mlos_core.spaces.adapters.llamatune

Implementation of LlamaTune space adapter.

LlamaTune is a technique that transforms the original parameter space into a lower-dimensional space to try and improve the sample efficiency of the underlying optimizer by making use of the inherent parameter sensitivity correlations in most systems.

See Also: LlamaTune: Sample-Efficient DBMS Configuration Tuning.

Classes

LlamaTuneAdapter

Implementation of LlamaTune, a set of parameter space transformation techniques,

Module Contents

class mlos_core.spaces.adapters.llamatune.LlamaTuneAdapter(*, orig_parameter_space: ConfigSpace.ConfigurationSpace, num_low_dims: int = DEFAULT_NUM_LOW_DIMS, special_param_values: dict | None = None, max_unique_values_per_param: int | None = DEFAULT_MAX_UNIQUE_VALUES_PER_PARAM, use_approximate_reverse_mapping: bool = False)[source]

Bases: mlos_core.spaces.adapters.adapter.BaseSpaceAdapter

Implementation of LlamaTune, a set of parameter space transformation techniques, aimed at improving the sample-efficiency of the underlying optimizer.

Create a space adapter that employs LlamaTune’s techniques.

Parameters:

  • orig_parameter_space (ConfigSpace.ConfigurationSpace) – The original (user-provided) parameter space to optimize.

  • num_low_dims (int) – Number of dimensions used in the low-dimensional parameter search space.

  • special_param_values_dict (Optional[dict]) – Dictionary of special

  • max_unique_values_per_param (Optional[int]) – Number of unique values per parameter. Used to discretize the parameter space. If None space discretization is disabled.

  • special_param_values (Optional[dict])

  • use_approximate_reverse_mapping (bool)

inverse_transform(configurations: pandas.DataFrame) → pandas.DataFrame[source]

Translates a configuration, which belongs to the original parameter space, to the target parameter space. This method is called by the register method of the BaseOptimizer class, and performs the inverse operation of transform() method.

Parameters:

configurations (pandas.DataFrame) – Dataframe of configurations / parameters, which belong to the original parameter space. The columns are the parameter names the original parameter space and the rows are the configurations.

Returns:

configurations – Dataframe of the translated configurations / parameters. The columns are the parameter names of the target parameter space and the rows are the configurations.

Return type:

pandas.DataFrame

transform(configuration: pandas.DataFrame) → pandas.DataFrame[source]

Translates a configuration, which belongs to the target parameter space, to the original parameter space. This method is called by the suggest() method of the BaseOptimizer class.

Parameters:

configuration (pandas.DataFrame) – Pandas dataframe with a single row. Column names are the parameter names of the target parameter space.

Returns:

configuration – Pandas dataframe with a single row, containing the translated configuration. Column names are the parameter names of the original parameter space.

Return type:

pandas.DataFrame

DEFAULT_MAX_UNIQUE_VALUES_PER_PARAM = 10000[source]

Default number of (max) unique values of each parameter, when space discretization is used.

DEFAULT_NUM_LOW_DIMS = 16[source]

Default number of dimensions in the low-dimensional search space, generated by HeSBO projection.

DEFAULT_SPECIAL_PARAM_VALUE_BIASING_PERCENTAGE = 0.2[source]

Default percentage of bias for each special parameter value.

property target_parameter_space: ConfigSpace.ConfigurationSpace[source]

Get the parameter space, which is explored by the underlying optimizer.

Return type:

ConfigSpace.ConfigurationSpace