sparta.common.tuning

class sparta.common.tuning.TunableItemCfg(_type: str, _value: Union[Dict, List], _is_nested: Optional[bool] = False)

TunableItemCfg is used to describe the search space .. rubric:: Examples

python

cfg = TunableItemCfg('choice', _is_nested=True, _value={
    'openai': {},
    'sparta': {
        'BM': TunableItemCfg('choice', [32,64]),
        'BN': TunableItemCfg('choice', [8,16]),
    }
})
nni_space = {
    'test': {'_type':'choice', '_value': [
        {'_name': 'openai'},
        {
            '_name': 'sparta',
            'BM': {'_type': 'choice', '_value': [32,64]},
            'BN': {'_type': 'choice', '_value': [8,16]},
        }]}
}

# converted to a `NNI` search space (See more in https://nni.readthedocs.io/en/stable/hpo/search_space.html)
assert search_space_cfg.to_nni_search_space() == nni_space

Parameters

• _type (str) – paramter type, allowed one of (‘choice’).

• _value (Dict | List) – options for paramter.

• _is_nested (bool) – whether this space is nested (default: False). If True, the _value should be Dict[str, Dict[TunableItemCfg]]

to_nni_search_space()

convert to nni search space

class sparta.common.tuning.Tunable

The wrapper of NNI tuners that supports nested choice search space.

static create_tuner(algo: str, search_space_cfg: Dict[str, TunableItemCfg], tuner_kw: Optional[Dict] = None)

create NNI Tuner

Parameters

• algo (str) –

  tuning algorithm, allowed algo values and their corresponding tuners are:

  algo	tuner
  grid	nni.algorithms.hpo.gridsearch_tuner.GridSearchTuner
  rand	nni.algorithms.hpo.random_tuner.RandomTuner
  tpe	nni.algorithms.hpo.tpe_tuner.TpeTuner
  evolution	nni.algorithms.hpo.evolution_tuner.EvolutionTuner

• search_space_cfg (TunableItemCfg) – search space config

• tuner_kw (Dict) – parameters passed to NNI tuner