,
)
}
}
else:
raise ValueError(
"Raw data has an invalid type. The data must either be in the form "
"of a dictionary of metric names to mean, sem tuples, "
"or a single mean, sem tuple, or a single mean."
)
data = Data.from_evaluations(evaluations, trial.index)
trial.mark_completed()
self.experiment.attach_data(data)
self._updated_trials.append(trial_index)
self._save_experiment_if_possible()
[docs] def log_trial_failure(
self, trial_index: int, metadata: Optional[Dict[str, str]] = None
) -> None:
"""Mark that the given trial has failed while running.
Args:
trial_index: Index of trial within the experiment.
metadata: Additional metadata to track about this run.
"""
trial = self.experiment.trials[trial_index]
trial.mark_failed()
if metadata is not None:
trial._run_metadata = metadata
self._save_experiment_if_possible()
[docs] def attach_trial(
self, parameters: TParameterization
) -> Tuple[TParameterization, int]:
"""Attach a new trial with the given parameterization to the experiment.
Args:
parameters: Parameterization of the new trial.
Returns:
Tuple of parameterization and trial index from newly created trial.
"""
trial = self.experiment.new_trial().add_arm(Arm(parameters=parameters))
trial.mark_dispatched()
self._save_experiment_if_possible()
return not_none(trial.arm).parameters, trial.index
# TODO[T42389552]: this is currently only compatible with some models.
[docs] def get_best_parameters(
self
) -> Optional[Tuple[TParameterization, Optional[TModelPredictArm]]]:
"""
Return the best set of parameters the experiment has knowledge of.
If experiment is in the optimization phase, return the best point
determined by the model used in the latest optimization round, otherwise
return none.
Custom type `TModelPredictArm` is defined as
`Tuple[Dict[str, float], Optional[Dict[str, Dict[str, float]]]]`, and
stands for tuple of two mappings: metric name to its mean value and metric
name to a mapping of other mapping name to covariance of the two metrics.
Returns:
Tuple of (best parameters, model predictions for best parameters).
None if no data.
"""
# Find latest trial which has a generator_run attached and get its predictions
model_predictions = get_best_from_model_predictions(experiment=self.experiment)
if model_predictions is not None: # pragma: no cover
return model_predictions
# Could not find through model, default to using raw objective.
parameterization, values = get_best_raw_objective_point(
experiment=self.experiment
)
return (
parameterization,
(
{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_recommended_max_parallelism(self) -> List[Tuple[int, int]]:
"""Recommends maximum number of trials that can be scheduled in parallel
at different stages of optimization.
Some optimization algorithms profit significantly from sequential
optimization (e.g. suggest a few points, get updated with data for them,
repeat). This setting indicates how many trials should be in flight
(generated, but not yet completed with data).
The output of this method is mapping of form
{num_trials -> max_parallelism_setting}, where the max_parallelism_setting
is used for num_trials trials. If max_parallelism_setting is -1, as
many of the trials can be ran in parallel, as necessary. If num_trials
in a tuple is -1, then the corresponding max_parallelism_setting
should be used for all subsequent trials.
For example, if the returned list is [(5, -1), (12, 6), (-1, 3)],
the schedule could be: run 5 trials in parallel, run 6 trials in
parallel twice, run 3 trials in parallel for as long as needed. Here,
'running' a trial means obtaining a next trial from `AxClient` through
get_next_trials and completing it with data when available.
Returns:
Mapping of form {num_trials -> max_parallelism_setting}.
"""
if not (self._experiment and self.generation_strategy):
# Auto-selected generation strategy is set on experiment creation.
raise ValueError(
"`get_recommended_max_parallelism` requires an experiment to be "
"set on AxClient first."
)
parallelism_settings = []
for step in self.generation_strategy._steps:
parallelism_settings.append(
(step.num_arms, step.recommended_max_parallelism or step.num_arms)
)
return parallelism_settings
[docs] def load_experiment(self, experiment_name: str) -> None:
"""[Work in progress] Load an existing experiment.
Args:
experiment_name: Name of the experiment.
Returns:
Experiment object.
"""
raise NotImplementedError( # pragma: no cover
"Saving and loading experiment in `AxClient` functionality currently "
"under development."
)
[docs] def get_report(self) -> str:
"""Returns HTML of a generated report containing vizualizations."""
raise NotImplementedError( # pragma: no cover
"Report generation not supported for `AxClient` yet."
)
[docs] def should_stop_early(self, trial_index: int, data: TEvaluationOutcome) -> bool:
"""Whether to stop the given parameterization given early data."""
raise NotImplementedError( # pragma: no cover
"Early stopping of trials not supported for `AxClient` yet."
)
# ---------------------- Private helper methods. ---------------------
@property
def experiment(self) -> Experiment:
"""Returns the experiment set on this Ax client"""
if self._experiment is None:
raise ValueError(
"Experiment not set on Ax client. Must first "
"call load_experiment or create_experiment to use handler functions."
)
return self._experiment
def _save_experiment_if_possible(self) -> bool:
"""[Work in progress] Saves attached experiment if DB settings are set on this AxClient
instance.
Returns:
bool: Whether the experiment was saved.
"""
if self.db_settings and self._experiment:
raise NotImplementedError( # pragma: no cover
"Saving and loading experiment in `AxClient` functionality "
"currently under development."
)
return False
def _get_new_data(self) -> Data:
"""
Returns new data since the last run of the generator.
Returns:
Latest data.
"""
return Data.from_multiple_data(
[self.experiment.lookup_data_for_trial(idx) for idx in self._updated_trials]
)
def _suggest_new_trial(self) -> Trial:
"""
Suggest new candidate for this experiment.
Args:
n: Number of candidates to generate.
Returns:
Trial with candidate.
"""
new_data = self._get_new_data()
generator_run = not_none(self.generation_strategy).gen(
experiment=self.experiment, new_data=new_data
)
return self.experiment.new_trial(generator_run=generator_run)
