Source code for nni.tuner

# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.

"""
Tuner is an AutoML algorithm, which generates a new configuration for the next try.
A new trial will run with this configuration.

See :class:`Tuner`' specification and ``docs/en_US/tuners.rst`` for details.
"""

import logging

import nni

from .recoverable import Recoverable

__all__ = ['Tuner']

_logger = logging.getLogger(__name__)


[docs]class Tuner(Recoverable): """ Tuner is an AutoML algorithm, which generates a new configuration for the next try. A new trial will run with this configuration. This is the abstract base class for all tuners. Tuning algorithms should inherit this class and override :meth:`update_search_space`, :meth:`receive_trial_result`, as well as :meth:`generate_parameters` or :meth:`generate_multiple_parameters`. After initializing, NNI will first call :meth:`update_search_space` to tell tuner the feasible region, and then call :meth:`generate_parameters` one or more times to request for hyper-parameter configurations. The framework will train several models with given configuration. When one of them is finished, the final accuracy will be reported to :meth:`receive_trial_result`. And then another configuration will be reqeusted and trained, util the whole experiment finish. If a tuner want's to know when a trial ends, it can also override :meth:`trial_end`. Tuners use *parameter ID* to track trials. In tuner context, there is a one-to-one mapping between parameter ID and trial. When the framework ask tuner to generate hyper-parameters for a new trial, an ID has already been assigned and can be recorded in :meth:`generate_parameters`. Later when the trial ends, the ID will be reported to :meth:`trial_end`, and :meth:`receive_trial_result` if it has a final result. Parameter IDs are unique integers. The type/format of search space and hyper-parameters are not limited, as long as they are JSON-serializable and in sync with trial code. For HPO tuners, however, there is a widely shared common interface, which supports ``choice``, ``randint``, ``uniform``, and so on. See ``docs/en_US/Tutorial/SearchSpaceSpec.md`` for details of this interface. [WIP] For advanced tuners which take advantage of trials' intermediate results, an ``Advisor`` interface is under development. See Also -------- Builtin tuners: :class:`~nni.algorithms.hpo.hyperopt_tuner.hyperopt_tuner.HyperoptTuner` :class:`~nni.algorithms.hpo.evolution_tuner.evolution_tuner.EvolutionTuner` :class:`~nni.algorithms.hpo.smac_tuner.SMACTuner` :class:`~nni.algorithms.hpo.gridsearch_tuner.GridSearchTuner` :class:`~nni.algorithms.hpo.networkmorphism_tuner.networkmorphism_tuner.NetworkMorphismTuner` :class:`~nni.algorithms.hpo.metis_tuner.mets_tuner.MetisTuner` :class:`~nni.algorithms.hpo.ppo_tuner.PPOTuner` :class:`~nni.algorithms.hpo.gp_tuner.gp_tuner.GPTuner` """
[docs] def generate_parameters(self, parameter_id, **kwargs): """ Abstract method which provides a set of hyper-parameters. This method will get called when the framework is about to launch a new trial, if user does not override :meth:`generate_multiple_parameters`. The return value of this method will be received by trials via :func:`nni.get_next_parameter`. It should fit in the search space, though the framework will not verify this. User code must override either this method or :meth:`generate_multiple_parameters`. Parameters ---------- parameter_id : int Unique identifier for requested hyper-parameters. This will later be used in :meth:`receive_trial_result`. **kwargs Unstable parameters which should be ignored by normal users. Returns ------- any The hyper-parameters, a dict in most cases, but could be any JSON-serializable type when needed. Raises ------ nni.NoMoreTrialError If the search space is fully explored, tuner can raise this exception. """ # FIXME: some tuners raise NoMoreTrialError when they are waiting for more trial results # we need to design a new exception for this purpose raise NotImplementedError('Tuner: generate_parameters not implemented')
[docs] def generate_multiple_parameters(self, parameter_id_list, **kwargs): """ Callback method which provides multiple sets of hyper-parameters. This method will get called when the framework is about to launch one or more new trials. If user does not override this method, it will invoke :meth:`generate_parameters` on each parameter ID. See :meth:`generate_parameters` for details. User code must override either this method or :meth:`generate_parameters`. Parameters ---------- parameter_id_list : list of int Unique identifiers for each set of requested hyper-parameters. These will later be used in :meth:`receive_trial_result`. **kwargs Unstable parameters which should be ignored by normal users. Returns ------- list List of hyper-parameters. An empty list indicates there are no more trials. """ result = [] for parameter_id in parameter_id_list: try: _logger.debug("generating param for %s", parameter_id) res = self.generate_parameters(parameter_id, **kwargs) except nni.NoMoreTrialError: return result result.append(res) return result
[docs] def receive_trial_result(self, parameter_id, parameters, value, **kwargs): """ Abstract method invoked when a trial reports its final result. Must override. This method only listens to results of algorithm-generated hyper-parameters. Currently customized trials added from web UI will not report result to this method. Parameters ---------- parameter_id : int Unique identifier of used hyper-parameters, same with :meth:`generate_parameters`. parameters Hyper-parameters generated by :meth:`generate_parameters`. value Result from trial (the return value of :func:`nni.report_final_result`). **kwargs Unstable parameters which should be ignored by normal users. """ raise NotImplementedError('Tuner: receive_trial_result not implemented')
def _accept_customized_trials(self, accept=True): # FIXME: because Tuner is designed as interface, this API should not be here # Enable or disable receiving results of user-added hyper-parameters. # By default `receive_trial_result()` will only receive results of algorithm-generated hyper-parameters. # If tuners want to receive those of customized parameters as well, they can call this function in `__init__()`. # pylint: disable=attribute-defined-outside-init self._accept_customized = accept
[docs] def trial_end(self, parameter_id, success, **kwargs): """ Abstract method invoked when a trial is completed or terminated. Do nothing by default. Parameters ---------- parameter_id : int Unique identifier for hyper-parameters used by this trial. success : bool True if the trial successfully completed; False if failed or terminated. **kwargs Unstable parameters which should be ignored by normal users. """
[docs] def update_search_space(self, search_space): """ Abstract method for updating the search space. Must override. Tuners are advised to support updating search space at run-time. If a tuner can only set search space once before generating first hyper-parameters, it should explicitly document this behaviour. Parameters ---------- search_space JSON object defined by experiment owner. """ raise NotImplementedError('Tuner: update_search_space not implemented')
[docs] def load_checkpoint(self): """ Internal API under revising, not recommended for end users. """ checkpoin_path = self.get_checkpoint_path() _logger.info('Load checkpoint ignored by tuner, checkpoint path: %s', checkpoin_path)
[docs] def save_checkpoint(self): """ Internal API under revising, not recommended for end users. """ checkpoin_path = self.get_checkpoint_path() _logger.info('Save checkpoint ignored by tuner, checkpoint path: %s', checkpoin_path)
[docs] def import_data(self, data): """ Internal API under revising, not recommended for end users. """ # Import additional data for tuning # data: a list of dictionarys, each of which has at least two keys, 'parameter' and 'value' pass
def _on_exit(self): pass def _on_error(self): pass