mlos_bench.environments.composite_env module

Composite benchmark environment.

class mlos_bench.environments.composite_env.CompositeEnv(*, name: str, config: dict, global_config: dict | None = None, tunables: TunableGroups | None = None, service: Service | None = None)

Bases: Environment

Composite benchmark environment.

Attributes:

children

Return the list of child environments.

parameters

Key/value pairs of all environment parameters (i.e., const_args and tunable_params).

tunable_params

Get the configuration space of the given environment.

Methods

new(*, env_name, class_name, config[, ...])	Factory method for a new environment with a given config.
pprint([indent, level])	Pretty-print the environment and its children.
run()	Submit a new experiment to the environment.
setup(tunables[, global_config])	Set up the children environments.
status()	Check the status of the benchmark environment.
teardown()	Tear down the children environments.

property children: List[Environment]

Return the list of child environments.

pprint(indent: int = 4, level: int = 0) → str

Pretty-print the environment and its children.

Parameters:

indentint

Number of spaces to indent the output at each level. Default is 4.

levelint

Current level of indentation. Default is 0.

Returns:

prettystr

Pretty-printed environment configuration.

run() → Tuple[Status, datetime, Dict[str, int | float | str | None] | None]

Submit a new experiment to the environment. Return the result of the last child environment if successful, or the status of the last failed environment otherwise.

Returns:

(status, timestamp, output)(Status, datetime, dict)

3-tuple of (Status, timestamp, output) values, where output is a dict with the results or None if the status is not COMPLETED. If run script is a benchmark, then the score is usually expected to be in the score field.

setup(tunables: TunableGroups, global_config: dict | None = None) → bool

Set up the children environments.

Parameters:

tunablesTunableGroups

A collection of tunable parameters along with their values.

global_configdict

Free-format dictionary of global parameters of the environment that are not used in the optimization process.

Returns:

is_successbool

True if all children setup() operations are successful, false otherwise.

status() → Tuple[Status, datetime, List[Tuple[datetime, str, Any]]]

Check the status of the benchmark environment.

Returns:

(benchmark_status, timestamp, telemetry)(Status, datetime, list)

3-tuple of (benchmark status, timestamp, telemetry) values. timestamp is UTC time stamp of the status; it’s current time by default. telemetry is a list (maybe empty) of (timestamp, metric, value) triplets.

teardown() → None

Tear down the children environments.

This method is idempotent, i.e., calling it several times is equivalent to a single call. The environments are being torn down in the reverse order.